TensorWave signed a six-year agreement to supply OpenAI with 457MW of dedicated, single-tenant AI capacity built on AMD Instinct MI455X GPUs in AMD Helios rack-scale systems, across multiple U.S. data center sites. Capacity is expected to begin coming online in the first half of 2027, with full deployment by early 2028. The deal falls under OpenAI's previously announced partnership with AMD, and TensorWave, which is backed by AMD Ventures and Magnetar, says it is its largest deployment to date. No contract value was disclosed (TensorWave)
Private Companies
Nscaleother
Nscale filed an amended IPO prospectus on Nov. 12 showing $307.5M of estimated revenue for the nine months to Sept. 30, up from $21.3M a year earlier, and a $1.69B net loss. The GPU cloud provider, which plans to list on the NYSE as NSCL, reports $105.8B of active and contracted value across roughly 1.37GW, up to about $44.6B of it from Anthropic. The price range is still blank, and a $1.0B Nvidia investment is expected to close around Nov. 16 (SEC filing)
Luminalfundraise
Luminal raised a $23M Series A led by Standard Capital, bringing its total funding to more than $28M, SiliconANGLE reported. Existing investors Y Combinator, SagaVC, Liquid 2 and Felicis also took part. The company pairs a search compiler, which automatically finds and writes the fastest code for a given chip, with a managed inference cloud. It processes more than 45 trillion tokens a month and runs significant workloads on AWS Trainium. Its seed round, about a year earlier, was $5.3M (SiliconANGLE)
DeepXother
DeepX plans to spend more than $100M developing its DX-M2 AI chip, Bloomberg reported. The South Korean startup has finished the design and sent it to Samsung's foundry unit, which will build it on Samsung's second-generation 2nm process. The chip is built to run AI models on robots, industrial equipment and AI PCs rather than in data centers. DeepX expects first samples by March 2027, followed by customer evaluations, and plans mass production in 2028 (Bloomberg)
Public Markets
Cerebras$108.72▲ 11.2%Mkt Cap: $26B
Cerebras shares rose 11.2% Friday after it reported third-quarter core revenue up 70% to $230.9M and raised full-year guidance to $940M–$960M from $880M–$890M. It expects to triple revenue in 2027 (SEC filing)
Emerging
Optical weight broadcast: A paper from Hannah Atmer, Yuan Yao, Thiemo Voigt and Stefanos Kaxiras, Broadcast, Don't Replicate: Eliminating Weight HBM in Datacenter LLM Serving, streams model weights from a central memory server over fiber to accelerators that read them through receive-only co-packaged optics, instead of storing a copy in each chip's HBM. Across 80,000 H100-class accelerators, it removes 4.0 PB to 8.4 PB of HBM. Its best layouts deliver 0.96x to 1.57x the throughput of an HBM baseline at 10% to 42% lower energy per token, with the largest gains on DeepSeek-V4-Pro. The results come from an analytical model, not hardware (arXiv)
Learning while serving: A paper from Yimin Gao, Liangtao Dai, Melika Morsali, Junting Huo and Mircea Stan, Mutable Inference: Architecting LLM Accelerators for Inference-Time Learning, designs an accelerator for models that keep training on their context during inference, which caps attention cost at a fixed recent window but adds backward passes and weight updates. Its Learning While Serving design shares compute between latency-critical prediction and deadline-flexible learning, reaching 1.6x the throughput of serial execution and 1.3x that of the best prior co-execution design across nine models, for 1.5% more NPU core area. Performance comes from modeled datacenter and edge NPUs and the area figure from RTL synthesis, not silicon (arXiv)