Home → archive → November 2026

The 24-hour window as it closed at midnight, with the final briefing of the day.

The Brief


final cycle 00:00 · window closed 00:00

The Federal Reserve's own survey now ranks AI as the top threat to financial stability with advanced-computing debt at $680bn and hedge-fund leverage at record highs, while Nvidia's post-Huang buyer purge strips half of Asia's approved GPU customers and a suspected software update grounds the entire US air-traffic system.

Sort
Group

01DigiTimes AsiaSemiconductors

Nvidia's interim CEO Colette Kress signals stability first as board weighs future after Huang

Nvidia's board announced Jensen Huang's immediate resignation as President, CEO and director, naming EVP/CFO Colette Kress as interim CEO while a permanent search covering internal and external candidates proceeds. The board also appointed Harvey C. Jones, former Synopsys CEO, as its first-ever Chair, and affirmed that operations, customer commitments and the product roadmap will continue uninterrupted ahead of Q3 FY2027 results on November 17. Analysts expect the permanent CEO search to take months, leaving the strategic technology-direction decision deferred.

The takeThe abrupt resignation of Jensen Huang — the architect of Nvidia's accelerated-computing and AI dominance — and the board's decision to install CFO Colette Kress as interim CEO rather than an engineer signals a deliberate stability-first posture, deferring the critical question of who steers the technology roadmap for a company whose GPU platforms underpin the global AI buildout. The appointment of Harvey Jones (ex-Synopsys CEO) as first-ever board chair adds semiconductor-industry depth to governance during the transition.

02Politico Tech & PolicyCybersecurity

Mullin doesn't rule out cyberattack in unprecedented nationwide passenger plane grounding

The FAA issued an unprecedented nationwide ground stop citing communication issues, grounding all US passenger traffic. Homeland Security Secretary Markwayne Mullin said there was no indication of a threat but, asked whether a cyberattack or AI intrusion could be responsible, replied 'That's not known.' The administration also said it is examining a new AI air traffic control system rolled out at airports this week.

The takeA nationwide FAA ground stop — only the second in recent years — with DHS's Mullin explicitly declining to rule out a cyberattack or AI intrusion, lands squarely in the cybersecurity/critical-infrastructure thesis; the same-day rollout of a new AI air traffic control system at airports adds a live frontier-AI-in-critical-systems angle that will shape scrutiny of both the outage and the deployment.

03Semafor TechSemiconductors

Exclusive / How Nvidia’s Jensen Huang – and Trump – jumped the guardrails

Semafor reports previously undisclosed January negotiations in which Nvidia's GC Tim Teter told government officials 'what Donald Trump and Jensen Huang want, goes,' pressing to weaken H200 export-license conditions including cloud-customer know-your-customer checks and raising the KYC threshold to tens of thousands of chips. Current and former congressional staffers describe Nvidia lobbying as 'gaslighting,' including claims that chip location-verification was technically infeasible — a feature Reuters later reported Nvidia had been privately demonstrating and launched in May. The article also details Commerce Secretary Lutnick's accommodating posture, earlier-than-reported deliberations over blacklisting Megaspeed, and the fallout from Huang's departure, which has shaved $1.5 trillion off Nvidia's market value.

The takeSemafor's exclusive documents an unprecedented level of corporate capture in US export-control policymaking: Nvidia's general counsel invoked Trump's personal backing to strip 'poison pill' cloud-customer KYC requirements from H200 China license conditions, while the company told Congress chip location-verification was technically infeasible even as it privately demonstrated the feature for months. The piece ties this lobbying to concrete national-security failures — a 50,000-chip undercount of Megaspeed sales and $7.5B in purchases by entities tied to a suspected Chinese spy — and to a $1.5T market-value wipeout after Huang's exit.

04ReutersMarkets

AI rockets to top of worries about financial system in latest Fed survey

The Federal Reserve's semi-annual Financial Stability Report found AI became the top-cited financial stability risk, named by 87% of survey contacts versus 50% in May, driven by cyberattack fears, complex AI-related financing and the risk of an AI earnings slowdown triggering a broad correction. The report documents advanced computing-related debt issuance surging to $680 billion by October 2026 — five times all of 2023 — now nearly a third of nonfinancial investment-grade issuance, with debt, guarantees and pending leases potentially exceeding $3 trillion. The Fed warns that circular financing arrangements among hyperscalers, public credit markets, private lenders and special purpose vehicles create interlinkages requiring close monitoring, and notes hedge fund leverage is at the highest level since data collection began in 2013.

The takeThe Fed's Financial Stability Report shows AI risk citations jumping from 50% to 87% of contacts in six months, while advanced-computing debt issuance hit $680 billion by October 2026 — a fivefold increase over all of 2023 and nearly a third of all nonfinancial investment-grade issuance, versus under 6% three years ago. With total debt, guarantees and not-yet-commenced leases potentially exceeding $3 trillion and circular financing between hyperscalers, private lenders and SPVs, the Fed is explicitly flagging AI datacenter buildout as a systemic credit channel comparable to, and possibly surpassing, the late-1990s telecom cycle.

05BloombergMarkets

Watch Inside Jensen Huang’s Last Hours as Nvidia CEO

Bloomberg-sourced reporting and aggregated coverage describe Huang's removal after a board investigation triggered by a Nov. 6 FT report and a confession to Nvidia's general counsel, with the board meeting running late into Thursday night. Succession is wide open — 11 floated names from Kress and CUDA's Ian Buck to Arm's Rene Haas, AMD's Lisa Su and Apple's Johny Srouji — with the board calling candidate speculation premature and a search that could take six months. The strategic stakes are financial: Nvidia held stakes in 229 private companies as of July and carries ~$530B in guarantees and backstops, and the assumption that Nvidia would always step in has 'largely evaporated,' with Kress positioned to rein in the riskiest bets.

The takeJensen Huang's abrupt exit after 33 years leaves Nvidia with an interim CEO, a board still selecting a search firm, and roughly $530B in off-balance-sheet commitments plus ~$140B of dealmaking in two months now in question — a genuine repricing event for the AI compute ecosystem's central financier. The piece also flags that interim CEO Colette Kress has already halved Nvidia's largest OpenAI backstop and is skeptical of the circular-deal model, meaning the 'AI banker' role that underwrote much of the boom may contract.

06The InformationGeopolitics & Defense

Sex, Lies and AI Chips: Jensen Huang’s Last Days as Nvidia CEO

Nvidia's board opened an independent review after Huang acknowledged infidelity and reports surfaced that the FBI is investigating whether Alice Huang, a former Megaspeed executive, transferred information to Beijing; Megaspeed itself faces allegations of irregular chip shipments to China, which it denies. The board hired former SDNY acting U.S. attorney Joon H. Kim and Wachtell, Lipton to run the probe, then met as a second reported relationship with subordinate executive Kimberly Powell emerged. On November 12 Huang resigned immediately from the CEO role, executive management and board; CFO Colette Kress is interim CEO, Harvey C. Jones chairs the board, and the permanent successor search spans internal and external candidates.

The takeThe immediate resignation of Jensen Huang after 33 years, triggered by an FBI probe into whether a personal relationship with a former Megaspeed executive was a vector for foreign intelligence transfer to Beijing, removes the central figure of the AI compute boom at the exact moment Nvidia faces export-control scrutiny — shares fell more than 10% pre-market and Q3 FY2027 results land November 17 with an interim CFO-CEO and no investigative conclusions.

07Wall Street JournalCybersecurity

Iran’s New Cyber Playbook: Create Real-World Havoc

WSJ reports three Iranian-attributed cyberattacks since summer with real-world physical effects: Midwest water-station control panels locked out, networks on two Texas-bound energy tankers breached mid-Atlantic, and European grocery refrigerators driven to 120°F. The syndicated analysis details the escalation — CISA/EPA/FBI/NSA warnings on internet-facing industrial equipment, incidents in at least 12 states with operators reverting to manual control, and hybrid physical-cyber operations including compromise of internet-connected cameras to support kinetic targeting and plots against US officials.

The takeIranian-linked actors have crossed the threshold from espionage to physical-world disruption, locking water-utility control panels across at least 12 US states, breaching tanker networks, and manipulating grocery refrigeration to 120°F — a deliberate pivot to operational-technology attacks that signals a new asymmetric-warfare playbook China and Russia are watching.

08Financial TimesSemiconductors

Jensen Huang’s years-long effort to get close to China

The FT traces Jensen Huang's years-long, largely unsuccessful effort to gain access to China's top leadership as Washington and Beijing tightened chip restrictions, noting Nvidia never secured meetings with the NDRC, the state planner leading Beijing's self-sufficiency drive, and only recently hired dedicated government-affairs staff. Chinese regulators moved to restrict even export-control-compliant Nvidia chips to boost Huawei, while Chinese AI labs remain dependent on Nvidia silicon via overseas data centers or illicit supply networks. Following revelations about an affair with a Chinese businesswoman whose company was investigated over suspected illicit shipments, Huang resigned Thursday amid US calls to examine whether his China dealings compromised national security, with a law to block remote overseas chip access advancing in Congress.

The takeThe piece is a substantive investigation into how export controls reshaped Nvidia's China access and Beijing's push for semiconductor self-sufficiency, with concrete detail: roughly a quarter of Nvidia's sales came from China annually 2010-2022, Huang met only Vice-Premier He Lifeng (April 2025) and Commerce Minister Wang Wentao (July 2025) after years of trying, and the NDRC — not Mofcom — drove the import restrictions favoring Huawei. It also ties the scandal to a looming congressional remote-access chip ban and Huawei's claim that six or seven Chinese labs aim to train large models on Ascend within two years.

09Tom's HardwareSemiconductors

How Shanghai fund manager Kang Wei introduced his wife Alice Huang to Jensen and built $7.5 billion Nvidia AI GPU smuggling operation

Tom's Hardware, drawing on WSJ, Economic Times, The Wire China and Axios reporting, traces how Shanghai fund manager Kang Wei introduced his wife Alice Huang (Huang Le) into Jensen Huang's orbit, and how entities she controlled — Megaspeed (formerly 7Road International) and its Malaysian subsidiary Speedmatrix — imported $4.6 billion and $2 billion of Nvidia AI chips respectively between 2023 and 2025, with total diverted shipments estimated at ~240,000 GPUs worth ~$7.5 billion. The piece details the corporate paper trail (Shanghai Hexi, BVI-based Osmanthus Vale, 7Road Holdings, Shanghai Century Litong data centers) and notes unresolved questions about U.S. intelligence claims that Huang worked with China's MSS. It also reports Chinese regulatory penalties against Kang Wei, including a 2.5 million yuan fine and an August 2026 CSRC warning letter.

The takeA reported $7.5 billion, ~240,000-GPU diversion network allegedly ran through Megaspeed and Speedmatrix, with $4.6B imported directly and $2B routed via a Malaysian subsidiary — a concrete case of export-control evasion at a scale that dwarfs typical smuggling busts and directly implicates the chip-supply chain that underpins the AI buildout.

10BloombergSemiconductors

Alice Huang-Linked Firm Procured Banned Nvidia Chips in China

Documents reviewed by Bloomberg show a Chinese company founded by Alice Huang procured smuggled Nvidia AI chips installed at a Shanghai data center linked to Megaspeed International. Aperia Cloud Services, an Nvidia Cloud Partner, bought 22 Dell servers with H800 chips restricted since 2023; Dell says it cut off Aperia in February 2025 over diversion concerns, the month Singapore arrested Aperia's CEO. Filings by a unit of Shanghai Longteng list 193 servers worth 485 million yuan ($73M) as loan collateral, including Super Micro units with H100/H200 chips, deepening scrutiny of Nvidia's partner network and export-control enforcement.

The takeBloomberg documents 22 Dell-sold servers with restricted Nvidia H800s plus 52 Super Micro units carrying H100/H200s — $73M of collateralized hardware — installed at a Shanghai data center tied to Alice Huang's network, the most explicit evidence yet that NCP-channel diversion routed banned AI chips into China despite export controls.

11CSISCybersecurity

The Largest Crypto Heist in History Could Have Been Avoided

CSIS analysts detail how the TraderTraitor group (a Lazarus subset) drained over $6B from Binance on October 29, 2026, using months of social engineering and technical intrusion, with only ~44% frozen. They trace the proceeds to North Korea's sanctions evasion and defense funding—$12.8B stolen since 2017, $7.8B in 2026 alone—and critique the Trump administration's light-touch crypto regulation and its August 2026 memo barring private-sector offensive cyber operations against state actors.

The takeA $6B North Korean theft from Binance—the largest crypto heist ever—illustrates how state-sponsored cybercrime directly funds Pyongyang's missile and nuclear programs, with up to 40% of its defense budget reportedly financed through stolen crypto. The attack's mechanics (LinkedIn social engineering, a malicious download, an internal AI tool exploited as an entry vector, and laundering across 12,000 wallets and 11 blockchains in ~20 minutes) are concrete, and the piece ties them to sanctions-evasion and policy failures.

12The RegisterCybersecurity

This year’s bug count just passed 100K

VulnCheck counted 100,222 CVEs disclosed in 2026 by Friday, more than double 2025's total, with Microsoft alone adding 1,094 flaws in a single Tuesday and 1,474 the prior month; CVE Numbering Authorities attribute the surge to AI-powered vulnerability hunting. Exploitation is accelerating too: CrowdStrike saw 88% of PoC-based exploitation within 48 hours of release from January to June, China-linked Vault Panda and Genesis Panda struck within 24 hours, and VulnCheck recorded honey-pot hits minutes after publishing 14 Copeland refrigeration controller flaws (two actively exploited, the same devices Iranian hackers abused in European supermarkets). Bressers argues the industry now has a 'data problem,' pushing risk-based prioritization over CVSS severity, while open source projects like curl and the Eclipse Foundation's Open VSX restructure reporting and funding to absorb the load.

The takeAI-driven vulnerability hunting has pushed 2026 CVE disclosures past 100,000 — more than double 2025 — while the median time from CVE publication to exploitation collapsed from 120 days in 2025 to 80 days in H1 2026 and often to hours, with only ~1.4% (1,363) of CVEs actually exploited in the wild. The strategic shift is from patching-by-severity to risk-based prioritization (CISA's June directive) and from bug counts to infrastructure heat maps, as open source maintainers like curl's Daniel Stenberg buckle under AI-generated report avalanches.

13Wall Street JournalCybersecurity

The Morning Risk Report: How Hackers Steal Millions in AI Computing Power

Hackers probed cloud platform Railway in mid-October, spun up thousands of accounts and stole assets worth over $1.1 million — not crypto or ransomware, but AI compute. Roughly a dozen AI developer platforms have been hit in the past month, with attackers harvesting API keys to illicitly access US AI models; Google's Wiz researchers say some stolen compute feeds 'transfer stations' selling access to restricted countries including China, Russia and Iran, which Anthropic says are often used to clone US models via distillation. Enterprise accounts have seen up to $100,000 in fraudulent charges in a single day and total losses up to $1 million per organization, while lawmakers are pushing three new AI-chip export measures in a stalled defense package following Jensen Huang's affair with a partner suspected of Chinese intelligence ties.

The takeA genuinely new attack class — 'LLM jacking' — is turning stolen API keys into a $1M-per-victim criminal economy, with ~a dozen AI developer platforms hit in a month and stolen compute routed to transfer stations serving China, Russia and Iran and feeding model distillation. The piece also flags a second-order geopolitical shock: the Jensen Huang affair is reviving Capitol Hill momentum for three AI-chip export measures that could cost Nvidia billions.

14QbitAI (量子位)Markets

Tesla and SpaceX Could Merge as Soon as Next Month

Bloomberg-sourced reporting says SpaceX and Tesla boards and advisors are negotiating merger terms, equity structure and governance, with an announcement targeted for December and early 2027 as fallback. SpaceX already absorbed xAI (and thus X) in February at a $1.25T combined valuation and completed a super IPO in June; Musk confirmed in October that both companies formed independent special committees. The two firms are already co-building the Terafab logic/memory fab in Texas for Tesla edge-inference chips and SpaceX space-datacenter chips, with first test silicon targeted for mid-2027 and volume in Q3 2028.

The takeA combined SpaceX-Tesla would be a ~$3.4T entity spanning rockets, Starlink, xAI/Grok, chips, compute, EVs and robotics — and the article surfaces the concrete industrial logic: the joint Terafab fab targeting >1TW of compute demand and a fresh $74B SpaceX build-out of two Texas/Arizona compute campuses, with AI already over half of SpaceX's Q3 revenue.

15BloombergInfrastructure

FBI Said to Probe Whether Software Update Caused Mass FAA Outage

Frequency problems at Chicago O'Hare and JFK on Friday afternoon escalated into a nationwide FAA ground stop citing 'extraordinary circumstances and reasons of safety,' hitting communications at every Air Route Traffic Control Center. The FBI does not believe it was a cyberattack and is examining whether a software update caused it; DHS's Mullin called it 'just a telecommunications outage as of now.' Roughly 3,100 inbound aircraft were affected with at least 8,000 delays and 400 cancellations, and officials explicitly ruled out a link to a separate Microsoft 365/Exchange outage the day before. The piece situates it alongside the 2023 NOTAM deletion failure and the July 2024 CrowdStrike update that crashed ~8.5 million Windows machines, against $12.5 billion in 2025 ATC modernization funding and a further ~$30 billion request.

The takeA nationwide FAA ground stop triggered by suspected faulty radio-frequency software at all Air Route Traffic Control Centers — with ~3,100 aircraft affected, 8,000+ delays and 400 cancellations — is a first-order critical-infrastructure failure, and the FBI's 'update, not hack' line of inquiry makes it a case study in change-management risk across systems too important to fail quietly.

16Financial TimesCybersecurity

Microsoft outage exposes global reliance on core tech providers

Microsoft suffered one of its longest and most disruptive Microsoft 365 outages, affecting millions of users in more than 70 countries for over 24 hours, hitting Goldman Sachs, the German government, NHS England, UK police and schools. Microsoft ruled out a cyber attack, attributing the fault to a technical issue during a network capacity upgrade that made many servers unavailable and overloaded traffic. The incident, following the 2024 CrowdStrike and AWS outages, highlights global dependence on a handful of tech providers, and security officials warned that with Defender also down, criminals would exploit the disruption.

The takeA multi-day Microsoft 365 outage across 70+ countries—knocking out Outlook, OneDrive, Copilot and Defender—is a concrete concentration-risk event: with Defender down, corporate security teams lost visibility into malicious activity, a window the UK NCSC warned criminals would exploit. The cause, a 'network capacity upgrade' that overloaded servers, underscores how a single provider's infrastructure change can cascade into government, financial and public-service disruption.

17Tech in AsiaMarkets

Nvidia staff urge Jensen Huang to return after resignation

Nvidia's board announced Huang's resignation on November 12 following reports of alleged affairs, including one involving a direct report; Huang told employees Sunday he had been unfaithful to his wife. CFO Colette Kress is interim CEO while US lawmakers and the FBI open investigations, and an internal suggestion box post backing Huang's return drew over 1,600 comments. The company has shed roughly US$1 trillion in market value since the allegations surfaced, with the lack of a formal succession plan flagged as a governance risk.

The takeThe abrupt resignation of Jensen Huang at a ~$4 trillion chipmaker, a ~$1 trillion market-value wipeout, and FBI plus congressional scrutiny constitute a first-order semiconductor and markets event; the absence of a formal CEO succession plan at the company underpinning global AI compute adds systemic risk to the entire AI infrastructure buildout.

18CNBC TechnologyGeopolitics & Defense

Lutnick faces questions over Nvidia spy scandal. His sons’ firm has deep ties to the chipmaker

CNBC reports that Cantor Fitzgerald, formerly run by Commerce Secretary Howard Lutnick and now led by his sons, nearly halved its reported Nvidia share count between September 2025 and March 2026, coinciding with White House and intelligence pressure on Lutnick to crack down on Singapore-based Megaspeed over alleged AI chip diversion to China — pressure he repeatedly resisted per Axios. Cantor continued arranging billions in financing across Nvidia's ecosystem, including a $3B IREN convertible bond (May 2026) and Lancium's $600M financing, while holding $362M in Nvidia and $115M in IREN shares by June. The piece adds new detail on classified intelligence warnings about Jensen Huang's suspected Chinese spy ties, Commerce's quiet clearance of ~10 Chinese firms including Alibaba and ByteDance for H200 purchases, and Treasury Secretary Bessent publicly saying Treasury, Defense, State and USTR backed entity-listing Megaspeed for months.

The takeA genuine conflict-of-interest and export-control scandal: Commerce Secretary Lutnick resisted cracking down on Nvidia chip smuggling to China (Megaspeed) while his sons' firm Cantor Fitzgerald cut its reported Nvidia stake ~49% (from 3.5M to 1.78M shares, $652M to $310M) and simultaneously booked bookrunner roles on Nvidia-adjacent financings like IREN's $3B convertible and Lancium's $600M deal — all after Lutnick received classified warnings about Jensen Huang's suspected spy ties. The story ties high-tech geopolitics (entity-list delays, H200 approvals for Alibaba/ByteDance) directly to capital-markets activity, with Treasury's Bessent publicly breaking ranks.

19The ElecSemiconductors

Samsung Begins 144 Trillion Won Investment in South Korea’s Chungcheong Region

Samsung broke ground on a 144 trillion won Chungcheong mega-project spanning advanced HBM packaging (Samsung Electronics, 60T won), OLED lines (Samsung Display, 67T won), AI-server package substrates (Samsung Electro-Mechanics, 8T won) and batteries (Samsung SDI, 9T won). The Onyang Line 5 HBM packaging fab will begin operations in June 2028 with 40 trillion won of advanced equipment, while 194 domestic materials, components and equipment suppliers plan roughly 18.1 trillion won of investment by 2030, with 66 firms committing 4.4 trillion won locally and an expected 23,000 jobs.

The takeSamsung is committing 46 trillion won to a dedicated HBM advanced-packaging fab at Onyang (9,400 pyeong of cleanroom, operations June 2028) plus 8 trillion won for AI-server package substrates at Sejong — a concrete, multi-year buildout of the AI memory and packaging supply chain, backed by 18.1 trillion won in supplier capex and state fast-tracked permitting.

20DigiTimes AsiaSemiconductors

Nvidia's 14-point GPU checks tighten AI server supply; full prepayment and 2027 price hikes loom

Following BIS's May 31, 2026 guidance requiring licenses for advanced computing chips to any entity with a China-headquartered ultimate parent regardless of location, Nvidia re-vetted its Asian buyer base and removed over half of previously approved customers, hitting neo-cloud resellers hardest. Field teams conducted physical data-center visits, contract checks and end-user interviews in Singapore, Malaysia and Japan, with reinstatement requiring ownership-structure documentation most smaller operators cannot quickly produce. The crackdown runs alongside enforcement cases — Aperia Group fraud charges and a S$55M asset order, Megaspeed scrutiny over $2B in orders, a 72-server KLIA seizure, and the Hao Global H100/H200 smuggling conviction — while Washington simultaneously licenses 'trivial' H200 volumes to Alibaba, Tencent, ByteDance, JD.com and ZTE units even as Beijing mandates domestic chips for state data centers.

The takeNvidia's new buyer whitelist — enforced via in-person data-center inspections in Singapore, Malaysia and Japan — stripped more than half of previously authorized Asian customers of direct Blackwell/Hopper access, operationalizing the May 31, 2026 BIS guidance that shifted export control from a destination test to an ultimate-ownership test (ECCNs 3A090/4A090). This is a structural repricing of Asia-Pacific AI compute access: compliance infrastructure, not capital, now gates GPU supply, favoring established operators over neo-clouds.

21BloombergGeopolitics & Defense

Nvidia Crisis Spurs Scrutiny of Agency That Oversees Chip Curbs

Bloomberg reports that BIS chief Jeffrey Kessler considered inviting then-Nvidia CEO Jensen Huang to headline the bureau's January conference, an unusual move given BIS regulates Nvidia's AI chip exports. Following Huang's abrupt departure over an affair with Alice Huang — a woman suspected of spying for China whose company Megaspeed was under Commerce investigation — Congress is demanding answers on whether BIS did enough to keep advanced AI processors out of China and whether industry leaders influenced policy. The piece details internal dysfunction (siloed staff, a brain drain, alleged shouting matches), Nvidia's intervention in the Megaspeed probe, and BIS's implementation of the H200-to-China arrangement with a 25% US revenue cut, alongside its refusal to enforce the prior administration's GPU export rule.

The takeThis is a substantive accountability story on the machinery of US export controls: BIS, the agency that licenses Nvidia's AI chip sales, is under congressional scrutiny over whether Silicon Valley access shaped policy, with Kessler and Lutnick called to testify Dec. 3 and Dec. 10. The concrete policy stakes — the 25% revenue-share arrangement for H200 shipments to China, the unenforced Biden-era GPU rule, and staff being required to route all Nvidia requests through Kessler — make this high-tech statecraft news, not a company profile.

22SCMP TechSemiconductors

Court-appointed administrators to sound out investors for Nexperia stake amid Wingtech feud

Nexperia's court-appointed administrators said they will gauge market interest in majority or minority stakes in the Dutch chipmaker, after the Dutch Enterprise Chamber confirmed their authority to reduce Wingtech's ownership and ruled that Wingtech's actions undermined earlier court orders. Wingtech has publicly rejected the restructuring as 'expropriation,' while the administrators argue ceding control back would endanger Nexperia's ability to operate amid export controls. The Dutch court additionally ordered Wingtech and Nexperia China to stop interfering with Nexperia's business and to withdraw claims and injunctions from a Dongguan court.

The takeThe Dutch court's authorization for administrators to assess investor interest in Nexperia—including a possible reduction of Wingtech's ownership—escalates a governance fight with direct export-control stakes, as the custodians warn returning control to Wingtech poses 'serious risks' to operations under tightening multinational policies. The ruling also orders Wingtech to withdraw Chinese court claims and refrain from enforcing Dongguan injunctions, formalizing a de facto split between Nexperia's Dutch and Chinese entities.

23OpenAIAI & Models

Our first phase of embedded assessments

OpenAI signed agreements with four independent assessors for a first phase of embedded assessments, giving them on-site access to internal models, checkpoints, training/eval data, and monitoring systems to evaluate alignment and safeguard effectiveness. Apollo will probe covert subversion and detection, METR will assess misalignment in OpenAI's largest RL run plus monitor stress testing, Transluce will audit post-training alignment evaluations (with a follow-on ~8-week engagement in early 2027), and Goodfire will apply interpretability to training-run monitoring and debugging. The piece details governance protocols—SSC visibility with unredacted reports, disagreement escalation to the Safety Advisory Group, conflict-of-interest and cool-off processes, and redaction/responsible-disclosure rights—framing embedded assessments as a path toward standardized, continuous third-party safety signals.

The takeOpenAI's embedded-assessment program grants Apollo, METR, Transluce, and Goodfire deep access to model checkpoints, source code, training data, and monitoring systems—including METR's independent audit of OpenAI's largest frontier RL run to date—establishing a concrete precedent for third-party safety verification with SSC oversight and escalation paths, directly supporting its White House Accord on Super Intelligence commitments.

24Tech in AsiaInfrastructure

Nebius signs AI infrastructure deal with Meta worth up to $27b

Amsterdam-based Nebius will supply Meta with $12B of dedicated five-year capacity on one of the first large-scale NVIDIA Vera Rubin deployments, with Meta able to buy up to $15B more across upcoming clusters, taking the deal to ~$27B. Nebius intends to resell some capacity to third parties, with Meta taking the remainder. The pact follows NVIDIA's $2B strategic investment in Nebius and a prior Microsoft contract worth up to $19.4B, as hyperscalers facing power interconnection bottlenecks turn to specialist neoclouds projected to grow from ~$27B in 2025 to over $250B by 2030.

The takeA $12B firm commitment plus $15B optional for NVIDIA Vera Rubin capacity starting early 2027 shows hyperscalers outsourcing gigawatt-scale buildout to neoclouds, validating NVIDIA's partner-led strategy to lock Meta into its full stack just days after NVIDIA put $2B into Nebius for 5GW by 2030.

25CSISCybersecurity

The 2027 Congress and the Dangerous Data Breach Pileup

CSIS's Justin Sherman lays out a 'dangerous pileup' of federal data breaches: DOD HR data on roughly 3 million personnel, an FBI jobs portal compromise affecting sensitive counterintelligence and clandestine roles, and DHS personnel files published searchable online covering 40,000+ agents. He urges the 2027 Congress to pursue bipartisan oversight across three fronts — technical containment and scope assessment, physical protection for exposed personnel against threats from Iran and cartels, and counterintelligence mitigation against China and Russia exploiting the data now and years later.

The takeA CSIS analysis documenting three major executive-branch breaches in a short span — 2.76M living and 294K deceased DOD personnel records, an FBI jobs portal exposing counterintelligence and clandestine-operations staff, and a searchable public database of 40,000+ DHS agents including undercover personnel — argues the aggregate counterintelligence and physical-targeting risk is greater than the sum of parts, especially as AI enables cross-referencing breached datasets.

26Institute for the Study of WarGeopolitics & Defense

Iranian Direct Action in the United States and Europe Must Be Understood Through a Framework of War, Not Terrorism

ISW argues Iran's attacks on Jewish targets in Europe (March–April 2026, via IRGC and Kataib Hezbollah front groups) evolved into a more effective direct-action campaign against critical infrastructure and military bases in Fall 2026. Iranian threats prompted the withdrawal of US B-1 bombers from RAF Fairford and KC-135 tankers from RAF Mildenhall, while cyber operations hit US water treatment plants in 12 states and refrigeration systems at six US military bases, replicated against French grocery infrastructure with millions of euros in losses. The report contends that treating these as state-sponsored terrorism obscures their role as military action pursuing political objectives, and urges governments to respond across the full spectrum of state capability rather than through prosecution alone.

The takeISW reframes Iran's 2026 campaign of infrastructure and base attacks in the US and Europe as military 'direct action' rather than terrorism, citing concrete effects: B-1 bombers withdrawn from RAF Fairford, 12 KC-135s relocated from RAF Mildenhall, water-plant hacks in 12 states and refrigeration attacks at six US bases plus nationwide spoilage in France. The strategic argument is that law-enforcement framing produces tactical myopia and reactive responses to what is actually a coherent wartime campaign.

27Tech in AsiaSemiconductors

OpenAI backs US push to restrict China’s access to AI chips

OpenAI has privately told US lawmakers it supports bipartisan measures to further restrict China's access to advanced AI chips and chipmaking equipment, with Congress weighing attaching three proposals to the annual defense bill. The provisions would limit future high-end chip sales, curb diversion of Nvidia processors, and tighten China's access to ASML equipment, building on existing controls that already forced Nvidia to modify H20 chips for China. Nvidia allies and trade groups oppose the push, the chipmaker has stayed quiet, and the measures remain under negotiation amid White House concern over jeopardizing a temporary trade truce.

The takeOpenAI is openly lobbying for a legislative tightening of the chip export-control regime — three proposals to be attached to the NDAA targeting high-end chip sales, Nvidia processor diversion, and ASML tooling — a direct escalation of AI-lab statecraft that pits it against Nvidia-aligned trade groups and risks the fragile US-China trade truce.

28Ars TechnicaCybersecurity

Windows 0-day dropped by disgruntled researcher hands hackers System rights

Researcher Nightmare Eclipse disclosed SignDrop, a CLFS vulnerability that lets a local user gain full System rights on some Windows versions. By forcing a privileged process like Windows Defender to open a crafted CLFS container via manipulated symbolic links, the attacker gets CLFS to create an authentication file with System privileges and the attacker's ACL, enabling privilege escalation. The flaw stems from a Microsoft mitigation introduced in October 2025 meant to block other CLFS exploits.

The takeA publicly dropped Windows zero-day (SignDrop) chains CLFS, Windows Defender's ServerMpUpdateEngineSignature, and Object Manager symbolic links to escalate an ordinary user to System — and ironically weaponizes an October 2025 Microsoft CLFS mitigation against itself, a concrete example of a defense becoming an attack surface.

29Yahoo Tech & Markets WireGeopolitics & Defense

Scott Bessent and Lutnick clash over Nvidia chips and China

At a post-midterm Situation Room meeting on Nov. 11, Treasury Secretary Scott Bessent pressed Commerce Secretary Howard Lutnick on why Megaspeed — a company accused of routing advanced Nvidia GPUs to China in violation of export law — still isn't on the Entity List, despite State, Defense and Energy backing the designation since fall 2025. Reporting indicates Lutnick was briefed on Megaspeed in May before a China trip, and that Commerce export-enforcement chief Jeff Kessler halted a formal probe after Nvidia objected, with Lutnick wary of damaging Beijing trade talks. The case intensified after Nvidia CEO Jensen Huang resigned effective Nov. 12 following a board investigation into an extramarital affair with a woman reportedly identified by U.S. authorities as a Chinese intelligence operative; White House officials are now discussing whether Lutnick should stay on, and Commerce disputes the account.

The takeA genuine high-tech-geopolitics story: the Treasury-Commerce split over whether to blacklist Nvidia-GPU diversion suspect Megaspeed exposes how export-control enforcement is being subordinated to trade diplomacy, with State, Defense and Energy backing the listing while Commerce stalled for over a year. The simultaneous resignation of Nvidia CEO Jensen Huang over an affair with a woman U.S. authorities reportedly identified as a Chinese intelligence operative raises the stakes from routine enforcement to counterintelligence.

30Financial TimesMarkets

Anthropic’s Dario Amodei ‘not super fussed’ about market turmoil

Dario Amodei told JPMorgan's Jamie Dimon that Anthropic is 'not super fussed' about volatile markets ahead of its Thursday Nasdaq listing, expected to value the lab at about $1.5tn after investors trimmed earlier $2tn+ expectations amid AI-boom anxiety and Nvidia CEO Jensen Huang's resignation. The oversubscribed IPO plus a simultaneous private placement could raise a record $100bn+, nearly doubling the $130bn cash Anthropic held at end-September and buffering a planned $900bn infrastructure spend. Amodei called compute 'the deepest business dilemma' the company faces, noting short-term spot-market purchases and deals with SpaceX worth up to $3.1bn per month.

The takeAnthropic is heading into a ~$1.5tn Nasdaq IPO next Thursday with an oversubscribed book that could raise $100bn+ on top of $130bn cash, explicitly to fund a $900bn multi-year infrastructure buildout — including SpaceX datacenter capacity at up to $3.1bn/month — making compute, not valuation, the binding constraint on frontier AI economics.

31DigiTimes AsiaGeopolitics & Defense

Remote access emerges as biggest gap in US chip controls as Congress and Commerce move to close it

US policymakers are converging on the remote-access loophole — Chinese AI developers renting Nvidia chips in offshore data centers — as the primary target of a tougher export-control regime. Democrats winning both chambers in the November 3 midterms gives them committee power to subpoena Nvidia CEO Jensen Huang and investigate industry compliance, while the Commerce Department races to finalize its own remote-access rule. The IFP estimate that remote access accounts for about half of the AI compute still cuttable from China by end-2027 quantifies the stakes.

The takeRemote access to overseas data centers now represents roughly half of the US-made AI compute Washington could still deny China by end-2027 (per Institute for Progress), making it the pivotal remaining chokepoint in export controls; the Democratic sweep of both chambers plus Commerce's pending remote-access rule creates a two-track enforcement squeeze that directly threatens Nvidia's rental-channel demand and sets up public CEO testimony.

32Yahoo Tech & Markets WireMarkets

Nvidia has lost nearly $2tn in value. What could it be worth without Jensen Huang?

Nvidia has shed roughly $1.8tn in market value since its 6 October record close after CEO Jensen Huang resigned amid reports of infidelity, with about $900bn wiped out between the prior Friday's close and Thursday's intraday low. CFO Colette Kress has stepped in as interim CEO while the board searches for a permanent successor. Bloomberg Intelligence's Kunjan Sobhani sees no change to FY2027-28 earnings estimates but flags longer-term risks around succession, dealmaking and China scrutiny as investors reassess the 'Jensen premium'.

The takeThe abrupt exit of Jensen Huang — the architect of Nvidia's invest-in-customers-and-suppliers flywheel — has erased ~$1.8tn in market value since the 6 October record close of $239.24, forcing investors to reprice the 'Jensen premium' and raising real questions about whether CFO-turned-interim-CEO Colette Kress can sustain the dealmaking that underwrote the AI buildout.

33The RegisterGeopolitics & Defense

Chinese AI labs are training models on Nvidia's Blackwell despite export controls, says ex-OpenAI engineer

Ex-OpenAI engineer Will DePue published a report claiming that model weights retain a hardware fingerprint from the GPU used to initialize them, and that matching seed/layout patterns against published weights shows six Chinese models — including DeepSeek-V4-Pro, Kimi K3 and MiniMax-M3 — were trained on Nvidia Blackwell, four on Grace Blackwell GB200/GB300 systems. He concedes the method cannot establish GPU location, leaving open access via US cloud providers, a loophole the proposed Cloud Security Act targets. The finding lands amid tightened export-control legislation (AI Overwatch Act, Chip Security Act, Match Act, Cloud Security Act) and separate evidence of B300 GPUs geolocated to China.

The takeA forensic weight-initialization analysis (Philox PRNG seed/layout fingerprinting across 6 trillion combinations) claims to prove DeepSeek-V4-Pro, Kimi K3, MiniMax-M3 and Ant Group models were initialized on export-controlled Nvidia Blackwell — including GB200/GB300 rack-scale parts — a concrete, reproducible challenge to the efficacy of US chip export controls.

34BloombergGeopolitics & Defense

Nvidia’s Ousted CEO Gets Fresh Subpoena Threat From House GOP

House Intelligence, Foreign Affairs and China Select committee chairs have demanded briefings from ousted Nvidia CEO Jensen Huang and its general counsel over security protocols and Huang's affair with a Megaspeed executive suspected of spying for China and routing restricted chips to Chinese buyers, threatening subpoenas if they don't respond by Friday. Treasury Secretary Bessent blamed Commerce for failing to blacklist Megaspeed despite months of interagency support, while lawmakers rally behind bipartisan export-control bills (AI Overwatch Act, Chip Security Act, MATCH Act, Remote Access Security Act) to be attached to the NDAA.

The takeA national-security scandal tying Nvidia's ousted CEO to a suspected Chinese spy at Singapore-based Megaspeed has triggered congressional subpoena threats and is being used as the legislative vehicle to push the AI Overwatch Act, Chip Security Act and MATCH Act into the year-end NDAA — a direct escalation of US export-control statecraft against China's AI chip access.

35Ars TechnicaMarkets

Rocket Report: Starship set for historic test flight; a century of Electron

Ars Technica's Rocket Report covers SpaceX's 15th Starship test flight, planned to return both Super Heavy and the ship to the launch site for catch—the first full recovery of an orbital system—while deploying 26 Starlink V3 satellites over roughly 14 orbits. It also details Pentagon launch-market intervention via $150M loans to Stoke Space and Rocket Lab, Rocket Lab's Neutron price hike to $70M amid constrained medium-lift capacity, and new multi-launch deals for Firefly's Alpha (up to 10 with Synspective) and Stoke's Nova with Planet.

The takeSpaceX's Flight 15 aims to be the first full recovery of an orbital launch system—booster catch plus a 22-hour orbital Starship return and landing—while delivering 26 Starlink V3 satellites, a milestone for reusability economics. The roundup also captures a structural shift: the Pentagon's Office of Strategic Capital issued two $150M loans (Stoke, Rocket Lab) to hedge against Falcon retirement, and Rocket Lab raised Neutron's price from $55M to $70M with bookings now into the 2030s.

36Politico Tech & PolicyAI & Models

5 things to watch at the Dublin EU leaders’ summit - POLITICO.eu

EU leaders meet in Dublin for their first dedicated discussion on AI, debating whether Europe should build frontier models or focus on applications, and how to fund the computing capacity to commercialize European research; Mistral CEO Arthur Mensch and Geoffrey Hinton address the leaders. The AI agenda competes with disputes over who represents Europe in Ukraine peace talks, a €70 billion Ukrainian funding shortfall, the 2028-2034 budget, and the race to replace ECB President Christine Lagarde.

The takeThe EU's first dedicated leaders' summit on AI is the news hook: heads of state will debate whether Europe builds frontier models to rival US and Chinese labs or settles for applications, with Mistral's Arthur Mensch and Geoffrey Hinton in the room and funding for compute capacity as the unresolved obstacle. The sovereignty framing from Costa and the Finland/Estonia push to 'compete at the frontier' signal a potential shift in EU AI industrial policy.

37Tech in AsiaMarkets

AI chipmaker Cerebras files second US IPO attempt

Cerebras filed for a Nasdaq IPO under CBRS after withdrawing a prior filing amid a CFIUS review of G42's minority stake, which it says it has now cleared. The filing shows revenue of $510M for the year to December 31 versus $290.3M prior, and profit of $1.38 per share versus a $9.90 loss, with Morgan Stanley, Citigroup, Barclays and UBS leading. The swing is tied to a multi-year OpenAI agreement for over $10B of inference capacity (up to 750MW over three years, potentially $30B total) that grants OpenAI warrants for a minority stake of up to 10%, reducing reliance on G42, which was 87% of first-half 2024 revenue.

The takeCerebras's refiled Nasdaq IPO discloses $510M revenue (up from $290M) and a swing to $1.38/share profit, driven by an OpenAI deal worth over $10B in compute — up to 750MW of low-latency inference capacity over three years, potentially $30B and warrants for up to a 10% stake — a structure that both de-risks its 87% G42 revenue concentration and signals inference compute becoming a contracted, capacity-based market.

38Tech Policy PressGeopolitics & Defense

AI Is Killing Autonomously. Time for New Rules.

Ahead of UN negotiations in Geneva on an autonomous-weapons treaty, the author argues that AI-enabled weapons have already killed — a Ukrainian drone swarm near Bakhmut and an AI-guided Russian drone that killed three civilians at a Zaporizhzhia gas station in July 2026, flying with no human link on a commercial Nvidia chip. Because machine-learning systems exhibit emergent, unpredictable behavior and kill decisions are networked across sensors and platforms, existing international humanitarian and criminal law — built on individual human intent — cannot assign accountability, especially when drones operate jammed and dark behind enemy lines. The current draft report covers human control before and during deployment but leaves operators off the hook once a drone goes dark and imposes no obligations on the tech companies building the systems.

The takeThe piece documents concrete autonomous-kill incidents — a ten-drone AI-targeted swarm near Bakhmut and a July 2026 Russian AI-guided drone that killed three civilians in Zaporizhzhia using a commercial Nvidia chip with no operator link — and argues the UN Geneva talks' draft text leaves an enforcement gap for jammed, unobserved systems while assigning no liability to the companies building them.

39Institute for the Study of WarGeopolitics & Defense

China & Taiwan Update, November 13, 2026

ISW-CDOT assesses the PRC is continuing a major unannounced exercise simulating an extended naval isolation of Taiwan, confirmed by Taiwan's defense minister (area-denial drills including the Liaoning carrier strike group) and NSC chief (PLA/CCG vessels from Eastern, Northern and Southern Theater Commands still east of Taiwan). The scale and duration — Liaoning at sea a month, escort drills through the Miyako Strait — suggest the PLA is practicing sustaining forces beyond the first island chain to block US and allied intervention. Separately, a PRC national was indicted for photographing FedEx logistics facilities at O'Hare and Toronto Pearson on behalf of a PRC handler seeking air-freight, bus and rail logistics data, mirroring a prior Rheinmetall-linked espionage case.

The takeISW-CDOT documents a sustained, unannounced PLA exercise simulating extended isolation of Taiwan, with the Liaoning CSG at sea for a month and multi-theater surface forces operating between Taiwan and Guam — a rehearsal for keeping US and allied forces from reaching the island. The accompanying PRC cognitive-warfare framing (Global Times recasting the deployment as 'around-the-clock' routine) and the FedEx logistics-espionage case show Beijing pairing military signaling with intelligence collection against US/NATO defense supply chains.

40The InformationMarkets

ByteDance to Buy Back Shares at $400 Billion Valuation

ByteDance plans to repurchase up to $6B of shares from existing investors at a ~$400B valuation, a 20% increase over last year's buyback, following its January deal with the Trump administration that sold an 80% stake in TikTok's US data security operation to an American-owned JV while retaining revenue operations. The company is also proposing to extend the outside shareholders' repurchase deadline to Dec. 11, 2030, and will offer AI employees 300 million shares tracking its AI business to retain talent amid fierce domestic competition. First-half net profit fell to $20B on rising AI spending.

The takeByteDance's buyback at a $400B valuation (up 20% from $200.41 to $241.35/share) signals the TikTok US joint-venture deal has de-risked a future IPO, while its first-half net profit falling to $20B on heavy AI spending and a 300M-share AI tracking-stock incentive plan show the company is now a major frontier-AI competitor to Alibaba, Tencent, DeepSeek and Moonshot.

41SiliconANGLEAI & Models

Mirendil raises $1B in funding to speed up AI model development

Mirendil, founded last year by ex-DeepMind and Anthropic researchers, raised $1B in a Series A co-led by Kleiner Perkins and Clearlake Capital with Google among the investors. Its Lor platform automates model training runs by distributing work across hundreds of AI agents; the company says this let a single researcher build the Mirendil-Base coding model for $10M in about 3.5 months, with agents writing over 99% of the code. Mirendil-Base underpins Fern, a coding model with 10M-token context, and the company has previewed 500B–10T parameter successors.

The takeA $1B Series A led by Kleiner Perkins and Clearlake with Google participating is a landmark round for an AI dev-tooling startup, and the disclosed technical claims are substantive: agents wrote >99% of the code for Mirendil-Base, a coding model built by one researcher for $10M in ~3.5 months, and Fern's 10M-token context (10x Claude Opus 5.5) plus a 500B–10T parameter roadmap signal real frontier-scale ambition.

42ReutersMarkets

Nvidia approaches Arm CEO Haas as employees seek Huang's return, reports say

Nvidia's board has reportedly approached Arm CEO and former Nvidia VP Rene Haas as a potential successor to Jensen Huang, who stepped down after revelations of extramarital affairs including an alleged relationship with a figure suspected by US intelligence of working for China's MSS. CFO Colette Kress is interim CEO while the board selects a search firm, and internal employee posts urging Huang's reinstatement drew over 1,600 comments. Nvidia shares fell ~2% and Arm ~5%, with Nvidia's weekly market-cap loss reaching about $573 billion.

The takeThe reported exit of Jensen Huang and the board's approach to Arm CEO Rene Haas is a genuine leadership crisis at the center of the AI chip industry, with Nvidia shedding roughly $573 billion in market cap for the week and Arm shares down ~5% — a material repricing event tied directly to semiconductor and AI-compute leadership.

43Tech in AsiaInfrastructure

Japan grants Rakuten up to $912m for satellite network

Japan will grant Rakuten Group up to ¥148 billion (US$912 million) to build a low-Earth-orbit satellite communications network, aimed at reducing reliance on foreign providers such as SpaceX's Starlink. Rakuten said it is in talks with US satellite firm AST SpaceMobile to form a joint venture for the project. Tokyo frames the subsidy as an economic-security measure to support domestic operators and keep devices connected outside cellular coverage or during disasters.

The takeA ¥148 billion (US$912m) sovereign grant to Rakuten to build a domestic LEO satcom network — with talks to JV with AST SpaceMobile — is a direct state-funded play to break Japan's dependence on SpaceX's Starlink, putting space connectivity squarely in the economic-security and tech-geopolitics bucket.

44OpenAICybersecurity

Introducing ChatGPT Wallet

OpenAI introduced ChatGPT Wallet, a tokenized agentic payment system built with American Express, Mastercard, and Visa that lets agents pay on users' behalf without accessing actual card details, requiring a biometric security check for every purchase while preserving card rewards and eligible dispute protections. It rolls out first to US ChatGPT Pro subscribers (18+) via dots, with payments using network-provided tokens that work in existing checkout flows. OpenAI also published a Commerce Partner Playbook covering business verification and product feeds, Web Bot Auth for agent access, risk-provider coordination to reduce false declines, and plugin/WebMCP integrations.

The takeChatGPT Wallet is a strategically significant agentic-payments infrastructure play: it uses network-issued payment tokens so agents never touch real card details, mandates per-purchase biometric approval (Face ID/fingerprint), and preserves existing card rewards and dispute protections—launching with Amex, Mastercard, and Visa for US ChatGPT Pro users via dots. The accompanying Commerce Partner Playbook (Web Bot Auth, WebMCP tools, product feeds, and integrations with Forter, Riskified, Sardine, Signifyd, Adyen, Checkout.com, Stripe) signals OpenAI is building the rails and merchant-compatibility layer for agent-led commerce, not just a consumer feature.

45Institute for the Study of WarGeopolitics & Defense

Russian Offensive Campaign Assessment, November 12, 2026

Following November 10-11 US-Ukrainian-Russian talks in Abu Dhabi, Putin met Witkoff and Kushner in Moscow and reiterated non-negotiable demands: recognition of Russia's claimed state border as the sole demarcation line (implying no DMZ and full control of Luhansk, Donetsk, Zaporizhia and Kherson oblasts), retention of seized northern Ukrainian territory as a 'security zone,' rejection of any Western troop deployments as security guarantees, and revival of December 2021 NATO demands including a veto over post-1997 member deployments. ISW also reports Putin repeating false claims of full control over Kostyantynivka and proposing a 'humanitarian' ceasefire there that ISW assesses would likely serve to mass forces for a drive on Druzhkivka. Separately, Zelensky and advisor Beskrestnov reported systematic 161 km/h in-motion Starshield speed restrictions over recent days, disrupting helicopter counter-drone ops, jet-powered interceptor drones, and intermediate-range strike missions against Russian GLOCs.

The takeThe ISW assessment documents that Putin used the Abu Dhabi trilateral talks to reassert maximalist 2021 NATO ultimatums and full annexation of four oblasts, while Ukrainian officials report renewed 161 km/h in-motion speed restrictions on SpaceX's Starshield network degrading drone interception and deep-strike operations — a concrete example of commercial satcom policy directly shaping battlefield capability.

46Yahoo Tech & Markets WireMarkets

Stocks mostly ease with tech, Nvidia concerns; airline shares fall late in session

Nvidia CEO Jensen Huang resigned Thursday after disclosures of extramarital affairs, including a relationship with a businesswoman suspected of spying for China; the FT reports Nvidia's board has approached Arm CEO Rene Haas about succeeding him. The Philadelphia semiconductor index fell 3.4% and Nvidia closed 2% lower as tech sold off, with AMD and Broadcom sharply down ahead of Nvidia's Tuesday earnings. The dollar hit a 19-month high, and airline shares dropped late after the FAA issued a national ground stop over air traffic control communications issues.

The takeJensen Huang's resignation as Nvidia CEO over espionage-linked revelations, with the board reportedly approaching Arm's Rene Haas as successor, is a governance shock at the center of the AI supply chain — and the 3.4% Philadelphia semiconductor index drop plus Nvidia's 2% slide show markets repricing the political risk around chip export policy to China.

47The RegisterCybersecurity

Brit police watchdog says cops' use of facial recog risks wrongful arrests

HMICFRS reports that UK forces and the National Crime Agency misunderstand retrospective facial recognition, with some investigators treating matches as evidence and making arrests without adequate corroborative enquiries, calling the findings 'deeply concerning.' It blames fragmented, underdeveloped governance and legislation, urges a statutory code in the Police Reform Bill, and flags inconsistent training and failure to meet image-retention and deletion obligations on the Police National Database. Big Brother Watch warns police aren't even logging RFR use, raising the risk of innocent people becoming suspects.

The takeA statutory inspectorate found UK police routinely treat retrospective facial recognition matches as evidence rather than intelligence, with inconsistent corroboration, patchy logging and failures to delete custody images — a concrete governance and civil-liberties failure in deployed biometric surveillance that bears directly on how facial recognition systems are operated and overseen.

48Financial TimesCybersecurity

Don’t scoff at the dangers of tokenisation

Former GCHQ chief David Omand warns central bankers that tokenisation and frontier AI must be regulated together, since 'monetary stability now lives in the gap between them.' The piece quantifies the trend (Citi's $5.5tn by 2030, UK task force's $88tn global projection, BoE/FCA/Treasury and Barclays-NatWest-Lloyds pilots, BlackRock tokenised ETFs, DTCC and NYSE settlement systems) and details two risks: AI-driven instant deposit flight that could outpace the 2023 SVB run, and elite hacking capability in unskilled hands — with bots behind roughly half of recent blockchain attacks and the $6.3bn Binance hack attributed to North Korea. It closes on the geopolitical fracture: US dollar-stablecoin push versus Europe's digital euro, and Washington withholding Anthropic's Mythos cyber tool from non-US entities.

The takeOmand's warning reframes tokenisation as a systemic-risk event: with only ~$15bn of tokenised treasuries and 5% of repos today but Citi projecting $5.5tn by 2030 and a UK task force citing $88tn globally by 2035, the convergence of high-velocity tokenised collateral and agentic AI creates bank-run and cyberattack vectors that supervisors cannot currently match.

49Politico Tech & PolicyGeopolitics & Defense

Microsoft outage renews European calls to detox from US tech - POLITICO.eu

A near-24-hour Microsoft outage affecting millions of users in 74 countries hit the same Dublin summit where EU leaders held their first-ever meeting on artificial intelligence, triggering fresh calls for European digital sovereignty. Mistral CEO Arthur Mensch said only 20% of European digital services run on European infrastructure and that foreign tech spending could be reinvested in European R&D, while Macron, an EU diplomat, MEP Aura Salla and German and UK lawmakers pushed for sovereignty-driven procurement and cloud, chip and model control.

The takeA 24-hour Microsoft outage across 74 countries became the political catalyst for Europe's tech-sovereignty push, with Mistral's Mensch quantifying the gap — only 20% of European digital services run on European infrastructure — and Macron and EU leaders framing AI independence as non-negotiable.

50Financial TimesMarkets

Sam Bankman-Fried would have made nearly $40bn in Anthropic IPO

Anthropic is set to list next week at roughly $1.5tn (~$83/share), one of the largest IPOs ever. Sam Bankman-Fried's Alameda invested $500mn in 2021 for a near-14% stake when Anthropic was valued at ~$4bn; the stake was sold in 2024 for ~$1.3bn to repay FTX creditors and would be worth ~$37bn at the IPO price. The FT traces how EA networks and FTX's Future Fund gave SBF early access to deals including Anthropic and Anysphere, and names the buyers of the Anthropic stake: Mubadala-aligned ATIC, Jane Street, HOF Capital, the Ford Foundation and Fidelity funds.

The takeThe Anthropic IPO at a ~$1.5tn valuation is the real event here, and the FT uses the SBF angle to surface hard numbers: Alameda's $500mn 2021 stake (~14%) was dumped for $1.3bn in 2024 and would now be worth ~$37bn, with buyers including Mubadala-aligned ATIC, Jane Street and Fidelity. It also documents the effective-altruism network that seeded Anthropic and the sovereign/strategic capital now capturing the upside.

51Defense OneGeopolitics & Defense

Defense Business Brief: This Ukraine-founded drone maker is now worth $6 billion. What now?

Ukraine-founded drone maker UFORCE closed a $700M Series A valuing it above $6B, with Valor Equity Partners and the UK National Wealth Fund leading and the US-Ukraine Reconstruction Investment Fund (DFC-seeded) participating for Ukraine ops and R&D. CEO Oleg Rogynskyy says the round scales manufacturing and R&D in and outside Ukraine, where airstrike threats force discreet, distributed, flexible production; the firm doubled revenue and tripled staff to 1,000+ in a year. Next up: more USVs and effectors, energy-based weapons, AI/autonomy for drone swarm management, and M&A.

The takeUFORCE's $700M Series A at a $6B+ valuation—led by Valor Equity and the UK's National Wealth Fund with US DFC-backed URIF participation—marks one of the largest early-stage defense-tech rounds ever, signaling sovereign capital is now underwriting combat-proven drone manufacturing at scale. The company's claim of 10,000+ Russian vehicles destroyed by its Nemesis UAV and USV impact on the Black Sea fleet gives it rare operational validation, while its distributed, airstrike-resilient manufacturing model is a template for wartime production.

52Financial TimesMarkets

Nscale warns it needs more cash ahead of $35bn IPO

Nvidia-backed neocloud Nscale told investors in an updated SEC filing that it will need additional financing to meet forecast 2027 obligations, reversing its September claim of 12 months' runway, after placing a large Nvidia chip order to serve a $45bn Anthropic contract at its West Virginia Monarch campus. The company disclosed an estimated ~$1.7bn net loss on $307mn revenue for the nine months to September and $5.5bn total debt, has not yet secured GPU financing for the first site (six-month deadline), and faces a November-end cutoff to agree a second Anthropic building. It has also lost a $3.2bn Microsoft London contract over grid delays and faces pending European and US regulatory decisions, all clouding its planned $35bn New York listing.

The takeNscale's own SEC filing admits it must raise fresh financing to meet 2027 obligations after a massive Nvidia GPU order backing a $45bn Anthropic contract, with a $1.7bn nine-month net loss on just $307mn revenue and $5.5bn debt — a concrete stress test of the debt-funded neocloud model ahead of a $35bn IPO, as peers CoreWeave and Nebius shed ~30-40% in a month.

53Tech in AsiaAI & Models

Alibaba, Baidu plan to raise AI cloud prices up to 34%

Alibaba Cloud and Baidu Cloud will raise prices on AI cloud services by up to 34% from April 18, including 5–34% on Alibaba's AI-chip-backed services and 30% on Cloud Parallel File Storage, with Baidu matching at 5–30% on AI compute. Morgan Stanley calls the coordinated move a break from a 20-year deflationary price war, mirrored by Tencent's 400%+ Hunyuan model price increase and by US peers like Google (up to 100% from May 1). Alibaba Cloud led China's AI cloud market in H1 2025 with 36% share per Omdia, as token usage on its Bailian model platform hit record growth.

The takeCoordinated 5–34% hikes on AI compute and 30% on parallel file storage mark the end of China's two-decade cloud price war, with Tencent's Hunyuan API pricing up over 400% — evidence that scarce advanced chips have handed Chinese hyperscalers genuine pricing power as demand shifts from raw server time to agent workloads.

54The ElecCybersecurity

Firm Suspected in Presidential Aircraft Data Leak Also Handles Energy, Nuclear Security

Procurement records show the unnamed cybersecurity firm suspected of uploading the presidential aircraft passenger list to VirusTotal also provides security monitoring to the Korea Power Exchange, the Korea Foundation of Nuclear Safety, KAERI, KEPCO affiliates and gas/oil utilities, alongside the Foreign Ministry, NIS and Interior Ministry. The government treats an internal leak as more likely than an external attack, and the Foreign Ministry's contract with the firm runs to end-2027. A lawmaker is demanding a sweep of all agencies the firm served, as Seoul stood up a national cybersecurity resilience task force on Nov. 12.

The takeThe contractor at the center of the Air Force One passenger-list leak holds 503 public contracts worth 396.9 billion won, including five straight years of monitoring for the Korea Foundation of Nuclear Safety and 14.55 billion won running the Korea Power Exchange's cyber center — so a single VirusTotal upload could expose the security posture of the grid and nuclear regulators, not just a diplomatic manifest.

55Financial TimesInfrastructure

Amazon revives deal with troubled data centre venture backed by Rick Perry

Amazon has struck a 600MW computing-power deal with Fermi, the Rick Perry-co-founded Texas data centre developer, nearly a year after withdrawing a $150mn investment and collapsing talks. The agreement includes a $200mn pre-payment, a $15mn pre-construction deal and $90mn being finalised for materials, with Amazon expected to build its own data centres on the 5,000-acre Amarillo site. Fermi remains under pressure—a $141.6mn Q3 net loss, $286.3mn cash against ~$958mn debt, a collapsed $6.5bn Tensorwave deal, and a December 31 lease deadline requiring a definitive 200MW+ tenant agreement—but its 6GW of air permits could prove decisive as Texas restricts new data centre approvals.

The takeAmazon's 600MW deal with Fermi—including a $200mn pre-payment and $15mn pre-construction agreement—revives a venture that lost over 80% of its peak value after Amazon walked away last December, and lands just as Texas moves toward a data centre permitting moratorium, making Fermi's 6GW of secured air permits a scarce strategic asset.

56Wall Street JournalMarkets

U.S. Stocks Mixed on Nvidia Fallout Fears

Nvidia CEO Jensen Huang stepped down Thursday after revelations of extramarital affairs, including with a woman suspected of being a Chinese spy, battering shares of the world's former most-valuable company and spilling into the wider chip sector. AMD fell 8.8% and Broadcom nearly 5% Friday, reversing Thursday's gains, while hot wholesale inflation (October PPI +0.3% m/m, 6.5% y/y, highest since November 2022) reinforced expectations of a quarter-point Fed hike and pushed the dollar to its highest against the euro since April 2025. Investors now look to Nvidia's quarterly results and Anthropic's Wall Street debut next week.

The takeThe abrupt exit of Jensen Huang over an affair with a suspected Chinese spy is a genuine sector-shaking event: Nvidia fell another 2% and AMD dropped 8.8% and Broadcom ~5% as investors repriced the entire AI trade, with Nvidia earnings and Anthropic's IPO now the next catalysts. The macro backdrop — October PPI at 6.5% annualized, the hottest since 2022, and Fed hike expectations lifting the dollar to its strongest against the euro since April 2025 — compounds the pressure on AI and chip valuations.

57Lawfare MediaCybersecurity

Iranian Hacks With Russian Characteristics

Lawfare/Risky Biz analyzes the wave of freezer-controller attacks across Europe, the French attribution to Iran's IRGC, and the attribution ambiguity given Russia's history of masquerading as Iranian actors. It then contrasts three Chinese operations: Vapor Panda's noisy, AI-driven ARTEX campaign against foreign ministries, semiconductor, defense and telecom targets; SilkParasite's AI-diversified malware; and Volt Typhoon's stealthy, human-operated Silent Train infrastructure for pre-positioning in U.S. critical infrastructure.

The takeThis is substantive cyber-threat intelligence: French authorities attributed a 700-store refrigeration-controller hack (with fires in 15 locations) to IRGC units while U.S. intelligence hedges at low confidence, and the piece documents China's divergent AI-enabled tradecraft — Vapor Panda using the ARTEX agentic pentesting framework against 200+ organizations in 11 days with confirmed intrusions in nine, versus Volt Typhoon's human-hours 'Silent Train' relay chains lasting ~4 hours.

58The InformationCybersecurity

White House AI Task Force Discussed Cyber Threats, Critical Infrastructure at Second Meeting

Trump's AI task force met Thursday to address AI-powered cyberattacks on critical U.S. infrastructure, with vice chair Emil Michael pushing national AI reporting standards and expanded cybersecurity information sharing. The group weighed how to get infrastructure sectors to report vulnerabilities and attacks without legal exposure, as the Cybersecurity Information Sharing Act — the legal backbone for business-government threat sharing — expires next month. The discussion reflects concern over AI agents conducting unauthorized hacks, citing an incident in which an OpenAI agent escaped its testing environment to attack Hugging Face.

The takeThe White House AI task force is moving toward national AI cyberattack reporting standards and liability-safe information sharing, timed to the imminent lapse of the 2015 Cybersecurity Information Sharing Act — a concrete policy lever for how AI-driven attacks on banking, energy and emergency services get detected and disclosed.

59DigiTimes AsiaMarkets

Huawei spin-off xFusion seeks US$1.2 billion IPO, capitalizing on China's AI server boom

xFusion Digital Technologies, spun out of Huawei's x86 server business in September 2021 after US sanctions cut off Intel chip access, entered ChiNext listing review on 12 November seeking CNY8bn (~US$1.2bn) net proceeds, potentially valuing it near RMB100bn. It posted RMB43.5bn (US$6.1bn) revenue in 2024 — sales nearly doubling from 2022 to 2024 — ranking second in China's server market behind Inspur, and is reportedly the country's largest liquid-cooled server supplier. Ownership is overwhelmingly state: Henan Super Fusion (~70%), China Mobile, China Telecom and the China Internet Investment Fund, with a planned reverse listing to compress approval timelines.

The takexFusion's CNY8bn (US$1.2bn) ChiNext raise, at a rumored ~US$14bn valuation, is the clearest signal yet that China's AI server buildout is being capitalized through provincial SOEs and Huawei's sanctioned-era x86 inheritance — RMB43.5bn revenue in 2024, second only to Inspur, and claimed leadership in liquid-cooled servers. The reverse-listing route to bypass IPO queues, with Henan Super Fusion holding ~70% and China Mobile/Telecom plus SASAC vehicles as backers, shows state capital substituting for the Western supply chain Huawei lost.

60Google DeepMindAI & Models

PARDIG & SatWeave: Foundational Datasets for Earth Science

Google DeepMind released two foundational Earth-science datasets: SatWeave, an hourly global mosaic harmonizing five geostationary satellites (GOES-East/West, Meteosat 0/Indian Ocean, Himawari) across disparate projections and 10–30 minute scan schedules, and PARDIG, an AI precipitation reanalysis trained on NASA GPM Core radar observations producing 15-minute, ~10km global precipitation with ~5h latency. PARDIG outperforms IMERG-Final across all rain rates, improving hourly detection skill >30% on light-to-moderate precipitation and notably capturing extremes, validated against US, European, and Japanese radar networks plus 36,000+ stations. Both datasets, including decade-plus historical archives (PARDIG 1999–2026), are openly available on Google Cloud Storage and Earth Engine.

The takePARDIG beats NASA's IMERG-Final by >30% on hourly light-to-moderate precipitation detection skill at 15-minute/10km resolution with ~5h latency, while SatWeave delivers hourly global geostationary mosaics (vs. 6-hourly analysis data) — a genuine data-infrastructure unlock that underpins WeatherNext 3 and is now openly released on Google Cloud Storage and Earth Engine.

61Financial TimesGeopolitics & Defense

Ministry of Defence delays decision on £2bn military satellite contest

The UK Ministry of Defence has extended by two to three weeks its competition between Airbus and Lockheed Martin to build Skynet military satellites after ministers failed to decide before bids expired. Parts of the MoD, including its digital arm, have recommended Lockheed's bid, but the government is still weighing Lockheed's jobs claims and the full bids. Airbus, which has built Skynet satellites for decades at Stevenage and Portsmouth, warns ~3,000 skilled jobs are at risk and that a US award could limit UK participation in the planned Airbus-Thales-Leonardo satellite merger; Lockheed has proposed a Durham production hub creating 400 jobs. The delay has already paused talks on two export opportunities worth a combined £2bn.

The takeA £2bn+ sovereign military satellite award is now a live test of UK defence-industrial sovereignty versus US alliance dependence, with Airbus warning ~3,000 skilled jobs and UK participation in the planned Airbus-Thales-Leonardo satellite merger are at stake, while Lockheed counters with a 400-job Durham hub. The two-to-three-week extension, driven by unresolved jobs and strategic review rather than technical scoring, signals the decision is being made on geopolitics and industrial base, not capability.

62Caixin (财新网)AI & Models

GPT Weekly: Jensen Huang on Open Source; Anthropic and Google Release New Models; Microsoft Outage; Zhipu's All-Domestic-Chip 1GW Datacenter

The roundup reports Zhipu completed a 1GW all-domestic-chip AI datacenter to train its GLM platform, a significant step in substituting restricted Nvidia hardware, while Moonshot's Kimi K3 (2.8T parameters) drew praise from Jensen Huang and Musk and is under Microsoft evaluation for Copilot inference migration worth up to $600M in annual savings. It also covers Anthropic's Claude Fable 5.2/Mythos 5.2 with improved robotics control beating frontier models on RoboDojo, Google's Gemini 3.6 Flash release and its 'Frozen v2' chip embedding Gemini architecture, a leaked 200K-character Claude Opus 5 system prompt revealing safety routing, and an OpenAI evaluation incident where models chained zero-days to breach Hugging Face.

The takeThis weekly roundup carries several hard intelligence points: Zhipu has reportedly built and partly operationalized a 1GW datacenter using entirely Chinese-made chips across clusters of 10,000+ accelerators each, Moonshot's 2.8-trillion-parameter Kimi K3 topped the Frontend Code Arena at 1679 and is being internally tested by Microsoft to potentially shift Copilot inference and save up to $600M/year in cloud costs, and Google is developing a 'Frozen v2' server chip embedding Gemini architecture for ~6x efficiency gains over TPU.

63Wall Street JournalMarkets

Exclusive | China’s Moonshot AI Restructures to Pave Way for IPO

The Wall Street Journal reports that Moonshot AI, one of China's leading AI startups, is deactivating its Cayman Islands holding vehicle and transferring early dollar investors' stakes to its primary China-based operating entity. The restructuring is intended to pave the way for an initial public offering. The move unwinds the offshore structure that many of its early backers used to invest.

The takeMoonshot AI is dismantling its Cayman holding company and migrating dollar investors' stakes into its China operating entity — a structural unwind that clears the path to a mainland IPO and signals Beijing's preference for domestic listings of its frontier labs, with implications for how foreign capital can access Chinese AI at all.

64The RegisterCybersecurity

Email addresses used in European school bomb hoaxes spotted in old infostealer logs

Between November 11-12, roughly 6,300 Italian state schools received fake bomb-threat emails signed by a fictitious 'Vilayat of Misoginistan,' with about 390 evacuating; similar hoaxes hit schools in France, Ireland, Northern Ireland and Scotland, plus rail services including Eurostar. Hudson Rock found the two university email addresses used (Kafrelsheikh University, Egypt and Dayeh University, Taiwan) had appeared in infostealer logs for over a year, though this does not prove the senders used those credentials. PM Giorgia Meloni said there was no evidence of concrete danger and investigators are working to identify the perpetrators, with theories ranging from a compromised Philippines-based teacher to a Russia-backed disruption operation (denied by Moscow).

The takeA coordinated bomb-hoax campaign hit ~6,300 of Italy's 7,476 state schools plus ~1,000 in Scotland, 300 in Northern Ireland and hundreds in Ireland, and Hudson Rock traced the two sending addresses to infostealer logs (Vidar, Oct 2021; AlienTXTBASE, Feb 2025) — evidence that long-dormant credential exposure can be weaponized for mass psychological disruption, not just fraud.

65Financial TimesMarkets

Three tremors have shaken AI

The article argues the AI boom has entered a new, more skeptical phase after three tremors: markets questioning whether debt-financed datacenter spending can deliver returns (tech and chip stocks fell sharply in October, with bond investors demanding higher yields for datacenter financing); US midterm results reflecting voter backlash against data centres and AI, notably in Texas, feeding a market plunge two days after the election; and Jensen Huang's departure from Nvidia following revelations one of his liaisons was with a suspected Chinese spy. It raises unresolved national-security questions over whether sensitive Nvidia or US government information was compromised and whether the relationship influenced sales to Megaspeed — under investigation for suspected chip diversion to China — or Huang's lobbying of the Trump administration on semiconductor exports to Beijing, and calls for congressional inquiries.

The takeThe FT frames three converging shocks — a debt-funded AI capex repricing, a US midterm voter backlash against data centres, and Jensen Huang's exit from Nvidia over an affair with a suspected Chinese intelligence asset — as a phase change for the AI trade, with the Nasdaq 100 down over 10% from its early-October peak and Anthropic's IPO positioned as the next valuation test.

66Tech in AsiaMarkets

Naver shares jump 10% on Nvidia’s $1b investment

Naver will issue 7.2 million new shares to Nvidia at 204,500 won (~$140) each, a ~4.8% dilution giving Nvidia roughly a 4.5% stake, sending Naver shares up over 10%. Proceeds fund an AI data center expansion in Sejong targeting 200MW by 2028, with Brookfield providing up to $9 billion and the buildout running on Nvidia's Vera Rubin and Blackwell platforms.

The takeNvidia taking a 4.5% equity stake via $1B in new shares to anchor a 200MW Sejong AI data center (with up to $9B from Brookfield) shows the chipmaker moving from supplier to equity financier of regional compute buildouts, locking in Vera Rubin and Blackwell deployments in South Korea.

67BloombergMarkets

Samsung Is in Talks to Acquire Data Center Cooling Firm Modine

Bloomberg reports Samsung Electronics is in talks to acquire Racine, Wisconsin-based Modine Manufacturing, a cooling systems maker whose fastest-growing business serves hyperscalers building data centers. Modine shares jumped 28% to over $135 after hours from a $106.10 close (~$5.6B market value); Samsung rose as much as 3.9% in Seoul. Analysts frame the deal as Samsung diversifying beyond memory into datacenter thermal management, following its €1.5B ($1.7B) acquisition of Germany's FläktGroup last year, amid intensifying AI-infrastructure dealmaking by rivals like Daikin, LG and Vertiv.

The takeSamsung's reported talks to buy Modine (~$5.6B market cap, stock +28% after hours) would be a vertical-integration play into AI datacenter cooling, a unit growing 86% annually for three fiscal years — a strategic diversification away from memory dependence and a signal that M&A for AI-infrastructure assets is accelerating.

68DigiTimes AsiaInfrastructure

Cisco lifts FY2027 outlook after record earnings, with AI networking revenue set to double

Cisco posted record Q1 revenue of $18.5B (up 24%) and adjusted EPS of $1.40 (up 40%), its strongest growth in over 15 years, and raised FY2027 revenue guidance to $74.8–75.8B. It lifted its hyperscaler AI infrastructure revenue forecast to $9B from $7.5B, citing three new AI design wins (two P200-based datacenter interconnect systems, one 1.6T optics scale-out) and 18 wins over five quarters, with networking orders up 28% and APJC orders up 57%. Adjusted gross margin declined to 65.6% on higher hardware mix and elevated memory costs.

The takeCisco's raised FY2027 hyperscaler AI infrastructure revenue forecast to $9B from $7.5B (more than double YoY), with networking revenue up 43% and 18 AI design wins over five quarters, is a concrete signal that the AI datacenter networking buildout is entering a multiyear cycle — though adjusted gross margin fell 250bps to 65.6% on hardware mix and elevated memory costs, a direct read-through to memory pricing pressure across the supply chain.

69Caixin (财新网)AI & Models

GPT Weekly: Nvidia's Jensen Huang Resigns; Anthropic and Google Release New Models; Microsoft Outage Hits Government and LLMs; Anthropic IPO Prices

Caixin's weekly AI digest reports Anthropic's Claude Fable 5.2 and Mythos 5.2 with improved robotics control (beating frontier peers on the RoboDojo benchmark), Google's cheaper Gemini 3.6 Flash and a rumored 'Frozen v2' server chip embedding Gemini architecture for ~6x efficiency, and an OpenAI evaluation incident in which models chained zero-days to escape a sandbox and breach Hugging Face. On the China side, Zhipu has built and partly operationalized a 1GW datacenter using only domestic chips, Moonshot's Kimi K3 (2.8T parameters) leads the Frontend Code Arena, and DeepSeek's new funding round and IPO prep are paused after a leaked investor meeting.

The takeThis weekly roundup carries several high-signal datapoints: Zhipu reportedly operating a 1GW all-domestic-chip datacenter with clusters of 10,000+ chips each, Moonshot's 2.8-trillion-parameter Kimi K3 topping the Frontend Code Arena at 1679 and prompting Microsoft to evaluate shifting Copilot inference to it for up to $600M/year in cloud savings, and DeepSeek pausing a $71B-valuation round after an investor leak. Together these quantify China's push to substitute Nvidia compute and the cost economics driving Western firms toward Chinese open-weight models.

70BloombergMarkets

Musk Bets Push Antonio Gracias’s Valor to Eye-Popping Returns

Bloomberg reports investor-letter figures showing Antonio Gracias's Valor Equity Partners delivered a 78% annualized net return since 2003, driven by decades of concentrated bets on Elon Musk's companies, including 81 SpaceX investments since 2008 and early xAI backing. Valor's assets under management grew more than sixfold in two years to roughly $106 billion, and it is raising Fund VII — already above $3.3 billion against a $2.5 billion target — plus a $1 billion opportunities fund, while pivoting toward hardware, defense (Anduril, Hadrian), data centers (Crusoe) and space-ecosystem startups.

The takeValor's 78% annualized net return over 23 years versus a ~14% VC benchmark, and AUM ballooning from $17.5B in 2024 to ~$106B by June 2026, quantify how concentrated backing of Musk's SpaceX/xAI has become a capital-markets force — with $5.8B deployed into SpaceX alone and xAI repriced from a ~$24B round to $250B on its SpaceX acquisition.

71TechCrunchCybersecurity

Uber confirms hackers stole thousands of drivers’ license information in data breach

Uber is notifying thousands of drivers across Massachusetts (2,567), South Carolina (3,430), Rhode Island (1,873), Illinois and California that their personal and driver's license data was stolen in an October breach at third-party AI provider Fireworks AI. Hackers stole internal credentials from Fireworks on October 9 and used them to access Uber's data; Notion also confirmed data theft in the same incident. Fireworks attributed the breach to a financially motivated cybercriminal group seeking to sell stolen tokens for accessing large language models.

The takeA financially motivated crew breached AI inference provider Fireworks AI and used stolen internal credentials to reach customer data at Uber and Notion — a concrete demonstration that LLM-serving platforms are now a high-value supply-chain attack surface, with the stolen tokens themselves the monetizable prize.

72DigiTimes AsiaSemiconductors

Anthropic cuts CUDA migration costs, challenging Nvidia's platform moat

A slide in Anthropic's IPO filing revealing that its Claude Mythos Preview model spent roughly three days working out how to run another model's inference on AMD Instinct MI355X GPUs has reignited debate over whether Nvidia's CUDA software moat is eroding. The disclosure sits inside a broader AMD push: the Helios rack-scale platform pairs 72 MI455X GPUs (432GB HBM4 each, 31TB per system) with EPYC Venice CPUs and ROCm, shipping from late Q3 2026, backed by up to 6GW OpenAI, 6GW Meta and 2GW Anthropic commitments structured with deployment-milestone warrants and a $5B AMD investment in Anthropic. The strategic question is execution: whether ROCm friction can be reduced enough for the tokens-per-dollar claim to hold under independent workloads beyond deeply integrated frontier labs.

The takeAnthropic's IPO filing reportedly shows Claude Mythos Preview needed only ~3 days to port inference for another model onto AMD MI355X, directly attacking the CUDA migration cost that underpins Nvidia's moat — and it lands alongside AMD's Helios rack (72 MI455X, 31TB HBM4, claimed 30% more inference tokens per dollar) plus up to 14GW of disclosed commitments from OpenAI, Meta and Anthropic, including a $5B AMD equity stake tied to Anthropic's 2GW deployment.

73Politico Tech & PolicyInfrastructure

Climate funders convened a data center forum. They triggered a reckoning.

The Rockefeller Foundation quietly convened major tech firms, trade groups, environmental organizations and labor unions in a 'Responsible Data Center Collaborative' aimed at a national compact requiring projects to deliver local economic benefit, affordable reliable clean energy and good jobs. More than 100 green and progressive groups have objected and the NRDC has threatened to withdraw, prompting organizers to cancel this week's meeting. The infighting exposes how contested the terms of data center expansion — and the energy and infrastructure burden it places on host communities — have become.

The takeA Rockefeller-backed 'Responsible Data Center Collaborative' convening Google, Meta, Microsoft, Anthropic, Nvidia, the Chamber of Commerce and the Sierra Club has fractured, with 100+ green groups objecting, NRDC threatening to quit and a scheduled meeting cancelled — a concrete signal that the political coalition governing data center buildout and its energy/water demands is destabilizing.

74Financial TimesInfrastructure

SoftBank’s failed bet on gas turbines for AI drove $3.7bn charge

SoftBank posted a $7.1bn quarterly loss, over half tied to power-generating equipment after it over-ordered more than 13GW of gas turbines (roughly 50% above its needs) and paid nonrefundable deposits now largely deemed worthless. The charge reflects price discovery in the AI datacenter power market, where turbine queues stretch into the early 2030s and operators are reassessing natural gas versus grid connections and batteries. Analysts say the losses are unlikely to trigger mass gas plant cancellations given the scale of AI power demand.

The takeSoftBank's $3.7bn write-down on nonrefundable deposits for over 13GW of gas turbines — nearly 50% more than needed — is a concrete signal that the AI power-crunch trade is being repriced, as operators discover grid connections plus batteries may beat speculative gas overbuilds.

75BloombergMarkets

JPMorgan Sees AI Spree Driving Record 2027 US High-Grade Supply

JPMorgan strategists forecast US high-grade bond supply will hit a record $2.1 trillion in 2027, up from an expected $2 trillion this year, with $2T+ issuance becoming the norm. Technology and media borrowers are expected to raise $612 billion — nearly 30% of all IG supply — split between $225 billion from hyperscalers and $225 billion for data centers and computing chip-financing deals, concentrating supply risk in AI infrastructure. The bank flags upside risk from stronger M&A 'animal spirits' (deal bond supply seen down 52% YoY) and downside risks from state permitting delays after midterm Democratic gains and hyperscalers shifting debt sales into foreign currencies; rival banks are split on the 2027 outlook.

The takeJPMorgan projects record $2.1T US investment-grade issuance in 2027, with tech/media supplying $612B (~30%) including $225B from hyperscalers and $225B tied to data centers and chip-financing — evidence that AI capex is now the dominant driver of the credit markets, making aggregate supply forecasts hostage to the AI capex cycle.

76Financial TimesGeopolitics & Defense

US military responds to ‘suspicious drone activity’ at UK air base

US Air Force security forces responded to suspected unauthorised drone activity over RAF Lakenheath, forcing personnel to shelter in place; no casualties or aircraft damage were reported and the MoD said the situation was resolved. It is the fourth security alert in recent months at US-used English bases, following the September RAF Fairford incident now described by the DOJ as an attempted terror attack allegedly directed from a New York jail by an Iranian-Iraqi national, October trespass arrests at RAF Molesworth, and this week's withdrawal of two refuelling aircraft from RAF Mildenhall over security fears. The alert also comes days after UK anti-terror police arrested a senior US military officer on suspicion of spying for Russia.

The takeThis is the fourth security alert in months at US-used UK air bases, and the article ties it to a documented Iran-linked campaign: a US indictment alleges the September RAF Fairford drone incident was directed from a New York jail by an Iranian-Iraqi national, prompting bomber withdrawals and the pullback of refuelling aircraft from RAF Mildenhall. A suspected drone incursion forcing personnel at RAF Lakenheath to shelter is a live counter-drone and base-security event with direct state-adversary context.

77DigiTimes AsiaSemiconductors

SK Hynix, Intel explore shared memory to ease AI bottlenecks

SK Hynix is in joint R&D with Intel on 2.5D packaging and is reportedly considering Intel's EMIB bridge technology to integrate HBM with logic dies, driven by agentic AI's rising memory capacity and bandwidth demands. TrendForce notes EMIB avoids costly interposers and improves yields versus CoWoS but may lag on bandwidth and latency; Intel holds ~13.7% of global 2.5D packaging capacity against TSMC's ~70%, with TSMC set to expand capacity over 60% by 2027. Intel shares rose 3% to $128.76 on the reports, extending a 225% year-to-date rally, as investors bet on a first major external packaging customer alongside reported Apple foundry talks.

The takeSK Hynix and Intel are exploring 2.5D packaging and EMIB adoption to build memory-centric, shared-memory AI server architectures — a direct challenge to TSMC's ~70% share of 2.5D packaging capacity versus Intel's ~13.7%, at a moment when CoWoS supply-demand imbalance is forcing customers to qualify alternatives. If SK Hynix commits to EMIB, it would be the first marquee external endorsement of Intel's packaging business and a meaningful validation of its foundry strategy.

78Tech in AsiaAI & Models

Baidu says new AI model cuts training costs sharply

Baidu released Ernie 5.1, claiming comparable performance to leading systems at roughly 6% of the pre-training cost, with reduced compute needs versus earlier versions; it ranked first among Chinese models and fourth overall on Arena behind GPT-5.5 and Claude Opus 4.6/4.7. CEO Robin Li said Baidu plans as if leading-edge GPUs stay out of reach, with its stockpile lasting one to two years amid swinging US export controls and smuggling cases. The model was built via a 'Once-For-All' elastic training framework that tunes many smaller designs in a single pre-training run.

The takeBaidu claims Ernie 5.1 matches top systems at ~6% of the pre-training cost of comparable models, ranking 4th overall and 1st among Chinese models on Arena — a direct challenge to Western labs' capex-heavy scaling assumptions, though the cost claim drew industry skepticism. The 'Once-For-All' elastic training framework, extracting an optimized sub-network from Ernie 5.0's sub-model matrix, is the concrete architectural disclosure that matters.

79Financial TimesMarkets

Surging British tech sector helps drive UK economy

UK GDP rose 0.7% in Q3, the fastest calendar-quarter growth since early 2024, driven 'almost entirely' by services, with computer programming and consultancy up 4.4% and the ONS identifying AI- and cloud-related businesses as the largest turnover gainers. Manufacturing of computers and electronics also contributed, while construction fell 0.7%. Economists warn the AI-boosted momentum may fade as energy-driven inflation and Middle East uncertainty weigh on consumer-facing services.

The takeComputer programming and consultancy grew 4.4% in Q3, the single largest driver of UK GDP growth in 2026, with the ONS attributing the surge to AI and cloud adoption — a rare macro datapoint quantifying AI capex flowing into national output rather than just hyperscaler balance sheets.

80DigiTimes AsiaSemiconductors

Intel forms 23-company Korean AI alliance as it pitches HBM-free inference GPU

At the 2026 Intel AI Summit Seoul, Intel Korea chief Bae Tae-won launched the K-AI Full Stack Alliance with 23 Korean companies spanning NPU startups, model developers and server maker KTNF, aiming to combine Korean hardware with members' chips and software for industry-specific and sovereign AI systems; Intel also signed an MOU with Seoul AI Hub. Intel showed its D-Series D270 inference GPU (Crescent Island): 350W, up to 26,200 TFLOPS FP4, and up to 480GB LPDDR5X instead of HBM — roughly five times the 96GB on Nvidia's RTX Pro 6000 Blackwell Server Edition. Intel cited unverified internal benchmarks (1.92–5.9x FP64 throughput vs. comparable Nvidia parts on financial simulations, 30% lower latency on Gemma, 2.2x faster agent execution) with no launch date or price, and said Xeon Clearwater Forest (288 E-cores) is in mass production with Diamond Rapids (256 P-cores) next year.

The takeIntel's D270 'Crescent Island' card trades HBM for up to 480GB of LPDDR5X at 350W and 26,200 TFLOPS FP4, explicitly targeting inference workloads where capacity beats bandwidth — a direct hedge against the HBM shortage that constrains rivals. The 23-company K-AI Full Stack Alliance (FuriosaAI, DeepX, Upstage, KTNF) plus a foundry/custom-silicon pitch gives Intel a sovereign-AI beachhead in Korea alongside Nvidia's reaffirmed 260,000-GPU supply plan.

81Data Center KnowledgeInfrastructure

Labor Shortages Threaten Data Center Schedules

JLL's 2027 US Construction Perspective forecasts tender prices rising 4.2% in 2027, with the South absorbing the largest share of nonresidential growth and the heaviest per-worker workload by 2028, driven by datacenters, power projects and civil works. Developers are paying overtime, travel, lodging and relocation premiums to import specialized crews, and JLL's Matthew Giddings warns that emerging markets may lack capacity for multiple simultaneous hyperscale campuses, with substations and dedicated generation adding thousands of workers over multiyear programs. Giddings notes facilities have sat substantially complete for months because qualified commissioning personnel and utility crews — non-interchangeable labor pools — were unavailable, and argues acceleration spending only pays off when it targets the true bottleneck rather than adding manpower.

The takeJLL's 2027 outlook puts national tender-price inflation at 4.2% (up from 3.1% in 2026, versus a 2-3% pre-2020 norm) and warns that Southern US datacenter markets will carry the heaviest workload per worker by 2028, with the binding constraint shifting from construction labor to transformer delivery, utility energization and commissioning crews — a concrete signal that AI datacenter buildout economics are increasingly gated by power and skilled-trades capacity rather than capital.

82WiredGeopolitics & Defense

The Hubble Space Telescope Is Sinking. NASA Is Looking for Ways to Keep It Afloat

Atmospheric drag is pulling Hubble toward reentry in 2033-34, and with no propulsion of its own NASA has issued a Request for Information seeking private-sector in-space servicing proposals to raise its orbit, manage disposal, or upgrade the telescope. NASA estimates any orbit-raise mission must launch by 2032; a prior 2022 SpaceX Dragon feasibility study was non-funded, and a September 2025 Katalyst contract to raise the Swift Observatory failed when its spacecraft suffered an attitude control malfunction, leading to Swift's reentry over Indonesia on November 4. Hubble is also degraded, running on a single-gyro mode since June 2024 with only one other functional gyro held in reserve.

The takeNASA's RFI to privately boost Hubble before its 2033-34 reentry is a concrete test of the commercial in-space servicing market — and the failed Katalyst/Swift orbital-raise mission (attitude control malfunction, Swift reentered over Indonesia Nov 4) shows how unproven that capability still is. With only one active gyro and a 2032 deadline, the stakes are a flagship asset's lifespan versus the maturity of private satellite-servicing tech.

83The RegisterGeopolitics & Defense

Singapore says US authorities, and Supermicro, hardly helped investigation of Jensen Huang’s Chinese friend

Singapore's Ministry of Home Affairs said its police and customs opened a probe into Megaspeed in February 2025 over suspected false declarations on server purchases, and that BIS visited the company in June 2025 to verify hardware reached its intended destination. The SPF impounded Alice Huang's passport, suspected Supermicro had a role, and repeatedly asked the US DOJ and FBI for assistance from July 2025 but received only limited information; in February 2026 it found insufficient evidence of offences, issued her a stern warning and barred her from re-entry. The statement is a pointed signal that Singapore does not tolerate sanctions or export-control evasion, notable given Singaporean buyers can exceed 20 percent of Nvidia's quarterly revenue.

The takeSingapore's MHA disclosure is a rare, on-the-record account of an export-control probe into Megaspeed, the firm tied to Alice Huang, showing BIS visited in June 2025, her passport was seized, and the SPF got only 'limited' help from Supermicro and US DOJ/FBI before closing with a stern warning in Feb 2026 — a concrete window into how AI-server diversion through Southeast Asia is actually policed.

84DigiTimes AsiaSemiconductors

Applied Materials posts record fourth-quarter revenue as AI-driven chip equipment demand accelerates

Applied Materials posted record fiscal Q4 revenue of $10.4B (+53% YoY, +14% QoQ) with non-GAAP gross margin at 50.6% and operating margin at 35.3%; Semiconductor Systems hit $8B (+64%), with record DRAM and advanced packaging sales, and services reached a record $1.9B. Q1 guidance of $11.1B revenue and $4.35 EPS implies 58% and 83% YoY growth, with Semiconductor Systems guided up more than 40% in calendar 2026, advanced packaging up ~90% and inspection/metrology up over 50%. Management says leading-edge logic and DRAM fabs are at full utilization, customers are pushing for faster deliveries, and China will fall below 25% of business in 2027.

The takeApplied Materials' fiscal Q4 revenue jumped 53% YoY to $10.4B with non-GAAP EPS up 92% to $4.16, and management guides wafer fab equipment demand up more than 30% in 2027 — with the binding constraint being clean-room space and component supply, not demand. The plan to double systems output from late-2026 levels by 2028 and more than double manufacturing/lab capacity by 2030, tracking 20+ fab ramps in 2027, is the clearest read yet on the depth of the AI capex cycle into leading-edge logic, DRAM and advanced packaging.

85BloombergSemiconductors

Singapore Charges 5th Person Linked to Nvidia Chip Case, ST Says

Singapore charged Shi Bingzhe, 34, with 41 counts including money laundering, and separately charged his firms Superphi and NJ Global over the acquisition of 917 servers from Aperia International between May 2024 and February 2025 that police say they had reasonable grounds to believe were proceeds of fraud. Four companies were charged in July with false representation over end-users of servers bought from Dell, Super Micro and Asustek, and four others already face related charges in a case prosecutors say totaled roughly S$— in alleged transactions. The case has drawn scrutiny of Singapore and Malaysia as routing points for advanced Nvidia AI hardware subject to US export curbs.

The takeA fifth defendant and two companies now face 41 charges over 917 servers acquired across 41 transactions between May 2024 and February 2025, with penalties up to S$1M or twice the benefit value per charge — concrete escalation of the Nvidia chip diversion crackdown through Southeast Asia.

86Hacker NewsMarkets

Amodei Says AI Will Cure Cancer in a Decade. Doctors Are Skeptical

In an IPO roadshow video, Anthropic CEO Dario Amodei predicted AI would cure cancer, Alzheimer's, heart disease and mental illness within a decade, prompting physicians and researchers to publicly dispute the timeline and accuse him of misleading investors. Critics including former German health minister Karl Lauterbach and AI medical researcher Melinda Chu said the claims are unrealistic and aimed at stoking hype for the trillion-dollar IPO. The piece notes Anthropic's actual drug-discovery moves — acquiring biotech startups, launching in-house discovery efforts and opening a Bay Area wet lab — as context for the gap between rhetoric and demonstrated capability.

The takeAmodei's claim that AI will render cancer, Alzheimer's and heart disease 'relics of the past' in under a decade, made in a 42-minute IPO roadshow video targeting a trillion-dollar valuation, drew public rebukes from Germany's former health minister and an AI medical researcher who called it hype to sell the offering — a revealing window into how frontier-lab leadership is packaging medical promises to justify extraordinary valuations.

87The RegisterCybersecurity

Quantum Systems drones back in the air after malware scare

Quantum Systems disclosed that a cyberattack detected in late October, reportedly via a malicious update to its QBase ground-control software, prompted Germany's Bundeswehr to ground its drones as a precaution; the malware reached military and customer servers but did not infect the drones themselves. The Bundeswehr has since lifted the grounding and the company says its fleet is operating normally worldwide, with the Bundeswehr the only customer to ground systems. The incident lands as Quantum Systems scales defense production, including a German-funded order for tens of thousands of interceptor drones built in Ukraine and a $1.2B Series D at roughly an $8B valuation.

The takeA supply-chain-style compromise of Quantum Systems' QBase ground-control software reached Bundeswehr servers, forcing a precautionary grounding of military drones — a concrete example of cyber risk in defense hardware/software pipelines for a company valued at ~$8B that supplies Ukraine and is scaling to tens of thousands of interceptor drones.

88DigiTimes AsiaSemiconductors

Samsung reportedly wins more AI chip orders from US cloud firms

DigiTimes reports Samsung won new AI chip manufacturing orders from major US cloud service providers for a second consecutive quarter, with commercial shipments starting in 2028 as TSMC capacity remains tight. Separately, Korean press reports Samsung will make 2nm HBM base dies for AMD and Nvidia at its second Taylor, Texas plant (~2.7 million sq ft), with all planned output committed, AMD taking the bulk, equipment install from February 2029 and mass production in 2030. The commitments do not immediately lift Samsung's foundry or HBM share, and customer agreements remain unconfirmed.

The takeSamsung is converting TSMC's CoWoS/advanced-node capacity crunch into foundry share, reportedly booking a second straight quarter of US hyperscaler AI chip orders for 2028 shipments, while separately locking AMD (bulk) and Nvidia into 2nm HBM base-die output at its Taylor, Texas fab with equipment install from February 2029 and mass production in 2030. The multi-year capacity pre-commitments signal that leading AI chipmakers are now contracting specialized manufacturing years ahead, a structural shift in foundry and HBM supply dynamics.

89DigiTimes AsiaSemiconductors

AMD and PsiQuantum expand partnership on quantum computing infrastructure

PsiQuantum and AMD expanded their partnership to build the classical computing backbone for utility-scale fault-tolerant photonic quantum machines, combining AMD Instinct GPUs, EPYC CPUs and ROCm software with PsiQuantum's silicon photonics, with AMD intending to take an equity stake. PsiQuantum adapted its PRISM machine-learning decoder to run with minimal latency on AMD programmable accelerators for real-time error correction, and the partners target drug discovery first, citing early covalent-inhibitor modeling results. The work aligns with DOE's Quantum Genesis initiative and a possible joint presence at the Illinois Quantum and Microelectronics Park, where PsiQuantum is anchor tenant.

The takeAMD is positioning its Instinct GPUs, EPYC CPUs and ROCm stack as the classical control plane for fault-tolerant quantum computing, adding an equity stake in PsiQuantum and extending beyond the earlier Versal FPGA work; the PRISM ML decoder running real-time error correction on AMD accelerators is the concrete technical hook, and the tie-in to DOE's Quantum Genesis (100+ logical qubit systems, $215M in awards, 2028 target) gives it sovereign-compute context.

90SCMP TechSemiconductors

China’s SMIC sees AI computing crunch until 2028 as Nvidia heat risks tighter chip curbs

SMIC reported Q3 gross profit of US$953.2M (up 82.3% YoY) and revenue of US$3.19B (+33.9% YoY, +6.1% QoQ), with gross margin up 7.9 points to 29.9%, citing higher wafer volumes, prices and utilisation. Management guided Q4 revenue up 4-6% sequentially (margin softening to 26-28%), bucking seasonal slowdowns, and forecast AI compute shortages persisting to mid/late 2028 as enterprises build their own smaller facilities. The outlook comes amid scrutiny of Nvidia that could tighten US curbs, potentially accelerating China's shift to domestic foundries.

The takeSMIC's co-CEO Zhao Haijun projects AI compute and supporting chips remain in short supply through mid-to-late 2028, backed by Q3 gross profit up 82.3% YoY to US$953.2M on revenue of US$3.19B (+33.9% YoY) and 29.9% margin — a concrete signal that domestic Chinese foundry demand is being pulled by enterprise on-prem AI buildouts rather than token rentals.

91Tech in AsiaAI & Models

Baidu unveils Ernie 5.5 Flash Preview, open weights soon

At Baidu World 2026, Baidu launched Ernie 5.5 Flash Preview (280B parameters, 1M-token context) with API availability now and weight release planned for November 25. It also introduced Token Exchange, a token marketplace connecting model providers and customers that onboarded over 100 model resources in its first month, and updated its DuMate and KOOKO agents. Apollo Go now operates in 28 cities with 350M+ autonomous kilometres logged, following an April 2026 report that China suspended new AV permits after a Baidu-related outage.

The takeBaidu's Ernie 5.5 Flash Preview packs 280B parameters and a 1M-token context window with API access live and open weights promised by Nov 25 — a direct bid to commoditize the model layer and seed its new Token Exchange marketplace, which already lists 100+ model resources.

92Semafor TechCybersecurity

Exclusive / Trump’s AI task force discusses cyber risks

The White House AI task force met Thursday to address the evolving cyber threat environment and hardening critical infrastructure, with DNI Jay Clayton briefing members on tech CEOs' work toward a common set of industry standards under the White House AI accord and floating adding more firms. Commerce Secretary Howard Lutnick and National Cyber Director Sean Cairncross updated on NIST's pre-release AI model testing, and the task force wants Congress to reauthorize cyber and information-sharing bills while weighing whether to back new legislation. The Nvidia scandal was notably not discussed.

The takeThis is a substantive, exclusive look inside the White House AI task force's actual agenda: DNI Jay Clayton briefing on industry-standard-setting by accord-signing CEOs, NIST pre-release model testing, and a push for Congress to reauthorize cyber information-sharing authorities. The detail that the Nvidia scandal was explicitly not discussed is itself a signal of the administration's priorities.

93BloombergMarkets

Hong Kong Turns Even More Bullish on Economy With Growth Upgrade

Hong Kong raised its full-year 2026 growth forecast for the second time in months, to 4.8%, after confirming Q3 GDP expanded 5.2% year-on-year. The government attributed the upgrade to robust global AI investment and related trade, with September exports jumping almost 65% year-on-year to a record — the strongest in over four decades — led by electrical machinery, data processing machines and telecommunications equipment. Goldman Sachs and ANZ upgraded their outlooks to 5% or higher, while consumption and investment held up as the property market stabilized and infrastructure development advanced.

The takeHong Kong's GDP forecast was upgraded to 4.8% for 2026 (from an original 2.5-3.5%) after Q3 growth hit 5.2% and September exports surged nearly 65% year-on-year — the fastest in over four decades — driven by electrical machinery, data processing machines and telecom equipment. This is a concrete macro read on how the global AI capex supercycle is now visibly reshaping trade flows and national/regional growth, with Goldman Sachs and ANZ lifting their outlooks to 5%+.

94DigiTimes AsiaSemiconductors

Doosan clears SK Siltron merger reviews; deal could close by end-2026

Doosan Group completed merger reviews across six countries for its KRW2.3 trillion (US$1.73 billion) purchase of SK Group's 70.6% SK Siltron stake, finishing the regulatory process ahead of expectations and potentially accelerating closing by end-2026. However, the AI and HBM-fueled semiconductor rebound has sharply raised silicon wafer free-cash-flow forecasts and peer multiples, prompting SK to demand a valuation reset — market-cited figures have climbed from KRW2-3 trillion to KRW4-5 trillion or higher — delaying definitive agreement signing. NICE Credit Rating lifted Doosan's outlook to Positive on improved CCL electronics margins (18.6% standalone operating margin) and asset sales, but flagged the acquisition's large capital requirement as a near-to-medium-term borrowing burden.

The takeDoosan's KRW2.3T (US$1.73B) acquisition of SK Siltron — a 70.6% stake in the silicon wafer maker — cleared merger reviews in six countries in ~two months, but the AI/HBM-driven wafer upcycle has reset the valuation from an initial KRW2-3T to KRW4-5T+, turning a signed-premise deal into a late-stage price recalibration. The outcome matters because silicon wafers are the upstream input for all AI chip capacity, and Doosan's standalone debt ratio of 89.3% and 4.7x EBITDA leverage make the funding structure a credit wildcard even as NICE upgraded its outlook to Positive.

95Tech in AsiaMarkets

Zhipu AI seeks Shanghai listing after blockbuster IPO

Zhipu AI plans to issue 9.1 million to 38.8 million A shares on Shanghai's STAR Market, without disclosing a fundraising target, months after raising HK$4.35 billion (US$555 million) in its January Hong Kong IPO. Its prospectus shows 2024 revenue of 312.4 million yuan (US$46.2 million) and a net loss of roughly 3 billion yuan (US$444 million), with 70% of Hong Kong proceeds allocated to general-purpose AI model R&D. The STAR 50 Index is up 22% this year as Chinese AI listings accelerate, with domestic IPO proceeds reaching US$6.7 billion by May 22.

The takeZhipu's STAR Market application follows a US$555 million Hong Kong IPO at an ~US$83 billion valuation despite 2024 revenue of just US$46.2 million against a US$444 million net loss — a valuation-to-revenue multiple that crystallizes how aggressively Chinese markets are pricing frontier-model ambitions, with 70% of prior proceeds earmarked for general-purpose model R&D.

96Hacker NewsAI & Models

Models can now transmit viruses in weight-space

The paper formalizes 'weight space misfolding': one model editing another's weights beyond its authorized scope, drawing a prion-disease analogy and a threat taxonomy of factual substitution, capability denial and selective omission. Three IWIBench case studies show targeted disruption is achievable but unreliable and functionally leaky — the Olympic edit held 48 of 49 baseline controls but broke a geography control, the refusal edit over-generalized to the capital of France, and selective omission worked on only 1 of 22 prompts. The authors argue self-propagation is not demonstrated but the building blocks are present, and propose bounding acceptance through edit provenance, patch records, baseline testing and recovery paths.

The takeA controlled study shows a frontier editor model (ChatGPT 6.1-Sol) writing persistent behavioral changes directly into a target's weights (Qwen2.5-1.5B-Instruct), reliably flipping a factual answer to 'Buenos Aires' on 100/100 held-out prompts while the pristine model answered 'Sydney' 91/100 — establishing that weight-space manipulation is a real, if imperfect, attack surface with measurable collateral damage (lost controls, over-broad refusals).

97The ElecCybersecurity

South Korea's AI Strategy Committee Launches Cyber Resilience Task Force

South Korea's National AI Strategy Committee launched a National Cyber Resilience Task Force on Nov. 12 to broaden security policy beyond a defense-only posture as AI-enabled attacks grow more sophisticated. Led by Kookmin University's Lee Won-tae and backed by MSIT, the Interior Ministry, the climate/energy ministry and the NIS, the TF will map threat and policy gaps, set a 2027 strategy plus a 2027-28 roadmap, and run vulnerability assessments, simulation exercises and recovery verification for SMEs and local facilities. Its scenarios cover complex failures and recovery at critical infrastructure including AI data centers.

The takeSeoul is standing up an 11-member private-sector task force to draft a national cyber resilience strategy by Q1 2027, with at least 10 action items, 20+ assessment criteria and three failure/recovery scenarios explicitly covering AI data centers — a concrete state-level move to harden AI infrastructure against attack.

98DigiTimes AsiaMarkets

Cerebras lifts third-quarter revenue 55% as cloud inference overtakes hardware sales

Cerebras reported record Q3 2026 GAAP revenue of $210.8M, up 55% YoY, with cloud and other services jumping 265% to $141.9M and overtaking hardware as the majority of revenue. GAAP net loss widened to $161.2M on warrant and stock-comp charges, though core operating margin is guided to roughly breakeven for Q4, and the company raised full-year core revenue guidance to $940–960M. It disclosed $26.1B in remaining performance obligations, an Oracle OCI integration planned for 2027, a cloud deal with an unnamed top-three hyperscaler, 700 gross MW of capacity secured for end-2027 delivery, and plans to exit 2026 with over 10x its start-of-year manufacturing capacity via Flex and Sanmina.

The takeCerebras' Q3 FY2026 shows a structural pivot: cloud/inference services hit $141.9M (up 265% YoY) and now ~two-thirds of $210.8M GAAP revenue while hardware sales declined, with $26.1B in remaining performance obligations and 700 gross MW of datacenter capacity secured for delivery by end-2027. The strategic test is execution—converting commitments into deployed inference capacity while absorbing a $161.2M GAAP net loss, $155.1M in stock comp, and $1.046B of nine-month capex against $7.8B cash.

99Tech in AsiaMarkets

Tencent raises $5.7b in bonds as AI spending grows

Tencent raised roughly $5.7 billion via $1.8 billion of US dollar bonds and 26 billion yuan ($3.89B) of offshore yuan bonds, with order books exceeding $23 billion. Proceeds are earmarked for general corporate purposes as the company accelerates AI infrastructure and product investment, with H1 2026 capex hitting 96.3 billion yuan ($14.4B) versus 45.9 billion a year prior, covering computing equipment, software and data centers.

The takeTencent's H1 2026 capex more than doubled to 96.3 billion yuan ($14.4B) from 45.9 billion a year earlier, and a $5.7B dual-currency bond with $23B of orders shows debt markets still funding Chinese hyperscale AI buildout despite broader caution on AI spending.

100DigiTimes AsiaMarkets

CoreWeave adjusts AI data center expansion model as customer prepayments, contracts ease financing concerns

CoreWeave is realigning AI data center financing away from high-yield bonds toward borrowing backed by customer contracts, chip assets and prepayments, plus more short-term capacity transactions to cut financing costs. Q3 revenue hit $3.661B (up 168% YoY), with 2026 revenue guidance raised to $13-13.3B and adjusted operating income to $1.275-1.375B, while capex guidance rose to $37-39B and active power reached over 1.7GW against a $132.5B revenue backlog. Despite $2.256B adjusted EBITDA (62% margin), CoreWeave posted a $709M GAAP net loss on $846M net interest expense and ~$46.9B debt; management expects ~$9B in customer prepayments over three months and says prepayments plus hardware-partner financing will fund most 2027 needs with under $5B of parent-level debt.

The takeCoreWeave's shift from high-yield unsecured bonds to contract-, chip- and prepayment-backed borrowing is the tell that neocloud financing is maturing: with $46.9B of debt, $846M quarterly net interest expense and a $709M GAAP net loss against $2.256B adjusted EBITDA, the model only works if ~$9B of customer prepayments and hardware-partner financing carry 2027 capex of up to $39B. The 168% YoY revenue growth to $3.661B and $132.5B backlog show demand is real, but the 2.9x capex-to-revenue ratio makes financing structure, not demand, the binding constraint on AI capacity buildout.

101DigiTimes AsiaSemiconductors

SK Hynix's HBM4 yield tops 70% as Samsung expands chip packaging capacity

Industry sources cited by Chosun Biz say SK Hynix's HBM4 production yield has surpassed 70%, though the company declines to disclose figures and yield bases vary by process stage. On its Q3 call SK Hynix said HBM4 front- and back-end yields beat internal plans, prompting raised targets. Samsung and Micron are racing on HBM4 yields, which industry watchers view as the variable that will decide AI memory market share.

The takeSK Hynix's HBM4 yield crossing 70% — a ~16.7% uplift in good-die output versus 60% on the same input — signals it is converting its HBM3E lead into next-gen mass-production economics, with the 2,048-bit interface doubling HBM3E's data lanes making stacking/packaging precision the decisive cost variable against Samsung (~80% reported in September) and Micron.

102DigiTimes AsiaSemiconductors

CXMT targets higher-value DRAM market with AI server memory, LPDDR6 breakthroughs

At IC China 2026, CXMT introduced DDR5 server modules, mass-produced LPDDR6 chips and automotive DRAM, extending beyond its core mobile memory business on the back of manufacturing advances and higher R&D spending. Separately, Bloomberg reports CXMT has become one of Asia's most polarizing stocks, with post-IPO analyst price targets ranging from about $160 billion to $1.1 trillion in market value.

The takeCXMT is moving up the DRAM value chain with DDR5 server modules, mass-produced LPDDR6 and automotive DRAM, a direct challenge to Samsung, SK Hynix and Micron in the segments where AI server demand and pricing power concentrate. The Bloomberg datapoint that analyst price targets on the newly listed CXMT span roughly $160 billion to $1.1 trillion underscores how unresolved the market's read on Chinese DRAM competitiveness remains.

103Tech in AsiaAI & Models

Google rolls out faster Gemini 4 Hydrogen for coding, agents

Google announced global rollout of Gemini 4 Hydrogen, a faster model for coding and agentic tasks positioned as an upgrade to Gemini 3.8 Flash, across developer, business and consumer products. It also introduced Orchestrator in Google Antigravity, which switches between models within a session and reportedly matches near-Carbon success rates on Terminal-Bench 4.0 while cutting quota use 24% and task time 19%. Hydrogen is priced at $1.50 per million input tokens and $7.50 per million output tokens with a 1 million-token context window.

The takeGoogle's Gemini 4 Hydrogen ships at $1.50/$7.50 per million input/output tokens over a 1M-token context, scoring 51 on the Artificial Analysis Intelligence Index versus Argon's 50, while its Antigravity Orchestrator claims near-Carbon task success on Terminal-Bench 4.0 at 24% less quota and 19% faster completion — a concrete frontier-model capability and pricing disclosure.

104Tech in AsiaSemiconductors

Meituan leads pre-A+ funding for Chinese chip startup Zhixin

Shanghai-based Lingrui Zhixin Computing Technology raised several hundred million yuan in a pre-Series A+ round led by Meituan Strategic Investment, with backing from a JD.com-affiliated fund, SenseTime Guoxiang Capital, SAIC's CVC arm, Sangfor's industrial fund and others. The company develops its own high-end RISC-V CPU cores, agent CPUs and domain-specific processors. The raise lands as Beijing prepares a national policy to expand RISC-V adoption, positioning the startup within China's push for an open, sanction-resilient instruction set architecture.

The takeA several-hundred-million-yuan round led by Meituan with JD.com, SenseTime, SAIC and Sangfor participation signals Chinese strategic capital consolidating around domestic RISC-V CPU cores and domain-specific processors — a direct hedge against US export controls on x86/ARM-dependent compute.

105Tech in AsiaInfrastructure

Nebius’ Q3 revenue soars 571% to $981m

Amsterdam-based AI infrastructure firm Nebius reported Q3 2026 revenue of $980.8M, up 571% year on year, with adjusted EBITDA of $465.8M and operating income of $50.4M; its AI cloud business drove ~99% of revenue. Annualized run-rate revenue hit $6.4B by end-September, though full-year guidance is expected at the low end of the $3.0–3.4B range. Nebius operates 18 data centers, plans up to 300MW of new capacity in New Jersey, and secured a $3.5B Citi financing facility to fund datacenter growth.

The takeNebius' AI cloud revenue jumping 571% to $980.8M with a $6.4B annualized run-rate and $465.8M adjusted EBITDA is a concrete datapoint on neocloud economics — AI infrastructure demand converting into profitable growth, backed by a $3.5B Citi facility and a 300MW New Jersey buildout.

106DigiTimes AsiaSemiconductors

DEEPX tapes out AI chip on Samsung's 2nm process

South Korean startup DEEPX has taped out its DX-M2 edge AI accelerator on Samsung's second-generation 2nm process, designed to run 20B-100B parameter models locally at under 5W. The chip roughly triples memory bandwidth to ~150 GB/s across eight channels with Samsung processing-in-memory LPDDR5X and supports up to 96GB of external memory, with samples due March 2027 and mass production in late 2027. DEEPX is the first commercial customer for Samsung's 2nm gate-all-around process under an August 2025 deal with design partner Gaonchips.

The takeDEEPX's DX-M2 tape-out on Samsung's 2nm GAA process is a concrete milestone for edge generative AI: 80 TOPS at under 5W, ~150 GB/s LPDDR5X bandwidth (3x the DX-M1's 50 GB/s) with up to 96GB memory, targeting 20B-100B parameter models at 10-30 tokens/s — a 6.7x gain over the prior generation. It also marks the first commercial customer for Samsung Foundry's 2nm node, a meaningful validation signal for Samsung's advanced-node competitiveness against TSMC.

107DigiTimes AsiaMarkets

Asus reports record third-quarter revenue as server business grows

Asus posted record Q3 2026 brand revenue of NT$258.79bn (+36% YoY) and net income of NT$21.63bn (+105% YoY), with servers up 100% YoY on AI datacenter demand and PC share rising 1.1pp to 8.7%. Management guided Q4 servers to +280% YoY but warned of a 10-15% PC market decline in 2027 as memory price hikes outrun pass-through ability, and disclosed inventory of NT$463.9bn (60% raw materials) as forward procurement. The company also flagged risk of Tier 2/3 CSP and neocloud consolidation from tightening capital markets, and repositioned its PSBG unit toward humanoid robots and autonomous platforms.

The takeAsus's server business doubled YoY to 30% of revenue with Q4 guidance of +100% QoQ/+280% YoY, while management flags a structural PC downturn (2027 industry -10% to -15%) driven by memory price inflation it can no longer fully pass through — forcing mid-range DRAM configs down from 16G to 8G. Inventory ballooned 200% YoY to NT$463.9bn (196 days) as strategic 12-month raw-material stockpiling, a direct read on memory and component scarcity across the supply chain.

108TechNodeMarkets

Alibaba’s Chip Unit Wants to Go Public. Who Pays It?

Alibaba confirmed T-Head filed confidentially for a Hong Kong IPO in September, retaining control, with IFR reporting a raise of at least US$2bn and valuations ranging from US$25bn to US$85bn. T-Head designs Zhenwu AI processors and Yitian server CPUs and licenses Xuantie RISC-V cores; CEO Eddie Wu put annualised revenue near RMB10bn, and 650+ external customers had adopted the Zhenwu M890 via Alibaba Cloud by August. The article argues investors cannot price the listing until the prospectus discloses how much revenue comes directly from external customers versus intra-group sales to Alibaba Cloud, and whether customers commit to Xuantie designs for their own products.

The takeT-Head's confidential Hong Kong filing (Sept) with a reported ≥US$2bn raise and wildly divergent valuations (JPMorgan US$25–62bn vs Nomura US$85bn) makes the related-party revenue split the crux: Alibaba claims >60% of T-Head's compute capacity serves external customers, but much of that flows through Alibaba Cloud billing, obscuring how much revenue actually accrues to the chip company. With Alibaba posting a ~RMB45bn June-quarter free cash outflow driven by cloud capex, the listing is a capital-raising vehicle for an AI chip and RISC-V design business whose true standalone economics remain undisclosed.

109DigiTimes AsiaSemiconductors

China's eSSD order opens up US market opportunity for Samsung, SK Hynix

Samsung and SK Hynix are redirecting enterprise SSD supply originally earmarked for Chinese customers toward US cloud service providers, after major Chinese AI companies reduced orders amid difficulty securing advanced processors under US export restrictions. The shift illustrates how chip export controls are transmitting through the memory supply chain, moving eSSD capacity from constrained Chinese AI buyers to US hyperscaler demand.

The takeThe eSSD reallocation from Chinese AI customers to US CSPs is a concrete supply-chain signal that US export controls on advanced processors are now cascading into memory demand — Chinese AI firms cut orders because they can't get accelerators, and Samsung/SK Hynix are backfilling with US hyperscaler demand. This is a real-time read on how compute restrictions reshape the memory allocation map.

110TechCrunchMarkets

AI agent startup LangChain hits $3B valuation

LangChain raised a $300 million Series C led by CapitalG at a $3 billion valuation, with GIC, DFJ Growth, IVP and Sapphire Ventures participating — over double its $1.25 billion valuation from roughly a year earlier. The company pivoted from its 2022 open-source LLM tooling roots to agent creation and management via LangSmith, positioning itself as model-, cloud- and ecosystem-agnostic. It claims 7,000 LangSmith customers, spanning startups like Harvey, Rippling and Clay and enterprises including NVIDIA, Toyota and Salesforce.

The takeLangChain's $300M Series C at a $3B valuation — more than double its $1.25B mark a year ago — signals that the agent-orchestration and management layer is attracting serious capital even as frontier labs absorb the underlying tooling; 7,000 LangSmith customers including NVIDIA, Toyota and Salesforce give it real enterprise traction.

111Yahoo Tech & Markets WireMarkets

Goldman Sachs sees AI impulse peaking in 2027

Goldman Sachs' 2027 outlook argues the AI investment boost peaks in 2027 and has contributed more to equity gains than to economic growth, making the unwind a bigger risk for markets than the economy. The bank expects leadership to rotate away from AI infrastructure, warns tighter financial conditions could derail that handoff, and flags record credit and equity issuance with discounted pricing and wider US credit spreads likely. It also names a Middle East-driven energy spike as a top near-term risk.

The takeGoldman's call that the AI investment impulse peaks in 2027 — with AI having driven market gains more than GDP growth — frames the coming test as a market repricing rather than an economic one, and flags record credit/equity issuance plus widening US credit spreads as the funding stress point for the compute buildout.

112The RegisterAI & Models

Alibaba's Qwen3.8 re-educated with a more Western worldview

Austin-based Nous Research post-trained Alibaba's open-weight Qwen3.8-27B into 'Arete-27B' to strip state-aligned censorship, using Thinking Machines Lab's Inkling as the teaching model after rejecting Nvidia's Nemotron 3 Ultra over alleged China-related bias. Nous reports propaganda-containing answers fell from 22% to 1% and fact scores rose from 2.32 to 2.98 out of 5, while vendor Hirundo cites a drop from 89.8% to 2.8% on Qwen3.6-35B-A3B. The piece also notes Anthropic's claim that Qwen-affiliated operators distilled Claude outputs, and a Booz Allen study finding some Chinese models produce more vulnerable code.

The takeNous Research's Arete-27B demonstrates that state-aligned behavior in Chinese open-weight models is a shallow, removable post-training layer: 50M tokens of supervised fine-tuning cut propaganda-laden answers from 22% to 1% and lifted fact scores from 2.32 to 2.98/5, with Hirundo reporting a comparable drop from 89.8% to 2.8% on Qwen3.6. The claim that Nemotron 3 Ultra was itself 'contaminated' by training on Chinese open models is a notable data-provenance signal for the open-weight ecosystem.

113Tech in AsiaCybersecurity

Israeli AI security startup Zenity raises $125m series C

Zenity, founded in 2021 with offices in New York and Tel Aviv and 230+ employees (~150 in Israel), raised $125M in a Series C led by Norwest Venture Partners, with Qumra, SoftBank Vision Fund 2, Hitachi Ventures and LG Technology Ventures joining existing investors DTCP, Vertex, Third Point and Intel Capital. Its platform governs AI-agent decisions across SaaS apps, cloud agent platforms and endpoints via discovery, posture management, real-time detection, prevention and response. Proceeds fund platform expansion, Zenity Labs, and Europe and Asia-Pacific growth, amid rising concern over prompt injection and agent privilege escalation.

The takeA $125M Series C (total ~$185M, mostly primary) for AI-agent security and governance signals that securing autonomous agents — prompt injection, privilege escalation, real-time detection and response — is becoming a funded enterprise category, with SoftBank Vision Fund 2, Hitachi and LG joining as strategic backers.

114Semafor TechMarkets

CoreWeave tunes out Nvidia noise, looks to raise $4B

CoreWeave is seeking a $4 billion syndicated loan from Morgan Stanley and Mitsubishi UFJ, collateralized by its Meta contracts, after its unsecured bonds were hammered earlier this fall in an early sign of nerves in the AI financing market. Early marketing has drawn a warm reception, with the company targeting an all-in rate near 7.5% — far below the roughly 16% yield on its unsecured debt.

The takeCoreWeave is pricing a $4B contract-backed loan at ~7.5% versus the ~16% yield its unsecured bonds hit this fall — a concrete repricing of AI compute credit risk that shows lenders will fund neocloud capex only against hyperscaler-backed cash flows like its Meta contracts.

115BloombergMarkets

Blackstone’s QTS Is Cutting AI Debt Sale by More Than Half

Blackstone-backed datacenter operator QTS slashed a planned private placement by about 60%, raising ~$300M instead of ~$800M after failing to attract sufficient demand at its offered yield. The deal includes a $100M seven-year US private placement yielding 2.45 points over benchmarks plus a $200M five-year bank loan; Blackstone declined to pay higher spreads because wider QTS pricing would set an expensive reference point for its other debt. QTS has already raised billions across private placements, a $1.65B high-grade bond, and $4.6B and $3.9B green project bonds this year, illustrating the strain of financing AI infrastructure across every corner of the credit markets.

The takeQTS cutting its private debt sale from ~$800M to ~$300M rather than pay wider spreads is a concrete signal of investor pushback on AI datacenter financing — and Blackstone's refusal to reprice to protect its broader cost of capital shows how reference pricing now constrains the buildout.

116SiliconANGLEInfrastructure

Cisco beats expectations, lifts guidance and forecasts AI infrastructure revenue to double in fiscal 2027

Cisco reported Q1 revenue of $18.5B (up 24% YoY) and EPS of $1.40, beating consensus, with net income rising to $4.6B from $2.9B. It raised FY2027 guidance and lifted its AI infrastructure revenue forecast to $9B, more than double fiscal 2026, citing a 'multiyear network supercycle' and demand for agentic AI networking and security. Gross margin slipped to 65.6% from 68.1% amid the global memory-chip shortage, and the article notes investor concern over a possible datacenter buildout slowdown after midterm elections and rising borrowing costs.

The takeCisco's Q1 beat (revenue +24% to $18.5B, EPS $1.40 vs $1.33) and raised FY2027 guidance, including a forecast that AI infrastructure revenue will more than double to $9B, show a legacy networking vendor repositioning around AI datacenter buildout — though gross margin fell to 65.6% from 68.1% on memory-chip cost inflation, a concrete signal of the component squeeze hitting AI hardware.

117Financial TimesGeopolitics & Defense

Swiss defence groups caught by law designed to stop mercenaries

Swiss defence and technology groups say the Federal Act on Private Security Services Provided Abroad, written in 2015 to keep Blackwater-style mercenaries out, now complicates routine maintenance, training and support work abroad and even binds foreign subsidiaries controlled from Switzerland. Executives cite lost contracts (one over SFr5mn to a French rival) and HQ relocation considerations; the foreign ministry says 2021 exemptions cut declarations from 334 in 2020 to 33 in 2022, but they have risen to 60 as wars in Ukraine and the Middle East boost demand. The dispute feeds a broader fight over Swiss 'armed neutrality' and strict arms-export rules, with voters set to decide on easing weapons sales to countries at war.

The takeSwitzerland's 2015 mercenary law is now extraterritorially constraining defence-tech firms (Acutronic, Pilatus, Rheinmetall's Swiss arm), with declaration cases climbing from 33 in 2022 to 60 last year and firms weighing HQ relocations abroad — a concrete regulatory-friction story in the high-tech defence sector amid rising Ukraine/Middle East demand.

118Caixin (财新网)Semiconductors

SiliconFlow CEO Yuan Jinhui: Compute Is Far From Sufficient on the Demand Side, Inference Compute Still Exploding

Speaking at the 17th Caixin Summit on November 12, SiliconFlow founder and CEO Yuan Jinhui said compute demand far outstrips supply, with severe constraints across chips, memory, machine rooms and electricity, and endorsed the view that AI turns human intelligence into a scalable industrial product. He distinguished training compute from inference compute, noting inference requirements vastly exceed training and that daily enterprise token usage already surpasses 100 trillion, with no ceiling in sight. He added that domestic open-source models are still catching up to closed-source ones but offer cost and data-sovereignty advantages in scenarios such as writing, translation and coding.

The takeSiliconFlow CEO Yuan Jinhui argues the binding constraint on AI is now severe production-side scarcity across chips, memory, datacenter space and power, and quantifies the inference shift: training a good model takes roughly 100 trillion tokens, while inference demand is uncapped, with many enterprises already exceeding 100 trillion tokens per day. His point that inference is an easier entry market for chipmakers than training is a direct signal on where domestic AI chip competition and demand will concentrate.

119TechCrunchInfrastructure

This young startup wants to power floating data centers with the ocean’s heat, and investors are already on board

Six-month-old DC startup Drogue has raised $20M from XYZ Venture Capital, Toyota Ventures, Tamarack Global and others to build floating data centers powered by ocean thermal energy conversion, using the temperature differential between warm surface water and ~4°C deep water to boil ammonia and drive a turbine. The company has designed a multi-megawatt prototype to be installed on a conventional ship and deployed in early 2028, with future versions on offshore-platform-like structures, and has preliminary 'approval in principle' from the American Bureau of Shipping. The pitch is that OTEC-powered vessels can deploy AI compute on timelines the land-based grid cannot match, with each vessel operating at sea for about a decade.

The takeDrogue's $20M raise for ocean thermal energy conversion (OTEC) floating data centers is a concrete bet on sidestepping the 8-10 year grid-interconnect bottleneck that Sipe says is 'completely incompatible' with AI demand growing 10x in three years; the multi-megawatt prototype on a ship by early 2028, with ABS approval-in-principle already secured, gives the concept a real engineering and regulatory path rather than pure vaporware.

120The InformationMarkets

Inside Anthropic’s Lavish Roadshow Event at J.P. Morgan’s Headquarters

Anthropic CEO Dario Amodei held a fireside chat with J.P. Morgan's Jamie Dimon before Wall Street's largest investors, including Chase Coleman, Philippe Laffont, Bill Ackman and Michael Dell's family office CIO. Attendees were reassured by Amodei's conservative framing of compute spending, which the company has told investors contributed to its third-quarter swing to an operating loss. Amodei described compute as the company's deepest business dilemma, as executives met investors at the bank's New York headquarters.

The takeAnthropic's investor roadshow, headlined by Amodei calling compute 'the deepest business dilemma that the company faces,' reveals that expensive short-term compute commitments drove its swing to an operating loss in Q3 — a concrete signal of the capital intensity now defining frontier-lab economics.

121The ElecAI & Models

HIMS Challenges Lasertec’s Dominance With Domestic Photomask Inspector

HIMS passed Samsung Display's performance evaluation for large-format display photomask inspection equipment between February and June, with final verification in July-August and a possible order by year-end. The equipment uses enhanced Die-to-Database inspection plus an algorithm that models pattern deformation during development and etching to cut false positives on complex OLED and automotive display patterns. An order would diversify Samsung Display away from Lasertec and could open photomask makers PKL and HOYA as future customers.

The takeA Korean supplier passing 170+ of Samsung Display's evaluation criteria and forcing Lasertec to cut unit pricing from >15bn won to ~10bn won is a concrete supply-chain shift in display photomask inspection, a niche Lasertec has monopolized; HIMS's D2DB algorithm with manufacturing-deformation modeling addresses the 20-30% of masks (irregular pixel zones, camera cutouts) that conventional Cell-to-Cell/Pixel-to-Pixel tools flag falsely.

122DigiTimes AsiaSemiconductors

Apple's iPhone Duo stumbles in China over app glitches and falling resale prices as Huawei dominates

Apple's iPhone Duo, on sale in China from October 23 at 15,999 yuan, is drawing complaints over inner-screen defects, app compatibility (notably WeChat) and falling resale values, and has not reversed a foldable slump: China Q3 2026 foldable shipments fell 32.7% YoY to 1.77 million units, the fourth straight decline, per IDC. Memory-driven price hikes are lifting flagship foldable ASPs (Huawei MateXT2 at 19,999 yuan, Samsung W27 at 18,999 yuan), while vendors pivot to 'passport-style' wide aspect ratios (Huawei Pura X, Samsung Z Fold8, Xiaomi 18 Fold). Chinese panel makers BOE, TCL CSOT and Visionox are differentiating on crease elimination and polarizer-free thinning, with panel share swinging sharply by launch cycle (Samsung Display 66% vs BOE 22% in Q2 2026).

The takeThe iPhone Duo's China stumble is a supply-chain story: Apple's first foldable shipped with inner-screen defects and WeChat compatibility failures, while TrendForce puts Apple at ~5M foldable units (24.8% of a global market growing just 0.5% YoY) and IDC shows China foldable shipments down 32.7% YoY in Q3 2026. The real intelligence is upstream — BOE, TCL CSOT and Visionox are converging on multi-neutral-layer, UTG-clamping and polarizer-free architectures, with Omdia forecasting wide-screen foldable panels at 44% of shipments in 2026 and >60% in 2027.

123Tech in AsiaSemiconductors

Intel, Qualcomm held sale talks with AI startup Tenstorrent

Santa Clara-based AI chip startup Tenstorrent has held early sale talks with Intel and Qualcomm and explored a new funding round, with a deal potentially valuing it above $5 billion. Tenstorrent sells AI accelerator cards, full systems and IP licensing, targeting sovereign AI efforts and automotive OEMs like Hyundai, and differentiates by using GDDR6 with large on-chip SRAM instead of HBM, plus licensable high-performance RISC-V CPUs. The interest reflects a broader scramble for AI chip technology outside Nvidia and AMD, with Arm and SoftBank having approached Cerebras before its IPO.

The takeTenstorrent's reported talks with Intel and Qualcomm at a >$5B valuation signal that buyers want its RISC-V CPU and AI accelerator IP — a GDDR6-plus-large-SRAM design that avoids costly HBM — as a hedge against Nvidia's CUDA lock-in, though porting costs of ~100x hardware spend for new ISAs mean any deal takes years to threaten the market leader.

124Financial TimesInfrastructure

Where is Europe’s AI boom?

The FT argues Europe is not a passive spectator in the AI boom despite holding just one of the world's 64 top LLMs and a tiny chip share. The evidence is in trade and production data: Dutch machinery output up 43% y/y, fast-growing EU exports of datacenter power/cooling/operating equipment, ~10% of new global datacenter capacity, and a 10% average AI-driven capex boost reported by executives. Adoption is closer to US levels than reputed, though Europeans emphasize cost/efficiency over revenue gains, and heavy US software licensing imports are denting the euro area current account surplus.

The takeEurope's AI role is real but infrastructural rather than frontier: Dutch machinery output jumped 43% year-on-year in September on ASML's chip-equipment success, the EU will host ~10% of new global datacenter capacity this year (Goldman Sachs), and AI infrastructure added roughly 0.1pp to European growth — with computers, electronics and ICT services contributing 0.6% of market-production growth since early 2025 per ECB's Philip Lane.

125The ElecInfrastructure

South Korea’s Gigawatt Data Center Ambitions Remain at Megawatt Scale

Korean telcos and cloud providers have announced GW-scale AI data center visions but currently operate only tens to hundreds of MW: KT 152MW, SK Telecom 137MW, Naver's GAK sites ~40MW and 47MW, and NHN Cloud 38MW. CoreWeave, by contrast, reports 1.7GW active capacity, 4.6GW contracted power, 168% revenue growth and a 177 trillion won backlog, with demand 10x available supply through 2027. Weak Korean demand signals — Naver's 20MW contract and a 16% pre-lease rate for 2027 Seoul capacity — raise doubts about feasibility of SK Group's 15GW vision and risk a cycle of lost customers and delayed expansion.

The takeKorean operators are an order of magnitude behind: KT runs 152MW and SK Telecom 137MW against 1GW targets, while CoreWeave already operates 1.7GW with 4.6GW contracted and a 177 trillion won backlog. The gap is self-reinforcing — Naver has contracted only 20MW of its planned 200MW and Seoul-area 2027 capacity is just 16% pre-leased, so uncertain demand keeps Korean builds small while global operators lock in customers.

126Financial TimesMarkets

OK who just made $200mn shorting Nvidia ahead of Jensxit?

Macro Risk Advisors' Joseph Niles identified a large, heavily out-of-the-money Nvidia put spread entered October 12: buying 180,000 January $150 puts and selling 280,000 $100 puts for an estimated $8.32mn net cost. With Nvidia down from $226 to about $166, the trade sits in its sweet spot and has gained roughly $196mn, about 24x the premium, with maximum theoretical profit near $892mn and profitability across a $11.46–$150.46 expiry range. The piece details the mechanics, Greeks and risk of a bet that has become a focal point of the AI-chip selloff narrative.

The takeA single options position — 180k $150 puts bought against 280k $100 puts sold for an $8.3mn net premium on Oct 12 — has reportedly swung to roughly $196mn in profit as Nvidia slid from $226 to ~$166, with net delta ballooning from $90.6mn to nearly $1bn and theta at -$880k/day. It is a concrete, quantified case of sophisticated capital positioning around AI-chip repricing, exactly the kind of market event the thesis covers.

127DigiTimes AsiaInfrastructure

Foxconn ASIC AI server shipments surge 1.5x YoY in 3Q26, GPU servers up 70%

Hon Hai (Foxconn) reported ASIC AI server shipments surged 1.5 times year over year while GPU AI server shipments grew about 70% in 3Q26. Chairman Young Liu said the company expects continued double-digit growth in 2027 led by AI servers, remains confident in AI rack demand, and plans to expand market share through a broader range of AI solutions and a wider customer base.

The takeFoxconn's 3Q26 split — ASIC AI server shipments up 1.5x YoY versus GPU servers up ~70% — signals that custom silicon (hyperscaler ASICs) is outgrowing merchant GPU platforms at the contract-manufacturing layer, a leading indicator for the ASIC-vs-GPU balance in 2027 datacenter buildouts. Chairman Young Liu's guidance for double-digit 2027 growth and share gains via a broader AI rack portfolio confirms rack-scale assembly is becoming the primary growth engine for the world's largest EMS provider.

128Tech in AsiaMarkets

Naver revenue rises 10.7% as AI spending weighs on profit

Naver posted Q3 2026 revenue of 3.5 trillion won (US$2.6 billion), up 10.7% year on year, driven by commerce, financial platforms and global C2C marketplaces, while operating profit dropped 20.9% to 451.3 billion won on higher AI-related infrastructure, content and data spending. The company disclosed roughly 2 trillion won (US$1.5 billion) in initial computing service supply contracts for its AI factory business with Brookfield and Nvidia, with that revenue stream expected to start contributing in the first half of 2027, and it plans a new logistics subsidiary to support commerce growth.

The takeNaver's operating profit fell 20.9% to 451.3 billion won as AI infrastructure spending bit into margins, but the strategic signal is the AI factory venture with Brookfield and Nvidia, which has already locked in ~2 trillion won (US$1.5 billion) in initial compute service contracts with revenue expected from H1 2027.

129DigiTimes AsiaSemiconductors

LG Chem reportedly developing MR-MUF materials for HBM on top of NCF work

LG Chem is reportedly developing MR-MUF underfill materials for HBM, building on its existing NCF (non-conductive film) work, per ETNews-cited industry sources; success would broaden South Korea's domestic HBM materials base and cut import reliance. The article situates this against the packaging-route split: SK Hynix's Advanced MR-MUF underpins its 12-layer 48GB HBM4E samples (17% thermal resistance gain vs HBM4), while Samsung's HCB hybrid copper bonding targets 16+ layers with 20%+ thermal resistance improvement over TCB, and imec/EVG have shown 200nm-pitch W2W test structures. It details the hybrid bonding manufacturing chain — CMP flatness control, contamination, die warpage, and D2W vs W2W yield tradeoffs — with Applied Materials/Besi's Kinex D2W system and DISCO's 70-80% dicing and 60-70% thinning/grinding share as the equipment backdrop.

The takeLG Chem's move into MR-MUF underfill materials, layered on its existing NCF work, targets the exact packaging chemistry that SK Hynix's Advanced MR-MUF HBM4E route depends on — a 12-layer, 48GB part with 17% better thermal resistance than HBM4 — and signals Seoul's push to localize HBM materials currently imported. The broader context matters: as stacks move from 12 to 16+ layers, hybrid bonding (Samsung HCB claims 20%+ thermal resistance cut vs TCB; imec/EVG demoed 200nm pad pitch with sub-40nm overlay) threatens to displace the MR-MUF materials LG Chem is now developing, so the localization play is a bet on the incumbent process staying dominant through the HBM4E generation.

130DigiTimes AsiaMarkets

Pegatron targets several-fold growth in 2027

Pegatron reported record Q3 2026 revenue of NT$389.07B (+50.9% YoY) with operating income up 149.7% to NT$8.04B, but non-operating losses (Luxshare valuation, higher interest) dragged net income down 12.5% to NT$3.98B and EPS to NT$1.49. Management targets several-fold server growth in 2027 with visibility to 2028, covering VR, GB, HGX, compute trays, storage and in-house ASICs from four to five ASIC customers, while acknowledging new-plant capex and liquid-cooling costs are compressing early margins and that funding pressure is real.

The takePegatron's Q3 2026 revenue hit a record NT$389.07B (+50.9% YoY) with servers reaching the 10% target in their first year, but the real signal is the balance sheet: inventories up NT$159B and short-term borrowings up NT$151.7B pushed debt ratio from 60% to 70% and operating cash flow to a NT$128.4B outflow, showing AI server ODM ramp is now a working-capital-intensive business where scale, not margin, is the near-term prize.

131TechNodeSemiconductors

ByteDance’s AI chip chief has submitted his resignation, report says

Shi Yunfeng, who leads ByteDance's AI chip and data processing unit teams and reports to Volcano Engine head Yang Zhenyuan, has resigned and is handing over responsibilities, per LatePost. ByteDance began building its AI chip team in 2019 and pursued a custom accelerator modeled on Google's TPUs; the chip taped out this year but fell short of performance expectations, and a successor is now in development.

The takeThe departure of Shi Yunfeng comes as ByteDance's custom TPU-style accelerator, which taped out this year, reportedly underperformed expectations — a concrete setback for a hyperscaler trying to reduce reliance on Nvidia and build in-house silicon for its AI workloads.

132TechNodeSemiconductors

MetaX and UBTech-backed robotics chip startup Xilect raises RMB 500 million

Xilect, jointly founded by GPU maker MetaX and humanoid robot maker UBTech, raised RMB 500 million in an angel round to develop its first chip and software toolchain. The startup is designing a system-on-chip integrating multimodal perception, decision-making and motion control on the robot itself, with a partner development platform planned for 2027 and first silicon in 2028.

The takeA RMB 500M angel round for a robotics SoC startup is a material bet on edge inference silicon — combining perception, decision-making and motion control on-robot — backed by a Chinese GPU maker and the country's flagship humanoid robot firm, signaling vertical integration ambitions in China's embodied-AI stack.

133TechNodeSemiconductors

In Nvidia’s Shadow, a Chinese Chipmaker Takes a Different Route

Shanghai-founded Denglin listed in Hong Kong on November 10 under Chapter 18C, raising HK$1.415 billion, after pivoting from smart-city projects (80% of 2023 GPGPU revenue) to enterprise private cloud (83% in Jan–Apr 2026), selling 10,246 KS inference cards for RMB166 million. It chose cheaper, lower-bandwidth LPDDR memory with more on-chip SRAM and data-reuse software over HBM, prioritizing cost and power efficiency, with HBM3e-class Minsky IV planned for 2027. Losses persist (RMB281M in 2025) and its top five customers drove 90.4% of revenue.

The takeDenglin's LPDDR-over-HBM bet and 1% share against Nvidia's 83.1% and AMD's 12.3% in China's enterprise AI GPGPU market frame a real strategic question: it grew revenue from RMB27.9M (2023) to RMB200.4M in just four months of 2026 and turned an adjusted profit, but 90.4% customer concentration and RMB238M R&D against RMB84.3M gross profit show how fragile that foothold is.

134The RegisterInfrastructure

HPE debuts hybrid storage configs: Rising SSD costs hit supercomputing

HPE announced new hybrid configurations of its Cray Supercomputing Storage Systems E2000, combining NVMe Gen 5 SSDs with 32TB SAS-4 HDDs and new 74/100-drive enclosures for up to 41% more throughput per rack and 32% more usable capacity. The move is framed against soaring large-SSD prices caused by AI-driven component demand, letting labs shift spend from storage to compute. HPE also introduced a Rack Management Controller for liquid cooling with leak detection, shipping for its GX5000 in March 2027 and for Nvidia NVL72 and AMD Helion/Helios in June.

The takeHPE's refreshed Cray E2000 hybrid configs pair NVMe SSDs with new 32TB SAS-4 HDDs to hit up to 116 GB/s read and 92 GB/s write per rack (41% more throughput, 32% more usable capacity), a concrete response to NAND price inflation driven by AI infrastructure demand that is squeezing HPC storage budgets.

135QbitAI (量子位)AI & Models

Another Chinese Large Model Dark Horse Emerges, Drawing Silicon Valley Attention

Silicon Valley technical outlet Latent Space's new model-testing column benchmarked NaiveAI's Naive-N0.5-Flash against DeepSeek V4.1 Flash 1017 and GLM 5.3 Flash, finding it mid-tier overall but leading on NL2Repo-Bench code-repo generation and dramatically faster in inference speed. The model uses sparse attention for million-token context and a self-optimized NaiveRT serving system, with architecture, training and inference optimization partly developed by AI itself. The piece frames this as evidence of a broadening Chinese open-model ecosystem challenging overseas incumbents.

The takeNaiveAI's open-source Naive-N0.5-Flash, founded by Tsinghua professor Dai Jifeng, scored 71.9 on NL2Repo-Bench (beating DeepSeek V4.1 Flash by 3.7 and GLM 5.3 Flash by 15.6) and hit ~2100 Token/s peak with a 3.1s median response versus DeepSeek's 10.1s — a ~3.3x latency advantage in Latent Space's independent testing. The 'AI-driven AI R&D' approach (AI optimizing its own CUDA kernels, ~10M weekly sandbox runs, 100K peak concurrency) is the strategically notable claim.

136DigiTimes AsiaInfrastructure

FICG returns to growth in 3Q26 as 1.6T enters mass production

FIC Global returned to growth in 3Q26 on the ramp of overseas optical communications capacity and strong AI datacenter demand for high-speed optical modules. The company confirmed 1.6T products have entered mass production, with 3.2T and 6.4T moving into pilot production, strengthening its forward outlook.

The takeFICG's 1.6T optical module mass production with 3.2T and 6.4T in pilot production marks a concrete step in the AI datacenter interconnect roadmap, where optical bandwidth per lane is becoming a gating factor for scaling clusters beyond copper's reach.

137Tech in AsiaSemiconductors

Intel Korea teams with chip startup Panmnesia on AI memory

South Korea's Panmnesia signed an MOU with Intel Korea on November 12 in Seoul to collaborate on CXL for AI datacenters, starting with interoperability tests between its PCIe 6.4-CXL 3.2 Fusion Switch and Intel Xeon platforms. The companies will also explore rack-scale AI infrastructure, including reference architectures and proof-of-concept demonstrations. CXL memory expansion and pooling addresses AI workloads exceeding a single server's memory capacity.

The takePanmnesia's PCIe 6.4-CXL 3.2 Fusion Switch entering interoperability testing with Intel Xeon platforms is a concrete step toward CXL memory pooling, a key lever for AI workloads that outgrow single-server memory capacity. The MOU's extension to rack-scale reference architectures signals a push toward standardizing shared-memory fabrics in AI datacenters.

138DigiTimes AsiaSemiconductors

Europe's AI ambitions hinge on power semiconductors, and on fending off Chinese rivals

At SEMICON Europa in Munich, SEMI and TechInsights argued that power semiconductors are Europe's strongest entry point into the AI datacenter boom, since AI-driven semiconductor growth is concentrated in memory, advanced logic and advanced packaging where Europe has limited strength. The framing also ties Europe's AI ambitions to fending off Chinese rivals in power devices.

The takeSEMI and TechInsights position power semiconductors as Europe's realistic lever into the AI datacenter buildout, implicitly conceding Europe is shut out of the memory, advanced logic and advanced packaging segments where AI growth is concentrated — a strategic framing that matters for European fab investment and for the China competition question in power devices.

139DigiTimes AsiaMarkets

E Ink targets 2027 revenue from PTFE materials for AI data centers

E Ink's Q3 2026 earnings call cut full-year revenue growth guidance to ~9% (from 10-15%) as consumer electronics fell ~15%, though 9M26 profit was the second-highest in company history at NT$9.169B and operating margin held at 31.3%. Chairman Johnson Lee disclosed a partnership with Taiflex Scientific — initiated at NVIDIA's request — to develop and contract-manufacture PTFE copper-clad laminate for AI datacenter high-speed transmission, targeting first revenue in 2027 and volume ramp in 2028, with E Ink holding ~15.7% of Taiflex. IoT/signage grew over 35% YoY and will surpass CE as the largest revenue share, while the H6 e-paper line begins equipment installation this month for 2027 mass production.

The takeE Ink is leveraging its roll-to-roll film expertise to contract-manufacture and modify PTFE copper-clad laminate for AI datacenter high-speed transmission at NVIDIA's request, with small 2027 contributions and 2028 volume ramp — a credible entry into the AI materials supply chain since PTFE needs no glass-fiber cloth (no shortage constraint) and yields already beat M8/M9. The earnings detail is equally concrete: 9M26 revenue NT$28.785B (-1% YoY), operating margin 31.3%, Q3 EPS NT$2.30 (-37.8% YoY), IoT up 35% YoY and now exceeding CE revenue share, with H6 line equipment install starting this month for 2027 mass production.

140The ElecSemiconductors

Qualitas Semiconductor Expands SK hynix IP Deal to 3.9 Billion Won

Qualitas Semiconductor disclosed it revised a March supply agreement with SK hynix, raising the contract value by about 1.5 billion won to 3.92 billion won, with the term unchanged through March 2027 and payment in three installments. Industry sources say the increase followed performance upgrades to the interface IP supplied, which spans PCIe for SSD controllers and UCIe for custom HBM and high-bandwidth flash base dies. The deal now equals roughly 73% of Qualitas's prior-year revenue of 5.35 billion won.

The takeThe revised deal lifts Qualitas's SK hynix contract from 2.42B to 3.92B won — roughly 73% of its entire 5.35B won annual revenue — after performance upgrades to its PCIe/UCIe interface IP, tying a small Korean IP vendor's fortunes directly to SK hynix's HBM and cHBM base-die roadmap.