← Back to research
Briefing · Policy & infrastructure· Updated 14 September 2026· 6 min read

The build-out: 33 gigawatts, a $5 trillion chipmaker, and the bubble argument

In three weeks of October 2025 OpenAI signed for several times the power of Ireland's whole grid, Nvidia passed $5 trillion, and the people running the boom said a bubble was likely. What was committed, who is paying, and what it looks like from a country where data centres already take a fifth of the electricity.

Coiled grey cables and a white power plug on a pale desk

The models described elsewhere in these briefings are cheap to use because somebody is spending a fortune to build them, and the size of that fortune became the largest single question in the world economy in the autumn of 2025. This briefing sets out what was committed in the six weeks to the start of November, by whom and with whose money; what the sceptics and the builders actually said; and what it looks like from Ireland, where the constraint is the grid rather than the share price. It is written on 3 November with the sources available then.

Stargate, and then the October deals

Stargate was announced at the White House on 21 January 2025 as $100bn immediately and up to $500bn over four years, backed by OpenAI, SoftBank, Oracle and the Abu Dhabi fund MGX, with a first campus in Abilene, Texas.[1] On 23 September OpenAI added five more American sites, in Texas, New Mexico, Ohio and the Midwest, bringing planned capacity to nearly seven gigawatts and committed investment over three years above $400bn.[2] The chips to fill it came in a rush. On 22 September Nvidia signed a letter of intent to invest up to $100bn in OpenAI as ten gigawatts of its systems were deployed.[3] On 6 October AMD agreed to supply six gigawatts, with OpenAI receiving warrants for up to a tenth of the company as milestones were met.[4] On 13 October Broadcom agreed to build ten gigawatts of custom accelerators to OpenAI's design.[5] Behind all of it sat Oracle's contract, reported at about $300bn over five years, and CNBC's tally of the fortnight put OpenAI's commitments at roughly 33 gigawatts.[6]

For scale, Ireland's entire electricity system peaks at a little over six gigawatts. Almost none of the 33 exists yet. They are commitments to build, financed by the chipmakers' own investments in their customer, by debt and by revenue that has not been earned.

The seller of shovels

$5tn
Nvidia's market value on 29 October 2025, the first company to reach it [7]
$500bn+
Orders for its chips through the end of 2026, per its chief executive the same week [7]
≈33 GW
Compute OpenAI committed to in three weeks, across Nvidia, Oracle, AMD and Broadcom [6]

Nvidia passed $4 trillion in July and $5 trillion on 29 October, with Jensen Huang telling a Washington audience that the company had visibility of more than $500bn in orders for its current and next generation of chips through 2026.[7] It is also now the industry's investor, with the OpenAI letter of intent the largest of a long list, which is where the argument about circular financing begins: money that leaves Nvidia's balance sheet as an investment returns to its income statement as chip revenue, through the customer it invested in. Defenders reply that vendor financing is old and ordinary. What is new is the scale.

Can it pay for itself?

Bain & Company's annual technology report, published on 23 September, did the arithmetic that everyone else has since quoted. Meeting projected demand means about 200 gigawatts of new global compute by 2030; funding it needs about $2 trillion a year in AI revenue; and on current trends the industry will fall roughly $800bn short.[8] The people best placed to know have been candid. Sam Altman told reporters in August that investors as a whole were overexcited about a kernel of truth, and compared the moment to the dot-com bubble, in the same month his company was raising at $500bn.[9] Jeff Bezos, at Italian Tech Week on 3 October, called it an industrial bubble rather than a financial one, meaning the capacity will be used even if the investors are not repaid, which is roughly what happened to the fibre-optic cable laid in 1999.[10] Jamie Dimon told the BBC on 9 October that a lot of the money going into AI would probably be lost, and that some of it would produce great returns.[11] Both halves of each statement are worth keeping.

Ireland: the grid

In Ireland the build-out is measured in megawatts. The Central Statistics Office reported in June that data centres consumed 22% of the country's metered electricity in 2024, up from 5% in 2015, a share unlike any other country's.[12] The International Energy Agency's report on energy and AI, published in April, put global data-centre demand at around 415 terawatt-hours in 2024 and projected it to more than double by 2030; Ireland is the extreme case of a worldwide trend.[13] Since 2021 the system operator's position has amounted to a moratorium on new connections around Dublin, and the Commission for Regulation of Utilities spent 2025 consulting on rules under which new data centres would have to bring their own generation or storage and source most of their power from additional renewables. As this briefing is written the decision is expected before the end of the year.

The capacity that does get built here will be built by the same companies filling the American sites, and priced in the same market. For an Irish business the practical consequences are three. Compute for inference will keep getting cheaper, whether or not the investors are repaid. Capacity in Ireland is constrained by the grid, not by demand, so an EU-hosted workload increasingly means Paris, Frankfurt or the Nordics as well as Dublin. And the whole stack, from chips to clouds to labs, is a handful of mutually invested companies, which is a counterparty risk that belongs in a procurement note.

Sources

  1. [1]Stargate LLC · Wikipedia, as compiled · 1 Nov 2025
  2. [2]OpenAI to build five more Stargate data centres in the US with Oracle and SoftBank · SiliconANGLE · 23 Sept 2025
  3. [3]OpenAI and Nvidia announce strategic partnership to deploy 10 gigawatts of Nvidia systems · OpenAI · 22 Sept 2025
  4. [4]AMD signs AI chip-supply deal with OpenAI, gives it option to take a 10% stake · RTÉ News · 6 Oct 2025
  5. [5]OpenAI taps Broadcom to build its first AI processor in latest chip deal · RTÉ News · 13 Oct 2025
  6. [6]OpenAI's compute deals now total about 33 gigawatts · CNBC · 14 Oct 2025
  7. [7]Nvidia becomes the first $5 trillion company · Fortune · 29 Oct 2025
  8. [8]Technology Report 2025: AI's $800 billion revenue gap · Bain & Company · 23 Sept 2025
  9. [9]Sam Altman's paradox: warning of an AI bubble while raising trillions · Fortune · 19 Aug 2025
  10. [10]Jeff Bezos calls AI an industrial bubble · Fortune · 4 Oct 2025
  11. [11]Jamie Dimon says a lot of AI investment will probably be lost · CNN Business · 9 Oct 2025
  12. [12]Data centres metered electricity consumption 2024 · Central Statistics Office · 10 Jun 2025
  13. [13]Energy and AI · International Energy Agency · 10 Apr 2025
  14. [14]The CRU publishes its decision on new electricity connection policy for data centres · Commission for Regulation of Utilities · 12 Dec 2025
  15. [15]Data centres consumed 23% of Ireland's electricity in 2025 · Data Centre Review, citing CSO · 16 Jul 2026
  16. [16]Data centres in Ireland: demand and forecasts · Enable Research, citing EirGrid · 2026
  17. [17]AI giants Anthropic and OpenAI seek more office space in Dublin to expand European operations · The Irish Times · 19 Dec 2025
  18. [18]Big Tech AI capex in 2025: Microsoft, Google, Meta, Amazon and the spending race · Value Add VC · Sept 2026
  19. [19]Nvidia · Wikipedia, financial results as compiled · 14 Sept 2026
  20. [20]Ilya Sutskever's Safe Superintelligence partners with Nvidia to scale its AI research · TechCrunch · 27 Jul 2026
  21. [21]Anthropic · Wikipedia, compute agreements as compiled · 14 Sept 2026
  22. [22]The circular deals powering the AI boom · Bloomberg · 2026

Begin your AI transformation.

Book a call with the founders. Thirty minutes to understand your business, your team, and where AI could actually help. No deck, no pitch.

Talk to us →