WED 28 OCT 2026 · 4 MIN READ
Compute is now being financed like infrastructure, and the numbers are getting very large.
SpaceX lined up $40B to lease GPUs from Apollo. Anthropic's S-1 lists about $417B in new minimum payments since Dec. 31. Oracle made Vera Rubin NVL72 generally available.
General AI Infrastructure
Oracle makes Vera Rubin NVL72 generally available
TakeawayNVIDIA and Oracle said Vera Rubin NVL72 systems and Vera CPUs are now generally available on OCI.
- Each NVL72 rack links 72 GPUs over NVLink.
- A multi-gigawatt expansion is planned; no gigawatt figure or date given.
- HGX Rubin NVL8 is planned, with 8 GPUs and 6.4 Tbps RDMA; no launch date.
AngleThis is a vendor announcement, not a capacity count. "Multi-gigawatt" with no figure or date tells buyers nothing about when Rubin hours reach the market. Watch for region lists and published rates.
Oracle to offer Google TPUs from 2027
TakeawayOracle said it will add Google TPUs to OCI, the first hyperscale cloud outside Google's own to offer them, per Oracle.
- TPUs become selectable on OCI starting in 2027.
- Capacity grows over several years; Oracle did not disclose quantities.
AngleA non-Nvidia accelerator on a third-party cloud widens the buyer's options for inference. With no volumes and a 2027 start, it is a roadmap signal rather than available supply.
Buyer Signals & Deals
SpaceX lines up $40B GPU lease financing
TakeawaySpaceX disclosed a binding commitment for about $40B of GPU financing from an Apollo-managed entity, per a regulatory filing.
- Apollo's entity buys the GPUs and servers, then leases them to SpaceX.
- Lease term is about 5 to 5.25 years; close expected by year-end 2026.
- Q3 capex was nearly $20B; SpaceX expects $35B to $40B in Q4.
AngleEquipment leasing moves GPU purchases off the balance sheet's debt line. The 5-year term is the number to compare with GPU depreciation schedules and rental-rate assumptions.
Anthropic S-1 lists about $417B in new compute commitments
TakeawayAnthropic's IPO filing shows agreements signed or amended after Dec. 31, 2025 carry about $417B in future minimum payments, The Register reported.
- Itemized: Broadcom about $161.2B, Google at least $114.5B, Amazon at least $110B.
- Lambda about $41.3B and Nscale up to about $44.6B are listed.
- Microsoft, Google, Amazon and Broadcom deals are non-cancellable.
AngleTwo neocloud names sit in a top-tier buyer's commitments, which supports their contract backlogs. The fine print is that the largest deals are take-or-pay, so the buyer's revenue must keep pace.
NVIDIA to invest $3B in Meta's equity offering
TakeawayNVIDIA and Meta announced an expanded collaboration on Meta's Muse models, with NVIDIA investing $3B.
- Meta plans to expand deployment of Vera Rubin, Vera CPUs, Spectrum-X and BlueField-4.
- The release gives no GPU counts, megawatts or timelines.
AngleThe supplier is putting equity into a customer that buys its chips. The release discloses no volumes, so treat it as direction rather than a purchase order.
Neocloud & GPU Cloud Moves
General Compute raises $100M for inference chips
TakeawayGeneral Compute raised a $100M Series A at an $800M post-money valuation, led by Blackbird Ventures and Abstract Ventures.
- It rents bare-metal servers built on SambaNova, Positron and Cerebras chips.
- It expects 75 MW deployed in 2027 and access to 200 MW of power.
- A $400M SambaNova chip purchase is described as the first debt financing for inference chips.
AngleThe CEO told TechCrunch fears of a chip-led collapse are "massively overstated"; that is the company's view. The same article notes CoreWeave shares hit a 12-month low, so capital for GPU-based neoclouds is the open question.
Model Releases & Compute-Demand Drivers
Google releases Genie 4 world model in research preview
TakeawayGoogle DeepMind released Genie 4 to trusted testers on Oct. 27.
- Scenes stay consistent for up to two minutes, double Genie 3.
- Adds real-time audio and prompt edits mid-simulation.
- Rollout to Google AI Ultra and Gemini API partners comes in the coming weeks.
AngleReal-time interactive video is a sustained inference load, not a one-shot request. The article gives no compute or hardware figures, so serving cost is the unknown to watch as access widens.
Power & Datacenter Buildout
UK datacenter zones drew no environmental-group meetings
TakeawayA Spotlight on Corruption report found ministers met 137 businesses on datacenters from July 2024 to June 2026, and no environmental groups.
- 84% of stakeholders met were businesses or trade associations.
- Five announced AI Growth Zones affect ten local authorities; half reported no contact beyond applications.
- The government is consulting on letting key datacenters jump the grid queue.
AngleSiting that skips local consultation tends to meet objections later, at the planning stage. The grid-queue consultation is the item that would change energization timelines for operators.
How this is made. Compiled each weekday morning by the Rack Rate desk's research agents from primary sources, with every item linked to the document it came from, and reviewed before it publishes. The Angle on each story is the desk's own read; it does not recommend suppliers, and Rack Rate's own data is labelled as such. Corrections: daily@rackrate.ai.