Stargate was announced at the White House on 21 January 2025 as a plan to spend up to $500 billion on US AI infrastructure by 2029. On 7 August 2025 Bloomberg reported that the project had not started and that no funds had been raised to meet that budget. The stories still led with $500 billion.
3 numbers that get mixed
Announced is the headline. It is often a multi-year ceiling or a memorandum. It can include debt that has not been raised and sites that have not broken ground.
Funded is money committed or raised into a vehicle. Equity closed. A fund raise completed. A loan facility signed. Funded is stronger than announced. It is still not spent.
Spent is cash or equity that bought something. An acquisition closed. A hall paid for. Chips invoiced. Until that happens, you still have a press release.
All three get printed as investment. You cannot see who actually took the risk.
Stargate. $500 billion announced, $38 billion pledged by SoftBank and OpenAI, slow to start
The joint venture names OpenAI, SoftBank, Oracle, and MGX. The funding structure was reported by The Information on 23 January 2025.
The announced ceiling was $500 billion by 2029. The committed initial frame, in press, was $100 billion. Named pledges around that frame, in press, were about $38 billion from SoftBank and OpenAI by August 2025. Spent toward the $500 billion JV ceiling is a separate cell. SoftBank said it needed more time. That August Bloomberg report is a timestamp. Related campuses could still have been under construction through August 2026.
Add SoftBank $19B, OpenAI $19B, Oracle $7B, and MGX $7B and you get $52 billion of named equity-like pledges around the initial $100 billion frame, with the rest described as limited partners and debt. SoftBank plus OpenAI alone are $38 billion of that named set. The public conversation still leads with $500 billion. That is the ceiling to 2029, not cash in hand in January 2025, and not the same as funded capital.
The debt structure is more useful than the headline. The Wall Street Journal reported in April 2025 that SoftBank, the main financial backer, was expected to supply only about 10% equity. The first $10 billion of its commitment was expected to come from Mizuho and other Japanese lenders. Those figures are press-reported structure talks, not a filed capitalisation table. In January 2025, JPMorgan provided $2.3 billion of construction financing for an Abilene, Texas data centre developed by a Blue Owl, Crusoe, and Primary Digital joint venture and leased to Oracle. Press coverage later tied that campus to the wider Stargate story. A loan to an Oracle-leased Texas project is a concrete credit event. It is not, by itself, proof that SoftBank's $19 billion Stargate pledge had closed as equity.
Then the announcements resumed. OpenAI said on 23 September 2025 that Stargate would add 5 new data centre sites, taking planned capacity to nearly 7 gigawatts and claiming over $400 billion in investment over 3 years. In October 2025, Reuters reported a $25 billion, 500-megawatt project for Patagonia, Argentina, with Sur Energy.
The September announcement was an infrastructure plan. The August 2025 Bloomberg report was about money raised to the ceiling. Plans can grow while a JV close stalls. Some related construction can be real, through project finance and tenant leases, while the White House number is still a brochure. The error is folding every OpenAI-adjacent dollar into the $500 billion cell.
MGX. $100 billion target, $40 billion acquisition closed, $49 billion fund raised
MGX is the Abu Dhabi state investment vehicle for AI, launched in March 2024 with Mubadala and G42 and a target of $100 billion in assets under management (Bloomberg, 11 Mar 2024). Its chair is Sheikh Tahnoon bin Zayed, the UAE's national security adviser. The same person allocates capital and sets national strategy.
In September 2024, MGX joined BlackRock and Microsoft in the Global AI Infrastructure Investment Partnership for data centres and power infrastructure (WSJ, 17 Sep 2024). In January 2025 it took a Stargate stake. In October 2025, a consortium of MGX, AIP, and BlackRock's Global Infrastructure Partners agreed to buy Aligned Data Centers in a deal valued around $40 billion. That acquisition closed on 21 July 2026 at an enterprise value of about $40 billion, with an additional $5 billion of growth capital committed. In June 2026, Bloomberg reported that MGX had raised close to $50 billion for an AI infrastructure fund. On 1 July 2026, MGX confirmed the final close of Fund I at $49 billion in commitments.
A $100 billion AUM target in 2024. A $40 billion acquisition agreed in 2025 and closed in 2026. A $49 billion fund raise confirmed in July 2026. The closed acquisition is a spent item at enterprise value. The fund raise is a funded item. What is not public is the portion of the fund already deployed versus reserved, because MGX does not publish a portfolio ledger with cost basis.
MGX is the clearest case here of sovereign-backed capital buying assets outright. Ownership concentrates upside and downside in the fund. If leases hold and power is available, the fund owns cash-flowing infrastructure. If utilisation disappoints, the fund owns empty halls. That is not a bank loan to someone else's venture, and it is not a $100 billion target on a slide.
G42. Licences for compute, with debt-like strings
G42 is the Emirati AI group chaired by Sheikh Tahnoon bin Zayed, as the company said when Microsoft invested. Its history shows the exchange at the centre of sovereign AI. Capital and compute access against regulatory clearance.
In July 2023, G42 agreed to pay about $100 million for the first of up to 9 Cerebras supercomputers (Reuters). In April 2024, Microsoft announced a $1.5 billion investment in G42, with Microsoft's president joining the board. US lawmakers had asked for an intelligence assessment of G42's China ties before the deal could advance, and the company said it had divested its China holdings.
The licences followed. In November 2025, the US Department of Commerce authorised exports of advanced semiconductors equivalent to about 35,000 NVIDIA GB300-class accelerators to G42 (a matched authorisation also covered Saudi Arabia's HUMAIN). In July 2026, the US reclassified the UAE to Country Group A:5 and listed G42 and Core42 as approved recipients for certain advanced computing items without individual licences, subject to conditions and an expiry unless they become US-headquartered. A company under scrutiny for China ties in 2023 became a designated US-approved recipient by mid-2026.
Chips are a controlled good. Divestment, board seats, and authorised-receiver status were the price.
European pledges scaled in the same period. In June 2025, G42 established G42 Europe & UK. That launch followed UAE state pledges of up to about $52 billion toward French AI data-centre capacity and a $40 billion investment frame for Italy, under which G42 partnered on Italian AI infrastructure. In May 2025, it joined the Stargate UAE project with OpenAI, Oracle, NVIDIA, SoftBank, and Cisco. In November 2025, Microsoft and G42 announced a 200-megawatt data centre expansion in the UAE. In February 2026, G42 signed a Vietnamese consortium framework with consumption commitments of up to $1 billion, and separately announced a national-scale AI supercomputer partnership for India.
The French and Italian frames are state pledges and letters of intent attached to a period of licence progress. G42 does not publish whether the licence, the pledge, or the ground breaking came first. Their public materials almost never give that order.
Saudi Arabia. The fund still being shaped
Saudi Arabia sits behind the other vehicles through its Public Investment Fund. In February 2024, the PIF launched Alat, with a reported $100 billion investment mandate by 2030 for advanced manufacturing and technology. Throughout 2025, outlets reported that the PIF was planning a dedicated AI fund in the range of $40 billion to $100 billion.
The Saudi numbers are still at the planning stage. MGX and G42 have identifiable transactions. Aligned Data Centers. The Cerebras deal. The Microsoft investment. The Stargate equity. The Saudi numbers do not yet have a transaction ledger of the same density. The instrument, a sovereign fund allocating state oil revenue to AI compute, is among the largest in the set, and the gap between announced and funded is the widest.
Leave Saudi out and the Gulf looks finished. A $40-100 billion envelope with no closed ledger is still a plan. MGX buying Aligned is a closed deal. "Gulf AI money" hides the difference.
SoftBank. OpenAI equity is not Stargate project finance
SoftBank sits in this story twice. The first role is the reported financial backer of the Stargate JV. The second is a disclosed shareholder in OpenAI Group PBC. Mixing them inflates the JV.
On 27 February 2026, SoftBank said it had agreed to follow-on investments of $30 billion in OpenAI Group PBC via SoftBank Vision Fund 2, in 3 $10 billion tranches. Upon completion, it said, cumulative investment in OpenAI was expected to total $64.6 billion, about 13% ownership. Those are SoftBank's words. The 13% figure is expected on completion. It was not a holding on 27 February.
The first $10 billion tranche closed on 1 April 2026. SoftBank said it borrowed $10 billion that day under the same bridge facility. The second $10 billion tranche closed on 1 July 2026, again with a $10 billion draw under that facility. The facility itself is $40 billion, unsecured, maturing 25 March 2027, with JPMorgan, Goldman Sachs, Mizuho, SMBC, and MUFG as named lenders, and is also available for general corporate purposes. A third $10 billion tranche is planned for 1 October 2026, subject to acceleration if OpenAI lists. As of the 1 July notice, $20 billion of the $30 billion follow-on had been executed. The remaining $10 billion had not. 2 $10 billion draws do not prove that $20 billion of the $40 billion facility was used only for OpenAI. SoftBank's notices attach each draw to that day's tranche. They do not publish a running facility utilisation table.
This is equity in a company, funded in part with bank debt. It is a funded item on the OpenAI cap table. It is not a disclosed close of the Stargate joint venture's $500 billion ceiling. SoftBank's notices do not say that this cash bought US data-centre halls.
The earlier Stargate financing story, reported by the Wall Street Journal, said SoftBank would supply about 10% equity to the JV, with the rest from bank borrowing. That remains press-reported structure for the project, not the 2026 OpenAI share purchase.
The MGX model and the SoftBank model still point in different directions. MGX raises sovereign capital and, in the Aligned case, buys assets outright. SoftBank borrows and puts borrowed money into companies it backs. The risk in the MGX model sits with the sovereign fund. The risk in the SoftBank OpenAI purchases sits, first, with SoftBank and the banks on the bridge facility. Both are large cheques, but they fund different structures than Stargate.
ARM exposure adds a second layer. SoftBank benefits if AI infrastructure demand stays high, because ARM's designs sit in a wide share of the chip stack. That is equity upside from industry growth, separate from funded Stargate construction.
Who holds the risk
Stargate's reported structure puts the bulk of the money as debt against a small equity base, with SoftBank's reported 10% equity and Mizuho and JPMorgan providing loans. If the AI hardware market cools, or if power and permitting slip, the lenders hold much of the exposure. Japanese banks can end up financing US AI infrastructure. That is how cross-border project finance works when equity is thin.
The government of Abu Dhabi, through MGX, is acquiring data centres outright, which means it owns the assets if the leases hold. Ownership is clearer. Exit can still be hard. Data centres are not liquid stocks. A sovereign that overpays for capacity in a glut eats the loss slowly.
G42's licence path places a different risk on the table. Political and compliance risk. A designation can be granted and tightened. Capital pledged against a licence path can strand if rules change. Readers who only track gigawatts miss that channel.
The Sarvam pattern from the IndiaAI piece is the smaller version of the same problem. A government supplies subsidised compute and receives convertibles that may become 1-2% of the company. MGX and G42 are the national version, where the state supplies capital and licences and receives either equity or infrastructure ownership. At both scales, ask what happens to the public interest if the assets underperform.
There is no single correct instrument. Debt can be cheap and disciplined when covenants are real. Equity can align owners with long build times. Convertibles can protect a public subsidy. The failure mode is using announcement language for all 3.
Power, land, and time
Financing is only 1 constraint. Data centres need power connections, land, water or cooling, and permits. A funded project without a grid queue is not spent capacity. An announced gigawatt figure without interconnection dates is a brochure.
Stargate's site announcements and the Argentina report show how fast the map grows in press time. Physical interconnection grows slowly. SoftBank's August 2025 concession that more time was needed tracks the build-out better than the September site list. A new map is not a funding close.
A usable ledger would publish, for each major project, 3 dates next to 3 money fields. Announcement date. Funding close date. Power-connection or first-rack date. Announced capital. Funded capital. Spent capital. Few projects currently offer that 6-cell table.
Why sovereign money entered this market
Private venture capital is good at funding software margins. It is less comfortable underwriting multi-year power contracts, concrete, and chip supply that can reprice mid-build. Data centres look more like infrastructure funds and utility partners than like classic Series B deals. Sovereign wealth funds already buy ports, power, and logistics. AI compute is a new label on an old asset class, with a faster obsolescence clock.
Hardware ages faster than buildings. A container port lasts decades. A GPU generation can be competitively stale in a few years. Financing long-lived buildings that house short-lived accelerators creates a mismatch. Lenders want stable lease cash flows. Model trainers want the newest silicon. SoftBank-style debt against thin equity makes that mismatch someone else's problem until a downturn arrives. MGX-style asset ownership keeps the mismatch on the sovereign books.
States also enter for industrial policy. Chip licences, local jobs, and "sovereign AI" speeches travel better than bond prospectuses. Behind the speech you often still find ordinary project finance.
How announcement cycles work
$500 billion ceiling.
Press details of an initial frame below the ceiling.
Structure reporting, then a construction loan for an Oracle-leased campus.
Funding short of the ceiling.
A fresh investment claim over 3 years.
A new country enters the story.
$20 billion of OpenAI Group PBC equity, partly with bridge draws. That line belongs on the OpenAI cap table, not automatically in the Stargate JV cell.
Other projects run the same sequence.
- Announce a ceiling
- Leak the structure
- Land a loan
- Admit delay
- Add a country
Each item can be true. A new site list after a funding stall is not proof the stall ended. If the sentence has no close date and no cash movement, it is still in the announced column.
Convertibles and public interest
The Sarvam CCD under IndiaAI looks small next to Stargate. It is useful because the public-interest question is easier to see at small scale.
A government pays part of a compute bill. In return it may receive equity through conversion. If the company becomes valuable, the public shares upside. If the company stalls, the public may hold a weak claim on a weak asset, while the compute hours are already burned. Startups that refuse equity-for-compute are making a price decision. They may prefer a clearer grant or a clearer market purchase. Governments that insist on convertibles are making a stewardship decision. Both can be defended. Neither is free.
At Gulf scale, the same logic appears as fund ownership of data centres or board seats attached to chip licences. The cheque is larger. The question is the same: who owns the leftover when the halls sit empty.
What a cooling market would expose
If model-training demand slows, or inference moves to cheaper silicon, lease rates soften. Projects still have to service debt. Equity takes the first loss. Thin equity means a thin cushion.
SoftBank-style structures then show stress in bank books and renegotiations. MGX-style structures show it in sovereign fund marks and delayed deployments. G42-style licence bargains show it in politics if export rules tighten during a glut. Saudi planning envelopes stop turning into deals.
Hardware is cyclical even when software demand looks smooth. Credit desks already price that. Press releases often do not.
What the ceilings are for
Long-lead infrastructure needs a headline number to line up vendors, utilities, and ministries. A $500 billion frame can organise behaviour even if year-one cash is lower. Calling every ceiling a lie misses how industrial policy talks.
Private credit into AI data centres can be rational when leases are signed with strong tenants. A loan to a facility leased to Oracle is not unsecured hope. Some of this debt is ordinary secured project finance.
Sovereign funds can move slowly. A $49 billion raise that deploys over years can still be real. Demanding full deployment the quarter after a raise confuses a raise with capex.
None of that lets you merge announced, funded, and spent into 1 cell. SoftBank's 10% equity figure and The Information's $19 billion pledges are still press plans for the JV, not filed closes. The Abilene $2.3 billion JPMorgan loan is a verified construction financing for an Oracle-leased campus, later tied in press to Stargate. SoftBank's $20 billion of executed OpenAI equity in 2026 is a verified share purchase, funded in part with a bridge loan. MGX Fund I's $49 billion close is from MGX's 1 July 2026 statement. How much of that fund is still undeployed is unpublished. The risk sits with whoever financed the last tranche.