NVIDIA is putting $5B into Safe Superintelligence, the research lab Ilya Sutskever built after leaving OpenAI. The company has been valued at $32B with no reported products or revenue. The deal is structured as equity plus compute rather than a straight check: SSI gets early access to NVIDIA's next-generation Vera Rubin platform, roughly a tenfold increase in the compute it can throw at its own research over the next year. SSI has now raised about $7B total on the strength of Sutskever's reputation and NVIDIA's willingness to buy relationships with labs it doesn't own outright — a hedge that only makes sense if NVIDIA believes the balance of power in frontier AI is still unsettled. Sutskever's own account of why he thinks scaling alone has stalled, and what SSI is doing instead, is worth the two hours in Dwarkesh Patel's interview with him. (Bloomberg)
Private Companies
Hiveware
Fundraise
Hiveware announced their $24M Series A, led by RTP Global yesterday. The company is building a next-generation SystemVerilog simulator, with the goal of removing verification from the critical path of modern SoC development. Verification is a key pre-tapeout bottleneck for new chips coming to market. Hiveware enters the race of other AI chip design startups looking to compress design cycles as semiconductor design complexity grows rapidly. (LinkedIn)
Public Markets
A report that a state-backed Shanghai firm has begun mass-producing domestic DUV lithography tools — with units earmarked for SMIC, CXMT, and Hua Hong — threatens ASML's near-monopoly and undercuts the export-control leverage behind Congress's pending MATCH Act. (Bloomberg)
China's CXMT surged 466% on its Shanghai debut to a $488B market cap after Asia's largest 2026 IPO, arming Micron's most direct DRAM rival with fresh public-market capital just as memory pricing tightens. (CNBC)
Emerging
A field study out of the Chinese University of Hong Kong, Shenzhen tried running two production-scale workloads — a mixture-of-experts alignment pipeline and a multimodal medical benchmark — on a 16-device Huawei Ascend 910 cluster instead of NVIDIA GPUs. The team needed twelve source-level patches to the vendor stack and had to disable several throughput features just to get numerically correct output, cataloging eight distinct platform limitations along the way. It's rare, reproducible evidence — rather than geopolitical narrative — on how much engineering cost still separates China's leading domestic accelerator from a genuine CUDA substitute. (arXiv)
Watch This Week
July 29
Microsoft, Meta, Qualcomm, and Arm all report quarterly earnings after the close. Microsoft's Azure capex and Meta's infrastructure spend are the two numbers compute investors will react to first. (Microsoft)
July 29
SK Hynix reports before the Korea market open — the first live HBM supply-and-pricing readout since Monday's double-digit selloff across Korean memory names. (SK Hynix)
July 30
Amazon and Samsung Electronics both report. AWS capex guidance and Samsung's memory-pricing commentary are the lines to watch. (Amazon)
August 4
AMD reports its first quarterly print since last week's Advancing AI event, with MI455X and CDNA5 shipment commentary expected. (AMD)