Research · Sector & thematic · November 10, 2026
Compute Strategies at the Frontier Laboratories
A profile-by-profile reading of how Anthropic, OpenAI, Google, Meta, xAI/SpaceX, and Amazon are securing chips and power—using filings and company announcements where they exist, and labelling everything else.
Anthropic figures are from Amendment No. 2 to the Form S-1, filed November 10, 2026 (updating the $895.4 billion aggregate in the November 2 amendment). Meta commitment items are from the company’s third-quarter 2026 Form 10-Q. Other figures are company announcements or reputable news reports and are labelled as such. Turn-on rankings in the closing table are the author’s qualitative assessment. This note is for information only and is not investment advice.
1. Anthropic
- Models. The Claude family, as described in the company’s registration statement.
- Compute. The S-1/A describes a multi-chip approach: Google TPUs, including five-year Broadcom equipment finance leases (aggregate rent about $161.2 billion excluding interest); an Amazon agreement of at least $110.0 billion, which the filing elsewhere links to Trainium and compute infrastructure; NVIDIA GPUs through Microsoft ($47.2–49.2 billion), SpaceX.AI LLC (up to $120.5 billion), Nscale, and Lambda; and an unfixed AMD hardware agreement. This is diversification as filed, not a slogan.
- SpaceXAI. Up to $120.5 billion of NVIDIA GPU capacity into 2029, about $3.1 billion a month once all capacity is delivered (expected December 2026), cancellable by either party on 90 days’ notice after initial periods ending in January and March 2027. See the contract note.
- Investors and suppliers. The S-1/A describes compute supply and, separately, financing or investment ties involving Amazon, Google, and Broadcom, and a concurrent Class N private placement that includes Google. The company has applied to list Class A shares under the symbol ANT; the offering is not completed.
- Framing, as analysis. A mix of hard minima and a flexible NVIDIA contract with a supplier that is building capacity to sell. Whether that produces faster turn-on is an open question, discussed in the related notes.
2. OpenAI
- Models. The GPT family and related reasoning models, as the company describes them. No registration statement is available.
- Compute. Public arrangements are still described as NVIDIA-centered, with large capacity leased through partners: incremental $250 billion of Azure services, disclosed by Microsoft on October 28, 2025; up to 4.5 gigawatts of additional U.S. capacity with Oracle, announced by OpenAI in July 2025 (a later computing agreement of about $300 billion was reported by the Wall Street Journal and Reuters in September 2025); up to $22.4 billion with CoreWeave, reported by Reuters and Bloomberg on September 25, 2025; and a $38 billion AWS agreement (Reuters, November 3, 2025) later expanded by $100 billion over eight years, including about 2 gigawatts of Trainium, announced by OpenAI and Amazon on February 27, 2026.
- NVIDIA lease guarantees. Reuters and Bloomberg, on August 17, 2026, describing a NVIDIA disclosure, reported a guarantee of up to $105 billion tied to leases on an Ohio data-center campus. This is a contingent residual-value guarantee, not a $105 billion commitment to buy chips.
- Investors. Microsoft is the closest disclosed partner (equity plus Azure). SoftBank appears in Stargate and related announcements. These are company or news sources, not an OpenAI S-1.
- Framing, as analysis. Absolute contracted scale is large, and much of it depends on partners building and powering sites.
3. Google (Alphabet)
- Models. The Gemini family.
- Compute. Custom TPUs, designed and used in-house. CNBC reported in November 2025 that Gemini 3 was trained entirely on TPUs. Google Cloud also sells NVIDIA capacity to customers (Business Insider, April 2026). “Low NVIDIA dependence” applies most cleanly to Google’s own frontier training, not to every Cloud product.
- Anthropic TPU contracts. The Anthropic S-1/A states a Google Cloud minimum of at least about $121.3 billion and, separately, additional unfixed TPU capacity that Anthropic estimates at $125–150 billion, not included in the $900.9 billion aggregate. Those estimates are Anthropic’s, for Anthropic’s own arrangements—not a Google company-wide sales total.
- Eighth-generation TPUs and Oracle. A Google Cloud blog post on October 27, 2026, described eighth-generation TPUs and a partnership with Oracle to offer Google TPUs on Oracle Cloud Infrastructure starting in calendar 2027, which the companies said would be the first time a hyperscaler other than Google offered TPUs. Pricing and initial OCI capacity were not disclosed. This is a company announcement, not a completed deployment.
- Investment in Anthropic. The S-1/A describes Google as a compute counterparty and as a purchaser of Class N shares in a concurrent private placement (up to $10 billion from Google in the placement as described). This note does not invent a current ownership percentage.
4. Meta
- Models and products. The Llama family. “Muse” is Meta’s AI agent; NVIDIA said on October 27, 2026, as reported by CNBC, that it would deepen its partnership with Meta on Muse, which had launched the previous month. That is a company/news description, not an appraisal of the product.
- NVIDIA partnership. On February 17, 2026, NVIDIA announced a multiyear partnership covering millions of Blackwell and Rubin GPUs (NVIDIA release; Reuters). Meta’s October 19, 2026 engineering post on Prometheus describes a gigawatt-scale supercluster of hundreds of thousands of GPUs, intended to scale toward about one million, with GB200 and GB300 systems and Vera Rubin NVL72 arriving. On October 27, NVIDIA also said, in connection with its equity investment, that Meta would expand Vera Rubin and Vera CPU deployment.
- AMD Helios. AMD and Meta announced on February 24, 2026, a multi-year agreement to deploy up to 6 gigawatts of AMD Instinct GPUs, with first-gigawatt shipments expected in the second half of 2026 on the Helios rack-scale architecture (AMD newsroom; Bloomberg). On October 27, 2026, AMD said initial Helios rack-scale systems were installed and being tested at Meta, ahead of a broader deployment (AMD newsroom). Testing is not the same as gigawatt-scale production.
- Equity plan. Reuters reported on October 27, 2026, that Meta increased the planned size of its equity offerings to $97.5 billion, including a common-stock offering, convertible preferred depositary shares, and a separate at-the-market program of up to $60 billion expected to begin after planned Class C shares start trading in the first quarter of 2027. This is a financing plan as reported, not an amount already received in full. The same day, NVIDIA and AMD announced investments of $3 billion and $1 billion, respectively, in the equity offering (Reuters, CNBC). Meta had said NVIDIA, AMD, and Broadcom were interested in purchasing up to $5 billion of shares.
- Lease and contract commitments. The company’s third-quarter 2026 Form 10-Q (filed October 22, 2026, Note 10) does not use a single “$825 billion” caption. It discloses, as of September 30, 2026, about $355.73 billion of operating and finance leases not yet commenced; $454.52 billion of non-cancelable contractual commitments (mostly cloud capacity, servers and network infrastructure, energy agreements, and Reality Labs hardware); and contingent obligations to purchase up to $14.72 billion of cloud capacity over five years. Those three items sum to about $825 billion, a total used in press coverage (including Bloomberg). The sum is a convenience, not a line in the filing. The contractual-commitments line also includes items other than data-center compute.
- Turn-on versus SpaceX. A claim that Meta is slower than SpaceX on permitting, power, or construction is the author’s assessment, not a finding in the 10-Q, and is not attributed to “many observers.”
5. xAI / SpaceX (SpaceXAI)
- Models. The Grok family, as the companies describe them.
- Compute. NVIDIA-focused, self-built clusters, with capacity also leased to Anthropic (up to $120.5 billion, S-1/A). SpaceX’s third-quarter 2026 results, as compiled on this site, showed Super Intelligence as the largest segment by revenue, with nameplate compute of 1.7 gigawatts and management commentary pointing to 2.5–3 gigawatts by year-end (outlook, not a historical result).
- Colossus 2 target. Data Center Dynamics has reported Elon Musk as saying that Colossus 2 would have around 1.2 million NVIDIA Blackwell GPUs by the end of 2026, including additional Southaven, Mississippi, sites. Bloomberg, on September 25, 2026, reported Musk as saying the cluster might more than double its NVIDIA chip count by year-end. These are executive targets and press reports, not a verified installed count or an SEC figure. This note does not treat 1.2 million GPUs as already in service.
- Power and construction speed. The claim that SpaceX sits at the top of the industry on behind-the-meter power and build speed is an analytical judgment in this note, not a measured ranking.
6. Amazon (AWS)
- Role. Cloud provider and designer of Trainium chips, while AWS also sells NVIDIA instances.
- Anthropic. A minimum of $110.0 billion, May 2026–April 2036, regardless of usage, with a true-up (S-1/A). The $110 billion contract does not break out chip type; the filing describes the Amazon collaboration as providing Trainium access.
- OpenAI. The $38 billion and subsequent $100 billion / ~2 GW Trainium announcements cited in the OpenAI section. Amazon has thus been announced as a supplier to both laboratories.
7. A compact comparison — including a qualitative ranking
The last two rows are the author’s assessment, not a filed or surveyed ranking.
| Lens | What it describes | Where it points, on the record or as assessment |
|---|---|---|
| Diversification and flexibility | Several chip families and a cancellable large NVIDIA contract | Anthropic, as filed in the S-1/A |
| Vertical integration | Custom silicon for own models and, in Amazon’s case, for cloud customers | Google (TPU), Amazon (Trainium) |
| Self-owned cluster scale | Large self-built installations | Meta (Prometheus, as Meta describes it); xAI/SpaceX (Colossus, as executives and the SpaceX segment results describe it) |
| Turn-on speed (analysis) | Ability to energize and use new capacity in the near term | Author’s order: xAI/SpaceX, then Anthropic via SpaceXAI, then OpenAI and Meta. Not a measurement, and not “the prevailing view.” |
| Reliance on the latest NVIDIA systems (analysis) | Blackwell / Vera Rubin weight | Author’s qualitative labels: high at xAI/SpaceX, Meta, and OpenAI; intermediate at Anthropic (about 39% of named contract dollars, derived); lower for Google’s own training. Not filed percentages except the Anthropic derivation, which itself is an estimate. |
This note is for information only and is not investment advice. Offerings and executive targets can change. Figures should be checked against the original filings and announcements.
Sources
- Anthropic, PBC, Amendment No. 2 to Form S-1, filed November 10, 2026; and Amendment No. 1, filed November 2, 2026.
- Microsoft, “The next chapter of the Microsoft–OpenAI partnership,” October 28, 2025. OpenAI, “Stargate Advances with 4.5 GW Partnership with Oracle,” July 2025. Reuters on the reported $300 billion Oracle agreement (September 10, 2025), the CoreWeave total (September 25, 2025), the $38 billion AWS agreement (November 3, 2025), and the NVIDIA lease guarantee (August 17, 2026). OpenAI, “OpenAI and Amazon announce strategic partnership,” February 27, 2026.
- Google Cloud, “Google and Oracle are partnering to bring Google TPUs to Oracle Cloud Infrastructure,” October 27, 2026. CNBC, November 25, 2025, on Gemini 3 and TPUs.
- Meta Platforms, Form 10-Q for the quarter ended September 30, 2026, filed October 22, 2026, Note 10, Commitments and Contingencies (Accession 0001628280-26-066773). Reuters, “Meta eyes record $97.5 billion equity raise,” October 27, 2026. CNBC, “Nvidia, AMD to invest $4 billion in Meta,” October 27, 2026. NVIDIA, “Meta Builds AI Infrastructure With NVIDIA,” February 17, 2026. Meta Engineering, “Inside Prometheus,” October 19, 2026. AMD, “Expanded Strategic Partnership to Deploy 6 Gigawatts,” February 24, 2026, and “AMD, Meta Advance Partnership,” October 27, 2026 (Helios systems installed and being tested).
- Data Center Dynamics, “SpaceXAI launching two additional data center campuses in Mississippi to further expand Colossus 2,” reporting Musk’s around-1.2-million-Blackwell year-end target. Bloomberg, “Elon Musk Aims to Double Colossus 2’s Nvidia Chips by Year-End,” September 25, 2026. This site’s SpaceX Q3 2026 results note.