Z.AI, the Chinese lab formerly known as Zhipu, activated a 1-gigawatt data center hub Monday built entirely on domestic chips — Huawei's Ascend accelerators, with zero Nvidia silicon — drawing power equivalent to roughly 750,000 homes across multiple clusters of 10,000-plus chips training its GLM models, Bloomberg reported. Z.AI has operated under a U.S. export blacklist since January 2025; this facility is the clearest evidence yet that the restriction is pushing frontier Chinese labs toward full domestic-stack buildouts rather than simply throttling their output. The binding constraint has shifted from logic dies — which Huawei and SMIC can already fabricate at scale — to HBM memory, where Chinese suppliers remain years behind; watch that supply line, not further export-list additions, for the next material shift in the balance.
Private Companies
Infinity
fundraise
Infinity, a 26-person startup building an AI agent that writes and optimizes low-level inference kernels for arbitrary chip architectures, raised a $15M seed at a $100M valuation from Touring Capital and individual researchers at OpenAI and Anthropic, TechCrunch reported. Founder Jeremy Nixon, a former Google Brain researcher, is positioning the product — called Ignition — as an alternative to hand-tuned CUDA kernels; the company already has several million dollars in ARR through a chip-partnership deal with d-Matrix. As chip diversity increases, the software gap between raw silicon and usable performance, not FLOPS, is becoming the scarcer resource.
Fluidstack
fundraise
Fluidstack co-founder Jamie Cox confirmed on X that the GPU-cloud provider raised $830M in December at a $7.5B valuation from Leopold Aschenbrenner's Situational Awareness fund — the first on-the-record confirmation of terms that had previously only circulated as anonymously sourced reporting. The disclosure lands as Fluidstack is separately in talks for a new $1B round at an $18B valuation, tied to its $50B, 20-year data-center lease with Anthropic.
SkyPilot
fundraise
SkyPilot, the open-source multi-cloud GPU-orchestration project out of UC Berkeley, launched as a company with $20M in seed funding led by Lux Capital, joined by Coatue, Amplify, Foundation, and Race Capital, plus angel checks from Databricks' Ali Ghodsi and Google's Jeff Dean. Co-founded by Databricks' Ion Stoica and Scott Shenker, its software already schedules workloads exceeding 10,000 GPUs across clouds, with usage up 6x in six months. The bet: the scarce resource in multi-cloud AI infrastructure is a neutral control plane, not more raw compute.
Public Markets
Microsoft will deploy AMD's Helios rack-scale system — Instinct MI455X GPUs, Venice EPYC CPUs, Pensando networking, and the ROCm software stack — inside Azure and Azure AI Foundry, making it the fourth flagship full-stack Helios customer after Meta, OpenAI, and Oracle. Helios racks are priced at $5-5.5M, roughly 40% above Nvidia's Vera Rubin, and analysts read Microsoft's willingness to pay that premium as evidence AMD's roughly 4.5% share of the data-center GPU market could climb toward 20-25% if the pattern generalizes to other buyers. Shipments begin in the second half of 2026; the open question is whether Microsoft is buying Helios on merit or because Nvidia allocation is simply full.
Emerging
Molex signed a $6.29B, 10-year supply agreement with Prysmian for the optical cabling that connects AI data centers internally, including a $550M upfront payment. Prysmian will double its U.S. fiber and glass-preform capacity, spending $1.25B through 2031 to add roughly 1,000 jobs, part of a broader roughly $11B hyperscaler-driven revenue build-out through 2035. Optical interconnect is now getting the multi-year, capex-locked supply treatment previously reserved for GPUs and power — a sign that fiber, not just chips, is a binding constraint on data center buildout timelines.
Watch This Week
Wednesday, July 22
SK Hynix and Texas Instruments report Q2 2026 earnings. AMD's Advancing AI 2026 event opens in San Francisco, where Lisa Su and ecosystem partners lay out the Helios roadmap following Monday's Microsoft deployment news.
Thursday, July 23
Intel reports Q2 2026 earnings, watched for whether 18A/foundry progress is translating into external customer wins. STMicroelectronics also reports Q2 earnings, a read on auto/industrial demand and SiC power capacity.
Tweet of the Day
"The Chinese open source AI panic is back, and I think it is aimed at the wrong target... cheap open models grow the inference market, the leading edge still requires massive training infrastructure, and the profit pressure lands on closed model labs, not the hyperscalers." — a structural read on why cheaper Chinese frontier models expand compute demand rather than shrink it.