The Pentagon is in talks to lend Fluidstack roughly $5B through its Office of Strategic Capital, its largest loan to date, to build domestic manufacturing capacity for the transformers and switchgear that sit between a signed compute lease and a live megawatt of delivered power — not to fund a new data center outright. The program has previously backed rare-earth and drone suppliers; extending it to a neocloud operator signals Washington now treats grid-interconnect hardware, not just chips, as AI's binding supply constraint. Terms remain preliminary. (Data Center Dynamics)
Private Companies
Anthropicother
Nvidia is weighing an investment of up to $10B as an anchor backer of Anthropic's planned IPO, Reuters reported, as the AI lab seeks to raise as much as $100B at a roughly $2T valuation — a deal that could rank among the largest in history. An anchor stake would extend Nvidia's pattern of taking direct equity in the AI labs and infrastructure firms that buy its GPUs, this time reaching into a public offering rather than a private round or licensing deal. Talks remain preliminary. (Bloomberg)
Public Markets
Shares jumped as traders positioned for possible S&P 500 inclusion at the September 21 rebalance, adding to a rally already underway after Q2 revenue rose 104% year-over-year to $392M and adjusted EPS beat consensus. (TipRanks)
Shares hit an all-time high after RBC initiated coverage at Outperform with a $640 target, citing a $95B AI-server order backlog and 26% growth in storage revenue as enterprises turn to Dell as a one-stop AI-infrastructure vendor. (CNBC)
Emerging
China's last dragon goes public: Shares in Tencent-backed Enflame surged 179% on their Shanghai trading debut, valuing the AI chipmaker at roughly $26B after a retail offering oversubscribed more than 4,000 times over. Enflame is the last of China's "four little dragons" — alongside Moore Threads, Biren, and MetaX — to list, completing Beijing's push to put domestic Nvidia alternatives on public markets even as the company still operates at a loss. (South China Morning Post)
Ising machines at scale: A new paper, A 28nm 27,648-Spin Multichip Digital Ising Accelerator with Pegasus Connectivity, describes a four-chip system built to solve NP-hard optimization problems like MaxCut and circuit factorization, delivering 11.5x the throughput of prior designs at a quarter of the transistor area per spin. The result adds physics-inspired, problem-specific silicon to the field of post-GPU architectures now competing for a slice of AI infrastructure spend, alongside photonic and analog in-memory approaches pursuing the same GPU-adjacent workloads. (arXiv)