Anthropic has signed $517B in compute contracts over the past eleven months, according to The Information — at least 14.8 gigawatts of capacity locked in through deals with Google, Amazon, Microsoft, Nscale, Lambda, Fluidstack, and SpaceX, among others, well past the roughly $180B in server commitments the company had previously guided investors toward. The pace comes months after CEO Dario Amodei warned rivals that they "don't really understand the risks" of overcommitting to compute, a caution his own company has now abandoned at scale. Anthropic still trails OpenAI's stated $750B build-out through 2030. (The Information)
Private Companies
Celerofundraise
Celero Communications raised a $275M Series C at a valuation above $3B, co-led by Atreides Management, Valor Equity Partners, and CapitalG — Alphabet's venture arm — bringing total funding to $415M two years after founding. The round follows validation of the industry's first 2nm coherent DSP silicon, built to move data over fiber at 1.6T today and 3.2T next-generation, aimed at the optical links binding AI data centers together. (Bloomberg)
Fab2fundraise
Fab2, the Jim Keller and Sam Zeloof-founded startup that builds machines to mass-produce small semiconductor fabs, raised a $500M Series A at a $3.7B valuation — up nearly 37x from its 2023 seed round, backed by the OpenAI Startup Fund at roughly $100M. Fundomo led the round, with Corner Capital, Paradigm, Duquesne Family Office, Maverick Silicon, StepStone Group, Protagonist, UDC Ventures, and Naval Ravikant participating. Formerly Atomic Semi, the company designs and builds its own fab equipment in-house, then replicates entire small fabs from it — a bet that factory-building know-how, not just process nodes, is the binding constraint on new chip capacity. (PitchBook)
Public Markets
Qualcomm issued Amazon a warrant for up to 25 million shares at $161.26 as part of a data-center partnership worth up to $60B, covering custom AI inference chips and optical connectivity built on Qualcomm's SerDes and DSP technology. (CNBC)
Corning and Verizon signed a multi-year fiber-optic supply deal running 2027–2032, covering more than 80 million miles of high-density fiber for broadband and AI data-center interconnects, adding carrier network spend to a growth story already re-rated on hyperscaler capex. (CNBC)
Emerging
Compute-as-proof: OpenAI says an internal model — more capable than GPT-6 Astra — produced a solution to the Navier-Stokes existence and smoothness Millennium Prize Problem, running roughly 10,000 agent instances over the effort at a cost the company puts in the millions of dollars. Whatever the mathematical verdict, the episode is a data point on how much compute frontier labs will now spend chasing a single open problem rather than a product feature. (OpenAI)
Photonic prefill: A new paper, Scaling Inference Prefill with High-Radix Photonic Interconnects, finds that 3D-integrated photonic interconnects cut inference prefill latency 2.1–3.2x in high-batch regimes and up to 5.8x over baselines in communication-limited configurations, letting clusters scale past the practical limits of copper GPU interconnects. The result adds to the case that photonic I/O, not just faster logic, is the next bottleneck between long-context reasoning models and cost-effective serving. (arXiv)