Nvidia is paying its suppliers in advance to expand capacity for Vera Rubin. The largest disclosed payment is a $1.1B advance to test house King Yuan Electronics for new US testing sites, targeted for high-volume production in April 2027. Nvidia also made a multibillion-dollar advance to IBIDEN for substrates, the base layer a chip package is built on, and plans to lift substrate capacity 2.5x over three years across IBIDEN, Unimicron and Samsung Electro-Mechanics. Packaging and optics are part of the plan too. Amkor will add more than 300% more interposer wafer capacity in South Korea by the end of 2028, and SPIL, a subsidiary of ASE Group, will expand optical-engine capacity 15x by mid-2027. Nvidia expects its total testing capacity to double from 2026 to 2027. "Capacity is the constraint," Jensen Huang said. (NVIDIA)
Private Companies
Crusoepartnership
Polar Data Centers is expanding its Norway site for Crusoe. The Tørdal campus already runs 12MW of IT capacity for Crusoe Cloud. Its next phase, DRA02, will add 40MW of liquid-cooled capacity in late 2027 for Crusoe's Nvidia Vera Rubin NVL72 systems, and both phases run on hydroelectric power. Separately, Crusoe introduced Spark Liquid, a modular data center unit of up to 3.5MW with closed-loop liquid cooling, built for GB300 and Vera Rubin. (Polar Data Centers press release)
Nscaleother
Nscale plans to deploy Nvidia Vera Rubin infrastructure across five European markets, according to an update Nvidia published at GTC Berlin. The same update included test results for DSX MaxLPS, Nvidia software that allocates power across GPUs. On GB300 NVL72 systems, it fit 37% more GPUs and delivered nearly 50% more aggregate inference throughput within the same power budget. The figures come from Nvidia's own test setup. (NVIDIA)
Photon Bridgefundraise
Photon Bridge, an Eindhoven startup formerly called Photon IP, raised €10M to build laser light sources for optical links in AI data centers. Amadeus APEX Technology Fund and 360 Capital co-led the round, with Innovation Industries, Faber and the Brabant Development Agency participating. The company combines several lower-power lasers at different wavelengths into one fiber, targeting up to 40 milliwatts per wavelength, and says weaker lasers fail less often. It uses outside chipmakers rather than its own fab. First samples are planned for 2027, with mass-production preparation from 2028, and no customers have been named. (The Next Web)
Public Markets
Supermicro raised its September-quarter revenue forecast by about $4B, to $18.5B–$19.5B, and lifted gross margin guidance to 12.0%–12.5%. The company cited AI platform demand. Dell rose 5.3% and HPE 3.7%. Full results are due Nov. 3. (Motley Fool)
Emerging
Scale formats for FP4: When a model is compressed to 4-bit numbers, values are grouped into small blocks that share one scale factor, as in OCP Microscaling and NVFP4. A new paper finds the format of that shared scale matters mainly under aggressive FP4 quantization. The authors' UE5M3 scale format cut perplexity degradation by up to 25.3% against an NVFP4-like baseline, at a cost of +0.06% chip area and +0.49% power. The paper, MX-Like Tensor Core for LLM Inference with NPU Integration via Hardware-Software Co-Design Exploration by Zijun Zhao, Ruichu Ao, Stef Cuyckens, Marian Verhelst and Chao Fang, reports results from a co-design flow on the open-source SNAX NPU platform, not fabricated silicon. (arXiv)