The Omniscient Bulletin · 2026-07-20
The Omniscient Bulletin — July 20, 2026
Xi pitches open-source AI in Shanghai, Apple briefly tops Nvidia, and Anthropic looks to rent compute from its rival Meta
Good morning! The weekend had a Shanghai theme. At the World AI Conference, Xi Jinping used his first keynote to cast China's open models as a global public good, Alibaba answered with a 2.4-trillion-parameter model it says trails only Anthropic's best, and the show floor filled with working humanoids. The counter-current ran through the United States: markets briefly crowned Apple over Nvidia as AI-spending doubts grew, Anthropic moved to rent computing power from its rival Meta, and Washington sketched a Wall-Street-style regulator for frontier models. Here's the weekend.
Policy
Xi makes his first AI-summit keynote a pitch for China's open-source order
For the first time in the World AI Conference's nine-year history, Xi Jinping gave the opening keynote in Shanghai on Friday, saying AI development "should not be a solo performance by a single country but a symphony of international cooperation," and urging the world to embrace open source. He paired it with pledges: 5,000 AI training slots for developing nations over five years, joint AI centers with blocs from ASEAN to the African Union, and access for 30 countries to China's MAZU weather model. It was Beijing's clearest bid yet to set global AI's terms around openness, against Washington's closed models.
Frontier
Alibaba shows a 2.4-trillion-parameter Qwen and says only Fable 5 beats it
At the Shanghai conference Alibaba unveiled a preview of Qwen3.8-Max, a 2.4-trillion-parameter model it calls "comparable to leading frontier AI models, second only to Fable 5," Anthropic's strongest. The caveat: Alibaba published no benchmarks and no date for the open weights it promises, so the ranking rests on its word. The preview is live through Alibaba's subscription and coding tools at a fraction of standard pricing, and its shares jumped as much as 5 percent on the news. It is another Chinese lab pitching an open-weight model at the American frontier.
Robotics
China's robot makers turn the Shanghai floor into a showroom for working humanoids
The conference's second act was embodied AI. AgiBot, one of China's best-funded humanoid startups, debuted four machines led by the A3 Ultra, spec'd at 1.74 meters and 60 kilograms with 51 joints, a five-kilogram payload per arm, eight hours of runtime and Nvidia's Thor chip onboard. More than 30 of its robots worked the floor guiding visitors, beside a wider fleet from Unitree, Fourier and others. The pitch has moved from parameter counts to labor: Chinese firms are racing to put humanoids on factory lines and in service jobs, chasing the same market as Tesla and Figure.
Industry
Apple briefly passes Nvidia as the market rethinks who wins from AI
For a few hours on Friday Apple overtook Nvidia as the world's most valuable company, touching about $4.88 trillion as Nvidia shares fell as much as 3.9 percent, cutting its market value to roughly $4.82 trillion, before both reversed by the close. Nvidia had held the top spot since June 2025, and Apple was long dismissed as an AI laggard for building no models of its own. With Wall Street uneasy about the payback on AI spending, money rotated toward the company selling AI to consumers over the one selling the picks and shovels. Apple is up 23 percent this year, Nvidia just 7.
Compute
Anthropic looks to rent computing power from Meta, a company it competes with
Anthropic is in early talks to lease as much as $10 billion of data-center compute from Meta over two years, in monthly installments, according to reporting confirmed Friday. Nothing is signed, and either side can still walk. The oddity is who sits on each end: Meta builds the Llama models that rival Anthropic's Claude, yet would become its landlord, renting out the infrastructure Mark Zuckerberg spent tens of billions to build. It shows how far the scramble for chips has gone, with a leading lab willing to rent compute from a direct competitor.
Policy
Washington weighs a Wall-Street-style regulator to vet the biggest AI models
The Trump administration is weighing an independent agency to vet frontier AI models, modeled on FINRA, the self-regulator for US brokerages, and reporting to the Securities and Exchange Commission, Bloomberg reported Friday. Developed with Treasury Secretary Scott Bessent and under review by chief of staff Susie Wiles, it would have labs voluntarily submit their top models about 30 days before release, to screen for cyber, bioweapon and deception risks. It is the firmest sign yet that a light-touch government is designing a standing federal gatekeeper for AI.
Research
Open models are only months behind the top labs at hacking, Britain finds
The UK's AI Security Institute measured how far open-weight models trail the best closed ones at running cyberattacks on their own, and found the gap has shrunk to four to seven months, from six to ten across 2025. The most capable open model, Zhipu's GLM-5.2, matched Anthropic's Opus 4.6 on narrow hacking tasks and the older Opus 4.5 on full simulated intrusions. DeepSeek's V4-Pro reached similar results for 28 cents per task against $12.50 for Opus 4.5. As Shanghai hailed open models as a public good, cheap downloadable systems are closing on frontier attack capability too.