AMD buys a world-model team, OpenAI books 457 megawatts, and the stack is the deal
AMD completed its World Labs acquisition, TensorWave signed a six-year 457-megawatt deal with OpenAI, and Goodfire embedded researchers inside OpenAI. The leverage has moved to the stack underneath the models.
- AMD completed its acquisition of World Labs, and Fei-Fei Li joins AMD as executive vice president and chief scientist, reporting directly to chief executive Lisa Su. Self-reported m:…3133
- TensorWave signed a six-year agreement with OpenAI to deliver 457 megawatts of dedicated AI infrastructure, with the first capacity online in the first half of 2027 and full deployment by early 2028. Self-reported m:…2559
- Preferred Networks and SEMIFIVE will collaborate on the MN-Core L2000 series of AI inference chips, planned for more than twice the memory bandwidth of the L1000 series, with tape-out scheduled for the first quarter of 2027. Self-reported m:…5158
- Goodfire researchers will embed within OpenAI, with deep access to OpenAI's internal models and systems, to monitor training runs and investigate model alignment and safety. Self-reported m:…3109
AMD buys the world-model team
AMD said it has completed its acquisition of World Labs, the company building AI models that create and simulate interactive 3D environments, along with technology for robotic learning and simulation. Fei-Fei Li joins AMD as executive vice president and chief scientist, reporting directly to chief executive Lisa Su. The World Labs team will continue advancing foundational model research inside AMD. The logic, as AMD puts it, is that delivering the infrastructure for the next era of superintelligence requires optimising compute platforms and models together. AMD is not buying revenue. It is buying a team that knows how to build models of the physical world, and putting that team next to its own silicon. AMD notes its technology powers billions of experiences across cloud and AI infrastructure. m:…3133
OpenAI books the power
The same week, TensorWave said it has signed a six-year agreement with OpenAI to deliver 457 megawatts of dedicated AI infrastructure across multiple US data center sites. The company will deploy AMD Instinct MI455X GPUs in infrastructure built for AMD Helios rack-scale systems, and says it is the first publicly announced neocloud infrastructure provider supporting the collaboration. TensorWave expects to bring the first capacity online in the first half of 2027 and complete the full deployment by early 2028. It says it has more than 2 gigawatts of long-term data center capacity secured. The deal is notable for what it is not: OpenAI did not buy chips. It bought delivered power and capacity, from a provider that was the first public cloud to deploy AMD's Instinct MI300X GPUs. The scarce resource in the agreement is megawatts, contracted years ahead. m:…2559
The challengers sign up too
Smaller chipmakers are signing their own versions of the same deal. Preferred Networks and SEMIFIVE said they will collaborate to develop the MN-Core L2000 series, chips designed specifically for AI inference, with more than twice the memory bandwidth of the earlier L1000 series. Tape-out is scheduled for the first quarter of 2027, the logic chip will be manufactured using Samsung Foundry's platform technology, and the work is being developed with support from Japan's Acquisition, Technology and Logistics Agency. The pattern is the same at every scale: nobody is waiting for a better model to arrive. They are contracting for the hardware that will run it. The design stacks logic chips with DRAM in a 3D structure, which PFN says achieves substantially higher memory bandwidth than GPUs and other chips used for AI processing. m:…5158
And someone gets to look inside
There is a fourth deal, and it is about visibility rather than volume. Goodfire said its researchers will embed within OpenAI, using its interpretability platform to monitor training runs, debug undesired behaviour and investigate model alignment and safety in training. The company says its researchers will have deep access to OpenAI's internal models and systems, and that it will publish learnings and findings as the collaboration progresses. Taken together, the four announcements describe where leverage now sits: in the team that builds world models, in contracted megawatts, in inference silicon, and in the right to inspect a training run. For a technology executive choosing a platform, the question is no longer which model is best. It is who controls the stack underneath it, and whether you can see inside. m:…3109
What is still unresolved
- The price AMD paid for World Labs was not disclosed in the announcement.
- Whether TensorWave's 457-megawatt build stays on its 2027 to 2028 schedule.
- What Goodfire will be allowed to publish from its embedded work, and when.
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.