The Omniscient Bulletin · 2026-07-28
The Omniscient Bulletin — July 28, 2026
Nvidia buys into Sutskever's lab, China starts building the machines it was blocked from buying, and Anthropic's chip argument meets its counterexample
Good morning! Monday put an argument and its rebuttal in the same news cycle. Anthropic published its case that China cannot build better models than the US without US chips, because its domestic production capacity is limited. Hours later, a state-backed Chinese group began mass-producing the lithography machines it has been barred from buying, and ASML fell to its lowest since June. Elsewhere the money kept moving: Nvidia put $5 billion into Ilya Sutskever's lab, which has never shipped a product. Microsoft shipped a security model it says finds flaws at half the cost. And Anthropic spent the day explaining that shared Claude conversations were turning up in Google. Here's Monday.
Industry
Nvidia puts $5 billion into Sutskever's lab, which has never shipped a product
Nvidia and Safe Superintelligence announced a partnership Monday morning built on a $5 billion equity investment. SSI also gets access to Vera Rubin, Nvidia's next-generation platform, in a deal both sides say will raise its computing capacity roughly tenfold over the next 12 months. Ilya Sutskever's lab had already raised roughly $7 billion, including from Andreessen Horowitz and Alphabet, and it has released nothing and says it does not intend to before it is finished. Nvidia is now funding both the largest buildout in the industry and the lab that refuses to productize.
Compute
China starts mass-producing the chipmaking machines it was barred from buying
A state-backed group of Shanghai Yuliangsheng, SiCarrier and Huawei has begun mass production of immersion DUV lithography systems, Bloomberg reported Monday. The plan is modest in units and not in meaning: about 5 machines delivered in 2026 and 20 in 2027, going to SMIC, Hua Hong and ChangXin Memory. ASML fell more than 6 percent to its lowest close since June. These are not EUV machines, and on throughput, overlay precision and reliability they sit a generation behind. They are also the tool export controls were built to withhold.
Policy
Anthropic says it never asked for an open-weights ban, and does want China cut off from chips
Anthropic published its position on open-weights models Monday, with Dario Amodei writing that the company has never advocated for a ban on them. Banning US businesses from using the models does nothing about the risk, he argues, because bad actors are unlikely to be legitimate US businesses. What Anthropic does back is the chip line: do not sell powerful chips or chipmaking equipment to China, and fight smuggling. The reasoning is that China's limited domestic production means it cannot build more powerful models than the US without US chips.
Research
Microsoft's new security model claims 96 percent on the standard benchmark at half the cost
Microsoft introduced MAI-Cyber-1-Flash on Monday, a compact code-heavy model from its MAI-Thinking-1 lineage that runs inside MDASH, its multi-agent vulnerability finder. It claims 96 percent on CyberGym, 12 points above Mythos, and a 50 percent cost saving against the GPT-5.4 stack it currently uses, on the theory that a small model can handle 90 percent of the work and the expensive ones only get the last 10. The demand is not hypothetical: NIST says CVE submissions to its vulnerability database ran nearly a third higher in early 2026 than the same period last year, with no sign of slowing.
Policy
Shared Claude conversations turned up in Google and Bing results
Conversations and Artifacts shared through Claude's link feature were being indexed by Google and Bing, Wired reported Monday. Anthropic's robots.txt has disallowed crawling shared chats since at least September 2025, but the company never added the noindex tag search engines require once a page is linked elsewhere, so both indexed hundreds of threads anyway. Reported contents included legal questions, engineering work, resumes and in some accounts credentials. Every one had been deliberately shared by its author, which is the distinction that matters and also the one nobody was told to think about.
Compute
Fifty Louisiana officials signed NDAs to land Meta's $50 billion campus
The New York Times reported Monday on how Meta's Hyperion campus in Richland Parish came together: months of confidential negotiation among Meta, Governor Jeff Landry's administration, local officials and the utility Entergy, with more than 50 state officials signing nondisclosure agreements while Louisiana raced to win the investment. Landry defended the secrecy, saying speed was necessary and that more transparency could have cost the state the deal. The $50 billion project turned farmland into some of Louisiana's most valuable ground before residents knew it was coming.
Labor
Big employers start hiring again, and blame the limits of AI
Companies from CSX to Alphabet told investors in recent days that they plan to add staff, the Wall Street Journal reported Monday, reversing a year of AI-era messaging about doing more with fewer people. Some executives said the costs and limitations of the technology now require more humans, not fewer; others are rehiring after layoffs. Alphabet's finance chief said the company expects to keep hiring in AI and cloud. The head of staffing firm Robert Half put it plainly: AI's effect on the job market is proving more benign than some feared.