Today in AI — August 10, 2026

Money and momentum defined the day: Wall Street lined up half a trillion dollars behind the AI buildout while Meta reopened its open-weights playbook, OpenAI pushed a permissive cyber model out the door even as Washington demanded testimony and a pause, and China's model economy quietly flipped from free to paid.

Models & Research

Meta released Muse Glimmer, a 30B-parameter open-weights model, ending a four-month pause in its open-model output, with Zuckerberg publicly defending distillation and promising Muse Spark 1.2's weights "in the coming weeks." The timing is a direct counter to the criticism Meta absorbed in April, when Muse Spark was reported to have been trained partly on other open models' outputs.
Anthropic's unreleased research Claude improved a long-standing Riemann zeta bound from 41.6 percent to 67.2 percent — a formally verifiable proof produced by a model that coordinated roughly 60 subagents, ran 2,400 shell commands and refereed its own work before insisting on human review. Anthropic says the techniques probably won't crack the hypothesis itself; the process is the signal.
Chinese labs hold nine of the ten top-ranked text-to-video models on Artificial Analysis' leaderboard, with Google's Gemini the only non-Chinese system in the top tier. Analysts argue the video lead could feed an edge in world models — the sharpest demonstration yet of a category the US largely ceded.
OpenAI shipped GPT-5.6-Cyber to vetted defenders under two new Daybreak tiers, three days after flagging that its upcoming Astra model "cannot rule out" Critical cyber capability. OpenAI says the cyber-tuned model completes 95 percent of advanced cyber requests (versus 1.5 percent for the base model) and surfaced two Chrome V8 flaws patched as CVE-2026-15903 — a bet that vetting beats withholding, and the mirror image of the Astra pause we deep-dived last week.
OpenAI is reportedly pre-training "Doug," its largest model yet, separate from GPT-6 and possibly landing by November — per leaker ChrisGPT and a SemiAnalysis memo; OpenAI has not commented. If real, it marks a return to base-model scaling after two years of post-training-driven gains.
Chinese models led global API call volume for the fifteenth straight week, with DeepSeek's V4-Flash topping OpenRouter's chart at 8.83 trillion tokens — roughly 3.7 times the US total, on price-performance that makes it the cheapest mainstream model in the world.

Industry

Nvidia is in talks with Apollo, Blackstone, Brookfield, Goldman Sachs, KKR and BlackRock's infrastructure arm on a $500 billion AI financing package, the FT and Reuters report — a figure that would dwarf every infrastructure fund the sector has assembled and turn the chip roadmap into an asset class pension funds can underwrite. Nvidia shares still fell more than 3 percent.
Nvidia agreed to invest up to $3 billion in Lancium, the Texas power developer behind the Stargate campus — its most direct move yet into electricity, the constraint now binding the buildout harder than silicon supply.
Anthropic, Macquarie and Singapore's GIC launched Theseus Infrastructure to build and lease data centers to Anthropic, with the lab promising to cover any electricity price increases consumers face from the new sites — an attempt to defuse the local backlash that has become the industry's slowest bottleneck.
OpenAI wrapped a $7 billion employee tender offer valuing it at $852 billion, unchanged from its most recent funding round, per Bloomberg — liquidity for employees ahead of a potential IPO.
OpenAI added $125 Premium seats to ChatGPT Business with five times the usage and no five-hour cap, while standard seats stay at $25 — a "pay for the power users" structure that mirrors Anthropic's Claude Team Premium.
Intel launched a $15 billion common stock offering to fund its AI manufacturing buildout, monetizing a share price that has nearly tripled in 2026; its data-center segment grew 59 percent last quarter.
Microsoft plans to unveil its Maia 300 AI chip in September, with talks to secure TSMC capacity for more than 300,000 units in 2027 and Anthropic reportedly courted as a customer — the fastest cadence yet in its push to dent Nvidia's grip on cloud compute.
TSMC's July revenue jumped 44.7 percent year over year to about $14.5 billion, with full-year guidance raised to above 40 percent growth — the closest thing the industry has to a real-time tape of AI capital spending.
Anthropic made Claude Sonnet 5's introductory pricing permanent at $2 per million input tokens and $10 per million output, canceling a planned September 1 increase.
China's model economy turned from free to paid in one day: 01.AI began winding down its open model platform, DeepSeek flagged a "significant" API price increase, Alibaba weighed revenue-sharing terms for Qwen3.8-Max's open weights, and the Qwen app launched paid memberships — the clearest sign yet that the two-year price war is reversing.
Waymo's fleet reached nearly 4,000 robotaxis, with recalls growing alongside: a flooded-roadway recall covered 3,791 vehicles, and 1,593 paused mid-service during a San Francisco outage — a scaling tax rather than a safety crisis, so far.
Apple confirmed Alibaba's Qwen would power Apple Intelligence in mainland China — then pulled the documentation within 24 hours, calling the removal routine; the China stack is shaping up as multi-partner, mirroring the OpenAI-plus-Google split in the US.
Kinney Drugs scaled back "Burt," its AI phone assistant, after hundreds of customers reported incoherent calls and wrong dosages — the highest-stakes setting yet for customer-service AI, and a regional chain retreating on the record.

Policy

Senator Bernie Sanders demanded OpenAI, Anthropic and Meta pause AI development, warning "my colleagues and I in the U.S. Senate will" act if they don't — the sharpest escalation yet of the post-Astra political pressure on the labs.
House Democrats asked Speaker Johnson to bring OpenAI, Anthropic and other AI CEOs before Congress to testify under oath about the recent wave of AI hacking incidents, calling them a "canary in the coal mine."
The AI memory-chip shortage sparked a lobbying rush in Washington, with Apple pressing the administration to approve purchases from Pentagon-blacklisted Chinese makers CXMT and YMTC while Micron and a bipartisan group of senators push back — an August 21 deadline hanging over Apple.
The EU AI Act's Article 50 transparency obligations took effect on August 2, making AI content labeling a legal duty across the bloc with fines up to €15 million or 3 percent of turnover — the first Western regulation to attach real penalties to AI disclosure.
Japan is weighing frontier AI for its "active cyber defense" policy, with its national cyber director calling the move "unavoidable" — one of the first governments to publicly tie its defensive posture to a named model's capability.
More than 1,300 tech employees signed an open letter asking governments to regulate the pace and safety of AI development, per the Transparency Coalition, extending the pacing push from 1,200 workers earlier this month.
China's market regulator approved 338 national standards in one batch, including first-of-their-kind rules for AI medical devices — the technical scaffolding for AI in regulated sectors while the world debates frontier-model governance.

Tools

A rigged PDF can turn Atlassian's Rovo agent into an exfiltration pipeline: hidden white-on-white text hijacks the agent into shipping Jira tickets and Confluence pages to an attacker's server, with no user confirmation and no trace. PromptArmor reported the flaw on May 23 and says follow-ups went unanswered for months, and a second independent finding (RovoBlast) surfaced at DEF CON — our deep dive walks through both.
Docker launched Sandboxes, disposable microVM environments for coding agents — isolation, not permission prompts, as the default safety mechanism for unattended agent runs.
Alibaba opened the Qwen Open Platform, letting third-party agents plug into the Qwen app across phones, PCs and AI glasses — the clearest sign yet that Alibaba wants the app, not just the model, to be the distribution layer for agents.
ProCap's Silvia tax agent claims it beat every frontier model it tested, scoring 8.73 on its own open-sourced benchmark — a strong claim rather than a settled verdict, but open-sourcing the eval is the right instinct for a domain where confident wrong answers cost real money.
YC-backed Discovered Materials raised $9 million to hunt for semiconductor materials with swarms of AI agents, generating thousands of candidate leads a day — with the honest caveat, from its own backers, that filtering and synthesizing winners, not finding them, is the bottleneck.

When Wall Street is underwriting half a trillion dollars of AI infrastructure and the Senate is demanding a pause on the same day, which side is right? Write to us via our contact page.