Sovereign AI trades as a single position. Underneath the label are five things a state can hold, at very different quality, and a question the prospectus does not ask about any of them: who can take it away. Andrea Dew developed the proprietary CAMEO framework to disaggregate strategic technology positions across Capability, Apparatus, Makers, Ecology, and Operations. Below is a July 2026 snapshot.

"Sovereign AI" has become a line item. A Gulf compute campus, a Paris lab, a London research bench, a Hangzhou model shop: one label, one thesis. The thesis is that states are buying their way into artificial intelligence, and that exposure to the buying is exposure to the outcome.

The label is doing more work than the thesis can support. Underneath it sit five different things a state can hold or fail to hold: what the models can do (Capability); the compute, power and physical plant they run on (Apparatus); the institutions that make them, with their governance and their entanglements (Makers); the research culture and talent that make the other elements productive (Ecology); and the deployment context in which any of it gets used (Operations). I use this decomposition, CAMEO, across strategic technologies. The sovereign AI trade is a good place to put it to work, because investors are pricing the label while the value sits in the parts.

Each element is held on some tenure. Owned means held outright. Borrowed means rented, hosted, or accruing to someone else’s balance sheet. Conditional means nominally held, with a foreign hand inside the terms. And each element is held at some quality: a world-class research bench and a serviceable one carry the same label in a prospectus and very different values in a portfolio. Tenure says who can take a position away. Quality says what it is worth in the meantime. Most of what follows is those two readings, taken five times, in four places.

The Emirates purchased the buyable element at the highest quality on offer. G42 sits at the center of the buildout: Microsoft has committed $15.2 billion to the UAE through 2029 and put $1.5 billion into G42 itself, Stargate UAE is a 1 gigawatt cluster being built for OpenAI inside a planned 5 gigawatt UAE–US campus, and a further 200 megawatts of Microsoft datacenter capacity is expected to start coming online before the end of the year. The hardware is top-shelf and the checkbook is sovereign. On paper this is Apparatus owned outright.

The tenure says otherwise. Chip access was approved in November 2025 under a framework called the Regulated Technology Environment, a compliance regime G42 designed and Washington approved, and the bilateral framework beneath it leaves chip access subject to continuing U.S. licensing oversight. The earlier price of entry was a divestiture of G42’s Chinese entanglements, though critics noted the divested holdings landed in a fund overseen by G42’s own parent; part of the ongoing price is Microsoft’s equity and a Microsoft seat on the board. The element the pitch decks lead with, sovereign compute, is precisely the element that has an American veto built into it. The advertised strength is also the point of foreign leverage.

The rest of the stack is thinner than the campus suggests. The frontier capability running on Emirati soil is mostly other people’s: OpenAI’s, Microsoft’s. The homegrown models are real and mid-table ones: Falcon from the state’s Technology Innovation Institute, Jais from G42’s stable. The research bench is largely imported on contract, and the 80,000 government workers now being trained to use agents rather than to build them. What Abu Dhabi genuinely owns is Operations: deployment across government at a pace few states attempt, with a target of AI support for half of federal operations within two years. Netted out, the position is conditional Apparatus of the highest quality, wrapped around borrowed Capability and borrowed Ecology, with owned Operations doing more work than the prospectus admits.

There is a scheduled moment when this gets priced. A G42 listing is being discussed for 2027 or 2028. Two of its subsidiaries already trade in Abu Dhabi; a group listing is the one that would force a public number onto conditional compute. Nothing listed today prices tenure cleanly, so the opening trade would be the first public quote on conditional compute.

Britain holds the highest-quality single element anywhere in this set. The research bench, DeepMind in London with the university benches at Cambridge, Oxford and Edinburgh behind it, is among the strongest outside the American frontier cluster. It also holds very little else. DeepMind’s Maker has been Alphabet since 2014, a £400 million sale that looked large at the time. The capability accrues to a California balance sheet. The compute is American hyperscale cloud, with a sovereign sliver, Isambard and the national research resource, that is real and rounding-error sized.

Tech Nation’s 2026 report put numbers on the netting: by its estimate, fifty-seven pence of every exit pound flows back to the United States, and about one venture dollar in two invested in British AI arrives from across the Atlantic. Where the rest of the pound lands is not published, and no public ledger of exit-value destinations yet exists; for a reader tracking ownership, the missing ledger is itself a finding. The lesson is blunt either way: the physical location of talent tells you almost nothing about where returns land. In London the stack is present, the work is British, and a disproportionate share of the gains is banked in America.

The state’s response is live and small. A £500 million Sovereign AI Unit opened in April, writing equity tickets up to £20 million with a million GPU hours attached; a £1.1 billion hardware plan followed in June, anchored by a £750 million machine in Edinburgh; a first sovereign frontier model, Lumen Sovereign, is slated for late this year. One American campus carries more debt than the whole fund; the state has bought a seed position in its own stack. The week the fund launched, OpenAI reportedly paused its Stargate UK project, which goes in the file as the year’s first data point on which way ownership is drifting. The bench keeps generating; the fund’s opening cohort is visibly staffed with DeepMind alumni. The open question is the one it has been since 2014: whether the next DeepMind is British-owned.

In the table, France is the row in motion. Mistral holds a genuine research culture at national scale, deep and narrow: roughly 860 people, founders out of DeepMind and Meta, the French mathematics pipeline behind them. The people are first-rate, the models are French, and the models generally trail the leading closed frontier systems.

The Apparatus story changed this spring, though less totally than the headlines suggest. Until recently Mistral was a tenant, running on Azure, Google Cloud and CoreWeave. This spring it raised $830 million in debt from a consortium of seven banks to finance 13,800 of Nvidia’s newest chips at a 44 megawatt facility south of Paris; training runs began there early in the year, with full operation expected by summer. The building belongs to a French operator, Eclairion, so the conversion is layered: Mistral owns the compute and leases the hall. It is in early talks to raise around €3 billion at a valuation near €20 billion, most of it earmarked for more of the same, with stated targets of 200 megawatts across Europe by the end of 2027 and a gigawatt by 2030. Reported revenue crossed $400 million annualized in February, from roughly $20 million a year earlier; reported is the operative word, since private frontier labs publish no audited numbers and the market trades on what leaks. Seven banks lending against that line is not an audit, but it is a market test, and for now the only one on offer.

Mistral is testing whether a first-rate lab can finance its own compute. Abu Dhabi is testing whether bought compute can grow a lab. The first answer arrives in quarters; the second waits for a cohort to mature.

Two hands rest inside the French position. The largest shareholder is ASML, which took eleven percent for €1.3 billion last September: Europe’s chokepoint company backing Europe’s lab, an ownership structure with no obvious American analogue, and a European hand, which is the point. The other is quieter. The largest planned campus, 1.4 gigawatts near Paris, is a joint venture with MGX, Abu Dhabi’s $100 billion AI fund, alongside Bpifrance and Nvidia. The tenure on French compute is improving. The capital underneath it is not entirely French, and the chips on top of it are allocated in Santa Clara.

China holds more of the stack than anywhere outside the United States. Investors obsess over the element Washington controls, the chips, and barely price the four China owns. DeepSeek’s V4, unveiled in April at 1.6 trillion parameters, sits near the frontier on several reported benchmarks while remaining behind the leading closed systems overall. The conditional element, Apparatus, is also the lowest-quality holding in this set: the Nvidia chips were bought before the ban and are aging on the racks, and Huawei’s Ascend line remains behind them, with projected shipments of the new 950 series this year, around 750,000 units, which one analyst’s arithmetic puts at roughly a week of quality-adjusted American chip production.

The rest of the ledger runs the other way. V4 was engineered top to bottom for hardware independence: its low-level code written in a portable language rather than Nvidia’s CUDA, its number formats chosen to run on Ascend, Cambricon and Biren. In June a Huawei-linked team reported completing full-parameter post-training of the largest V4 model on a cluster of about a thousand Ascend chips, a class of work domestic silicon had mostly been kept away from; the chips’ prior role, by the project’s own description, was inference. DeepSeek is reportedly designing an inference chip of its own. And the pricing remains the weapon: $3.48 per million output tokens against $30 and $25 at the leading American labs, with a stated intention to cut further as domestic supply scales.

Washington’s posture changed in January, and the change is stranger than a loosening. Exports of H200-class chips to China moved from a presumption of denial to case-by-case review, and the White House imposed a 25 percent tariff on covered chips routed through the United States before re-export. By May, roughly ten Chinese buyers, reportedly including Alibaba and Tencent, had been cleared on the American side. The wall has become a tollgate. The far side of the gate has its own keeper. Beijing has so far kept the approved chips from moving, steering demand toward Huawei instead, with exemptions under discussion for university research.

Chinese companies had reportedly ordered more than two million of the chips before the freeze, so the demand is real and the blockage is policy. One estimate put a year of permitted shipments at a lift of roughly 250 percent to China’s installed compute over a domestic-only path; as of late spring, reported shipments remained blocked. The estimate deserves its hedge, because no public census of installed AI compute by country exists and the denominator is analyst arithmetic. A condition with a published price schedule can be modeled. A condition with two governments’ hands on it is a negotiation.

The Apparatus column, read top to bottom: three conditionals and a borrowed, with no clean, frontier-quality owned position in the set. The element money buys quickly is the element with the most foreign terms attached, and in each of the three conditional cells the terms are written by the same counterparty. Abu Dhabi’s compute runs under a Washington-approved compliance regime with scheduled reviews. China’s ceiling is set in Washington and, for now, padlocked in Beijing. Britain’s ownership question runs through American acquirers and American capital. The French recalculation is the one being driven from home, and even there the largest campus has Gulf money underneath it and Nvidia allocation above it.

The elements also run on different clocks. Apparatus moves at the speed of a purchase order; a determined treasury can stand up a campus in two or three years. Ecology moves at the speed of a generation; no treasury in this set has managed to buy one, though several have priced the attempt. The durable question for any position here: which element does this actor hold, on what tenure, at what quality, and is it the fast element or the slow one. The next sovereign-AI deck will open on a render of the campus at dusk and the megawatt count. The deck will not say who can take it away, and that is the question the price turns on.

This snapshot carries a date because it is built to expire. The cells below are the ones wired to move, and what moves them.

Abu Dhabi, Apparatus. The licensing reviews are periodic, so the veto reprices on a schedule; the thing to watch is whether any review tightens deployment terms or the veto stays dormant. The forcing event is the listing window in 2027 or 2028, when a public market prices conditional compute for the first time. The size of the discount, or its absence, will be the cleanest reading on offer.

London, Makers. The bench is stable; ownership is the live variable. The signals are Maker events: a restructuring at Alphabet that touches DeepMind, a national-security intervention in a British AI acquisition, and whether the sovereign fund’s first cohort exits to American acquirers on the usual schedule or is held. April’s paused data center was the first data point of the year; the direction of the second will say whether fifty-seven pence is a floor or a ceiling. The number that does not yet exist is the full destination ledger for British exit value. Until someone publishes one, Tech Nation’s annual estimate is the only gauge, and its next revision is worth the diary entry.

Paris, Apparatus. Near-term checkpoints: the round closing at or above €20 billion, ground breaking on the MGX campus in the second half, and progress toward the 200 megawatt European target for 2027. Each one advances or stalls the conversion from tenant to holder. The quieter variable is allocation. Owned racks filled on someone else’s allocation put Mistral back in the tenant’s chair, one layer down.

Hangzhou, Apparatus, from both sides. On the American side: whether licenses keep issuing, and whether a Congress with legislation moving to re-tighten what January loosened lets the arrangement stand. On the Chinese side, currently the binding one: whether Beijing relents on H200 imports, and underneath that, Ascend production against the 750,000-unit projection and whether DeepSeek’s chip effort graduates from design to silicon. The cell moves toward owned on domestic supply or back toward tolled dependence on imports, and which government blinks first is the reading that matters.

A pattern runs through the soft spots above, and it is worth stating as a market rather than a caveat. The load-bearing numbers in this piece are estimates because the measurement layer beneath them does not exist: no public census of installed compute by country, no audited revenue from private frontier labs, no destination ledger for exit value, no instrument that prices tenure. Each absence is currently filled by one analyst’s arithmetic or one report’s estimate. Building those ledgers is somebody’s venture; until they exist, the working substitute is a reading like this one, taken on a schedule and keyed to the same cells.

Somewhere in the table a cell is already moving. The July reading says which cells are load-bearing and what they are attached to; the next quarter says which one repriced first. That is what the table buys while sovereign AI trades as a single position. The next reading is October’s.

Andrea Dew developed the proprietary CAMEO framework to disaggregate strategic technology positions across Capability, Apparatus, Makers, Ecology, and Operations.

The figures come from company and government disclosures, the labs’ own technical reports, Tech Nation’s 2026 estimate, and press reporting. “Reported” marks claims not independently disclosed. Estimates are identified in the sentence that uses them.

—Andrea