Anthropic signed a $10B, six-year deal with Volta Infra Holdings — a compute-leasing startup founded eight months ago by two former Brookfield infrastructure executives — to secure a 133MW data center in Tydal, Norway running Nvidia's next-generation Vera Rubin GPUs on Dell hardware. Volta doesn't build or operate anything itself: it raised $300M from Nvidia and Dell to buy chips on behalf of customers who don't want the capital outlay on their own balance sheet, then pays Bitdeer Technologies roughly $4.7B over 16 years to run the facility, with J.P. Morgan arranging a $1.3B credit backstop underneath the structure. The deal unbundles compute procurement into four separate risk-bearers — buyer, financier, operator, lender — which is precisely the appeal for a company with no credit rating of its own heading toward an IPO. (Bloomberg)
Private Companies
d-MatrixOther
d-Matrix acquired Wallaroo.ai on August 3, absorbing its inference deployment and orchestration software into d-Matrix's stack. The deal lands weeks after d-Matrix's Corsair accelerator entered full production, and it signals that inference chipmakers are done competing on silicon alone — the orchestration layer that moves a workload from a single server to a multi-rack cluster is becoming a differentiator investors should track alongside raw throughput. Terms were undisclosed. (PR Newswire)
Public Markets
Amazon closed above $3T in market cap for the first time on August 3, extending a rally from Thursday's Q2 print. AWS revenue grew 36.7% year-over-year to $42.2B, its fastest growth in eighteen quarters, and Amazon raised 2026 capex guidance to roughly $220B, citing rising memory prices as a direct input cost. (Bloomberg)
An FT investigation values the financing web behind Anthropic's chip supply at roughly $200B: a $35B Apollo- and Blackstone-backed vehicle buys Google's TPUs and leases them back to Anthropic, Broadcom's credit guarantee cuts the borrowing rate to 5.75%, and Google marked up its own Anthropic stake by $28.7B in the first quarter alone. (Financial Times)
Emerging
Yale researchers released Celty, a GPU kernel and microarchitecture co-design for LLM inference that combines weight pruning with activation sparsity — a workload combination existing kernels handle poorly because reconstructing sparse indices in software erases the theoretical speedup. The paper reports up to 5.3x throughput over cuBLAS at 70% combined sparsity with a proposed sparse SIMT core, evidence that meaningful inference-cost reduction is still available through hardware-software co-design rather than simply scaling cluster size. (arXiv)
Reuters reports the Trump administration is drafting an FCC rule adding new Chinese optical transceivers to its Covered List, targeting Zhongji Innolight, which controls roughly 27% of the global transceiver market supplying fiber links inside AI data centers. The rule isn't finalized — officials say it could still be modified or shelved — but US optical-component stocks rallied sharply on the report, with Coherent up 13% on bets that Chinese suppliers get displaced from a critical piece of the data-center supply chain. (Reuters)
Watch This Week
Tue Aug 4
FMS 2026 (Future of Memory and Storage) opens in Santa Clara. SK Hynix and SanDisk present the newly ratified High Bandwidth Flash standard, with a panel on breaking the memory wall running later in the week alongside Google DeepMind. (PR Newswire)
Tue Aug 4
AMD and Astera Labs both report Q2 2026 results after the close, a combined read on GPU shipments and PCIe/CXL interconnect demand inside the same buildout. (AMD Investor Relations)
Tue Aug 11
Super Micro Computer reports fiscal Q4 2026 results. A July 21 preliminary update already flagged gross margin well above guidance and more than $60B in backlog. (Supermicro Investor Relations)