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

Qatari mediators race a third day of Tehran talks against a US strike window and Iran's explicit threat to Gulf energy infrastructure, while South Korea's 259% semiconductor export surge and a bipartisan US datacenter backlash expose the physical and political limits of the AI buildout.

Sort
Group

01DigiTimes AsiaSemiconductors

India roundup: India accelerates semiconductor push as states, chipmakers and equipment suppliers expand investment

India's chip push broadens across states and supply-chain tiers: Tamil Nadu adds an independent subsidy route, KLA weighs India manufacturing under Semicon 2.0, and Spintronics AI evaluates a ~INR28B (US$294M) GaN fab in Ujjain. Engineering depth grows via HCLTech-Rapidus development centers in Bengaluru and Hubli, while Tata Communications plans US$400-450M for ~10,000 km of long-haul fiber linking AI hubs and data centers.

The takeKLA evaluating one or two Semicon 2.0 applications marks a potential shift from India as a demand market to an equipment manufacturing base, while Spintronics AI's ~US$294M GaN fab proposal in Madhya Pradesh signals widening beyond silicon into compound semiconductors.

02Yahoo Tech & Markets WireAI & Models

Reddit Stock Crashed 29%: Is Google AI Coming for More Stocks?

Reddit's stock crashed 29% to $91.73—its worst day ever—after CEO Steve Huffman confirmed Google's AI Mode 'does not drive traffic' and searches sent over 1M fewer weekly visits in October, with daily users projected to fall by a low double-digit percentage (~13M+) this quarter despite sales up 53% to $895M. The damage extends across the web: People publisher traffic fell 40% YoY, Tripadvisor and Pinterest cite AI Overviews, and Chegg's shares collapsed from $113 to ~$1. Reddit's defense rests on 80%+ of revenue from logged-in/app users and its role as a licensed AI training source, though its Google deal nears expiry.

The takeGoogle's AI Mode and Overviews are structurally cannibalizing referral traffic across the open web—Reddit lost 1M+ weekly visits in October and its stock fell 29% ($7B+), while Alphabet itself shed ~$500B as search growth slowed to 12%. The strategic signal is that AI answer engines are repricing the entire click-dependent internet, even as Reddit's content remains a licensed input (~$60M/yr from Google, ~$70M/yr from OpenAI).

03ReutersInfrastructure

EXCLUSIVE: Australian outback data centre pulls timeline after regulator stalls

Reuters reports that Australian startup Energy North scrapped the June 2028 target start date for its 1GW gas-and-solar data centre campus after the Northern Territory EPA suspended its statutory review deadline to consult federal officials developing AI infrastructure policy. The regulator flagged 10 potential impacts — from water strain to Aboriginal sacred sites — and expressed low confidence in proposed mitigations. The setback, alongside NextDC's Melbourne delay and Goodman's Sydney pullback, illustrates that even the sparsely populated north's land-and-sunshine pitch faces planning, social-license, and energy-supply hurdles in the global race for AI factory investment.

The takeEnergy North's 1GW 'Project Ares' — four times the size of NextDC's delayed Melbourne expansion — lost its June 2028 first-power target after the NT EPA suspended its seven-week review, citing 'low confidence' in mitigation measures across 10 impact areas. Combined with Beetaloo Energy's proposed 2GW facility, the two projects would total ~3GW, roughly 10x the entire Territory grid's peak demand, exposing how AI infrastructure ambitions are colliding with permitting, water, Aboriginal consultation, and gas-supply realities.

04BloombergGeopolitics & Defense

Billionaire Friedland to Build Major US-Backed Rare Earths Mine

Sunrise Energy Metals, led by Robert Friedland and backed by up to $400M in US Department of Defense long-term debt, has greenlit the Syerston scandium mine and refinery in New South Wales targeting 2028 production. The project hosts one of the world's largest high-grade scandium deposits, with initial output of 60 tons/year and potential to triple, aimed at displacing Chinese supply that has 'all but ceased' to Western markets since 2025 export controls. Scandium is critical to defense, advanced manufacturing, AI infrastructure, and wireless spectrum technologies.

The takeA US DoD-backed $400M debt facility for Sunrise's Syerston scandium project (60 t/yr initial, tripling potential) is a concrete step in Western critical-minerals de-risking, directly targeting China's near-monopoly that has already choked Western scandium supply since 2025 export controls.

05DigiTimes AsiaInfrastructure

Foxconn expands AI deployment with smart manufacturing across sites worldwide

Foxconn and Intel will co-develop and commercialize rack-scale AI infrastructure spanning Xeon-based CPU racks, AI accelerator architectures, high-speed interconnect, thermal management, liquid cooling, and system telemetry, plus custom ASIC/SoC design services. The collaboration extends to physical AI—robotics, edge intelligence, agentic systems—under Foxconn's 3+3+3 strategy and Intel's full-stack push from silicon to rack-scale systems. Foxconn is separately deploying AI across injection molding, CNC, SMT, and production scheduling, with digital twins and physical AI planned for process optimization and factory design.

The takeThe Foxconn-Intel pact targets rack-scale AI infrastructure—Xeon CPU racks, accelerator architectures, high-speed interconnect, liquid cooling, and custom ASIC/SoC design services—signaling Foxconn's move up the value chain from assembly to systems integration while giving Intel a systems-level channel against hyperscaler custom silicon. Intel's Q1 2026 revenue of US$13.6bn (+7% YoY) frames the partnership as a demand-capture play in inference and agentic workloads.

06The RegisterAI & Models

Linus Torvalds says Linux kernel is out of ‘stabilization phase’ and using ‘much healthier model’ for AI contributions

In the first RC announcement for Linux 7.4, Linus Torvalds noted an unusually large merge window of nearly 20,000 commits against a typical ~15,000, attributing much of it to AI-triggered work and calling it healthier that AI fixes arrive during normal development rather than flooding the stabilization phase. He acknowledged AI-suggested patches are increasingly common but 'often get rewritten,' and flagged that 7.4's expected LTS status may also explain the larger release. Notable inclusions cover AMD, Apple, and Google processor support, VM security tweaks, and Qualcomm Snapdragon X2 enablement.

The takeTorvalds reports ~20,000 commits in the 7.4 merge window versus the usual ~15,000, and frames AI-generated patches shifting from late-stage -rc fixes into normal development as a 'much healthier model' — a concrete signal that AI codegen is maturing into upstream kernel workflow rather than destabilizing it. His caveat that AI patches 'often get rewritten' and that 7.4's LTS status may independently inflate volume keeps the claim appropriately hedged.

07DigiTimes AsiaSemiconductors

The 20% problem: why Taiwan's imported equipment parts still carry 70% of the cost

US Q3 2026 GDP growth of 3.4% was 47% inventory restocking (1.61 pts), with equipment investment adding 1.16 pts but imports subtracting 2.64 pts; real final sales grew only 1.7%. BEA data show computer imports at $601B annualized versus $420B computer investment, and Taiwanese server exports to Mexico hit $28.4B in H1 2026, nearly double all of 2025, as Mexico briefly overtook Taiwan as the top US server supplier. The analysis implies AI hardware cyclical risk sits with Taiwanese and Mexican producers, while only unimportable inputs — structures, power, cooling, software — capture domestic value.

The takeThe BEA's own series show $1.43 of imported computer hardware per $1 of US computer investment in Q2 2026 (vs $0.91 at the dot-com peak), with marginal imports running $1.71 per $1 of added investment — confirming the AI capex boom's production value accrues to TSMC, Taiwanese server ODMs, and Mexican assembly, not US GDP. Mexico supplied 34.5% and Taiwan 36% of US server imports in H1 2026 while China's share collapsed from 60.4% to 2%, a hard supply-chain reallocation signal.

08DigiTimes AsiaSemiconductors

Analysis: Samsung's own memory unit pushes its foundry toward a 2027 capacity crunch

Samsung's foundry unit, long underutilized, now projects demand to outstrip capacity across all nodes below 8nm by 2027 at over 100% utilization. Its memory division posted Q3 revenue of 140.7 trillion won ($98.5 billion), up 427% YoY, driven by HBM4 ramp, with plans to raise high-value AI memory mix in Q4.

The takeSamsung's claim that sub-8nm demand will exceed capacity by 2027 with >100% utilization signals a structural foundry tightening that, combined with HBM4-driven memory revenue up 427% YoY to 140.7 trillion won, gives Samsung dual leverage across logic and AI memory.

09DigiTimes AsiaSemiconductors

Seoul Viosys begins mass production of rare earth-free white micro LEDs

Seoul Viosys has begun mass production of rare-earth-free white micro LEDs (WICOP 3), the first white opto-semiconductor below 100µm, with initial 25µm x 10µm chips emitting 0.015lm at ~100lm/W and costing roughly half a conventional white LED. The package-free, GaN-only architecture removes the rare-earth phosphor bottleneck that previously kept white chips above 1000µm, enabling embedded lighting and AR display applications. The company is also pitching WICOP arrays as optical interconnects for AI data communications and is recruiting micro-LED supply-chain partners, citing Omdia forecasts of a $60bn combined AR/optical-link market within five years.

The takeSeoul Viosys' WICOP 3 eliminates the YAG rare-earth phosphor and package entirely, using GaN-only white emission to break the 1000µm floor that has blocked white micro-LED scaling, claiming ~50% manufacturing cost reduction and ~100lm/W efficacy (targeting 200lm/W in two years). The strategic hook is the AI data-interconnect angle: 1000 10µm chips in ≤1mm² delivering 10x speed at one-tenth the power versus copper, targeting a projected $60bn AR/optical-link market.

10DigiTimes AsiaSemiconductors

Nytex targets 35% of sales from high-end materials by 2027

Taiwan specialty plastics compounder Nytex targets high-end semiconductor materials at 35% of revenue in 2027, up from ~30% in 2026 and 3.59% in 2023, driven by wafer carriers, EUV photomask pods for 3nm, FOUPs, and square-panel carriers for panel-level packaging. The company, a Gudeng partner supplying TSMC's EUV pod chain, reported 9M26 revenue up 17.7% YoY to NT$2.14B with 1H26 net profit nearly doubling to NT$127M, and plans a TSE main board listing at NT$88/share.

The takeNytex's high-end materials mix jumping from 3.59% in 2023 to ~30% in 2026 and a 35% target for 2027, alongside 17.7% YoY revenue growth to NT$2.14B in 9M26, signals real demand pull from EUV photomask pods, FOUPs, and CoWoS panel-level packaging consumables tied to TSMC's capacity expansion. The IC tray entry adds a second advanced-packaging growth vector.

11DigiTimes AsiaSemiconductors

Column: Samsung's memory boom reshapes its investment strategy

Samsung and SK hynix are redeploying record memory-cycle cash into sovereign-scale domestic fabs and non-China sites as Washington's termination of the VEU waiver subjects their Chinese memory and packaging lines to annual U.S. equipment licenses. Samsung's Taylor, Texas foundry comes online next year alongside a $4B Vietnam test/packaging plant, while SK hynix breaks ground on a $4B West Lafayette, Indiana packaging facility and scouts Japan and the U.S. Both firms publicly disclaim capacity expansion in China, retaining existing fabs only.

The takeSamsung's KRW107.4T Q3 2026 operating profit at ~55% margin—the first tech company past KRW100T quarterly—funds a KRW800T Jeonnam-Gwangju cluster plus KRW1.6 quadrillion Yongin buildout, while the 2025 loss of the U.S. validated end-user waiver forces annual licensing for China equipment and pushes capacity toward Taylor, Texas ($37B) and Vietnam.

12The InformationCybersecurity

Trump Administration Refuses to Release Information on AI Testing Framework

ProPublica reporting details that Trump political appointees at DHS asked CISA to secretly acquire and hack a voting machine and report findings to the White House without notifying the vendor, alarming career staff who feared the results could be used to invalidate elections; the request stalled partly because the agents capable of such an operation had been laid off under DOGE. The administration has shed over 1,000 CISA employees (about a third of the agency) and proposed a further $700M cut that could eliminate 900 more jobs, while state election officials report CISA was effectively 'nonexistent' during critical midterm planning and now distrust the agency after sensitive state-provided records were declassified. States are funding their own sensor technology and contingency plans, but cannot replicate federal foreign-actor intelligence.

The takeThe piece reveals a concrete insider account of DHS political appointees asking CISA to secretly procure and hack a voting machine while suppressing disclosure to the vendor—a potential political-interference vector—set against CISA's ~1,000-employee (one-third) reduction and a proposed $700M budget cut that could force 900 more layoffs, gutting federal election-security capacity ahead of the midterms.

13BloombergInfrastructure

Colorado Votes on Right to Buy Natural Gas in US Climate Fight

Colorado voters are deciding a constitutional amendment backed by a conservative nonprofit that would guarantee the right to buy and sell natural gas, countering years of state clean-energy policy and PUC mandates forcing utilities like Xcel and Black Hills to cut emissions. Opponents, including Earthjustice and a University of Denver law professor, warn it would trump utility regulation, invite litigation against electric-only building codes (e.g., Boulder), and let gas utilities pass data-center-driven infrastructure costs to consumers. The measure needs 55% to pass and has no public polling; supporters argue it protects existing home infrastructure and consumer choice.

The takeA constitutional amendment to enshrine a right to buy natural gas would override the Colorado PUC's decarbonization mandates and create a litigation vector for gas infrastructure—including new pipelines and plants justified by data-center load growth. With ~70% of state homes on gas and a 55% passage threshold, this is a genuine test of whether ballot-box energy policy can lock in fossil infrastructure against regulatory phase-downs.

14Nikkei AsiaInfrastructure

Nuclear reactors under construction rise to 79 globally on AI demand

The global nuclear construction pipeline has reached 79 reactors, the most in 40 years and the highest since 1986, driven by rising fossil fuel prices and surging electricity demand from AI. The piece frames this as an acceleration of new plant plans worldwide, citing Egypt's first nuclear plant built with Russian assistance as an example.

The take79 reactors under construction is the highest count since 1986, with AI-driven power demand cited as a primary accelerant alongside fossil fuel prices — a concrete signal that hyperscale compute buildout is now reshaping global baseload generation planning.

15Semafor TechMarkets

AI trade falters after disappointing tech earnings

The Nasdaq entered correction territory as investors unwound AI bets, driven by Intel's 29% crash, disappointing Amazon and Google earnings that raised hyperscaler spending concerns, and S&P Global revising Meta's outlook to negative over high capex. Analysts describe 'sober questions' emerging about the AI investments underpinning US economic growth, with one portfolio manager noting that a reassessment of AI means a reassessment of growth expectations even as economic data shows no slowdown yet.

The takeIntel's 29% single-day plunge—its worst since the 1970s—combined with S&P's negative outlook revision on Meta over capex and hyperscaler earnings misses signals a genuine market repricing of AI buildout sustainability, with strategists warning that an AI reassessment directly translates to a growth-expectations reassessment.

16Institute for the Study of WarGeopolitics & Defense

Iran Update, November 1, 2026

ISW/CTP reports an unidentified projectile struck a vessel in the Strait of Hormuz on November 1, assessed as an Iranian effort to disrupt civilian shipping and pressure the US via elevated oil prices. In Yemen, Houthi fighters advanced onto elevated terrain in al Maqatirah District, Lahij Governorate, positioning for a potential siege of Aden City, while Saudi forces struck 'vital resources' in Houthi-controlled Dhamar and Hudaydah and intercepted a Houthi ballistic missile and drone aimed at southern Saudi Arabia.

The takeIran's apparent projectile strike on a vessel in the Strait of Hormuz—the sole threat to civilian shipping there per ISW—signals a deliberate coercive strategy to spike oil prices and force a favorable US diplomatic settlement, while Houthi advances toward Aden City and Saudi strikes on Houthi oil infrastructure in Hudaydah/Dhamar widen the regional escalation.

17Tom's HardwareCybersecurity

New AI-driven Windows malware feeds stolen data to Claude to rank victims’ assets by value, then deletes itself — payload builds custom target assessment, can switch models if one refuses, or skip the AI step entirely

Cisco Talos documented a self-deleting Windows infostealer that harvests credentials, cookies, tokens, wallets, and Wi-Fi keys, then enriches the haul via live session abuse (Gmail receipt searches, DoorDash/Uber Eats histories, Discord profiles, Rolimons valuations) into an Identity_Report.txt. It sends that bundle to OpenRouter with a 'forensic intelligence analyst' prompt to rank the victim's assets by cash-out value, exfiltrating AI_Analysis.txt alongside the data over the victim's own connection. The malware is sold as a build-time template with six swappable credential slots, and Chrome App-Bound Encryption is bypassed via headless Chrome launch.

The takeVOLTA-AUTODOXER operationalizes LLMs as an automated post-exploitation analyst: it builds a structured identity dossier, POSTs it to OpenRouter (Claude Opus 4.6 at temperature 0.2 by default), and falls back to Kimi K2 then Gemini 2.5 Flash on refusal — a concrete signal that refusal resilience is becoming a malware design requirement and that model routers are an attack chokepoint.

18ReutersGeopolitics & Defense

Iran, Qatari mediators hold third day of talks to end US-Iran war

Qatari mediators held over 12 hours of talks in Tehran and will continue in Doha, pushing reciprocal US-Iran steps — UN inspector access to bombed nuclear sites in exchange for sanctions relief — to restart direct negotiations and avert a US military operation days away. Sticking points are Trump's demand for nuclear concessions versus Iran's demand that the US lift its naval blockade, release frozen funds and permit oil sales. Regional states (Egypt, Turkey, Iraq, Pakistan, Saudi Arabia) are urging Iran to deal before the midterms, while Israel lobbies for a joint strike and Iran vows to hit Gulf energy infrastructure in retaliation.

The takeHigh-stakes geopolitical brinkmanship: Qatari mediators logged 12+ hours in Tehran to avert a US strike days away, with Iran threatening retaliation against Saudi, UAE, Qatari, Kuwaiti and Bahraini oil/gas facilities — a direct tail risk to global energy supply. Internal Trump-circle split (Hegseth/Rubio favoring further strikes vs. Vance/Witkoff/Kushner pushing talks) and Netanyahu lobbying for a joint post-midterm operation underscore how fragile the off-ramp is.

19SemiAnalysisInfrastructure

The Datacenter Backlash Goes Bipartisan: Next Week's Midterms and the 2028 Buildout

The article maps how bipartisan anti-datacenter politics could delay the 2028 US buildout, noting ~three-quarters of the 33.8 GW due in 2027 is already under construction and largely insulated, while most of the 62.2 GW 2028 pipeline has not broken ground and depends on state permitting. It details escalating restrictions—Texas's TCEQ permit freeze, Oregon's expanded air/water permit hold through July 2027—and finds 9 of 20 gubernatorial nominees back a construction pause while 11 favor conditions, with the conditions-only candidate favored in Texas, Oklahoma, Pennsylvania and Iowa. It concludes hyperscalers can absorb higher costs but not necessarily schedule slippage, since moratoria block projects still awaiting approvals and capacity can't easily shift states.

The takeSemiAnalysis quantifies a concrete political risk to the AI buildout: of 62.2 GW of US net datacenter additions modeled for 2028, 43.5 GW (70%) sits in 13 key states, 10 of which elect governors on Nov. 3, and pro-moratorium candidates could put ~30 GW of that pipeline at risk versus 5.7 GW if conditions-only candidates win. The single biggest lever is Texas—~30% of 2028 US GW—where a TCEQ permit freeze since Sept. 21 could extend 5–9 months under a Hinojosa win versus a possible December restart under Abbott.

20Nikkei AsiaSemiconductors

South Korea exports nearly double in October on AI investment boom

South Korea's October exports rose 93.7% year-over-year to $115.28B, the strongest gain in over 50 years, driven by AI-related demand for chips and servers. Semiconductor shipments jumped 259.4% to $56.44B and computer sales spiked 496.6%, with exports to the U.S. up 185.8% and China up 133.6%. The trade ministry projects annual exports will cross $1 trillion within November, a milestone reached by only the U.S., China, and Germany.

The takeSemiconductor exports surged 259.4% to $56.44B and computer shipments spiked 496.6% as global AI capex drove South Korea's sharpest export jump since January 1974 (93.7% YoY to $115.28B), putting the country on track to join the U.S., China, and Germany as the fourth to hit $1T in annual exports.

21Politico Tech & PolicyGeopolitics & Defense

Russian strikes leave thousands without power in Ukraine, says Zelenskyy - POLITICO.eu

Russian strikes overnight targeted substations, transformers, and power lines in Dnipro, Vinnytsia, Zhytomyr, and Khmelnytskyi, leaving hundreds of thousands of Ukrainian households without electricity, per Zelenskyy. Russia's Nizhny Novgorod governor reported 100+ Ukrainian drones destroyed and significant damage to an industrial facility in Kstovsky district, where the NORSI oil refinery sits. The WSJ reported Kyiv agreed to pause refinery strikes until after the Nov. 3 U.S. midterms under Trump pressure, despite continued Russian bombardment of Ukrainian energy infrastructure.

The takeMutual energy-infrastructure warfare is now the core escalation dynamic: Russia struck substations and transformers across four regions leaving hundreds of thousands without power, while Ukraine reportedly hit an industrial site (likely the NORSI refinery) in Nizhny Novgorod with 100+ drones. The reported U.S.-brokered pause on refinery strikes until after Nov. 3 midterms reveals Washington trading Ukrainian leverage for domestic political timing—a strategically consequential constraint on Kyiv's most effective deep-strike campaign.

22The InformationInfrastructure

Google Energy Chief: Space Data Centers Cost Up to 100 Times More Than Those on Earth

Google is launching an experimental satellite carrying TPUs capable of answering simple AI queries, the first step of Project Suncatcher and what it calls the first attempt by a tech company to put a micro data center in orbit. The effort targets AI's energy bottleneck by tapping uninterrupted solar power in space, though Google's energy chief concedes orbital data centers cost up to 100 times more than Earth-based ones. Google has not disclosed spending, and the economics hinge on falling launch costs and future parity projections.

The takeGoogle's own energy chief pegs orbital data centers at up to 100x the cost of terrestrial ones, yet Google is still launching TPUs on a SpaceX rideshare next week—a signal that frontier labs view space as a strategic hedge against terrestrial power constraints (natural gas supplied >40% of US data center electricity in 2024, per IEA) rather than a near-term economic play.

23Wall Street JournalAI & Models

The Shopping Bots Are Coming for Amazon

Meta's AI shopping agent Muse rapidly scaled to over 10 million US/Canada users in its first six weeks, unseating ChatGPT as the top free App Store app, with shopping the dominant use case and average purchase size above $200. Amazon blocked Muse from its store citing terms-of-use violations, and CEO Jassy dismissed outside shopping agents as 'not great' while noting Amazon's own Alexa for shopping has 400M users and drives 50% higher spend per order. The standoff pits Meta's agent-plus-partner ecosystem (Shopify, Walmart, Best Buy) against Amazon's ad-monetized direct funnel.

The takeMeta's Muse agent hit 10M+ US/Canada setups in six weeks with >25% of users shopping at $200+ average order size, directly threatening Amazon's ad-driven funnel—prompting Amazon to block Muse and Jassy to publicly downplay agent commerce while touting Alexa for shopping's 400M users and 570% YoY interaction growth.

24Hacker NewsCybersecurity

We Took over ChatGPT Accounts

Parameter Labs found that ChatGPT's sandboxed HTML preview iframes could call a parent-page postMessage handler (from the MathLive virtual keyboard) that used innerHTML, allowing injection of an unsandboxed iframe. OpenAI's CSP blocked inline scripts but permitted YouTube, whose oEmbed endpoint reflected JavaScript, yielding code execution in the ChatGPT origin. This let a crafted share URL exfiltrate private data or act on behalf of any signed-in visitor; OpenAI patched it and paid a $5,000 bounty.

The takeA genuine, technically detailed vulnerability disclosure: a postMessage/innerHTML injection chain in ChatGPT's HTML-preview sandbox escalated via a YouTube oEmbed CSP bypass to full JS execution, enabling cross-user data exfiltration through a shared link — for a $5,000 bounty, which is notably low for a stored XSS-class flaw on a major consumer AI platform.

25Wall Street JournalSemiconductors

Memory Crunch Cements CXMT as China’s Chip Champion

ChangXin Memory Technologies posted a more than eightfold revenue surge for the first nine months of the year, capitalizing on the AI-driven memory supply gap to cement its position as China's leading DRAM maker. CXMT reportedly plans limited HBM3E production within two months despite yields of only 20-30% versus ~70% at SK Hynix, and has shipped samples to domestic AI chip designers as Beijing seeks to offset US export restrictions tightened in December 2024.

The takeCXMT's reported 20-30% HBM3E yields versus SK Hynix's ~70% underscore how far China's DRAM champion remains from cost-competitive frontier memory, yet its eightfold sales jump and imminent HBM3E sampling to domestic AI chip designers show US export controls are accelerating a parallel supply chain rather than halting it.

26Hacker NewsAI & Models

Brazilian Court Sanctions Lawyers for Prompt Injection in Court Filing

A Brazilian labor court fined two law firms 10% of a R$842,000 claim after they hid white-on-white prompt-injection text in a filing instructing the court's Galileu AI to respond superficially and not challenge documents. The court invoked Civil Procedure Code good-faith and anti-obstruction provisions, treating the conduct as undermining the dignity of justice. The Galileu system detected and blocked the injected content, underscoring the need for guardrails against adversarial inputs in enterprise AI.

The takeThis appears to be the first judicial ruling globally sanctioning intentional prompt injection against a court's AI system, with a concrete 10% fine (~$16,000) on a R$842,000 claim. It establishes a legal precedent that machine-readable filing content carries the same integrity obligations as human-readable text, directly implicating AI governance and adversarial-input defenses in high-stakes institutional deployments.

27Wall Street JournalMarkets

AI Investors Are Racing to Get Insurance on the Biggest Market Bet in History

Investors are increasingly buying downside protection on the AI trade via put options and credit-default swaps, with CoreWeave put volume hitting record highs and hyperscaler CDS spreads widening to 2021-era levels. Trading firms including Jump Trading and Susquehanna are building compute-derivative and OTC hedging products, with Nvidia B300 rental rates up ~56% this year to $7.56/hour serving as an underlying index. The piece documents concrete market mechanics—not vendor promotion—reflecting growing institutional skepticism about AI infrastructure economics.

The takeThis is a substantive markets story: CDS spreads on all five hyperscalers (Amazon, Alphabet, Microsoft, Meta, Oracle) widened to their highest since 2021 with no exceptions, and Oracle's spreads blew out over 1 percentage point—signaling broad credit-market concern about AI capex financing rather than idiosyncratic risk. The emergence of compute-derivative desks (Jump, Susquehanna) and record CoreWeave put volume shows institutional hedging infrastructure forming around the AI trade, a genuine structural signal.

28Tom's HardwareSemiconductors

Sam Altman meets with Samsung leadership to secure more RAM for data centers amid crushing shortages — OpenAI CEO travels to South Korea to meet Samsung leadership to discuss future AI data centers

OpenAI CEO Sam Altman met ~50 senior Samsung executives in Yongin, South Korea, to discuss AI infrastructure cooperation tied to the October 2025 letter of intent covering memory semiconductor supply for OpenAI's Stargate project, though no new supply agreement was announced. The visit follows AMD CEO Lisa Su's recent trip to secure additional HBM from Samsung, reflecting intensifying competition for constrained memory capacity as AI data centers compete with PCs and smartphones for supply. Samsung and SK hynix dominate the market, holding ~83% of HBM and ~63% of DRAM revenue, with shortages projected to worsen through 2027 and potentially persist into 2028.

The takeAltman's direct lobbying of Samsung's top brass underscores that memory—not compute—is becoming the binding constraint on AI scaling: Samsung and SK hynix control ~83% of HBM revenue and ~63% of global DRAM, while TrendForce projects DRAM supply falling 1-2% short of demand this year and Micron warns shortages could persist to 2028. With AMD forecasting DRAM revenue up 177% to ~$416B in 2026, securing allocation is now a board-level strategic priority for frontier labs.

29CNBC TechnologyAI & Models

For JPMorgan, Amazon, fighting AI bots like Meta Muse may not be a winning long-term strategy

Major banks and retailers are moving to block third-party AI personal agents like Meta Muse, citing security and data-control concerns, while Walmart is opening its site to agents. Experts and Accenture data (65% consumer openness to GPT-like financial assistants; 82% want to approve every AI action) suggest blocking is a short-term defense that will lose to customer choice, pushing incumbents toward authenticated agent infrastructure. The core fight is over who controls customer data and transaction flow as agents reduce switching costs to near zero.

The takeThe piece surfaces a concrete market-structure conflict: a Billions Network audit of 51 large U.S. retailers found only 11 have a deliberate AI-agent policy, 5 explicitly restrict agents (Amazon, Kohl's, Neiman Marcus), and 37 block them accidentally via legacy scraper rules—while JPMorgan's Dimon confirms active blocking of external agents. This is the emerging access-control battle over agentic commerce and financial data, with real liability, identity, and 'agentic flooding' implications.

30Lawfare MediaGeopolitics & Defense

Russia’s Partnership With Iran: Short-Term Gains, Long-Term Losses

Lawfare analysts Kimberly Marten and Carol Saivetz argue the Russia-Iran alignment is opportunistic rather than durable, tracing it from 2015 Syria cooperation through Iran's Shahed drone transfers to Russia in 2022 and the $1.75B Tatarstan drone plant. By 2026 Russia reversed the flow, supplying Iran with improved Geran-2 drones, satellite imagery, and targeting data used against U.S. and British facilities, reportedly as leverage over U.S. support for Ukraine. The article details short-term Russian gains—depleted U.S. interceptor stockpiles, ~$150M/day in added oil revenue, sanctions relief—against long-term costs to Moscow's Middle East influence, including deteriorated ties with Israel and Gulf states.

The takeThe piece quantifies a concrete strategic shift: Russia has reversed the drone pipeline, now supplying Iran with Geran-2 drones, satellite targeting data, and intelligence that enabled precision strikes on U.S./British bases, while the Gulf war has drained an estimated 60% of U.S. Patriot interceptors and 50% of THAAD missiles and netted Moscow ~$150M/day in extra oil revenue. This is a substantive geopolitical analysis of how a partnership of convenience is reshaping Middle East deterrence and Ukraine's air-defense math.

31Tom's HardwareAI & Models

Open-source USB tool unlocks Nvidia's half-speed GPU AI limiter on Windows — lifts GPU's FP16 math performance up to 65%, UEFI unlocker covers RTX 20, 30, and 40-series desktops and laptops

The nv-uefi-unlocker project boots from a FAT32 USB before the Nvidia driver loads, using the community V67 payload with read-back verification to trick signed Booter firmware into opening write locks on two speed-selector registers, running tensor units at full rate. It rebuilds firmware locally from Nvidia's public 610.43.03 driver packages rather than shipping firmware, and supports RTX 20/30/40 desktops and laptops (RTX 50 unsupported; some RTX 2080 variants fail). The unlock does not persist through power loss, requires re-running each boot, needs Secure Boot off or trusting a self-signed EFI, and gains are contingent on power and thermal limits.

The takeAn open-source UEFI pre-driver tool now unlocks Nvidia's half-rate tensor limiter on RTX 20/30/40 desktop and laptop GPUs from a bootable USB, with verified FP16 gains of 1.45x–1.91x (e.g., RTX 4080 BF16 100.37→192.18 TFLOPS, RTX 3060 25.37→45.79 TFLOPS). This materially lowers the barrier to recovering pro-tier tensor throughput on consumer silicon without VBIOS flashing or patched drivers, though the per-boot reapplication, Secure Boot trade-off, and power/thermal caveats constrain practical deployment.

32Hacker NewsAI & Models

Proving and Improving the Platonic Representation Hypothesis with JEPAs

The authors prove a Platonic Representation Theorem showing JEPAs trained separately on multiple modalities converge to the same representation up to rotation when positive pairs are 'good views' of one scene, and show next-token prediction is a JEPA with 'bad views' (document fragments), explaining why LLMs align with vision only through scale. They validate across 145k captions to 940M web images, derive a scaling law of alignment, and show a 109M-param text JEPA beats every language model up to 72B params and CLIP's text encoder on cross-modal alignment, with a single rotation yielding an image captioner (CIDEr .838 on COCO).

The takeThis paper converts the Platonic Representation Hypothesis into a provable theorem for JEPAs and demonstrates a striking efficiency result: good views need 42x fewer training samples than next-token prediction to reach equal cross-modal alignment, and a 109M-parameter text JEPA outperforms language models up to 72B parameters at aligning with DINOv2/CLIP image encoders. If validated, this reframes scale as a brute-force substitute for proper positive-pair design, with direct implications for how frontier labs should architect multimodal training.

33The VergeInfrastructure

Prop 45 in California could be hiding a loophole for data centers

California's Prop 45 would fast-track environmental review for nine categories of 'essential' projects, including 'broadband Internet access,' which legal experts argue could be exploited by data center developers to bypass CEQA scrutiny. The measure caps public comment at 45 days, requires EIRs within 365 days, limits tribal consultation to federally recognized tribes, and prevents courts from halting projects outright even when agencies fail to comply with CEQA. Proponents deny data centers qualify, but the text doesn't explicitly exclude them, and data center developers have funded the campaign.

The takeProp 45 would impose hard 365-day EIR and 270-day litigation deadlines, cap public comment at 45 days, and bar courts from fully halting noncompliant projects—procedural levers that could materially accelerate data center buildout in the nation's largest state economy. The ambiguity over whether 'broadband Internet access' encompasses data centers, combined with developers' campaign spending, makes this a genuine test of whether AI infrastructure demand can override environmental review.

34Tom's HardwareCybersecurity

Lenovo's 'encrypted' fingerprint USB drive is just an unencrypted SD card in a fancy case — AI-assisted follow-up finds a backdoor and hardcoded keys in other ‘secure’ USB products

Lenovo's ThinkPlus FU100 'encrypted' fingerprint USB drive was found to be an unencrypted SD card whose fingerprint reader merely toggled the USB connection; Lenovo issued high-severity CVEs and requires owners to call a China service line for replacement or on-site firmware fixes. The same research found Lenovo's TSD303 SSD transmitted the correct password in plaintext over USB, and an AI-assisted follow-up using OpenAI Codex broke copy-protection schemes from Netac, KakaSoft, and Nexcopy in 7–26 minutes, plus found two hardcoded 'backdoor' accounts in SecZure's YutianDun.

The takeA security researcher used OpenAI's Codex to reverse-engineer 'secure' USB drives in minutes — Netac's tool fell in ~7 minutes, Nexcopy's in ~10, KakaSoft's in ~26 — exposing hardcoded keys, plaintext password transmission, and hidden backdoor accounts across multiple vendors' products. This is a concrete demonstration that AI agents are collapsing the cost of hardware/firmware reverse-engineering, turning 'AES-256 encrypted' marketing into a liability.

35Financial TimesGeopolitics & Defense

US critical minerals push delivers a gold rush — for Washington lobbyists

The FT documents how Trump's critical-minerals push has spawned a lucrative lobbying economy, with firms like David Bernhardt's (60+ clients, $16mn in fees) securing policy wins such as the Senate's 50-49 vote overturning Minnesota's Twin Metals mining ban. Despite billions in pledged loans, contracts and conditional funding, US mineral production has barely moved, and defense contractors warn they cannot meet a January 1 deadline to stop sourcing Chinese rare earth magnets. The piece also flags conflict-of-interest scrutiny, including a $1.4bn Vulcan/ReElement deal signed months after a Trump Jr-backed VC invested in Vulcan.

The takeThe article quantifies a lobbying gold rush: 14 mining companies with Trump-era deals spent $11.5mn on federal lobbying in 2025/H1 2026 versus ~$2.5mn in 2024, while USGS data show 400+ mineral deposits with no active production (up from 160) and 2025's $38.1bn output gain driven mainly by prices, not volume. This exposes a widening gap between announced critical-minerals policy and actual domestic supply-chain capacity.

36CNBC TechnologyMarkets

Why AI is boosting Meta's ad business while hurting Google search

CNBC analyzes how generative AI is bifurcating the ad duopoly: Meta's AI ad tools (used by ~10M advertisers, driving ~5% more conversions and 9% higher ad spend among adopters) are accelerating its ad growth, while chatbots like ChatGPT, Claude, and Meta's Muse threaten Google's core search-click economics. Both firms are pouring historic capex into AI infrastructure ($205B Google, $145B Meta) with negative free cash flow, raising investor anxiety about returns—Alphabet's stock plunged 12%+ post-earnings, and Meta fell 14% in October.

The takeMeta's ad revenue grew 26% to $63.1B vs Google Search's 12% to $63.4B—a 14-point gap, the widest since 2021—while Alphabet shed ~$500B in market cap on its worst trading day ever. The divergence signals that AI-driven ad targeting is monetizing faster than AI-summarized search, with Google's AI Overviews cannibalizing its own click-based revenue model.

37SCMP TechAI & Models

Self-improving AI the key step towards AGI but safety is key: Z.ai founder

At the World Laureates Forum in Shanghai, Tsinghua professor and Z.ai co-founder Tang Jie called recursive self-improvement the main research problem for achieving AGI, noting Z.ai's GLM-5.3 helped optimize its own inference infrastructure for Chinese chips. He acknowledged loss-of-control risks and urged globally shared safety governance, while the article situates his remarks alongside OpenAI and Anthropic's reported early RSI signals.

The takeTang Jie frames recursive self-improvement as the core AGI research problem, citing Anthropic's data that AI led 26% of its internal AI R&D in August versus ~1% in March — a concrete signal that the RSI race is now a stated priority for both US and Chinese labs.

38Hacker NewsCybersecurity

Domain hacks with unusual Unicode characters (2018)

Microsoft Defender for Office 365 researchers, while building prompt-injection hunting logic, detected a high-volume finance-themed phishing campaign inserting invisible Unicode tag characters (e.g., U+E0020) inside lure words like 'funding' to evade literal keyword matching and potentially subvert ML/NLP tokenizers. The activity ran on ~150 finance-themed sender domains (~96% of flagged volume), peaked at multi-million weekday messages with a strict weekday-only cadence, and dropped sharply after May 15, 2026; it connects to a broader ActiveCampaign-delivered SBA-themed campaign previously documented by Fortra. The same rarity that makes tag-character insertion an effective evasion also yields a low-false-positive detection signal for defenders.

The takeMicrosoft telemetry shows a phishing campaign weaponizing the deprecated Unicode Tags block (U+E0000–U+E007F) to split financial keywords, jumping from ~21,000 signature hits on Feb 8, 2026 to over 1.3 million the next day and sustaining 1–2.37M weekday messages for ~3 months. It's a concrete demonstration that AI-era evasion primitives (ASCII smuggling from prompt-injection research) are being repurposed against ML/NLP email classifiers, exposing a tokenization/normalization blind spot defenders must patch.

39Tom's HardwareGeopolitics & Defense

USAF successfully tests dropping SpaceX Starfall cargo drop pods from space — 2,200-pound capsule drops military supplies in the Pacific to test StarCraft-like supply drops

The Air Force Research Lab tested Rocket Cargo point-to-point delivery using a SpaceX Starfall capsule launched on a Falcon 9 Bandwagon-5 rideshare, splashing down in the Pacific with MREs, water, a SATCOM tracker, and batteries aboard. The ~10-foot carbon-fiber-heat-shielded capsule carries up to 2,200 pounds and uses compressed-gas thrusters and parachutes for controlled descent, with AFRL eyeing tens-of-tons capacity via Starship for ammunition, weapons, and vehicles. The capability aims to resupply forces globally without risking aircraft and aircrews, a priority against Pacific adversaries, while Sierra Space's Ghost reentry vehicle offers a competing option.

The takeAFRL's successful Starfall drop—a 2,200-lb capsule splashing down in the Pacific via Falcon 9 rideshare—validates point-to-point space resupply that could eliminate vulnerable airlift logistics in contested theaters like the Pacific; the roadmap toward tens-of-tons payloads hinges on Starship, with Sierra Space's Ghost as a competing reentry vehicle.

40Tech in AsiaSemiconductors

Hong Kong-listed IBS to acquire QuantumCompute for US$71.7M, expanding into AI chip design and edge computing

International Business Settlement Holdings agreed on October 30 to acquire all of QuantumCompute Solutions for HK$563 million (US$71.7 million), which indirectly holds ~85.94% of Shanghai Tianjing Fulian, a designer of low-power AI system-on-chip and edge-computing solutions. Foxconn Industrial Internet owns 5.56% of Tianjing Fulian, which is described as an authorized independent chip design house for MediaTek. IBS also appointed former Intel, Hua Hong, and SMIC executive Tang Tianshen as chief strategic consultant for its fintech businesses.

The takeThe HK$563M (US$71.7M) deal gives IBS control of Shanghai Tianjing Fulian, an authorized independent MediaTek chip design house with Foxconn Industrial Internet holding 5.56%, signaling continued consolidation in China-adjacent AI SoC and edge-computing silicon despite export-control headwinds.

41Wall Street JournalCybersecurity

The Morning Risk Report: Binance and Polymarket Suspend Trading After Hacks

Binance suspended deposits and withdrawals after a hack it initially pegged at ~$6.3B (later revised to $3B), with TRM Labs and Pangolin Intelligence attributing it to North Korea's TraderTraitor group; Tether and Circle froze ~$1.7B in stolen assets. Polymarket paused trading after a security flaw in third-party monitoring vendor DebugBear let attackers drain user accounts (a suspected attacker wallet showed >$2.5M), with the firm promising to make victims whole and referring the matter to UK law enforcement. Polymarket is separately probing suspicious well-timed trades on the Israeli election amid broader insider-trading scrutiny of prediction markets.

The takeA $6.3B (revised to $3B) Binance heist attributed to North Korea's TraderTraitor group, plus a supply-chain compromise via DebugBear's embed code draining Polymarket wallets, signals escalating state-linked crypto targeting and third-party vendor risk. The rapid Tether/Circle freezes of ~$1.7B show stablecoin issuers are now de facto incident responders.

42BloombergAI & Models

DeepSeek Drops Enhanced Coding Agent With Aim to Rival Anthropic

DeepSeek released a prototype open-source coding agent called 'Super-Agent' via its Harness team, designed to autonomously monitor developer group chats and suggest fixes without being summoned, aiming to rival Anthropic's Claude Code. The move follows DeepSeek's earlier formation of the Harness team to build AI agents competing with US labs, and comes as the company raises roughly 100 billion yuan ($15 billion) ahead of its IPO.

The takeDeepSeek's open-source 'Super-Agent' harness directly targets Anthropic's Claude Code, signaling China's push into autonomous coding agents backed by a ~$15B (100B yuan) fundraising round ahead of an IPO. The proactive, unprompted group-chat monitoring design is a concrete technical differentiator in the agentic coding race.

43Politico Tech & PolicyGeopolitics & Defense

How Merz’s five-day Asia trip ended up being mostly about Europe - POLITICO.eu

German Chancellor Merz toured Singapore, Japan and South Korea to pitch collective security and a rules-based trade alliance among middle powers, citing North Korean missile launches (~10 short-range ballistic missiles fired near the DMZ visit), Russia-North Korea tech transfers, and eroding international law. Asian capitals welcomed German defense cooperation but pushed back hard on Franco-German 'Made in Europe' car subsidy rules as protectionism, while a Japanese request for Patriot missiles for Ukraine fell flat and Tokyo's future Ukraine loan contribution remained uncertain.

The takeMerz's push for a 'middle powers' trade/security alliance collided with concrete friction: South Korea and Japan protested EU 'Made in Europe' car subsidy rules, and a Japanese diplomat bluntly asked 'How exactly will you help us if China attacks Taiwan?' — revealing that defense-industrial courtship (Airbus, Diehl, Quantum-Systems in tow) doesn't translate into strategic alignment.

44QbitAI (量子位)AI & Models

Tsinghua's Andrew Yao Team Unveils 'Algebraic World Model' to Make AI Think Like Newton

Tsinghua researchers led by Turing laureate Andrew Yao and Yang Yuan propose an Algebraic World Model that abstracts observed game objects and rules into an algebraic world, then uses LLMs to explore, condense, and formally prove hypotheses in Lean 4. On ARC-AGI-3 it clears all 183 levels at 99.9% with 43% of human action counts, and unlike baseline models it can formally prove a level is unsolvable and identify the minimal additional step needed. The work targets the gap between Tycho-style observation logging and Kepler/Newton-style law discovery, with acknowledged limits as a 'local truth' bounded by exploration coverage.

The takeThe team's AWM combines LLM-driven exploration with Lean 4 formal verification via category-theoretic projection/induction limits, achieving 99.9% on ARC-AGI-3 (183/183 levels) with only 43% of human-baseline actions and correctly proving impossibility in 2 attempts where GPT-6 Astra and Claude Opus 5.5 failed 7 of 12 times. This is a substantive research contribution toward verifiable, non-hallucinating world models rather than a product plug.

45Financial TimesGeopolitics & Defense

Military briefing: what would another US-Iran conflict look like?

Eight months into the US-Israel war with Iran, Trump and Netanyahu are threatening renewed strikes without having achieved stated objectives — Tehran retains control, has made no nuclear commitments, and can still hit US assets and allies. Options range from limited maritime 'mowing-the-lawn' strikes to a multi-day campaign against Pickaxe Mountain, IRGC commanders, and civilian infrastructure, but analysts warn broader targeting would trigger like-for-like attacks on Gulf energy infrastructure, Iraqi militia and Houthi escalation, and possible Iranian strikes on European bases and undersea cables. Depleted US interceptor and long-range missile stockpiles — with Tomahawk and PrSM replenishment taking one to five years — constrain escalation options, particularly given Pentagon concerns about a China contingency.

The takeThe piece quantifies a genuine strategic bind: US Patriot interceptor stocks have fallen below 1,000 from a prewar 2,330 and Thaad below 280 from 452, while 81 US aircraft have been lost or damaged and 21 personnel killed — meaning any renewed Iran campaign would strain munitions the Pentagon is husbanding for a potential China contingency. The explicit Iranian threat to seabed cables and data centers, plus possible strikes on US bases in Europe, elevates this from regional conflict to infrastructure and alliance risk.

46ReutersCybersecurity

Malaysia Airlines says customers may be affected by Amadeus breach

Malaysia Airlines disclosed that customer data processed between August 2023 and September 2025 may have been compromised via a cybersecurity incident at Amadeus, its reservation-technology provider, though its own IT systems were not breached and no financial data exposure was evidenced. The disclosure follows Cathay Pacific (53,000 bookings), Japan Airlines (92,000 customers), Asiana Airlines, and others reporting exposure from the same Amadeus vulnerability affecting a travel booking display environment. Amadeus declined to disclose the total number of affected customers.

The takeA single vulnerability in Amadeus's travel booking display environment has cascaded across at least five major carriers, with Cathay Pacific reporting 53,000 bookings and Japan Airlines 92,000 customers affected — a supply-chain concentration risk in airline reservation infrastructure that scales far beyond any single carrier's security posture.

47Hacker NewsCybersecurity

Leaking Secrets from Air-Gap Computers by Spelling Radio Signals from RAM

Researchers present an attack in which malware on a compromised air-gapped computer generates radio signals directly from memory buses (RAM), encoding files, images, keylogging data, biometrics, and encryption keys. An attacker with software-defined radio hardware and a simple antenna intercepts and decodes the raw emissions back into binary. The paper details design, implementation, and fast modification methods achieving 1000 bits per second, and proposes countermeasures against the out-of-band air-gap threat.

The takeRAMBO demonstrates a software-only exfiltration channel from air-gapped systems at 1000 bits/sec by modulating RAM bus emissions into decodable radio signals, undermining the core assumption that physical network isolation prevents data leakage. This is a meaningful escalation in the air-gap covert-channel literature because it requires no special hardware on the compromised host—only malware and an off-the-shelf SDR receiver at range.

48ServeTheHomeSemiconductors

NVIDIA Updates Rubin GPU Specs as First Production Hardware Arrives

As Vera Rubin hits production ramp, NVIDIA's locked shipping specs reveal meaningful regressions from preliminary numbers: per-GPU HBM4 bandwidth falls to 19.2TB/s (9.375Gbps/pin, still above AMD's 7.6Gbps/pin MI455X), NVLink drops to 3.0TB/s, and NVL72 rack memory bandwidth falls 1,580→1,400TB/s with scale-up 260→216TB/s. HGX Rubin NVL8 takes the deepest cut, losing 300GB of HBM4 (one stack) and ~26% memory bandwidth, with FP4/FP8 training throughput also trimmed. Compute targets (50 PFLOPS FP4 inference) and 288GB capacity per full GPU remain intact.

The takeNVIDIA quietly cut shipping Rubin specs versus preliminary figures: HBM4 bandwidth down ~11% (22→19.2TB/s per GPU), NVLink scale-up down 17% (3.6→3.0TB/s), while scale-out rose 0.4→0.45TB/s — a signal that HBM4 stack yields/controller speeds couldn't hit aggressive 10.75Gbps/pin targets, with NVL8 losing a full HBM stack (2.3→2.0TB, ~26% bandwidth cut).

49Tech in AsiaInfrastructure

India’s MeitY unveils guide to train AI without sharing data

Bharat1.ai has opened the first phase of a from-scratch AI-native city in Bengaluru, a 500,000 sq ft facility already hosting ~25 startups with 400 Gbps connectivity to Neysa, E2E, AWS, and GCP for LLM training and fine-tuning. The full multi-phase project is estimated at ~$650M, with Nvidia as a core technology collaborator, university partners including IIT Kanpur, IIT Ropar, IIT Bombay, and BITS Pilani, and a $100M fund slated for H2 2026. Energy strategy hinges on a hybrid solar plus nuclear microreactor/SMR model, with the city targeting 25,000 researchers and thousands of humanoid robots by 2030.

The takeBharat1.ai's B1 AI Superpark in Bengaluru is a $650M, 500,000 sq ft AI-native city with 400 Gbps links to four AI clouds (Neysa, E2E, AWS, GCP), Nvidia as core GPU collaborator, and a planned hybrid solar/SMR energy model — a concrete sovereign-AI compute and ecosystem play targeting 25,000 practitioners by 2030.

50Tech in AsiaAI & Models

DeepSeek builds ‘super-agent’ for autonomous coding

Hangzhou-based DeepSeek said its V4.1-Flash model uses less key-value cache than the prior generation, reducing cache-related costs in AI agent workloads. Partners including CodeBuddy and OpenCode now support the model via the deepseek-flash setting, and DeepSeek plans to work with the open-source community on inference support and explore large-scale setups with 2,000 GPUs and a storage cluster. The company said the update lets it serve more users at lower cost and that it will pass savings to users.

The takeDeepSeek's V4.1-Flash cuts KV-cache overhead to lower inference costs for agentic coding workloads, and the disclosed plan to explore large-scale deployments of 2,000 GPUs plus a storage cluster signals a concrete push toward cheaper, higher-throughput frontier-model serving.