Looking back at April to June 2026, the public record shows buyers committing more money to AI compute than suppliers could deliver. The pressure on cost also moved. Capacity stayed short, and memory became the component that buyers named when they raised their budgets.
This review sorts the quarter into six threads: spending, supply, rental prices, financing, market structure and policy. Every figure comes from a public source listed in the full report.
Spending plans rose again
Meta and Amazon reported first-quarter results on 29 April [9][10]. On its own first-quarter call, Alphabet guided 2026 capital expenditures to $180 billion to $190 billion and said 2027 would increase significantly [7]. Meta raised its 2026 range to $125 billion to $145 billion, from $115 billion to $135 billion, and gave the main reason as "higher component pricing this year" [9]. Amazon's purchases of property and equipment were $44.2 billion in the quarter to March, and its trailing free cash flow fell to $1.2 billion from $25.9 billion a year earlier [10].
On 1 June, Alphabet announced $80 billion of equity offerings, including a $10 billion private placement with Berkshire Hathaway, as part of its plan to fund AI compute [7]. Its stated reason was demand "at levels that are exceeding the company's available supply" [7]. FactSet estimated that new debt funded 32% of capital spending at five large hyperscalers in the twelve months before June, up from 9% in fiscal 2024 [24].
Supply: accelerators and memory
NVIDIA reported data center revenue of $75.2 billion for its quarter ended 26 April, up 92% from a year earlier, with networking revenue up 199% [1]. Its outlook for the next quarter was $91.0 billion and assumed no data center compute revenue from China [1]. At Computex on 1 June, it said Vera Rubin was ramping into full production [2]. AMD's data center segment reported $5.8 billion, up 57% [3].
Memory was the sharper story. TrendForce expected conventional DRAM contract prices to rise 58% to 63% in the second quarter, and later reported DRAM industry revenue of about $154.7 billion for the quarter, up 59.5% [5][6]. Micron's revenue for its quarter ended 28 May was $41.46 billion, against $23.86 billion the quarter before [4]. TrendForce tied the increase to AI inference deployments at North American cloud providers and to long-term supply agreements [5].
Rental prices split by tier
Silicon Data publishes daily GPU rental indexes and keeps the hyperscaler and neocloud tiers apart. Its figures for April and May show two different markets for the same chip [17].
| Series (USD per GPU-hour) | 1 April | Range | 31 May | Annualized volatility |
|---|---|---|---|---|
| H100 hyperscaler on-demand | $7.44 | $7.40 to $7.55 | $7.46 | 3.8% |
| H100 neocloud | $2.63 | $2.47 to $2.76 | $2.73 | 18.4% |
The hyperscaler series barely moved: 43 of 60 daily changes were zero [17]. The neocloud series fell 6% in the first week of April, then rose 7.1% in May [17]. In June, the neocloud H100 index peaked at $2.79 on 5 June and slipped to $2.59 late in the month, while the H200 premium over the H100 swung from near parity to about 11% [18].
Operators reported firm pricing. Nebius said older-generation GPUs priced more than 30% higher in the quarter than in the first quarter, and that four large contracts carried annual contract value of $20 million to $25 million per megawatt [14]. That is a contract price per unit of power, a different measure from a posted price per GPU-hour.
Backlogs and the debt behind them
CoreWeave reported a revenue backlog of $99.4 billion, more than 1 GW of active power and over 3.5 GW contracted [12]. It also reported an $8.5 billion non-recourse, investment-grade delayed draw term loan and a $2 billion equity investment from NVIDIA [12].
Oracle reported remaining performance obligations of $638 billion and negative free cash flow of $23.7 billion for its fiscal year ended 31 May [11][25]. It planned to raise about $40 billion of debt and equity in fiscal 2027, and said prepaid and customer-supplied hardware within large AI contracts totaled $75 billion [11].
On 29 May, an IREN subsidiary signed about $3.6 billion of financing, secured by the GPUs it buys and by the cash flows of a contract to provide GPU services to Microsoft, with a minimum debt service coverage ratio of 1.05 [15]. Debt is following specific hardware and specific customers.
Exchanges moved toward GPU price risk
Two exchange groups announced plans for futures on GPU rental prices. On 12 May, CME Group and Silicon Data announced a compute futures market based on Silicon Data's on-demand rental indices, pending regulatory review [20]. On 19 May, Intercontinental Exchange and Ornn announced cash-settled GPU compute futures on Ornn's Compute Price Index, which their joint release describes as built only from printed transactions [21].
The difference matters. One index starts from published rental rates; the other from completed trades. Anyone hedging with either needs to know which of the two its own contracts resemble.
Policy and power
On 31 May, the Bureau of Industry and Security reaffirmed that exports of advanced computing items to entities headquartered in Country Group D:5 or Macau, or owned by a parent headquartered there, need a license wherever the entity sits [22]. On 18 June, FERC ordered six grid operators to justify or change how large loads such as data centers connect to the grid [23].
What the quarter means
For buyers, the tier mattered: the two H100 indexes differed by a factor of roughly 2.7 to 3 and moved differently. For operators, memory became a cost that buyers named in their guidance. For lenders, the quarter's deals tie debt to identifiable GPUs and named contracts, and rental income behind that collateral behaves differently by tier. The segmentation in this review follows the Rillor Compute Index methodology v1.0, which shows hyperscaler and neocloud segments separately in every series.
Read the full report
The full report, with tables for each thread, the method, the limits of the analysis and the 25 numbered sources behind every figure, is available as a PDF.