Who Will Run Tomorrow's AI Infrastructure?

Part 2 of this Convergence found organizations that describe hiring as broken — frozen requisitions, recruiters underwater, a résumé that no longer means what it used to. Those same organizations, often in the same quarter, are committing capital to AI infrastructure at a scale with no peacetime precedent. That isn't a contradiction. It's one balance sheet making two different bets at once: caution on people, conviction on compute.

This publication has already measured the physical side of that compute bet — the concrete, the megawatts, the water. This investigation asks a different question: when the buildout settles, who actually owns and controls it? Not who hosts the servers in a press release, but who sits at the chokepoints — the handful of companies and governments positioned to capture the economics of AI infrastructure rather than merely rent space inside it.

The semiconductor industry didn't consolidate by accident — decades of capital intensity and physics pushed it there. Extreme ultraviolet lithography, the technology required to print the most advanced chips, is manufactured by exactly one company on Earth: ASML, whose EUV machines are the sole tool capable of producing the most advanced logic and memory used in AI accelerators [1][2]. One layer down, the fabrication of those chips is itself concentrated: in the first quarter of 2026, TSMC's 7-nanometer-and-below processes accounted for 74 percent of the company's total wafer revenue, with the newest 3-nanometer node alone at 25 percent [3]. And at the design layer sits Nvidia, holding an estimated 80 to 92 percent of the AI accelerator market and, in mid-2026, a market capitalization near $5 trillion on $215.9 billion in annual revenue [4][5].

None of this is new in kind — chip manufacturing has always rewarded scale and punished anyone who can't keep pace with node transitions. What's new is the stakes: a technology chokepoint that used to matter mainly to electronics buyers now sits underneath the infrastructure every government and hyperscaler on the planet is racing to build.

Industrial policy used to mean subsidies and export rules. In 2026, it increasingly means equity. The US government converted roughly $8.9 billion in CHIPS Act grants and Secure Enclave funds into a 9.9 percent ownership stake in Intel — 433.3 million shares at $20.47 each — a passive, non-board position later reported worth approximately $36 billion as Intel's stock rose [6][7][8]. Sovereign wealth is moving the same direction from the other side of the world: Saudi Arabia's HUMAIN, a wholly state-owned national AI champion chaired at Crown Prince level, is targeting 1.9 gigawatts of compute capacity by 2030, while Stargate UAE — a $30-billion-plus, 1-gigawatt campus built with OpenAI, G42, Oracle, Nvidia, Cisco, and SoftBank — has its first 200-megawatt phase due online in the third quarter of 2026 [9][10][11].

Not every government bet is landing. The European Court of Auditors concluded the EU Chips Act is "very unlikely" to reach its own target of a 20 percent global market share by 2030; the European Commission's own forecast puts the real number closer to 11.7 percent, a shortfall traced to Brussels directly controlling only about €4.5 billion of the funding while the rest depends on uncoordinated national and private spending [12][13]. China, moving under direct state direction, has pushed its AI-chip self-sufficiency ratio from roughly 20 percent in 2023 to over 40 percent in 2026, with SMIC serving as the exclusive foundry for Huawei's Ascend accelerator line — even as reported yield rates on its most advanced domestic process remain a documented 20 to 40 percent [14][15].

This ownership question doesn't sit apart from the rest of this Convergence. Part 2 of this Convergence found the same organizations pouring capital into this buildout describing hiring, in the same quarter, as frozen or unworkable — one balance sheet making two different bets at once. The government-equity moves documented above also connect industrial policy directly to labor and trade politics: a passive, non-board stake is a different kind of intervention than a tariff or a subsidy, and it puts sovereign balance sheets directly on the AI infrastructure cap table rather than at arm's length from it [6][9]. This publication has separately measured the physical footprint of this buildout — the concrete, the megawatts, the water — that sits underneath every chokepoint and every neocloud contract described here.

Three groups of actors — chipmakers, governments, and the newest layer of cloud middlemen — are all making moves toward owning this buildout right now. The evidence below doesn't pick a winner; it lays out what each group's claim actually rests on.

The ownership question resolves into layers rather than one line: chokepoints stay singular at the top of the stack [1][2][3][4], governments are converting policy leverage into direct ownership at an uneven pace [6][9][12], and the newest tier — neoclouds — remains the least settled of all [16][18]. "Hosting" AI infrastructure and "capturing its economics" are turning out to be two different jobs, increasingly held by two different sets of entities.

Public, trackable indicators worth watching over the coming months and years:

• Corporate announcements — Whether Meta Compute is formally launched, and whether Nebius or CoreWeave disclose material contract renegotiations or cancellations tied to it [18][19].
• Manufacturing yields — Whether SMIC's 5-nanometer process reaches disclosed mass-production yields, and whether China's self-sufficiency ratio tracks toward or away from Morgan Stanley's projected trajectory [14][15].
• Government equity actions — Whether the US government's Intel stake converts from passive to any form of governance right, and whether other CHIPS-adjacent equity deals (quantum computing firms already followed this pattern) extend into additional sectors [6][7]. Whether the EU narrows or widens the gap between its 20 percent target and its Commission's own 11.7 percent forecast — the clearest public scoreboard for whether industrial policy without ownership can still work [12][13].

Chokepoint manufacturers, sovereign-capital governments, or a hyperscaler layer that just absorbs everything beneath it — which one do you think ends up holding the actual economics of AI infrastructure five years from now, and why? Tell us. We read every message.