Cloud & Infrastructure
AI compute got dearer in September, and data centres are being handed the grid bill
7 min read
CoreWeave is signing short GPU contracts at about $40m a megawatt, Oracle's $664bn backlog came with negative free cash flow, and US regulators moved to make data centres pay for grid upgrades. What that means for capacity plans.
In brief
The clearest price signal of September came from a specialist GPU cloud, not a hyperscaler. CoreWeave said it is signing three-to-six-month contracts at about $40 million per megawatt of annualised revenue. Short terms at high prices mean the seller sets the price. Oracle showed the other side of the same market: a $664 billion contract backlog, record deliveries of capacity, and free cash flow of about minus $5.4 billion for the quarter.
US regulators, meanwhile, spent the second half of the month making data centres pay for the grid upgrades they need. For anyone planning AI capacity, the result is higher unit prices, shorter terms, more risk concentrated in a few suppliers and customers, and new sites that cost more and take longer.
Price: a seller's market
- CoreWeave (16–17 September). The company says that since 30 June it has kept contracting capacity "at higher prices". That includes short-dated contracts at about $40 million per megawatt, measured as annualised revenue divided by the power required. It added more than $25 billion of new customer commitments early in the quarter, and its contracted power rose from about 3.7 gigawatts at 30 June to about 4.2 gigawatts at 11 August. It named no customers.
- Microsoft (reported 10 September). Bloomberg reported that Microsoft plans to take its capacity from about 12 gigawatts to more than 38 gigawatts by 2032, after shortages cost it business. Secondary accounts of that reporting say Temu signed with rival clouds and GitHub routed AI-agent traffic to AWS. Microsoft has not confirmed the figures.
- Broadcom (2 September). Chief executive Hock Tan listed the bottlenecks on the earnings call: "leading-edge silicon", "substrates", "HBM memory", and "land, power, and shell".
ByteNib found no change to hyperscaler list prices in September. The pressure is showing up in contract terms and availability first.
Balance sheets: who is carrying the risk
Oracle (10 September). Revenue rose 30% to $19.3 billion, and its infrastructure cloud grew 121%. Remaining performance obligations, the contracted revenue not yet earned, reached $664 billion, up $209 billion in a year. Oracle delivered 850 megawatts of data-centre capacity and more than 300,000 GPUs in the quarter, and spent $28.5 billion on capital expenditure. Its chief financial officer, Hilary Maxson, said about half the backlog should turn into revenue within 36 months, and that Oracle had not given a timetable for returning to positive free cash flow. A secondary reading of its quarterly filing adds $11.4 billion of customer prepayments and $19.9 billion raised by selling new shares. The filing does not say who paid.
Nscale (18 September). The neocloud's filing for a stock-market listing reports $103.4 billion of contracted value. That includes up to about $44.6 billion from Anthropic and about $43.8 billion from Microsoft, which Quartz puts at roughly 85% of the total. Against that, first-half revenue was $140.6 million and the net loss $1.02 billion. Nvidia has lent the company $3.1 billion in convertible notes. Fortune found that ByteDance supplied 73% of Nscale's 2025 revenue, using Nvidia chips at a facility in Norway through a Singapore entity. Fortune calls the arrangement legal but a source of "regulatory and reputational risks".
Anthropic and Akamai (24 September). Bloomberg reported an $11.6 billion, seven-year compute agreement. Secondary outlets add that Akamai granted Anthropic a warrant that vests as contract milestones are met, and that Akamai's shares rose 15%.
The pattern across all three is suppliers taking stakes in their customers' deals: Akamai's warrant, Nvidia's notes to Nscale, and Nvidia's backing of Lambda. That concentrates counterparty risk in a handful of AI labs. It also explains why headline totals mislead. Adding up backlogs is how "a trillion dollars of committed compute" gets written about a fortnight in which new deals with stated values came to about $65 billion. One outlet's "$1.03 trillion" headline about Nscale was a tenfold misreading of its $103.4 billion filing.
Power: data centres are handed the grid bill
| Date | Action | Effect |
|---|---|---|
| 16 Sep | The US House passes the Ratepayer Protection Act, 417 votes to 3 | States would have to consider standards making loads above 100 megawatts pay the "full, incremental" cost of grid upgrades and provide financial assurances. Senate passage before the midterms is seen as unlikely. |
| 18 Sep | North Carolina regulators reject Duke Energy's $584 million gas turbine | The commission found Duke's forecast growth, "based upon anticipated data center customer additions", not reliable enough. |
| 18 Sep | Texas regulators adopt softer interconnection rules | A proposed $50,000-per-megawatt fee was dropped, and study fees became a flat $100,000. |
| 21 Sep | California signs a seven-bill data-centre package | Data centres pay their "fair share" of grid upgrades and report energy use monthly. |
| 22 Sep | FERC refuses to let ComEd cancel a 1.8-gigawatt data-centre agreement | The deal was secured with a $1 letter of credit, which one commissioner called "an embarrassing legal fiction". The dispute moves to federal court. |
Clean firm power is coming, but not soon. Google's 396-megawatt geothermal agreement with Fervo, signed on 1 September, is due online in 2028. That does nothing for 2026 and 2027.
What leaders should do next
- Budget for dearer, shorter capacity. Where you need GPU capacity next year, price it at today's market, not last year's contracts, and compare the premium for short terms with the risk of a long one.
- Put suppliers' balance sheets in due diligence. For every AI capacity supplier, check free cash flow, customer concentration and who is financing the build. Oracle's negative cash flow, Nscale's two main customers and the new vendor-financing structures are all public.
- Design for the outage you will actually get. Google's us-central1 outage on 1 September lasted four hours and eleven minutes. Technicians unplugged every fibre path during routine maintenance, taking out two zones at once. Critical AI and data workloads need to survive losing a region, not just a zone.
- Expect the grid cost in your own projects. Anyone building or leasing new US data-centre capacity should expect demands to pay for grid upgrades and post real financial security, with Texas the partial exception for now.
- Do not add backlogs together. When a vendor, analyst or board paper cites trillion-dollar AI commitments, ask how much is new this quarter and how much is counted twice.
Sources and scope
CoreWeave's figures are from its release. Oracle's are from its results exhibit filed with the SEC, plus a secondary reading of its 10-Q by FinanceFeeds and call coverage from Yahoo Finance. Nscale's are from its S-1 filing, with Quartz and Fortune. The Akamai agreement is from Bloomberg and 24/7 Wall St.. Microsoft's plan is from Bloomberg, via TechTimes. Broadcom's comments are from an earnings-call transcript.
The power actions are from Utility Dive's reporting on the House bill, North Carolina, Texas, California and FERC, and from Fervo. The outage is from Google Cloud's incident report.
Bloomberg's Microsoft and Akamai reports were read through their headlines and other outlets' accounts, and are attributed as reported. The analysis and recommendations are ByteNib editorial analysis. ByteNib uses AI tools, including Anthropic's Claude, in research and drafting. Anthropic appears here as a customer in several of the deals described.
Continue exploring: Cloud & Infrastructure analysis, the related implementation tutorial, and the structured learning path.