Crusoe closed a $3.9B Series F co-led by Atreides Management, Mubadala Capital and Valor Equity Partners, pushing its valuation to $30.9B — more than double the level set by its $1.375B Series E last October — with Nvidia, Founders Fund and Qatar Investment Authority joining as new backers. The round funds a Colorado factory that will eventually turn out a gigawatt a year of Crusoe's modular "Spark" units, roughly a megawatt apiece, built indoors and trucked to wherever power already exists rather than waiting years for transmission capacity to reach a purpose-built site. Three new board members joining the raise — Cloudflare CFO Thomas Seifert, Primary Digital Infrastructure's Bill Stein and Redwood Materials founder JB Straubel — read as preparation for a public listing, and the pivot to manufactured, portable capacity is a bet that siting and power, not chip supply, are now the binding constraint on how fast AI infrastructure can scale. (TechCrunch)
Private Companies
GMI Cloudother
GMI Cloud, an Nvidia cloud partner, is in talks with banks and private-credit funds for a $300M loan to buy the GPUs for a new site in Thailand, with Tencent lined up as the anchor tenant for the capacity. The financing follows GMI's $12B sovereign AI buildout in Japan and marks a shift toward debt rather than equity to fund the chips themselves rather than the buildings around them — a structure closer to asset-backed lending than venture funding, and one more neoclouds are likely to copy as GPU purchase orders outrun what equity investors will underwrite. (Bloomberg)
Firmusother
Firmus, the Nvidia-backed Australian AI-data-center operator, began investor roadshow meetings in Sydney and Melbourne this week ahead of a planned October listing on the ASX that bankers are marketing at up to a $30B valuation and as much as $7B raised — a target Firmus has so far declined to confirm, with no prospectus yet lodged. The float would fund the build-out behind "Project Southgate," which crossed 900MW of contracted capacity and added OpenAI as an anchor customer earlier this month across sites in Tasmania and a new push into Malaysia. Public-market appetite for a roughly $30B AI-infrastructure valuation, tested before the numbers are locked into a prospectus, is the thing to watch into October. (ABC News)
Rebellionscustomer
Rebellions will supply more than 100 RebelRack NPU racks to ai&'s new Tokyo data center, the South Korean chipmaker's first confirmed hyperscale deployment in Japan and a bet that sovereign-AI buyers will choose an inference-specialized alternative over Nvidia now that capacity, not raw benchmarks, is the constraint. Rebellions declined to disclose contract value, citing an NDA with ai&, but a full 100-rack build would draw roughly 2.8MW — a meaningful commitment for a company that closed its $400M pre-IPO round and launched RebelRack only months ago. The deal extends Rebellions' push beyond its home market just as Japan leans on outside vendors to build compute capacity it considers strategically necessary but cannot yet supply domestically. (Seoul Economic Daily)
Public Markets
Applied Materials pledged $5B over the next decade, including a 57-hectare R&D park meant to double its India headcount, under an "India Vision 2035" plan unveiled at the opening of SEMICON India. Lam Research separately committed roughly $1.2B to build its first India-based silicon-component plant at the same event. The dueling pledges treat India's semiconductor-equipment ecosystem as a talent and supply-chain hedge rather than just a subsidized fab customer, arriving as Delhi pushes to backfill its chip value chain beyond the fabrication capacity it has already funded. (ANI)
Emerging
RISC-V gets a real AI workload: AMD and SiFive demonstrated AMD's ROCm software stack running a Gemma4-E2B model on SiFive's BigSky RISC-V server platform at the AI Infra Summit, pairing SiFive's Performance P870-D chips as the head node with AMD's Radeon AI PRO R9700 GPUs for inference. It is a demo, not a shipping product, but it is the first sign an open, royalty-free instruction set could sit in the head node of an AI server rack instead of the x86 or Arm CPUs occupying it today — the kind of proof point RISC-V's backers need before hyperscalers spec it into a real deployment. (SiFive)