The AI compute arms race has reached the point where even the robots are reserving capacity. Before we dig in, quick catch-up: the AI infrastructure race has moved from quarterly capex lines to named power and GPU commitments. The backdrop includes AWS planning 2 million additional Nvidia GPUs in 2027-28, Alibaba raising AI capex, and ByteDance reportedly discussing five to six gigawatts of compute in Ulanqab by early 2028. This is AI Daily Briefing. Today: who’s actually building the capacity, who’s buying it, and whether anyone’s utilization spreadsheet survives contact with reality. Here's TechNode:
DeepSeek plans to deploy at least 160,000 Huawei Ascend 950DT accelerators at a data center it is building in Inner Mongolia, according to people familiar with the matter. The proposed installation would be among the largest known clusters of Huawei AI chips, although its deployment schedule depends on the chipmaker’s production capacity.
We covered ByteDance’s buildout, but DeepSeek has a much more concrete reported plan: 160,000 Huawei Ascend 950DTs in Inner Mongolia. “Plans to deploy” matters here. Bloomberg says the schedule depends on Huawei’s production capacity, so capital clearly isn’t the bottleneck—chip output is. And DeepSeek wants these for inference only, not training. A 160,000-accelerator serving fleet is a serious bet that domestic models will have domestic demand behind them. At gigawatt scale, the chip-substitution story gets very real. Huawei has to make enough silicon, and DeepSeek has to keep enough users hitting it to make that power bill sensible. Jung Mok-hee, writing in The Herald Business:
AI startup Anthropic has signed computing capacity contracts totaling $517 billion over the past 11 months, according to a new analysis. The Information reported Sunday (local time) that, based on company announcements and contract details, Anthropic secured at least 14.8 GW of additional computing capacity over the 11-month period — on top of the 1 to 2 GW it had already locked in before October last year.
Anthropic signed $517 billion in compute contracts over 11 months, tied to at least 14.8 gigawatts. Contracted capacity isn’t delivered power, but The Information is working from announcements and contract details—not a benchmark extrapolation. Fourteen-point-eight gigawatts against $65 billion in annualized sales. Somebody has to keep those clusters busy at an almost absurd scale for a very long time. It’s nearly triple the $180 billion in server leasing Anthropic showed investors last December. The 11-gigawatt Amazon-and-Google deal alone runs above $300 billion over a decade—early investors are now deeply embedded in the supply side. Set DeepSeek’s 160,000 Huawei chips in Inner Mongolia beside Anthropic, and you get two separate infrastructure races. Capital is plentiful. The bottlenecks are dependable hardware, power, and keeping it all utilized. This one's from Converge Digest:
Gimlet Labs has raised $300 million in Series B funding as it scales an unusual approach to AI infrastructure: instead of building an inference cloud around one accelerator architecture, the company is assembling heterogeneous data centers in which GPUs, CPUs, near-memory compute and dataflow processors can work together on different portions of the same inference workload.
Gimlet raised $300 million five months after an $80 million Series A to route one inference workload across GPUs, CPUs, near-memory compute, and dataflow chips. Great—now prove the scheduler doesn’t spend the savings arguing with four kinds of hardware. a16z leading this one makes sense. After the DeepSeek cluster and Anthropic’s $517 billion contract figure, somebody is funding the hedge against getting trapped in a single chip supply chain. Gimlet says it has billions in contracted revenue and a gigawatt-scale pipeline, but the Series B announcement doesn’t name those customers. The architecture gets interesting when it cuts the dollar-per-token number without wrecking latency or debugging. “Multi-silicon cloud” is a lovely label. The useful part is simpler: software deciding which slice of a model belongs on which machine before the GPU bill eats the margin. Brian Beckmann, writing in Beckmann:
The US robot manufacturer Figure secures computing power worth $3.5 billion from the cloud provider Nscale, with an option for over six billion. In return, Nscale receives an equity stake in Figure. From the second half of 2027, up to 100,000 Nvidia graphics processors of the Vera Rubin platform are to run in Barstow, Texas.
Figure has raised about $1.9 billion total and just committed $3.5 billion to Nscale for compute. That’s a robotics company putting nearly twice its lifetime funding into a GPU tab that doesn’t come online before the second half of 2027. Up to 100,000 Vera Rubin GPUs in Barstow, Texas—and Nscale gets equity in Figure. Both sides clearly want exposure to Helix succeeding; a straight cash contract alone wasn’t enough comfort. Set this beside Anthropic’s $517 billion figure: the industry is signing capacity commitments far ahead of delivered hardware. Figure says Helix is running short of data and compute; okay—now it has to turn that very expensive runway into robots people actually buy. Nscale was already serving Microsoft from Barstow. Figure is buying into the same infrastructure neighborhood. Robotics is now competing for frontier-scale capacity before it has frontier-scale revenue. The Next Web writes:
Anthropic published the proof on Friday. Dozens of Claude agents wrote 13 million lines of Lean code and proved 30,300 intermediate theorems. They used 29,500 of those in a complete, computer-checked proof of a conjecture Pierre de Fermat scribbled in a margin around 1637.
The headline is “Claude did Fermat.” Kevin Buzzard, who actually compiled the code, says it adds nothing to mathematics—and he’s right. The feat is dozens of agents producing 13 million lines of Lean that a checker can verify. It took 11 days to produce 30,300 intermediate theorems, with 29,500 used in the final proof. That’s serious orchestration throughput. And unlike a typical agent demo, Lean gets to say no when the chain of reasoning breaks. Right—and after the $517 billion compute-contract number, this gives a much clearer picture of what Anthropic is buying capacity for: parallel pipelines that turn out machine-checkable results, not another cinematic chatbot video. I wouldn’t generalize from a closed formal system to an open-ended production agent. But if formalization starts happening alongside new research, every claim can arrive with a proof artifact instead of a pile of citations. That changes the workflow. Have feedback, story ideas, or a correction? Email us at aidailybriefing at lantern podcasts dot com. Your notes help us make the briefing sharper and more useful every day.
We’ll be watching as Nscale brings the first Vera Rubin systems for Figure online, no earlier than the second half of 2027. Links to every story are in the show notes, so take a look at whichever ones you’d like to explore further.
That’s AI Daily Briefing for today. We’ll be back tomorrow. This is a Lantern Podcast.