SpaceX has signed Google to pay $920 million per month, per Tom's Hardware, for approximately 110,000 Nvidia GPUs at xAI's Memphis data centers, running from October 2026 through June 2029 for roughly $30 billion — the second major external compute deal SpaceX has disclosed since merging with xAI in February, following Anthropic's $1.25 billion-per-month agreement announced in May. Both deals share a common origin: xAI built Colossus 1 with a mixed H100, H200, and GB200 GPU architecture, found that the heterogeneous cluster made efficient large-scale model training impractical, and migrated training workloads to Colossus 2, leaving Colossus 1 available for external leasing; the two agreements combined give SpaceX roughly $26 billion in annualized compute revenue. That Google — with $190 billion in planned 2026 capital expenditure and its own custom TPU fleet — is paying a competitor $11 billion per year for bridge capacity is the clearest signal yet of how far near-term supply has fallen behind frontier AI demand. This second deal is a major move on the path to SpaceX becoming a dominant compute provider, an important narrative as they head into their IPO this week.
ASML spinout founded in 2025 in Eindhoven, raised €20 million in seed led by Hitachi Ventures, with Samsung among the investors. The company uses High Harmonic Generation to produce soft x-rays across multiple wavelengths, generating richer 3D inspection signals than conventional single-wavelength laser optics — targeting chip inspection at the 2nm node, where optical metrology approaches its resolution limit. Samsung's participation ties directly to its 2nm process yield development. First commercial system under development; customer demonstrations planned at a new Eindhoven cleanroom.
Session: Friday, June 5, 2026
Broadcom's Q3 AI chip revenue guidance of $16 billion missed the $17.2 billion consensus, and management declined to raise its full-year AI revenue target — the first guidance reset from a major custom ASIC supplier in the current upcycle. Macquarie downgraded the stock Thursday, forecasting Broadcom's share of Google's TPU-related revenue falling from roughly 95% this year to roughly 65% by 2028 as MediaTek gains a custom ASIC design role; the two-day sell-off erased more than $320 billion in market value, the third-largest two-day wipeout in US public stock history.
Nvidia fell below the $5 trillion market cap threshold as Broadcom's guidance miss spread sector-wide: ARM fell 12.8%, SOXX closed down 10.4%, and the Nasdaq fell 4.2% — its worst session since April 2025, erasing roughly $1 trillion from AI-related equities. Nvidia has no direct exposure to the custom ASIC dynamics behind Broadcom's miss; the decline was sector contagion. Vera Rubin shipments remain on schedule for second-half 2026.
Macquarie's Thursday downgrade of Broadcom contains a forward-looking signal worth tracking independently of Friday's price move. The firm forecasts Broadcom's share of Google's TPU-related revenue falling from roughly 95% in 2026 to roughly 65% by 2028 as MediaTek gains a custom ASIC design role and Google advances in-house chip development. TrendForce separately reported in April that MediaTek doubled its 2026 AI ASIC revenue target to $2 billion, targeting 10–15% hyperscaler custom silicon market share. If that trajectory holds, it marks the first material fragmentation of the hyperscaler ASIC supply base — a market Broadcom has largely owned since 2023.
Oracle Q4 FY2026 earnings after close. Consensus: $1.96 EPS, $19.1 billion revenue; OCI AI infrastructure grew 243% last quarter, with management guiding 46–50% growth for Q4. The company carries $553 billion in remaining performance obligations, the majority AI-related. Following this week's sector sell-off, Oracle's characterization of AI compute demand and OCI capacity constraints will be a key read-through for the sector heading into summer.
Compute100 | Monday, June 8, 2026