Skeleton Technologies, a startup building high power systems for AI, has raised a round led by Siemens at a valuation above $1B. Siemens will also sell Skeleton's supercapacitor systems to data center customers worldwide under an exclusive partnership. The Estonian company is building a solution for the GPU's shifting power state, which produces sharp load swings and harmonics inside the facility. Skeleton's Curved Graphene cells sit close to the workload and absorb those swings, and the company claims energy savings of up to 40%, a figure it has not had independently tested. It plans to spend more than $200M on US scale-up and R&D alongside hyperscalers. By placing the buffer inside its grid-to-rack electrical portfolio, Siemens turns load smoothing from custom engineering into a standard line item for data center builders. (Siemens)
Private Companies
OpenAIleadership
OpenAI has named Uday Ruddarraju, previously its chief technology officer for compute, to run both its compute and data center businesses as Sachin Katti steps back from infrastructure. Katti wrote on LinkedIn that his last day is November 15, that he will stay on as an advisor, and that he is leaving to build something new. He pointed to double-digit gigawatts of new capacity secured across Ohio, Portsmouth, Georgia and other projects, with partners including Nvidia, Broadcom, AMD, Oracle, Microsoft, Google, AWS, Cerebras and CoreWeave. Ruddarraju, who ran infrastructure at xAI before joining OpenAI, told Bloomberg he now has to spend more time on the commercial side. One executive will now both build and negotiate the capacity behind a compute budget the WSJ put near $750B through 2030. (Bloomberg)
3C Datalecother
The liquid-cooling company 3C, founded by former CoolIT Systems chief executive Steve Walton, has merged with the London data center engineer Datalec Precision Installations, with new investment from Longbow Capital, S2G Investments, Apogem Capital and QIC. Terms were not disclosed. The combined company, 3C Datalec, forecasts more than $400M of revenue this year, has about 1,000 employees and active projects in 16 countries, and says it works with four of the five largest hyperscalers. It plans to hire 200 engineers and open sites in Warsaw, Calgary and Austin. Walton said rack densities have outgrown air, and the merger pairs a cooling specialist with a firm that installs and runs the facilities around it. (PR Newswire)
Public Markets
TSMC guided fourth-quarter revenue to $52.3B to $54.3B after a $46.7B third quarter, up 41.2% in dollars, and held 2026 capex at $60B to $64B. The 2nm ramp cut gross margin to 66.4%, and C.C. Wei said demand still exceeds supply. (TSMC)
CoreWeave closed up 5.9% a day after a drop of about 15%, as Treasury yields fell and AI debt concerns eased. Its bonds sold off across maturities Wednesday, and the company said the market "acts irrationally." Nvidia holds 47.2M shares, about 10%. (Yahoo Finance)
Emerging
Hardware diversity: A paper from Haoyu Zhang and 57 co-authors at Meta, Architecting for Hardware Diversity: Meta's End-to-End Co-Design for High-Performance AI Training on AMD GPUs, trains recommendation models on AMD Instinct MI350X and compares them with NVIDIA B200. Revising model structures to use the MI350X's 1.6x larger HBM capacity, instead of porting the model directly, lifts end-to-end training throughput by up to 1.5x. The authors call it the first hyperscaler report of productionizing AMD GPUs for large-scale training and say the work was upstreamed to PyTorch, Triton and ROCm. Model and software changes, not the silicon alone, decide whether a second GPU vendor is usable at scale. (arXiv)