Sovereign AI Debt: Announced, Funded, Spent

The largest AI compute projects of the past two years are financed with sovereign money, convertible instruments, and bank debt. The public discussion treats the announcements as investment. Three separate numbers tell different stories. Announced. Funded. Spent. They rarely match. The instrument in each case shows who holds the risk when the gap stays open.

This piece follows five actors that sit at the centre of the sovereign AI financing story. Stargate, the US joint venture with a $500 billion ceiling. MGX, Abu Dhabi's AI investment vehicle. G42, the Emirati AI group that trades capital and compliance for chip licences. Saudi Arabia's Public Investment Fund, still shaping a dedicated AI envelope. SoftBank, the debt engine behind much of Stargate's reported structure. A smaller Indian case, the Sarvam compute subsidy under IndiaAI, appears at the end as a scaled-down version of the same accountability problem.

Where only press reporting exists, the text says so. Where a figure is a ceiling, a commitment, a raise, or a completed purchase, the text keeps those words distinct.

Three 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 an asset or paid a contractor. An acquisition closed. A data centre paid for. Chips delivered and invoiced. Spent is the only number that shrinks the physical gap between speech and silicon.

Most coverage collapses the three into one word, investment. That word flatters every party. Sponsors look decisive. Governments look industrial. Vendors look busy. Readers lose the ability to see who actually took risk.

Stargate. $500 billion announced, $38 billion pledged by SoftBank and OpenAI, slow to start

Stargate was announced at the White House on 21 January 2025 as a joint venture of OpenAI, SoftBank, Oracle, and MGX, with plans to spend up to $500 billion on US AI infrastructure by 2029. The funding structure was reported by The Information on 23 January 2025.

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 matters more 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.

On 7 August 2025, Bloomberg reported that the project had not started and no funds had been raised to meet the $500 billion budget. SoftBank conceded it needed more time.

Then the announcements resumed. OpenAI said on 23 September 2025 that Stargate would add five new data centre sites, taking planned capacity to nearly 7 gigawatts and claiming over $400 billion in investment over three years. In October 2025, Reuters reported a $25 billion, 500-megawatt project for Patagonia, Argentina, with Sur Energy.

The September announcement was about an infrastructure plan. The August Bloomberg report was about money raised. Plans can grow while funding stalls. The instruments between them, a thin equity base with thick bank debt, mean the lenders hold much of the exposure when construction has not started.

Large infrastructure programmes often look quiet between announcement and ground breaking. The August 2025 Bloomberg line is a timestamp, not a final verdict. Later claims still need funded capital and spent capital, not only site maps.

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. That dual role is not a footnote. Capital allocation and national strategy sit in the same chair.

The vehicle's transactions are a ledger. 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 example in this set 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 different from a bank loan to someone else's venture. It is also different from a press release about a $100 billion target. Targets are not assets.

G42. Licences for compute, with debt-like strings

G42 is the Emirati AI group chaired by Sheikh Tahnoon bin Zayed (Wikipedia). 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 nine 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.

The record is a bargain. Access to advanced chips is a controlled good. Divestment, board seats, and authorised-receiver status are the price of that access. Agree with the bargain or not. See it plainly.

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. Sequence matters for risk. A pledge before a licence is a hope. A licence before a build is a permit. A build before a lease is a bet on demand. Public materials rarely order those events cleanly.

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.

Saudi Arabia belongs in the set because the story is unfinished. Omitting it would make the Gulf picture look complete. A large envelope without closed deals is a different risk object from MGX's completed acquisition. Treating both as "Gulf AI money" erases that difference.

SoftBank. The debt engine

SoftBank is the financial backer of Stargate and, through its ownership of ARM, a company deeply exposed to the AI infrastructure build-out. The Stargate financing, reported by the Wall Street Journal, said SoftBank would supply about 10% equity, with the rest from bank borrowing. The first $10 billion was expected from Mizuho and other Japanese lenders.

The MGX model and the SoftBank model point in opposite directions. MGX raises sovereign capital and deploys it directly, acquiring assets outright. SoftBank borrows and puts the borrowed money into ventures it chairs. The risk in the MGX model sits with the sovereign fund. The risk in the SoftBank model sits with the banks that lent to SoftBank. Both are governments and government-backed firms financing AI compute with instruments that sit closer to credit markets than to simple cash equity.

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. It is not the same as funded Stargate construction. Mixing ARM's strategic position with Stargate's project finance creates a halo that project lenders do not share.

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 three.

Power, land, and time

Financing is only one 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 geographic scope expands in press time. Physical interconnection expands slowly. SoftBank's August 2025 concession that more time was needed tracks the physics of build-out more closely than the September site list. Report both without picking a side.

A usable ledger would publish, for each major project, three dates next to three money fields. Announcement date. Funding close date. Power-connection or first-rack date. Announced capital. Funded capital. Spent capital. Few projects currently offer that six-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 all travel better than bond prospectuses. The financing stack behind the speech can still be ordinary project finance. Separating the speech from the stack is the whole point of this piece.

How announcement cycles work

Watch the calendar pattern around Stargate.

January 2025. White House announcement, $500 billion ceiling.
January 2025. Press details of a smaller initial equity frame.
April and May 2025. Debt structure reporting and a Texas loan.
August 2025. SoftBank says more time is needed, funding short of the ceiling.
September 2025. New sites and a fresh investment claim over three years.
October 2025. An Argentina project enters the story.

The pattern is not unique to Stargate. Announcement, structure leak, credit event, delay admission, new geographic chapter. Each step can be factually reported. The error is treating every step as equal evidence of progress. A site list after a funding delay is not a rebuttal of the delay. It is a different object.

A simple filter helps. 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 unit of account changes. The question does not. Who owns the residual when utilisation disappoints.

What a cooling market would expose

Assume, for argument, that model-training demand slows or that inference shifts toward cheaper silicon. Lease rates soften. Projects still need to service debt. Equity cushions absorb first losses. Thin equity means thin cushions.

In that world, SoftBank-style structures surface stress in bank books and renegotiations. MGX-style structures surface stress in sovereign fund marks and delayed deployments. G42-style licence bargains surface stress in politics if export rules tighten during a glut. Saudi planning envelopes simply stop converting into deals.

None of this requires predicting a crash. It only requires admitting that AI infrastructure is cyclical in hardware even when software demand looks smooth. Credit markets already know this. Press releases often do not.

Replies that improve the reading

Defenders of large announced ceilings make fair points.

Long-lead infrastructure needs a headline number to align vendors, utilities, and ministries. A $500 billion frame can organise behaviour even if year-one cash is lower. Treating every ceiling as a falsehood misses how industrial policy communicates.

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 the same as unsecured hope. Some of the debt in this story is ordinary secured project finance.

Sovereign funds are allowed to move slowly. A $49 billion raise that deploys over years can still be real. Demanding full deployment the quarter after a raise confuses fund-raising with capital expenditure.

These points improve the reading. They do not erase the need to label announced, funded, and spent as different cells.

What the public record still lacks

Deployment schedules with ground-breaking and power-connection dates next to site announcements. Equity-to-debt ratios at close. The WSJ report on SoftBank's 10% equity stake is still the most important capital-structure number in this piece, and it remains press reporting. A published list of MGX Fund I deployments at cost would separate raised from spent.

Press releases are cheap. Concrete and transformers are not. Until the funds publish what they spent rather than what they pledged, the AI build-out reads like a credit market with marketing attached, and the risk sits with whoever financed the last tranche.

Sources and limits

Figures in this piece were checked against the linked Reuters, Bloomberg, WSJ, FT, OpenAI, and MGX statements in August 2026. SoftBank's reported 10% equity structure and The Information's $19 billion SoftBank/OpenAI pledges remain press-reported capital plans, not filed closes. The Abilene $2.3 billion JPMorgan loan is a verified construction financing for an Oracle-leased campus tied in later press to the Stargate story. MGX Fund I's $49 billion final close is from MGX's own 1 July 2026 statement. Undeployed shares of that fund remain unpublished. India figures for Sarvam come from the IndiaAI essay, which flags the CCD terms as press-reported via Economic Times and MediaNama.