Huawei unveiled the physical Atlas 950 SuperPoD in Shanghai on July 16, a day ahead of WAIC 2026's opening address by Xi Jinping — a rack-scale system linking up to 8,192 Ascend NPUs via Huawei's UnifiedBus interconnect, which the company says delivers 6.7x the compute of Nvidia's NVL144. Huawei Vice Chairman Xu Zhijun conceded the company still trails Nvidia chip-for-chip but is betting the contest is now decided at the supernode level — how many accelerators can be lashed into a single low-latency memory pool — rather than on individual silicon. Watch whether Huawei's claimed specs hold up under independent testing and whether open-sourcing the UnifiedBus interconnect spec pulls other Chinese chipmakers into a shared standard, the way NVLink anchored Nvidia's own ecosystem.
Public Markets
Nebius disclosed a $20–25B 2026 capex plan with most new capacity not revenue-generating until 2027, as insiders sold over $140M in stock and analysts flagged Meta's new compute-resale plan as a direct threat to neocloud economics.
Emerging
NetApp acquired AI-data-infrastructure startup DataPelago, whose Nucleus engine runs GPU-accelerated processing at the storage layer instead of moving data to external compute clusters, claiming 80% lower infrastructure cost and up to 10x faster throughput for AI and analytics pipelines. It's a structural bet that AI economics increasingly favor moving compute to data rather than data to compute — the same thesis behind Vast's Cloudera partnership above — and gives a legacy storage vendor a credible entry into a stack it might otherwise have ceded to specialists.
ChangXin Memory Technologies (CXMT) went public on Shanghai's STAR Market in the largest Chinese semiconductor listing ever, an implied valuation near $85B and roughly $8.5–9.8B raised, oversubscribed 212 times by retail investors. Its capacity ramp — approaching Micron's wafer-start run rate — was cited as a direct catalyst in Thursday's memory-sector selloff, with Micron, SK Hynix, and SanDisk all down sharply. A fourth global DRAM player, state-backed and now public, is entering just as the AI memory supercycle is at its tightest — watch SK Hynix's earnings next Wednesday for the first read on whether that changes pricing power.
Watch This Week
Wednesday, July 22
SK Hynix reports Q2 2026 earnings, the first full look at whether this week's memory-sector selloff reflects a real demand air-pocket from CXMT's new supply or a valuation reset after outsized 2026 gains. Texas Instruments also reports Q2 earnings the same day.
Thursday, July 23
Intel reports Q2 2026 earnings, watched for whether the 18A/High-NA manufacturing proof point from earlier this week shows up in foundry revenue or guidance.