OpenAI's compute plan: Stargate, the 26 gigawatts and who supplies what

OpenAI has announced chip deals with Nvidia, AMD and Broadcom that add up to a planned 26 gigawatts of capacity, on top of the Stargate data centre programme with Oracle and SoftBank. The chip deals were reported in October 2025, deployments are due to begin in the second half of 2026, and full rollout runs to 2029. Several of the financial terms have since changed or remain unsettled.

This post separates what was announced, what has been built, and what is still uncertain. It also explains the strategy behind the deals, because the shape of the plan says as much about OpenAI's direction as the numbers do. Every figure is labelled by source and date, and disagreements between sources are stated plainly. OpenAI is a private company and its figures are unaudited. This is not investment advice.

Key numbers at a glance

• Planned chip capacity: 26 gigawatts. At least 10 from Nvidia, at least 6 from AMD, 10 from Broadcom. Company announcements, 22 September, 6 October and 13 October 2025, summed by TrendForce on 15 October 2025.
• Nvidia investment: up to $100 billion, released as each gigawatt is deployed. Nvidia statement, 22 September 2025, a letter of intent. The first gigawatt was slated for the second half of 2026 on Vera Rubin.
• AMD warrant: up to 160 million AMD shares. AMD statement, 6 October 2025. Roughly 10 percent of AMD if fully exercised, per TrendForce.
• Broadcom: 10 gigawatts of custom accelerators, rollout by 2029. Reported 13 October 2025. No financial terms stated.
• Stargate headline: $500 billion and 10 gigawatts over four years. Announced 21 January 2025.
• Stargate active plan: more than $400 billion and about 7 gigawatts over three years. Reported 22 and 23 September 2025, restated by a tracker in May 2026.
• Oracle contract: $300 billion over five years, up to 4.5 gigawatts. Tracker, May 2026, announced summer 2025.
• Microsoft: an incremental $250 billion of Azure services. Microsoft statement, 28 October 2025. Microsoft holds about 27 percent on an as-converted diluted basis.
• Amazon Web Services: $38 billion. Amazon announcement, undated on the page. An analyst gives 3 November 2025 as the signing date.
• CoreWeave: $22.4 billion across three contracts. Analyst figure, 3 November 2025.
• Analyst total of public commitments: about $1.15 trillion for 2025 to 2035. Tomasz Tunguz, 3 November 2025, an estimate with a 30 to 50 percent error margin.
• March 2026 funding round: $122 billion at an $852 billion valuation. Tracker figures, May 2026, not tied to a primary source.
• 2024 cloud compute spend: about $7 billion. Epoch AI, 10 October 2025, from September 2024 investor documents.

Terms explained in plain language

• Gigawatt. A thousand megawatts of electrical power. In these deals it describes the power draw of the data centres, not a number of chips. TrendForce cites industry estimates of 4 to 5 million high end GPUs for 10 gigawatts, so one gigawatt is roughly several hundred thousand, depending on the chip.
• Letter of intent and memorandum of understanding. A statement that two parties plan to do a deal and will finalise terms later. It is not a binding purchase contract. Nvidia's own release calls its September 2025 document a letter of intent and said details would be finalised in the coming weeks. TrendForce calls it a memorandum of understanding. No primary source published by 19 October 2026 confirms that a final agreement was signed.
• Tranche. One instalment. Nvidia's money is described as released in steps as each gigawatt comes online.
• Warrant. The right to buy shares at a set price later. AMD gave OpenAI a warrant that vests (becomes usable) as milestones are reached.
• Accelerator. A chip built for AI work. GPUs are one kind. Custom accelerators are chips designed for a particular customer's workloads.
• Circular financing. A phrase critics use when a supplier invests in a customer who then spends the money on the supplier's products. The commitment tracker cited in this post applies it to Nvidia. It is a characterisation and not a legal finding.
• Owned versus leased capacity. Owning means OpenAI or its partners hold the buildings. Leasing means renting from a developer or cloud. The tracker says OpenAI reframed Stargate in mid 2025 as an umbrella term that covers both, and that most new capacity is leased.

The three chip deals

A trade press summary from 15 October 2025 sets out the three deals side by side. In plain terms:

• Nvidia, at least 10 gigawatts. A memorandum of understanding from 22 September 2025 covering Nvidia's Vera Rubin platform, with the first gigawatt-scale system expected online in the second half of 2026. Nvidia was reported to plan to invest up to $100 billion in OpenAI, released in tranches as each gigawatt comes online, with the first tranche possibly about $10 billion.
• AMD, at least 6 gigawatts. Announced 6 October 2025, based on AMD Instinct MI450 GPUs with rack-level systems. The first 1 gigawatt phase starts in late 2026. AMD granted OpenAI warrants for up to 160 million AMD shares, vesting as deployment milestones are met. Fully exercised, that would give OpenAI roughly 10 percent of AMD.
• Broadcom, 10 gigawatts. Announced 13 October 2025. OpenAI and Broadcom co-develop custom AI accelerators, and Broadcom supplies Ethernet networking. Deployment starts in late 2026, with full rollout by 2029. No financial terms were stated.

Ten plus ten plus six gives the headline 26 gigawatts. For scale, a gigawatt is roughly the output of a large power station. Remember that this is planned capacity, with the first systems due only in the second half of 2026, so most of it is plan and not yet hardware.

Some extra detail comes from the primary releases. AMD's release says the warrant vests in tranches: the first with the initial 1 gigawatt deployment, later ones as purchases scale toward 6 gigawatts, and vesting also depends on AMD reaching certain share price targets and on OpenAI reaching technical and commercial milestones. The release does not state the price targets, the shares per tranche or any dollar value. Nvidia's release says OpenAI will treat Nvidia as a preferred strategic compute and networking partner. TrendForce quotes Sam Altman saying the Broadcom teams had collaborated for 18 months, working from the transistor level up. TrendForce says each Nvidia funding round is valued at OpenAI's market valuation at the time, and that the first tranche of about $10 billion comes from unnamed sources.

Stargate: the buildings

Stargate is the programme to build data centres with Oracle and SoftBank. A tracker of OpenAI's commitments lists two different headline versions. The original one, announced in January 2025, is $500 billion and 10 gigawatts over four years. The active programme as of May 2026 is described as more than $400 billion and about 7 gigawatts over three years, with confirmed sites including Abilene in Texas, five newer sites and an Oracle expansion. The tracker says the headline commitment is largely intact but shifted in timeline. OpenAI's own Stargate announcements are the place to check site level detail.

The tracker is also internally inconsistent: its introduction cites more than 10 gigawatts, its table about 7 and its FAQ about 10 over four years. It does not say whether the more than $400 billion includes the $300 billion Oracle contract, which runs five years.

A separate Oracle contract is reported at $300 billion over five years, up to 4.5 gigawatts. An independent analyst, adding up the public announcements, counts about $1.15 trillion of commitments across seven partners for 2025 to 2035: Broadcom (estimated $350 billion), Oracle ($300 billion), Microsoft ($250 billion, incremental), Nvidia ($100 billion), AMD ($90 billion), Amazon Web Services ($38 billion) and CoreWeave ($22.4 billion). That analyst's year by year spend profile, which rises from about $6 billion in 2025 to about $295 billion in 2030, is labelled as a guess with a 30 to 50 percent error margin. It is not a disclosure by OpenAI.

What is safe to say is that Stargate is a build programme for buildings and power, while the chip deals are about what goes inside them.

A timeline of Stargate and the compute deals

Wikipedia's compilation of Stargate, which cites press reports, and company pages give this sequence. It is a secondary source, so check dates against primary pages before quoting.

• 21 January 2025. Stargate is announced at the White House by President Trump with Sam Altman, Larry Ellison of Oracle and Masayoshi Son of SoftBank, with a plan of up to $500 billion by 2029. Ellison said ten data centres were being built in Abilene, Texas.
• 23 January 2025. The Information reports SoftBank and OpenAI each committed $19 billion, with Oracle and MGX at $7 billion each. This is a report, not an announcement.
• 22 May 2025. JPMorgan Chase agrees to lend $2.3 billion for the Abilene project, developed by a joint venture of Blue Owl Capital, Crusoe and Primary Digital Infrastructure and leased to Oracle. The post on how AI data centers are financed explains that structure. In May 2025 partners also announced a Stargate UAE project expected to open in 2026.
• 7 August 2025. Bloomberg reports the project had not yet started and that no funds had been raised toward the initial $500 billion. Wikipedia lists this as the last funding status it cites.
• 22 and 23 September 2025. OpenAI announces five new sites, in Shackelford County and Milam County in Texas, Doña Ana County in New Mexico, Lordstown in Ohio and an undisclosed midwestern location. Planned capacity reaches nearly 7 gigawatts and more than $400 billion over three years. Nvidia's letter of intent is dated 22 September.
• 6 October 2025. AMD deal.
• 10 October 2025. A letter of intent for Stargate Argentina, up to $25 billion and up to 500 megawatts. In May 2026 Argentina's Economy Ministry said it had received no formal application and that nothing had progressed beyond the letter.
• 13 October 2025. Broadcom deal.
• 28 October 2025. The Microsoft and OpenAI restructuring: an incremental $250 billion of Azure services, and Microsoft no longer holds a right of first refusal to be OpenAI's compute provider.
• 3 November 2025. The $38 billion Amazon Web Services agreement, per Tunguz. Amazon says OpenAI began using the capacity immediately and that all targeted capacity is due by the end of 2026, with room to expand in 2027 and beyond, on hundreds of thousands of Nvidia GB200 and GB300 GPUs.
• March 2026. A funding round reported by the tracker at $122 billion and $852 billion valuation.
• May 2026. The tracker's most recent restatement of Stargate at about 7 gigawatts over three years.

What the Microsoft and Amazon deals add

Chips and buildings are not the whole picture. Microsoft's statement of 28 October 2025 says OpenAI contracted to buy an incremental $250 billion of Azure services, that Microsoft's stake is about 27 percent on an as-converted diluted basis, valued at roughly $135 billion, and that its rights to OpenAI models and products run through 2032. It also says Microsoft no longer has a right of first refusal as OpenAI's compute provider. That change is what freed OpenAI to sign the other deals in this post.

The Amazon deal is smaller at $38 billion but notable because it is Nvidia hardware in a cloud that sells a rival chip family. It means one lab can buy the same chips through several routes.

What has changed since the announcements

Three cautions apply.

First, the Nvidia terms. The commitment tracker lists a $30 billion Nvidia contribution in a March 2026 funding round, described as securing GPU supply. That does not match the original plan of up to $100 billion released in tranches as each gigawatt is deployed. Treat the Nvidia financial terms as unsettled, and look for the latest primary statement before quoting them. The tracker lists the full March round as Amazon $50 billion, Nvidia $30 billion, SoftBank $30 billion, Microsoft undisclosed and about $12 billion from others. The tracker is run by a company that sells AI visibility tracking, and it does not tie its figures to specific sources, so the round's composition should be treated as unconfirmed.

Second, the difference between contracted and built. No public source states how much of the 26 gigawatts is operating today. Assume the answer is very little until a company says otherwise.

Third, the sources disagree with each other. The tracker's two Stargate figures conflict, and the analyst's dollar values for Broadcom and AMD are estimates, not terms either company published. The analyst's own year by year table sums to $650 billion, not $1.15 trillion, because the rest falls after 2030, and some of its column totals do not match its stated contract totals.

Why three different chip suppliers

The mix tells you the strategy.

• Nvidia is the incumbent merchant supplier. Its September 2025 release names it OpenAI's preferred strategic compute and networking partner, and the separate Amazon deal shows its chips can also be bought through a cloud.
• AMD provides a second merchant source and negotiating leverage, plus an equity link through the warrants.
• Broadcom is the custom silicon route: chips designed around OpenAI's own models and serving patterns, which can lower the cost of each unit of work once volumes are high enough.

This mirrors what other labs do. Anthropic runs on Google TPUs, Amazon Trainium and Nvidia GPUs, and Google splits its own newest chips between training and inference. Spreading across suppliers reduces the risk of a shortage or a price rise at any one of them. The chip side is explained in the guide to AI chip market share and the training and inference split, and compare the custom designs in Trainium, Maia and Jalapeño. For the other side of the same story, see Anthropic's compute commitments.

The money side: who pays and how

A plan of this size needs financing as well as power. Equity from suppliers (Nvidia, AMD warrants), partner capital (SoftBank), project loans for individual buildings, and long leases all play a part. The analyst's rough guide of about $35 billion per gigawatt for Broadcom hardware is an assumption in his model, not a published price. The cost of the buildings themselves is separate, and the guide to what an AI data center costs per gigawatt lists the published estimates. Smaller specialist clouds also carry part of the load, as covered in GPU as a service economics. For what OpenAI earns and burns, see OpenAI's reported revenue, losses and cash burn and its reported gross margin.

What training compute versus inference compute looks like at OpenAI

An independent research group estimated OpenAI's 2024 cloud compute spend at about $7 billion. Roughly $5 billion went to research and development compute, and about $1.8 billion to inference. Within the research figure, the final training run for one large model was estimated at a few hundred million dollars, while most research compute went to experiments and models that were never released. The authors note the inputs came from September 2024 investor documents and should not be treated as fully reliable.

More detail from that analysis, with its caveats. The group estimated the final training run for GPT-4.5 at about $400 million, with a 90 percent confidence range of $170 million to $890 million, based on an assumed 40,000 to 100,000 H100 GPUs for 90 to 165 days at $1.50 to $3 per GPU hour. Other released models were put at about $80 million combined, and about $4.5 billion of research compute at experiments and unreleased models. It notes the figures are cloud rental costs, not the capital cost of data centres, that OpenAI rented most of its compute from Microsoft, and that the two year amortisation of research compute is the group's own assumption. The group also cites OpenAI's reported 2024 revenue of about $3.7 billion, which it says fits compute costs near $6 billion. It estimates the inference figure implies roughly a 50 percent margin on inference.

The takeaway is that frontier labs spend heavily on trying things that do not ship. Compute plans therefore have to cover experimentation as well as the models customers see. Those 2024 numbers are also small beside the 2026 plans. The tracker says Greg Brockman testified under oath that 2026 compute spend would be about $50 billion, but it gives no date or document for that, so treat it as unverified.

What is firm and what is uncertain

• Firm, as announcements: the three chip deals, with their gigawatt figures and timing, are company announcements. The Microsoft restructuring and its $250 billion Azure commitment are in Microsoft's own post. AMD's warrant structure is in AMD's release.
• Firm only as intent: the Nvidia up to $100 billion is a letter of intent. Broadcom's deal states no financial terms.
• Changed: Stargate has been described as $500 billion and 10 gigawatts, and as more than $400 billion and about 7 gigawatts. The Nvidia financial terms may have changed.
• Estimated: every dollar value for Broadcom and AMD, and the year by year spend, are analyst estimates.
• Unknown: how much capacity is operating, what share is leased, and the actual price terms.
• Unverified: the March 2026 round composition and the Brockman testimony figure.

What to watch next

• The first gigawatt milestones. Nvidia's first gigawatt and AMD's first MI450 phase were both slated for the second half of 2026. Confirmation they are live would also trigger the first investment tranche and warrant vesting.
• A final Nvidia agreement, or a statement that terms changed.
• Broadcom's first custom racks, due to begin in late 2026.
• Stargate site progress, including Abilene and the five 2025 sites, and any change in the three year timeline.
• Financing news. Debt raised against leased sites and fresh equity rounds show how the plan is paid for.
• Power. Gigawatt campuses need grid connections, which is where schedules often slip. See how long an AI data center takes to build.

How to read the next announcement

• Count gigawatts and dollars separately. A gigawatt is capacity. A dollar figure may be a purchase, a lease or an investment.
• Ask when it starts. Deployments beginning in late 2026 and finishing in 2029 are a multi-year ramp.
• Check for equity. Warrants and stakes tie suppliers' fortunes to the customer's, which can reduce or hide the true cost of capacity.
• Look for the latest version. Terms in this field have moved within months.
• Check the document type. A letter of intent, a signed contract and a press release headline carry different weight.

What it means for teams that depend on OpenAI models

A vendor with several chip sources and a large build programme is less exposed to a single shortage, but it also carries large fixed obligations. For security and procurement teams, the practical step is to know which AI services your software calls. Safeguard's free scan at https://gold.safeguard.sh/scan can help you see what your code depends on.

Quick check

Sources

• OpenAI ramps up global compute power with Nvidia, AMD and Broadcom, securing 26 GW, TrendForce, 15 October 2025.
• OpenAI and NVIDIA announce strategic partnership to deploy 10 gigawatts of NVIDIA systems, Nvidia, 22 September 2025.
• AMD and OpenAI announce strategic partnership to deploy 6 gigawatts of AMD GPUs, AMD, 6 October 2025.
• The next chapter of the Microsoft and OpenAI partnership, Microsoft, 28 October 2025.
• AWS and OpenAI workloads announcement, Amazon, undated on the page.
• Stargate LLC, Wikipedia, for the timeline, a secondary compilation of press reports.
• OpenAI compute commitments tracker, Presenc, May 2026.
• OpenAI hardware spending 2025 to 2035, Tomasz Tunguz, 3 November 2025.
• Most of OpenAI's 2024 compute went to experiments, Epoch AI, 10 October 2025.
• Anthropic compute plans, 1 October 2026, for comparison.

Last reviewed: 19 October 2026.