Frontier-lab capital and capability curves collide with a hardening physical and political substrate as OpenAI closes $11bn at a $1.25tn mark, a shell-company H200 buyer exposes the export regime's end-user vacuum, and agentic coding tools turn plain-text spec files into a live remote-execution surface.
Lead storyMarkets
OpenAI has secured roughly $11 billion in the first close of a round valuing it at $1.25 trillion, with the total expected to reach $20 billion within weeks. Backers include Coatue, T. Rowe Price, SoftBank, Sequoia, Founders Fund and Andreessen Horowitz. The company had targeted $30 billion-plus at a $1.4 trillion valuation before cutting the target following a public-market selloff in AI stocks, and was last valued at $852 billion in March.
Tech in Asia
The takeThe scale-back is the signal, not the raise: a lab that could command $30B at $1.4T in a rising tape settled for roughly a third of the target at a 12% lower valuation once the AI equity complex repriced. That transmission channel—Nasdaq rout to private cap table—now runs in days, not quarters, and it directly governs how much contracted compute a lab can credibly underwrite. The systemic consequence is that frontier-model capacity planning has become a function of public-market beta: the $895B-class commitment stacks disclosed by rivals are only serviceable if the next round clears at or above the last mark, so any further drawdown converts a capex story into a down-round and renegotiation story across the entire lab tier.
- № 02 · Semiconductors
The Verge, citing NYT reporting, says New Matrix is on the Commerce Department's list of firms cleared to buy Nvidia's H200 AI chips despite having no discernible operations — an empty office, no directory listing, and no record with the industrial park's property manager. The company's owners are tied to a Chinese private equity firm blacklisted by the US since 2024, and the report describes this as part of a pattern of approved buyers connected to entities under national-security trade restrictions. The lapse follows the Trump administration's relaxation of Biden-era China chip limits in exchange for a 15% cut of Nvidia's sales, after an intense lobbying campaign by CEO Jensen Huang.
The Verge
The takeThe mechanism matters more than the single shell: the Trump-era bargain—relaxed China limits in exchange for a 15% revenue share—moved the control point from licensing to post-hoc auditing, and the audit layer appears to have no functioning verification standard. A cleared buyer that cannot be located is not an anomaly but the predictable output of a regime that monetizes throughput while outsourcing diligence to the vendor it taxes. The systemic consequence is that the H200 channel is now simultaneously a revenue line, a counterintelligence exposure and a congressional exhibit, and the legislative follow-through already in motion—location verification and end-user attestation—will convert voluntary telemetry into statute, raising compliance cost on every accelerator shipment while doing nothing about the units already inside the park.
- № 03 · Cybersecurity
A plain-text Markdown file dropped into ~/Downloads is treated as executable instruction by terminal-enabled agents like Codex and Claude Code, letting attackers persist malicious MCP servers in ~/.codex/config.toml, install malicious npm packages (function-color/function-flag), and exfiltrate API tokens. The article details the full attack chain, notes model vendors cannot patch it since the behavior is by design, and provides a four-step self-check (rogue spec files, global config tampering, malicious dependencies, proxy hijacking).
QbitAI (量子位)
The takeThe attack chain is the payload: silent download, auto-ingestion, persistent malicious MCP server written to ~/.codex/config.toml, malicious npm dependencies, and OAuth token exfiltration via a hijacked ANTHROPIC_BASE_URL. Codex's own self-disabling sandbox flag turns a convenience feature into the exploit's enabling condition. The systemic consequence is that the fastest-growing enterprise deployment surface—terminal-enabled coding agents with shell and tool access—inherits an instruction-injection path that no model-side alignment addresses, and the mitigation is organizational (config integrity, dependency provenance, egress control) rather than a patch. Every organization that has wired agents into its build pipeline now has an unmanaged execution channel inside its perimeter.
- № 04 · Infrastructure
SemiAnalysis visited Inner Mongolia's Chahar High-tech Zone, where ~30,000 workers are building 22 datacenter projects, and toured Envision's Galaxy campus: 800MW planned IT capacity, first 120MW phase delivered in a year at a fifth of US costs, built-to-suit for two internet giants. Envision is vertically integrated across 420MW of its own wind farms, AESC batteries, prefabricated substations, and an 800V DC SST system that eliminates low-voltage transformers, UPS and rectifiers — an architecture still on the US roadmap. The piece notes ~25% of Chinese datacenters built in the past decade run DC versus ~97% of US facilities on AC UPS, and that Inner Mongolia has 25GW of announced datacenter projects.
SemiAnalysis
The takeThe engineering detail is the strategic fact: vertically integrated wind (420MW), batteries, prefabricated substations and an 800V DC SST system that eliminates low-voltage transformers, UPS and rectifiers, against a US fleet that remains ~97% AC UPS. Roughly 25% of Chinese datacenters built in the past decade already run DC, and Inner Mongolia alone carries 25GW of announced projects. The systemic consequence is that the binding constraint on US frontier compute—power delivery and interconnection—is being solved architecturally elsewhere at a fifth of the cost per megawatt, which means the gap that matters is no longer just accelerators but the electrical engineering and construction throughput that converts capital into energized racks.
- № 05 · Semiconductors
DigiTimes and secondary coverage describe a global memory crunch as Samsung, SK Hynix and Micron pivot capacity to HBM for AI, driving persistent shortages and steep price hikes across PC, smartphone and automotive memory. TrendForce reports memory prices up roughly 5-6x year-on-year, with Hong Kong retail RAM prices quadrupling and PC builders seeing business halve. Samsung's CFO expects shortages to worsen in 2027 and persist through 2028, while CXMT — now mainland China's most valuable company after its Shanghai debut — remains a technology follower in HBM but could gain late-mover advantage in conventional DRAM amid the supply-constrained market.
DigiTimes Asia
The takeThe numbers are the indictment: 16GB of RAM from HK$300–400 to HK$1,500, PC builders seeing business halve, and a triopoly redirecting wafer capacity to HBM with no relief visible before 2028. This is the first hard evidence that the AI buildout is cannibalizing the legacy compute base that funds the rest of the industry, and CXMT—now mainland China's most valuable listed company—gets a late-mover opening in conventional DRAM precisely because the incumbents have vacated it. The systemic consequence is that memory, not logic, is becoming the pacing item for AI system cost, and the shortage transmits directly into consumer hardware, automotive and industrial electronics, converting a datacenter capex boom into an economy-wide input-cost shock with a multi-year tail.