Silicon Lien · Research · Data & trackers

Research · Dated note

Who Consumes the Compute — Q2 2026

Dated note · August 3, 2026

Summary

The question

Four large US compute sellers with reported backlog now show roughly $2.33T of contracted revenue, about $1.52T more than a year ago. Underwriting the infrastructure behind it requires a view on where the demand comes from. Concentration in a few frontier labs carries different risk from demand spread across enterprises. Disclosures only provide a partial answer to this question.

What four large US sellers disclose

Latest reported backlogs: Microsoft $678B commercial RPO (Jun), Oracle $638B (May), Alphabet $519.5B (Jun), and Amazon $496B (Jun, primarily AWS).

USD billions; total contracted services, not GPU-only demand. The OpenAI segment uses its December 2025 dollar amount and assumes it remained unchanged through June 2026.

Disclosed neocloud relationships

Terms and recognition timing differ. No quantified direct neocloud compute-purchase contract was identified for Oracle, Alphabet, or Amazon. Alphabet’s disclosed neocloud exposure takes a different form: quantified lease backstops for Fluidstack (credit support, not a compute purchase), disclosed in TeraWulf’s and Cipher Mining’s SEC filings.

Scope: seller backlogs include CPU, storage, databases and, for Microsoft, software seats — not GPUs alone.

Outside the four-seller perimeter

This is not a global supplier census. SpaceX/xAI reports gigawatt-scale AI compute capacity; Alibaba committed more than $53B to AI and cloud infrastructure over three years; and Tencent spent RMB79.2B on 2025 capital expenditures. Planned sovereign capacity includes HUMAIN at up to 500 MW and Mistral at 200 MW by 2027. These are investment or capacity measures, not comparable RPO, and are excluded from the $2.33T total.

How the backlog got here

The combined balance rose from about $807B in June 2025 to roughly $2.33T. Oracle jumped in September. Microsoft disclosed 110% year-over-year RPO growth in December. Alphabet’s total revenue backlog reached $467.6B in March, of which $462.3B was Google Cloud. Amazon rose from $244B in December to $496B in June.

Backlog trend for Microsoft, Oracle, Alphabet, and Amazon from March 2025 through June 2026 Microsoft rises from 315 billion dollars in March 2025 to 678 billion in June 2026. Oracle rises from 130 billion in its fiscal quarter reported in March 2025 to 638 billion in its fiscal quarter reported in June 2026. Alphabet rises from 92.4 billion in March 2025 to 519.5 billion in June 2026. Amazon long-term performance obligations, primarily related to AWS, rise from 189 billion in March 2025 to 496 billion in June 2026. $0B$200B$400B$600BMar-25Jun-25Sep-25Dec-25Mar-26Jun-26 Microsoft Mar-25: $315BMicrosoft Jun-25: $368BMicrosoft Sep-25: $392BMicrosoft Dec-25: $625BMicrosoft Mar-26: $627BMicrosoft Jun-26: $678B Oracle fiscal Q3 ended Feb-25; reported Mar-25: $130BOracle fiscal Q4 ended May-25; reported Jun-25: $138BOracle fiscal Q1 ended Aug-25; reported Sep-25: $455BOracle fiscal Q2 ended Nov-25; reported Dec-25: $523BOracle fiscal Q3 ended Feb-26; reported Mar-26: $553BOracle fiscal Q4 ended May-26; reported Jun-26: $638B Alphabet Mar-25: $92.4BAlphabet Jun-25: $106BAlphabet Sep-25: $155BAlphabet Dec-25: $242.8BAlphabet Mar-26: $467.6B total revenue backlogAlphabet Jun-26: $519.5B Amazon long-term performance obligations, primarily AWS, Mar-25: $189BAmazon long-term performance obligations, primarily AWS, Jun-25: $195BAmazon long-term performance obligations, primarily AWS, Sep-25: $200BAmazon long-term performance obligations, primarily AWS, Dec-25: $244BAmazon long-term performance obligations, primarily AWS, Mar-26: $364BAmazon long-term performance obligations, primarily AWS, Jun-26: $496B Microsoft 678Oracle 638Alphabet 519.5Amazon 496
Microsoft commercial RPOOracle RPOAlphabet revenue backlogAmazon performance obligations (primarily AWS)
USD billions; definitions differ. Oracle points use the calendar-quarter column containing the reporting date; its fiscal quarters end in February, May, August and November.

Two buyer channels

Disclosures reveal two overlapping paths to demand: frontier labs such as OpenAI, Anthropic, SpaceX/xAI and Mistral, including labs inside Google and Meta; and enterprises buying through clouds, APIs and applications. An enterprise may consume compute through a frontier lab’s API, so the categories are not additive. Sellers are intermediaries, not end demand.

Channel 1: what the labs account for

Microsoft alone quantifies lab exposure. In December, OpenAI was 45% of $625B of RPO, or roughly $281B. In June, Microsoft said all sequential growth came from customers “outside of frontier model companies” and ex-OpenAI RPO grew 25% year over year. If OpenAI’s December dollar balance is assumed unchanged, it would represent approximately 41.5% of June RPO.

Oracle rose from $138B to $455B in one quarter after naming OpenAI, xAI (now part of SpaceX), Meta, NVIDIA and AMD as customers, then reached $638B. The list mixes labs, integrated companies and suppliers; Oracle disclosed no split.

Amazon reports $496B of performance obligations, primarily AWS, with a 6.4-year weighted-average remaining life. The filing identifies a $100B OpenAI expansion and a greater-than-$100B Anthropic expansion, but does not allocate the balance by customer. Amazon also recorded $53.4B of other income, net, primarily from Anthropic-related valuation adjustments.

Channel 2: what the enterprise base measures

Ramp’s June spend index, covering 70,000+ US businesses, shows 55.0% paying for AI tools, with Anthropic at 42.4% and OpenAI at 39.5%. Menlo Ventures estimates 2025 enterprise generative-AI spend at $37B, triple the prior year. Microsoft’s ex-OpenAI RPO grew 25% year over year.

Menlo’s $37B is annual spending; Microsoft’s roughly $281B OpenAI balance is a multi-year stock. The figures are therefore not directly comparable, but together they show broad enterprise adoption and large lab commitments. They do not reveal how seller backlog divides between the two.

Where neoclouds sit

Neoclouds complicate attribution because they supply capacity to hyperscalers, frontier labs and other buyers. Their contracts reveal how capacity moves through the system, but not necessarily who ultimately consumes it. Adding neocloud contract values to seller backlog would therefore risk counting the same underlying demand twice.

Microsoft has a $17.4B Nebius contract, with an option to $19.4B. CoreWeave reported $5.131B of FY2025 revenue and said 67% came from Microsoft, which implies approximately $3.44B for the year. Lambda disclosed a multibillion-dollar Microsoft agreement without a value. These are evidence of capacity procured through neoclouds, not separate end-user demand.

The pattern extends beyond Microsoft. Meta committed up to about $35.2B to CoreWeave across a September 2025 order form of up to approximately $14.2B and a new approximately $21B March 2026 commitment, and committed up to $27B to Nebius. OpenAI committed about $22.4B to CoreWeave, while NVIDIA’s CoreWeave order form had a $6.3B initial value. The chart sums these disclosed headline values; its Microsoft marker is an annual-revenue estimate derived from CoreWeave’s disclosed revenue and customer concentration and is excluded from the total.

SpaceX/xAI blurs the boundary further because it is both a frontier lab and a compute provider. It operates gigawatt-scale capacity for Grok and has agreed to supply roughly 325,000 GPUs to Anthropic and 110,000 to Google, while Oracle names xAI among its largest OCI customers. The customer agreements can be terminated on 90 days’ notice after short initial periods, so multiplying monthly fees into firm multi-year values would overstate committed demand.

These relationships help trace capacity, but they still do not reveal the ultimate mix of labs, hyperscalers and independent enterprises. No quantified direct neocloud compute-purchase contract was identified for Oracle, Alphabet or Amazon. Alphabet’s disclosed neocloud exposure takes a different form: quantified lease backstops for Fluidstack (credit support, not a compute purchase), disclosed in TeraWulf’s and Cipher Mining’s SEC filings.

Another view of demand: open versus closed models

None of the disclosures above quantify self-hosted open-model demand, but one venue publishes usage. OpenRouter, a model-routing marketplace, routed 245 trillion provider-reported tokens in the thirty days through August 3; about 72% went to open-weight models, and eight of the ten highest-volume models were open-weight. That is roughly 5.7 billion reported tokens per minute in total, including about 4.1 billion from open-weight models. Separately, OpenAI reported more than 15 billion API tokens per minute in March, while Alphabet said in late July that its model APIs process approximately 22 billion tokens per minute, up from 16 billion a quarter earlier. Providers use different tokenizers and counting conventions, so these figures show scale rather than standardized compute shares.

This establishes a floor on open-model demand: consumers can also call open models directly from hosting providers, or run them on owned or rented hardware, and much of that activity does not pass through a common public meter. The comparator rates are floors of their own: the OpenAI and Google figures count API traffic only, so OpenAI’s ChatGPT and Google’s consumer surfaces sit outside them, and we did not identify a comparable Anthropic disclosure. The readings are also months apart in a fast-growing market. Hosted open-model demand is becoming a neocloud product: Crusoe and Nebius both sell managed inference on open models. Their cited product announcements do not quantify inference revenue separately from GPU rental, so those disclosures are insufficient to derive the neocloud share of open-model hosting.

What this supports, and what it does not

Supported: rapid backlog growth, very large lab commitments and broad enterprise adoption. Amazon now discloses $496B in total and the two $100B-plus lab commitments, but not their shares. Not supported: that all growth is lab-driven, an AI-only backlog split, or exact customer contracts as known RPO components. The key missing datapoint is concentration by end user.

Sources

Accessed July 30–August 4, 2026.

  1. Microsoft FY26 Q3 metrics and FY26 Q2 earnings call — the $315B–$627B trend through March 2026 and OpenAI’s 45% December share.
  2. Microsoft FY26 Q4 earnings release and earnings call — $678B, +84%, ex-OpenAI +25%, duration and recognition timing.
  3. Oracle releases for Q3 FY25, Q4 FY25, Q1 FY26, Q2 FY26, Q3 FY26, and Q4 FY26 — the full $130B–$638B trend.
  4. Alphabet filings for Mar. 2025, Jun. 2025, Sep. 2025, Dec. 2025, Mar. 2026, and Jun. 2026 — the full $92.4B–$519.5B trend.
  5. Amazon Q1 2025 10-Q and filings for Jun. 2025, Sep. 2025, Dec. 2025, and Mar. 2026 — the $189B–$364B trend, primarily related to AWS.
  6. Amazon Q2 2026 10-Q — $496B performance obligations, 6.4-year weighted-average remaining life, and OpenAI and Anthropic commitment expansions.
  7. Ramp AI Index and Menlo Ventures, 2025 State of Generative AI in the Enterprise — first-party published panels; estimates not independently audited.
  8. CoreWeave 10-K FY2025 — customer concentration.
  9. CoreWeave–OpenAI initial agreement and expansion — company announcements.
  10. Nebius 6-K, Sep 8 2025 — Microsoft capacity agreement, term, value, expansion option and financing condition.
  11. Lambda announcement, Nov 3 2025 — multibillion-dollar Microsoft agreement; exact value not disclosed.
  12. CoreWeave–Meta September 2025 8-K, March 2026 8-K, and Q1 2026 release — an initial order form of up to approximately $14.2B plus a new approximately $21B commitment that includes an exercised option.
  13. Nebius–Meta announcement, Mar 16 2026 — up to approximately $27B over five years.
  14. CoreWeave 8-K, Sep 15 2025 — $6.3B initial-value NVIDIA order form.
  15. SpaceX prospectus disclosure, Anthropic agreement disclosure, and Google agreement disclosure — acquisition of xAI, compute scale and customer terms.
  16. Alibaba infrastructure investment, Tencent 2025 annual report, NVIDIA–HUMAIN announcement, and Mistral Compute — non-RPO global infrastructure scale.
  17. OpenRouter public rankings — provider-reported token volumes, trailing 30 days through August 3, pulled August 4, 2026; open-weight versus closed classification is ours.
  18. Alphabet Q2 2026 earnings remarks, July 22 — “Our model APIs are now processing approximately 22 billion tokens per minute. That’s up from 16 billion just a quarter ago.”
  19. OpenAI, March 2026 — “Our APIs now process more than 15 billion tokens per minute.”
  20. Alphabet Q1 2026 earnings remarks — “more than 16 billion tokens per minute via direct API use.”
  21. Crusoe Managed Inference launch, Nov 2025; Nebius Token Factory launch — managed inference on open models; the cited product announcements do not quantify inference revenue separately from GPU rental.

General information from cited public sources, provided without warranty and not investment, legal, accounting, tax, appraisal, valuation, rating, or credit advice. Backlog definitions differ by issuer, and commitments may be conditional, amended, or recognized over different periods. Company and product names are used only for identification; no affiliation or endorsement is implied. Corrections: info@siliconlien.com.