The Compute 100 — Monday July 6, 2026

Crusoe is in talks to raise roughly $3B at a $30B valuation — tripling the mark it set nine months ago, per Bloomberg — the same week Meta, one of Crusoe's largest compute customers alongside Oracle, Microsoft, and Google, was reported to be building a cloud business to sell its own excess AI capacity externally. Public markets responded: CoreWeave and Nebius, whose disclosed Meta commitments run to $21B and $27B, fell a cumulative 18% and 22% over two sessions, even as private investors price Crusoe at a record multiple. Watch whether that gap holds — a $30B print says investors see Meta's move as optionality, not intent; a stalled or downsized round would confirm the public market's skepticism and tighten the Nvidia-backed vendor financing the sector leans on.

Private Companies

Firmus is the first named partner — alongside publicly traded Sharon AI — in a new Nvidia revenue-sharing model that has Nvidia acting as financial backstop for GPU-cloud buildouts, unveiled July 1. Firmus is building a 360MW, up to 170,000-GPU campus on Indonesia's Batam Island with Singapore's DayOne, targeting Q1 2027, and expects $25B–$30B in committed customer offtake over six years. If Nvidia's vendor-financing template spreads, it changes how sub-scale neoclouds get funded industry-wide.

Anthropic is in early talks with Samsung Electronics to manufacture a custom AI inference accelerator on Samsung's 2nm process with advanced chip-near-memory packaging, The Information reported July 2. No chip spec or timeline has been disclosed; Anthropic has hired dedicated silicon engineers, including OpenAI chip-team alum Clive Chan. Samsung backed Anthropic's $65B Series H in May alongside SK Hynix and Micron. Anthropic joins OpenAI and Google in pursuing custom silicon — each additional frontier lab diversifying away from Nvidia chips erodes a slice of Nvidia's captive-demand premium.

Samsung SDS is launching Korea's first NPU-as-a-Service offering on Samsung Cloud Platform this month, built on FuriosaAI's second-generation RNGD chip paired with LG AI Research's Exaone 32B model. FuriosaAI claims RNGD handles 7.4x more concurrent users than Nvidia's RTX Pro 6000 at equal power draw, at roughly 40% lower total cost of ownership. The underlying deal was announced in April; the commercial go-live is what's new. A named hyperscale-adjacent customer running production inference on non-Nvidia silicon is a concrete revenue proof point competitors can point to.

Tenstorrent CEO Jim Keller publicly denied acquisition talks with Qualcomm, in response to mid-June reporting that the two were in advanced negotiations on an $8B–$10B deal. Neither side has offered a fuller account. Qualcomm has been the more aggressive acquirer among mobile-chip incumbents diversifying into AI compute — worth tracking as either story develops.

Public Markets

SK Hynix led a two-day, sector-wide memory selloff — Samsung, Micron, and chip-equipment makers all fell double digits — after reports it is deliberately slowing its HBM4 ramp to redirect capacity toward higher-margin commodity DRAM, adding to fears about softening Nvidia Rubin production forecasts.

Emerging

Microsoft launched Frontier Company, a $2.5B unit that embeds roughly 6,000 engineers directly inside enterprise customers to design, deploy, and tune AI systems, led by former Microsoft Asia president Rodrigo Kede Lima. It mirrors OpenAI's Deployment Company and Anthropic's Goldman Sachs/Blackstone-backed services venture, both launched in May — three frontier players independently concluding that the adoption bottleneck is now implementation, not model capability or compute access. Watch for competitive differentiation to shift from silicon access toward enterprise-deployment capability.

Watch This Week