Core Scientific and AMD signed a 15-year infrastructure deal worth more than $14B in base contracted revenue, with AMD leasing 529MW of data-center capacity across five Core Scientific campuses in Texas, Oklahoma, Alabama and Georgia and holding options on another 1.9GW through 2028 that would push total capacity to 2.5GW. The sites will run AMD Instinct GPUs, EPYC CPUs and the ROCm stack, phased in from 2027 through 2028 — a gigawatt-scale hosting commitment AMD secured independent of the NVIDIA/CoreWeave/Oracle ecosystem, and the clearest signal yet of how much capacity AMD is willing to underwrite to build a ROCm installed base. Core Scientific's same-day Q2 print showed colocation/AI revenue of just $137M, a reminder of how far the pivot from bitcoin mining to AI hosting still has to run relative to the contracts now on its books. (Coindesk)
Private Companies
Rebellions
Launch
Rebellions is moving its Rebel100 inference accelerator from ISSCC validation to commercial shipment in the second half of 2026 — a quad-chiplet design on Samsung's SF4X process with UCIe-Advanced interconnect and 144GB of HBM3E, rated near Nvidia H200-class performance at lower power. The South Korean chipmaker, which raised $400M at a $2.34B valuation in March and is preparing a 2027 KOSPI listing, is pairing the chip with SK Telecom and Arm's Neoverse-based "AGI CPU" for a sovereign-AI buildout in Korea — a real customer attached to the non-NVIDIA inference silicon thesis. (Digitimes)
Fireworks
Launch
Fireworks AI launched Nexus on July 26, a routing and cost-control layer that diverts routine agentic coding work to open-weight models via a difficulty-aware router, paired with FireConnect, an Apache 2.0 connector that installs into Claude Code, Codex and OpenCode with one line. Fireworks claims a 3-5x cost reduction and a 33% drop in cost per merged pull request in early testing. The launch moves Fireworks from pure inference hosting into the cost-governance layer enterprise agentic coding spend now demands. (Fireworks AI)
Public Markets
A Shanghai state-backed firm has reportedly begun limited production of domestic immersion DUV lithography tools for SMIC, CXMT and Hua Hong — a credible dent in ASML's tool monopoly, and the reason chip-equipment names sold off together. (Bloomberg)
Corning beat Q2 estimates but Q3 guidance landed in-line rather than accelerating, and investors read it as evidence AI-datacenter capex growth may be leveling off — dragging photonics peers Coherent and Lumentum down as well. (Yahoo Finance)
Emerging
Mixture-of-experts models are cheap to serve in theory, but speculative decoding — using a small model to draft tokens a big model verifies — breaks down on MoE architectures because draft tokens scatter across disjoint experts, spiking memory traffic. This paper fixes that by weighting draft-token selection toward experts already loaded for verification, recovering up to 1.62x faster decoding on DeepSeek-V3.1, Qwen3-235B-A22B and GPT-OSS-120B. Since nearly every frontier model shipping now is MoE, a technique like this moves the needle on inference cost per token — the variable that ultimately sets GPU and HBM demand across the stack. (arXiv)
Watch This Week
Wed Jul 29
Qualcomm reports fiscal Q3 2026 results after market close. Its data-center AI ambitions ride alongside the core handset business this quarter. (Marketchameleon)