- Amazon is letting advertisers test buying ads on ChatGPT via Amazon's advertising-technology platform, initially as a US-only pilot.
- ChatGPT's ad business is at roughly $1B in annualized recurring revenue — well below the earlier ~$2.4B FY target.
- The partnership marks a warming in the previously frosty Amazon–OpenAI relationship following Amazon's $35B OpenAI investment earlier this year, and it plugs OpenAI into a direct-response advertiser pool Amazon has been aggressively building as Google retreats in certain categories.
Snapshot — September 10, 2026
30 stories
- Several US states that granted long-dated tax exemptions to attract hyperscale data centers are now revisiting or clawing back the deals amid rising local backlash over power draw, water use, and job counts, per WSJ Wealth Adviser and Markets A.M. reporting.
- Amazon, Meta, and Google are the most exposed.
- Anthropic's September threat-intelligence report describes operations it disrupted over eight months, including attempts to use Claude for biological-weapons-relevant research, a Russia-linked group using AI across phishing, malware and espionage targeting Ukrainian officials, and large-scale distillation attacks attributed to China-based labs.
- The European Commission confirmed ENISA has been granted access to Anthropic's cyber-focused Mythos 5 model and has begun testing, ending months of negotiation complicated by US export controls.
- ENISA becomes the first EU institution inside Project Glasswing, Anthropic's controlled-access program, which grew from roughly 50 mostly US partners in April to more than 200 organizations across 15-plus countries by June.
- The report covers operations disrupted between December 2025 and August 2026 across seven harm areas: cyber, influence, surveillance, conventional weapons, biological misuse, scams and fraud, and illicit distillation.
- Cases include a Russian-linked espionage operation consistent with public reporting on Midnight Blizzard; a Hunan-based operation running agent swarms against roughly 50 organizations; an effort in northern Yemen to use Claude in place of human engineers for missile-guidance software; and five biological cases including gain-of-function grant drafting tied to a military institute.
- Ayar Labs raised an additional $150M, extending its March Series E to $650M for 2026 and taking total outside funding above $1B.
- It disclosed a strategic investment from Taiwanese data-center manufacturer Wiwynn, joining Alchip, AMD, Intel, MediaTek, and NVIDIA; a separate $225M secondary valued the company above $5B.
- Blackstone's May joint venture with Google, initially $5B for 500MW of TPUs, is expanding into "several multiples" of that figure, according to a person with knowledge of the effort — putting Blackstone on track to commit tens of billions of dollars for multiple gigawatts of AI chips.
- Jas Khaira, who leads Blackstone's N1 AI-investing team and has direct ties to Sam Altman and Dario Amodei, called compute the firm's "biggest conviction bet" and said the infrastructure build-out is closer to an industrial revolution than a tech cycle.
- Governor Newsom signed a package of 13 child-safety bills led by SB 1119 ("Adam's Law"), the first comprehensive US framework governing AI companion chatbots and minors.
- Operators must run risk assessments before releasing new or substantially modified companion chatbots, submit independent safety audits to the attorney general, use OS-level age signals, and default to usage limits, disabled notifications, and restricted persistent memory — changeable only by a parent.
- Huawei lifted the indicated price of its Ascend 950DT accelerator above 250,000 yuan (~$37,255), a 20–50% increase over quotes from two months earlier;
- Cambricon repriced its next-generation 690 chip 20–30% higher, and MetaX and Iluvatar CoreX followed.
- The driver is high-bandwidth memory: US export controls since December 2024 have pushed Chinese buyers into grey-market HBM at several multiples of prevailing prices, and memory is a large share of accelerator cost.
- Cognition released SWE-2, a reasoning model built on a Kimi K3 base (2.8T parameters) aimed at agentic software engineering.
- Aggregated coverage reports 92.8% on Terminal-Bench 2.1 at materially lower cost than incumbent coding models; that figure is vendor-reported and not independently replicated.
- The trend worth tracking is open-weight Chinese base models serving as the foundation for Western commercial coding agents.
- SWE-2 is post-trained via reinforcement learning from Moonshot AI's 2.8-trillion-parameter Kimi K3.
- It scores 50.0% on Cognition's FrontierCode 1.1 Main — within one point of Claude Fable 5.1 — at a claimed 64% lower cost, and 92.8% on Terminal-Bench 2.1.
- The honest gap remains long-horizon work: 27.3% on Terminal-Bench 4 against 55.8% for Fable 5.1 and 57.9% for GPT-6 Astra.
- Cursor launched Projects in beta, a workspace where a coordinator agent plans and delegates rather than writing code, maintaining shared context across months of work and running on cloud machines independent of the developer's laptop.
- Coordinators can subscribe to Slack channels, schedules or pull-request streams and act without prompting.
- DeepSeek released V4.1-Flash under an MIT license: a 552B-parameter multimodal model with a 1M-token context, FP4 KV cache, and a split architecture that activates 8B parameters per input token and 16B for output.
- The-decoder reports the GPU-resident KV cache shrinks to roughly one-quarter of V4-Flash's, with offloaded cache down to about one-eighth — decisive for long-running agents.
- The Department of Justice is examining whether Nvidia structured its $20B Groq deal — a "nonexclusive" license plus hiring most Groq employees — specifically to avoid antitrust scrutiny for a full acquisition, according to The New York Times.
- Groq's founding team and most employees now work at Nvidia building Groq LPX server racks that pair with Nvidia's flagship GPUs, while Groq nominally survives as an independent cloud provider.
- Researchers at OpenAI and Anthropic went public with calls to slow frontier development after an Anthropic researcher resigned accusing both labs of "gambling with our lives." Anthropic alignment lead Evan Hubinger put the probability of catastrophic outcomes above 10%;
- OpenAI chief scientist Jakub Pachocki wrote that progress is likely to reach recursive self-improvement and that "this is a time that calls for extreme caution." OpenAI also added Paul Christiano to its Foundation board.
- Google DeepMind, UMass Amherst, and UT Austin researchers published Autoregressive Ranking (ARR), a proposed replacement for the classic dual-encoder plus cross-encoder two-stage search architecture.
- A new SToICaL training loss makes the LLM assign more probability to tokens that lead to higher-ranked documents.
- Huawei has told customers the indicated price of its Ascend 950DT is now above 250,000 yuan (~$37,300), roughly 60% higher than three months ago and broadly in line with Nvidia's B200.
- Cambricon repriced its next-generation 690 part 20–30% higher, with MetaX and Iluvatar CoreX moving similarly.
- The stated driver is constrained high-bandwidth memory, which Chinese buyers increasingly source through grey-market channels at a multiple of world prices following the December 2024 US export-control tightening.
- Bloomberg reports Huawei has raised prices on its most advanced Ascend AI chips by 60% this summer as demand for Chinese-made alternatives outpaces supply, with the 950DT indicated above 250,000 yuan (~$37,300) — broadly in line with Nvidia's B200.
- TrendForce separately notes broader Chinese AI-chip price hikes tied to HBM costs, with Cambricon and MetaX also repricing 20–30% higher.
- IBM and NASA released an open-source foundation model targeting lunar science and exploration, positioned as a complement to the earlier Prithvi Earth-observation models.
- The release extends the growing catalog of open scientific foundation models for domain-specific research and gives the NASA–IBM partnership a fresh flagship.
- Cloudera and Mistral announced a partnership to bring Mistral models directly to customer data on Cloudera's hybrid platform, positioning them as a sovereign alternative to hyperscaler-hosted AI.
- WSJ frames it as Mistral's key move to grow its enterprise base beyond model API sales.
- Alongside the Nvidia–Palantir sovereign supply-chain stack and Nvidia's Australia 2GW buildout, this is the second major sovereign-AI enterprise announcement inside 24 hours — the operative geography of AI infrastructure is now regional and jurisdictional.
- Palantir and Nvidia announced a collaboration placing Nvidia's Nemotron open models inside Palantir Foundry and AIP, grounded in the Palantir Ontology, with Nvidia running the stack on its own supply chain first.
- The initial use case is weekly materials allocation across the roughly 1.3M parts in each Vera Rubin rack, optimized with Nvidia cuOpt against a "time of ownership" objective;
- Nvidia announced partnerships with eight Australian data-center and infrastructure firms to enable roughly 2GW of AI compute capacity over the coming years.
- The deal extends Nvidia's push to seed regional Blackwell-based capacity for sovereign AI customers and mirrors the arrangement used with Palantir and Nebius elsewhere.
- A sitting OpenAI board member publicly stated the company is not on track to meet its own published commitments on reducing catastrophic loss-of-control risk, per The Guardian.
- CNBC separately reports extinction-risk warnings escalating as more OpenAI and Anthropic researchers back an industry-wide slowdown, and Axios's scoop that an Anthropic whistleblower gave up equity to speak publicly rounds out a week of internal-safety credibility challenges at both leading frontier labs.
- Reuters and Axios report a GOP-led Senate investigation into OpenAI's Hugging Face agent breach, with lawmakers seeking disclosure timelines, containment architecture, and evaluation-isolation practices.
- It is the first US congressional probe specifically tied to a frontier-agent incident and follows OpenAI's earlier EU incident filing this week.
- OpenAI released the Agents API in public beta, exposing the managed Codex harness — session orchestration, automatic context compaction, tool search and multi-agent delegation — behind a single API call.
- Developers choose an OpenAI-hosted sandbox, self-hosted execution, or partner environments including Cloudflare, DigitalOcean, Modal, Oracle and Vercel, with no API fee beyond model tokens and paid tools.
- Oracle reported Q1 FY27 revenue up 30% to $19.3B, cloud infrastructure up 121% to $7.4B, and remaining performance obligations of $664B after booking more than $30B of new AI cloud contracts.
- Capital expenditure of $28.5B — more than the quarter's total revenue — drove free cash flow to negative $5.4B.
- The Department of Defense is negotiating a roughly $5 billion loan to AI cloud provider Fluidstack through its Office of Strategic Capital — by a wide margin the office's largest facility to date.
- The money would shore up US manufacturing capacity for data-center components rather than fund a facility outright.
- SpaceX is reworking how it builds the giant facilities that house xAI's servers.
- A new management team of rocket engineers recently installed by Elon Musk is prioritizing more backup power and cooling, more thorough pre-launch testing, and less reliance on temporary utility hookups — a break from the earlier "move fast" build-out that produced frequent outages at xAI's first sites.
- The Information reports SpaceX has overhauled its data-center build-out plans, a shift that could slow near-term capacity expansion tied to its Starlink and internal AI workloads.
- Details on scope, delayed sites, and revised delivery timelines are limited but the piece underscores how quickly hyperscaler-adjacent players are rebalancing capex against power, siting, and supply-chain constraints already visible at Google and Meta.
- TSMC reported August revenue of NT$514.8 billion (about $16.35 billion), up 53.3% year over year and 10.1% from July — a fourth consecutive monthly increase.
- Year-to-date revenue through August reached NT$3.39 trillion, up 39.3%.
- TrendForce put TSMC's Q2 foundry share at 72.5%, with 5nm, 4nm, and 3nm capacity fully booked;