Newsletter · October 4, 2026
Weekly Digest 40
AMD acquires Fei-Fei Li's World Labs for about $8.2 billion in stock to build the models and simulation tools that shape demand for its chips. Plus, Tencent reportedly leases 100,000 AI chips at Oracle data centers in Southeast Asia, and US hiring slows to 29,000 jobs in September.
Topics we are tracking
AMD Acquires World Labs
Source: AMD — AMD to Acquire World Labs to Advance the Future of AI Compute
AMD announced the acquisition of World Labs for approximately $8.2 billion in stock, which signals a broader ambition in AI for AMD to develop models and software that will shape demand for its hardware. The acquisition brings Fei-Fei Li’s spatial-intelligence research team into AMD, with Li becoming executive vice president and chief scientist. Robotics and simulation feature explicitly in AMD’s rationale for the deal, which is expected to close by the end of 2026.
World Labs was founded on the belief that language alone is insufficient for many problems in science, engineering, and robotics. Addressing them requires models that can represent and reason about the physical world. Li argues that advancing this work requires closer integration with hardware to improve efficiency and scale. The companies already began collaborating on model training and inference optimisation on AMD GPUs last year.
Bringing these teams together could create a tighter feedback loop between model research, software, and chip development. World Labs gains access to AMD’s engineering resources, while AMD gains direct insight into the computational requirements of emerging applications. Optimising models for existing GPUs is the starting point; using those requirements to influence future hardware is the larger opportunity.
The acquisition also fits AMD’s broader investment in software. Its ROCm platform supplies the compilers, libraries, and development tools needed to turn GPU specifications into useful application performance. There have been significant improvements in AMD’s software quality, support for popular inference frameworks, and speed of optimisation for new models. We are increasingly optimistic about AMD’s ability to compete with Nvidia on the software front.
In July, World Labs acquired SceniX, which highlights its ambitions in robotics simulation. SceniX is building technology to reconstruct physical environments and interactions as simulations for robot training and evaluation. Unlike language models, robotics lacks a comparable repository of training data on physical actions and interactions. Simulation allows developers to generate data, vary environments, repeat tasks, and investigate failures before deploying changes on physical machines.
Nvidia has already assembled a broad platform around this process. Cosmos provides world models, Omniverse and Isaac Sim support simulated environments, Isaac Lab supports robot learning, and Isaac GR00T supplies robot foundation models. Jetson Thor provides onboard computing for deployment. Long before ChatGPT popularised language models, Nvidia had spent years supplying GPUs to the scientific community and universities around the world, building relationships and expertise in these fields.
We see AMD’s acquisition as an effort to compete across more of that process. Developing useful models and simulation tools can attract developers and increase demand for the underlying computing platform. World Labs gives AMD a stronger research base for this strategy, although it does not immediately reproduce Nvidia’s entire ecosystem. Delivering the benefits of co-design will require sustained investment in the less visible engineering work alongside frontier model research.
Chinese Companies Access More Compute Abroad
Source: Reuters — China's Tencent leases 100,000 chips from Oracle to accelerate AI push, FT reports
Tencent has reportedly agreed to spend approximately $7 billion over five years to access around 100,000 advanced AI chips at Oracle data centers in Southeast Asia. The agreement includes an upfront payment of roughly 30%. The reporting cited people familiar with the arrangement; Reuters could not independently verify it.
Under current rules a Chinese company can, in principle, train a model abroad and deploy it domestically. Processing Chinese users’ prompts or private documents on overseas servers raises additional questions, because those transfers may trigger China’s data-protection requirements. There is no blanket requirement that all domestic data remain in China, but personal information, designated “important data,” and critical-infrastructure data can require safeguards or a security assessment. Training abroad on an appropriately sourced dataset can therefore be more straightforward than processing sensitive domestic user data offshore.
For cloud providers, this creates a way to serve Chinese demand without exporting the chips to China. Restrictions on hardware shipments do not necessarily prevent Chinese companies from accessing the same technology through overseas computing services. This can limit the effectiveness of export controls in restricting access to advanced computing, while directing revenue and infrastructure investment towards foreign providers. It does not mean that every arrangement is permitted: restrictions on particular customers and end uses, alongside the conditions of export licences, can still apply.
Chinese companies access computing resources through four sources: legally imported chips, smuggled hardware, domestically produced processors, and remotely accessed capacity abroad. A ChinaTalk central estimate from April put Chinese compute resources at approximately 2.8 million H100-equivalents, with an estimated 90% range of 1.8–4.8 million. Overseas access accounted for roughly one million H100-equivalents, making it the largest individual category in the calculation.
Overseas access is China's largest single source of AI compute
Estimated AI chip stock by source, H100-equivalents: central estimate and 90% range
ChinaTalk's April 2026 supply-side estimate by source: remote access abroad about 1.03 million H100-equivalents (range 0.6 to 1.8 million), domestic chips about 0.90 million (0.56 to 1.1 million), legal imports about 0.46 million (0.40 to 0.57 million) and smuggled chips about 0.45 million (0.19 to 1.35 million). The total is about 2.8 million, with a 90% range of 1.8 to 4.8 million.
ChinaTalk supply-side estimate, April 2026. The four sources total about 2.8 million H100-equivalents (90% range 1.8–4.8 million). Ranges are per source and do not add up to the range of the total. Newer estimates differ by source: Epoch AI puts smuggled chips at about 660,000 through 2025 (April 2026), and The Substrate puts remote access at about 670,000 for 2026 (May 2026).
Source: ChinaTalk, Epoch AI, The Substrate · 2026-04-07
The Oracle deal is not the only one of its kind. Reporting the previous week linked ByteDance to computing capacity at Nscale’s Glomfjord facility in Norway. A publicly filed financing agreement names Singapore-based Spring as a cloud customer and describes 2,304 Nvidia B200 GPUs. The FT has identified Spring as a ByteDance subsidiary.
US Hiring Slows, but Rate-Hike Odds Persist
Source: U.S. Bureau of Labor Statistics — The Employment Situation, September 2026
Friday’s US employment report showed payrolls increasing by 29,000 in September, down from a revised 133,000 in August. July and August were also revised down by a combined 60,000 jobs, leaving average monthly job growth over the third quarter at approximately 51,000. The revisions weaken the picture of summer hiring, even allowing for volatility in individual monthly reports.
The unemployment rate edged up from 4.1% to 4.2%, but remained within the 4.1–4.3% range observed since March. Average weekly working hours were unchanged, while average hourly earnings increased just 0.1% over the month and 3.0% over the year. Taken together, these figures are consistent with the low-hiring, low-firing environment observed in recent months.
The labour market data remain difficult to interpret. So far, they do not establish a general collapse in labour demand or identify AI as the cause of losses in particular industries. This is consistent with a recent review of the evidence by Alex Imas and Jacob Schaal. Imas, Google DeepMind’s Director of AGI Economics, and Schaal find little evidence of widespread AI-driven job displacement, although some studies suggest reduced hiring among junior workers in exposed occupations. Our position remains that the absence of a clear economy-wide effect today does not rule out more substantial disruption as adoption deepens and capabilities increase.
The immediate bond-market reaction was straightforward. Treasury yields fell after Friday’s jobs report as expectations of an October Fed increase receded. On Polymarket, the implied probability of a 25-basis-point October hike fell from around 26% on Thursday evening to approximately 16% on Friday morning. Traders still assigned a roughly 68% probability to a 25-basis-point increase in December, suggesting expectations were shifting towards delaying further tightening. However, Treasury yields reversed their initial decline during the day, with two-, ten-, and thirty-year yields all ending the session above Thursday’s closing levels.
Where the market stands now: Fed decision odds
Implied probability of each outcome, daily
Where the market stands now: Fed decision odds. October meeting: 50+ bps decrease <1%, 25 bps decrease <1%, No change 83%, 25 bps increase 18%, 50+ bps increase <1%. December meeting: 25 bps decrease 1%, No change 25%, 25 bps increase 74%.
Live from Polymarket, refreshed every 5 minutes. Latest reading: 4 Oct, 10:41 UTC. Outcomes that never reached 5% are not shown.
Source: Polymarket
Seen on X
Other interesting stories
- Derek Thompson: You Are No Longer Invited to Dinner
- Mustafa Suleyman: A Warning About ‘Model Welfare’
- New York Times: What Would John Maynard Keynes Say About A.I.? The Keynes Bot Told Us
- AI Safety in China: China Is Open-Sourcing Its Best AI Models. Is That Safe?
- Financial Times: Japanese and Korean Shipbuilders Deploy Robots to Take On China
