Looking back at July to September 2026, two things stand out. Buyers raised their spending again while saying they still could not get enough capacity. And the price of renting a GPU, until now a matter of provider pages and private contracts, moved toward regulated futures markets, with the regulator asking hard questions first.
This review follows the quarter through spending, supply, rental prices, financing and the review of compute futures. Every figure comes from a public source listed in the full report.
Spending rose, capacity stayed short
Large buyers reported April to June results in late July. Alphabet's purchases of property and equipment were $44.9 billion, double the year before, and it raised 2026 guidance to $195 billion to $205 billion [8][9]. Its chief financial officer said, "We're still in a supply-constrained environment" [9]. Amazon guided to about $220 billion of 2026 cash capital expenditures, up from about $200 billion, and named the higher cost of memory as the reason [12]. Meta raised the low end of its range, narrowing it to $130 billion to $145 billion [10]. Microsoft's additions to property and equipment were $35.8 billion, against $17.1 billion a year earlier [13].
Cash flow showed the cost. Alphabet's free cash flow for the quarter was negative $5.9 billion [8]. Amazon's trailing twelve-month free cash flow was an outflow of $7.6 billion [11].
Supply: chips ramp, memory sets the pace
NVIDIA reported data center revenue of $89.0 billion for its quarter ended 26 July, up 117% from a year earlier, and guided the next quarter to $108.0 billion [1]. It said Vera Rubin was ramping into full production with racks running at partners [1]. Its supply commitments rose to $279 billion from $119 billion in one quarter, primarily for memory and manufacturing [2]. AMD said its Helios rack was in production and reported data center revenue of $6.7 billion, up 107% [4][5].
Memory prices kept rising. TrendForce expected server DRAM contract prices to rise 13% to 18% in the third quarter, with multi-year agreements limiting increases for several U.S. cloud providers [6]. Micron's quarterly revenue reached $54.23 billion, against $11.32 billion a year earlier [7].
Rental prices: newer chips firmed, B200 went quiet
Silicon Data's neocloud indexes, in U.S. dollars per GPU-hour, show the spread by generation [15][17].
| Series | Earlier point | Later point | Change |
|---|---|---|---|
| H200 | $2.70 (4 May) | $3.09 (27 July) | +14.4% |
| H100 | $2.57 (4 May) | $2.75 (27 July) | +7.0% |
| B200 | $5.42 (1 June) | $5.62 (1 September) | +3.7% |
The H200 premium over the H100 averaged 12.1% in July [15]. The B200 index held between $5.58 and $5.69 through August, with volatility about a quarter of July's [17]. On 7 September, a twelve-month B200 commitment priced about 5% below spot [17].
Posted reservation prices rose. AWS raised EC2 Capacity Block prices by about 20% from 1 July for its P6-B300, P6-B200, P5, P5e, P5en and P4de instance families, according to a press report of AWS's documentation [18]. Nebius reported four second-quarter deals at $20 million to $25 million of annual contract value per megawatt, said it saw a price opportunity of $40 million to $50 million per megawatt and had signed its first such contract, and estimated the payback period on its second-quarter deals at one year and ten months [20].
Silicon Data also reported that the hyperscaler premium over neoclouds for H100, averaged across seven regions, narrowed to about 2 to 2.5 times in the third quarter to date (its data ran through July), from 3 to 4.5 times in the first quarter of 2025 [16].
Financing widened, and so did the warnings
CoreWeave reported a revenue backlog of about $104 billion at 30 June, before more than $25 billion of net new customer commitments added early in the third quarter, and net interest expense of $640 million for the quarter [21]. NVIDIA announced partnerships with six large asset managers and banks intended to mobilize more than $500 billion of third-party capital for AI infrastructure, subject to definitive agreements [1].
Ratings and filings also flagged risk. S&P lowered Oracle to BBB- on 9 July [22]. In September, Oracle reported remaining performance obligations of $664 billion and negative free cash flow of $5.4 billion for its quarter [23]. SEC staff concluded on 29 July that data center securitizations do not issue asset-backed securities; a law firm summary notes that GPU securitizations still have to be assessed on their own facts [24]. Nscale's registration statement for a planned New York Stock Exchange listing, filed 18 September, warned that "compute is trending towards commoditization, placing increasing pressure on unit economics" [25].
Compute futures meet the CFTC
The quarter's main structural change was regulatory.
| Date | Event |
|---|---|
| 1 July | ICE and NATIVX announced cash-settled futures on an energy-normalized compute index [26] |
| 11 August | CME Group and Silicon Data set 5 October, pending review, for H100 and B200 rental index futures [27] |
| 19 August | The CFTC requested public comment on compute derivatives, with comments due 20 October [28][29] |
| 23 September | GuruFocus reported that the CFTC had extended its review of the CME contracts by 45 days, to 9 November [31] |
| 29 September | ICE said its Ornn-based H100 and B200 futures would list after the comment period, with no date set [32] |
The CFTC's chairman said, "America cannot win the AI race without a robust derivatives market for compute" [28]. The questions are precise. What proportion of compute transactions occur at publicly disclosed prices? Would it be appropriate to permit a contract settling to a price computed from data the Commission may not be able to observe, verify or surveil [37]? The Commission also wrote, as a preliminary view, that compute may not yet show the fungibility, standardization and liquidity that usually underlie a commodity derivatives market, and that the underlier could be rented capacity or another form, such as inference tokens [37].
Those are index questions as much as futures questions. Every GPU price series has to say what evidence it rests on, how it compares unlike offers and who can check it.
What the quarter means
For buyers, generation and term mattered alongside tier: H200 rents rose twice as fast as H100 rents, and twelve-month B200 commitments priced below spot. For operators, memory costs kept rising. For lenders, financing reached new scale while ratings and filings flagged negative cash flow and pressure on unit economics. For index users, the CFTC's questions set the standard. The Rillor Compute Index methodology v1.0 records the evidence class of every observation and keeps hyperscaler and neocloud segments apart; its series are for research and planning use and are not offered for settlement of any contract.
Read the full report
The full report, with the complete regulatory timeline, tables for each thread, the method, the limits of the analysis and the 38 numbered sources behind every figure, is available as a PDF.