Vista Equity Partners and Cambium Capital launched Vector Core Compute at Computex, opening a production disaggregated inference facility in Los Angeles backed by a $3.5 billion compute commitment to SambaNova. The system routes a single inference workload across three silicon architectures simultaneously — Intel Xeon 6 for orchestration, SambaNova SN40 RDUs for decode, Nvidia Blackwell for prefill — with Together AI as the first commercial customer and Chicago, Seattle, and Phoenix sites in development. The structural argument is that inference has distinct computational phases with different profiles, and matching silicon type to each phase reduces cost and latency in ways that homogeneous GPU deployment cannot — this is the first production deployment that attempts to prove it at scale.
Ayar Labs joined NVIDIA's NVLink Fusion ecosystem on June 3, making its co-packaged optics products directly compatible with Nvidia's optical and SerDes fabric for rack-scale AI infrastructure. Ayar is NVLink Fusion's first CPO-native integration partner; the move builds on a $500 million Series E closed in March 2026 with Nvidia participation. The NVLink Fusion endorsement places CPO inside Nvidia's production architecture specifications rather than on a future roadmap — the bandwidth-per-watt constraint in scale-out clusters now has an Nvidia-endorsed resolution path.
A Fortune profile published today surfaces the operational reality behind Nscale's $14.6 billion valuation: OpenAI exited the Stargate Norway project, with Microsoft absorbing the capacity; the Essex Loughton facility in the UK has slipped from end-2026 to 2027 on construction delays; European grid connection queues run 7-10 years. These are not Nscale-specific problems — they describe the structural constraint on European neocloud buildout. The gap between contracted GPU counts and powered, connected, production-ready clusters is the central risk in European AI infrastructure timelines.
Q2 FY2026 earnings reported after market close: $10.8 billion in AI semiconductor revenue (+143% year-over-year) against $30 billion in quarterly bookings — a book-to-bill near 3x — with Q3 guided above $16 billion AI revenue (+200%+ year-over-year) on $29.4 billion total. CEO Hock Tan confirmed six XPU customers: Anthropic, Google, Meta, and OpenAI are named; Anthropic scales from 1 gigawatt in 2026 to 5 gigawatts from 2027, and OpenAI's first custom chip targets 1.3 gigawatts in 2027 under a 10-gigawatt-by-2029 program. The $73 billion committed backlog anchors a $100 billion 2027 AI revenue target.
The counter-tension: Google is actively diversifying away from Broadcom as its sole XPU design partner, developing in-house chip capabilities and working with additional ASIC vendors — the backlog represents current lock-in, not permanent market share.
Intel rallied on the Computex Xeon 6+ launch — the first data center CPU built on Intel 18A, validating the manufacturing turnaround at a product level. Lip-Bu Tan cited agentic AI as what's returning CPUs to data center prominence: "All CEOs are calling me, saying I need more CPU." Wells Fargo, Barclays, and Mizuho all raised price targets. At $114 and up 250% YTD, the stock is pricing in a near-complete foundry and product turnaround before the free cash flow has materialized.
DeepSeek's first outside round. Reuters and CNBC reported today that DeepSeek is raising $7.4 billion — its first-ever outside capital — led by founder Liang Wenfeng ($2.9 billion from personal funds), Tencent ($1.5 billion), and CATL, the EV battery conglomerate, at a $52-59 billion valuation. The round breaks DeepSeek's founding no-outside-capital policy; stated use is AI infrastructure expansion. Where the capital flows — domestic accelerators (Ascend, Biren), NVIDIA-compatible GPU clusters, or custom silicon programs — is a direct read on China's compute independence timeline and on whether DeepSeek's inference efficiency model can scale to hyperscaler-level hardware at competitive cost.
Goldman raises HBM estimates. Goldman Sachs raised its 2027 HBM market target from $75 billion to $116 billion ($168 billion in 2028), projecting DRAM undersupply of 5.0% in 2026 and 5.9% in 2027 — calling it the deepest shortage on record. HBM pricing is expected to rise 44% in 2027. Micron and SK Hynix have both crossed $1 trillion in market cap — the first time all three major DRAM suppliers have simultaneously hit that threshold. The bottleneck is structural: producing one HBM wafer displaces the output of 3-4 standard DRAM wafers, and no capacity build resolves the gap before 2028.
GTC Taipei Day 4 — final day. Post-Broadcom analyst notes circulate; watch for Hock Tan's detailed commentary on per-customer XPU ramp schedules. Anthropic's 5 GW from-2027 commitment implies a delivery timeline that will be mapped against TSMC CoWoS and advanced packaging capacity.
Computex ends. Final Intel product showcases; watch for additional Intel-SambaNova rackscale deployment announcements and any new Vector Core Compute partnership disclosures.
TSMC May 2026 monthly revenue report due within the first 10 business days — sets the quarter's trajectory read. Anthropic S-1 registration details may surface ahead of the October 2026 target listing; the S-1's disclosure of compute infrastructure as a cost line item will be the document to read in this space.