FRI 30 OCT 2026 · 5 MIN READ
Compute demand is outrunning every part of the stack, not just GPUs.
Amazon spent $75.9B on capex in one quarter, and Intel says it cannot fill datacenter CPU orders. Small GPU operators are borrowing at about 9% against signed contracts.
General AI Infrastructure
AWS spends $75.9B in a quarter on AI build
TakeawayAmazon reported Q3 capex of $75.9 billion, mostly for AWS, and kept 2026 cash capex near $220 billion.
- AWS revenue was just under $46 billion, up 39.3%.
- The Register reports AWS installs an AI accelerator about every 14 seconds.
- Management said AWS could grow 40% or more in 2027 or 2028.
AngleCapex ran about 1.65 times AWS revenue in the quarter (computed from $75.9B and roughly $46B; capex covers all of Amazon, so it is a rough ratio). Suppliers of power, memory and networking are the ones who feel that pace first.
Intel warns it cannot meet datacenter CPU demand
TakeawayIntel said Q3 datacenter and AI sales rose 78% and it will likely not fully meet customer demand.
- CFO Dave Zinsner cited wafer and other supply shortages; first half of 2027 looks harder than the second.
- Intel is moving some PC chip capacity to datacenter products.
- Shares fell about 24% after hours.
AngleInference and agentic workloads are pulling CPUs, not just GPUs, into the bottleneck list. Buyers sizing 2027 clusters should treat host CPU supply as a lead-time item.
Oracle sets $90B-$95B capex for fiscal 2027
TakeawayAt its Oct. 28 investor day, Oracle laid out a fiscal 2027 build plan and a move to usage-based AI pricing.
- Capex: $55.7 billion in fiscal 2026, planned $90 billion to $95 billion in fiscal 2027.
- Oracle hopes to raise about $40 billion in debt and equity in 2027.
- Customers would pay per AI task completed, not per seat.
AnglePer-task billing ties Oracle's revenue to inference volume, which is the same volume its capex is built for. The $40 billion raise is the number to watch for how that build gets funded.
Seagate: nearline supply sold under contracts through 2028
TakeawaySeagate posted record quarterly revenue of $4.8 billion as cloud customers kept buying capacity drives.
- It shipped a record 198 EB of nearline drives to data centers.
- Most nearline supply is under long-term agreements through 2028, some to 2030.
- HAMR should be about 50% of nearline exabytes exiting 2026.
AngleStorage for training data and inference logs is being locked up the way GPUs were. Operators who have not contracted capacity for 2027 are bidding for what the hyperscalers left over.
Buyer Signals & Deals
OpenAI paid Microsoft $10B in a quarter
TakeawayMicrosoft's filing shows $10 billion of revenue from OpenAI in the quarter ended Sept. 30.
- That is nearly 11% of Microsoft's $92.3 billion quarterly revenue.
- OpenAI receivables rose to $7.2 billion from $6 billion at end of June.
- OpenAI committed to an additional $250 billion of Azure in October 2025.
AngleOne lab is now a double-digit share of a hyperscaler's revenue, and its receivables are growing. Capacity planning at Azure depends on that one customer's ability to keep paying.
Neocloud & GPU Cloud Moves
HIVE's BUZZ borrows $129M at 9.0% for GB200s
TakeawayBUZZ HPC closed a GPU-secured loan to build a cluster for a three-year Cohere contract.
- Fixed 9.0%, three-year term, amortizing; about 80% of cost.
- Funds 2,303 NVIDIA GB200 GPUs; delivery expected toward the end of Q4 2026.
- Lender is described only as an institutional lender.
AngleNine percent fixed on an 80% advance, repaid from a contracted customer, is the going price for small-operator GPU debt. Compare it with the 9.95% SharonAI paid earlier this month.
USD.AI leads $106M loan for NexGen Cloud's B300s
TakeawayNexGen Cloud secured a four-year, $106 million senior secured term loan against new Blackwell capacity.
- Collateral: 216 NVIDIA HGX B300 servers, or 1,728 GPUs, in a Finland data center.
- Capacity is fully contracted under a take-or-pay agreement; rate not disclosed.
- Co-lenders: Silver Point Capital and Victory Park Capital.
AngleLenders are underwriting the contract, not the chips. Take-or-pay is what gets a GPU loan done, and the undisclosed rate is the term to ask about.
Model Releases & Compute-Demand Drivers
Mistral opens weights for trillion-parameter model
TakeawayMistral released the weights of Mistral Large 4, a model of about one trillion parameters, on Hugging Face and its API.
- Aimed at cybersecurity work and on-premises deployment.
- Mistral claims it uses 47% fewer tokens per task than GLM-5.3.
- License and hardware requirements are not stated in the article.
AngleA trillion-parameter open model gives enterprises a reason to buy their own multi-GPU nodes. The token-efficiency figure is Mistral's own claim and needs independent testing.
Power & Datacenter Buildout
Quantum DC proposes 1.2GW campus near Adelaide
TakeawayQuantum DC plans a 1.2GW data center and energy hub at Tungkillo, South Australia, with no application lodged yet.
- Three stages: first operations in 2028, full build by 2031.
- About 80% wind and solar, 20% on-site gas turbines to firm supply.
- Local reports cite 400MW of gas; the developer has not confirmed a figure.
AngleFirming with gas rather than running it continuously is a design that other grid-constrained regions may copy. Approval by the planning minister, after at least 15 business days of public comment, is the gate.
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.