The Compute 100 — Monday July 27, 2026

Anthropic Launch

Anthropic shipped Claude Opus 5 on July 24, holding list price flat at $5/$25 per million input/output tokens — the same as Opus 4.8, and roughly half a rival flagship's price — while posting new highs on Frontier-Bench and GDPval-AA and adding a user-selectable "effort" dial that trades cost for compute. Holding price while adding a cost lever pushes the inference-economics tradeoff onto the customer rather than the vendor, a sign flagship-model pricing is compressing. Coverage also reiterates Anthropic is preparing an IPO later in 2026. (Anthropic)

Hailo Other

Microchip Technology signed a definitive agreement on July 24 to acquire Hailo, the Israeli edge-AI processor maker, after a spiral that saw its implied valuation fall well below the over-$1B mark it carried on roughly $340M raised. Terms are undisclosed, with closing expected by the end of Q3. (GlobeNewswire)

Baseten Partnership

Moonshot AI opened public weights for Kimi K3 — a 2.8-trillion-parameter, 896-expert open model with a 1M-token context window — on July 27, with Baseten and Nebius among the first third-party platforms hosting it, per Moonshot's own release announcement. The rollout follows a demand shock: GPU capacity ran out within 48 hours of K3's initial subscriber launch on July 16, forcing Moonshot to freeze new sign-ups. Spreading inference across independent neoclouds is how a lab without a hyperscaler-scale GPU fleet meets frontier-model demand. (Kimi/Moonshot AI)

CXMT ¥49.00 ▲ +466% Mkt Cap: $484.6B

ChangXin Memory Technologies closed its Shanghai STAR Market debut up 466% from an 8.66-yuan IPO price, valuing China's largest DRAM maker at roughly $484.6B and making it mainland China's most valuable listed company on day one. The $8.6B float is Asia's largest IPO this year, funding capacity expansion as CXMT's global DRAM share has climbed from 4% to 8% over the past year. It lands the same week Nvidia locked in HBM4 supply from SK Hynix — direct evidence Chinese memory capacity is scaling into the exact supply chain the West has been trying to keep controlled. (SCMP)

ThAME: 3D Memory-Enabled Heterogeneous Accelerator for LLM Mixture of Experts, from researchers at Washington State University, proposes a multi-chiplet accelerator built specifically for Mixture-of-Experts inference, which standard hardware handles poorly because of scattered expert weights, unpredictable token-dependent routing, and synchronization delays when expert outputs are aggregated. ThAME pairs FeFET non-volatile memory with conventional DRAM and a custom network-on-chip tuned for variable routing traffic, reporting up to 15.7x speedup and 9.8x better energy efficiency than existing solutions. As MoE becomes the default architecture for frontier models, that gap between general-purpose GPUs and workload-specific silicon is exactly where the next generation of accelerator startups is trying to compete. (arXiv)

Nvidia, Microsoft, Meta, and 22 other companies signed a joint letter on July 24 — "Open Weights and American AI Leadership" — arguing the US keeps its AI edge by building an open-model ecosystem rather than restricting open-weight systems, as Washington weighs curbs on Chinese models. The list doubled to 50 signatories within a day, adding OpenAI and Google, while Amazon and Anthropic stayed off it; Jensen Huang amplified it in his first-ever post on X. The signatory list is itself a map of who benefits from open-weight distribution — chip and cloud vendors — versus who has more to protect by keeping frontier weights closed. (Tom's Hardware)