> ## Content Index
> Fetch the complete content index at: https://www.ffpurpose.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# Testing Zuckerberg's Infrastructure Argument
- URL: https://www.ffpurpose.com/testing-zuckerbergs-infrastructure-argument/
- Published: 2026-09-09T00:40:37.000Z
- Updated: 2026-09-09T00:45:30.000Z
- Author: tanner stahl
- Tags: Essay

## The claim

On August 11, 2026, Mark Zuckerberg published a note arguing that the United States is losing the AI infrastructure race to China. His core evidence: "Countries like China are bringing online 1GW+ of nuclear capacity every other week, so we will need to accelerate building both energy and data centers to remain competitive." \[1\]

I checked that number against the U.S. Energy Information Administration's numbers, which draw on the International Atomic Energy Agency's Power Reactor Information System. China added 1.1 GW of nuclear capacity in all of 2025, and 2.2 GW through May of 2026\. \[2\] "1GW every other week" implies something north of 25 GW a year. The actual number is roughly a tenth of that, even after annualizing the partial-year figure generously.

I want to be careful here, because the easy move is to catch a rich and powerful person in an exaggeration and declare the whole argument dead. That is not what the evidence supports. Zuckerberg's broader claim, that infrastructure buildout speed rather than model quality will decide near-term AI competitiveness, holds up well against everything else I found. His specific number just isn't one of the things holding it up. Separating those two claims, and figuring out where the real constraints actually sit across the countries racing to build AI infrastructure, is the point of this piece.

## The actual question

Most coverage of "the AI infrastructure race" treats it as a single resource problem: whoever has more power, or more capital, or more water, wins. I started this piece assuming something similar, that I could rank countries by which single resource was scarcest for them. That framing turned out to be wrong once I looked at the evidence country by country. The countries racing to build AI infrastructure are not short the same thing. They are short different things, and the type of shortage matters more than its severity, because it determines whether the constraint is something a country can buy, build, or negotiate its way out of.

I looked at four regions: the United States, China, the European Union, and the Gulf states building the most aggressively, the UAE and Saudi Arabia. For each, I looked at what's actually stopping capacity from coming online today.

![](https://storage.ghost.io/c/e7/13/e713053b-8704-4cf5-bc37-27c031e3b55e/content/images/2026/09/ai-capacity-by-region--1--1.svg)

## United States

The U.S. is not short capital. Hyperscaler capex is at record levels and shows no sign of slowing. It is also not primarily short water, despite periodic regional stories about drought and cooling. What it is short is a functioning process for connecting new generation and new large loads to the existing grid.

As of early 2026, U.S. interconnection queues held somewhere between 1,500 and 2,600 gigawatts of proposed generation and storage capacity, depending on which tracker you use and when it was published. Lawrence Berkeley National Lab's 2025 figure was around 1,500 GW; several 2026 trade reports put the number closer to 2,300 to 2,600 GW. \[3\] I'm flagging that range rather than picking one number, because the sources disagree by more than a rounding error and none of them cite a single reconciled methodology. What every source agrees on is the direction: the queue is larger than the country's entire installed generating capacity, and it's growing faster than anyone is clearing it.

PJM Interconnection, the grid operator covering thirteen states and D.C., is the clearest single data point. The average time from interconnection application to commercial operation there rose from under two years in 2008 to more than eight years in 2025, according to RMI's analysis. PJM's own 2025 forecast projects 32 GW of peak load growth from 2024 to 2030, with about 30 GW of that coming from data centers alone. \[4\]

Even for projects that clear the queue, there's a second, more physical bottleneck: large power transformers, the equipment that steps voltage up and down along the grid. Wood Mackenzie found a 30 percent supply shortfall in large power transformers and a 10 percent shortfall in distribution transformers in 2025, with lead times for power transformers averaging 128 weeks and generator step-up transformers averaging 144 weeks. \[5\] A separate CNBC report in September 2026 cited somewhat lower 2026 shortfall figures (15 percent and 8 percent), attributing part of the gap to restrictions on China-made transformer units. \[6\] The exact percentage moved between reports; the fact that transformer supply is a hard, multi-year physical constraint did not.

The U.S. government has noticed. FERC issued a show-cause order in mid-2026 directing six grid operators to explain how they'll manage queue backlogs, cap ratepayer exposure to data center-driven transmission costs, and report spare capacity, with initial filings due by mid-July 2026\. \[7\] That's a real regulatory response, but it addresses process, not the transformer supply chain, which nobody can order into existence faster.

The upshot: the U.S. constraint is capital-solvable in the sense that money can eventually buy more transformer manufacturing capacity and more transmission steel in the ground. It is not solvable on the timeline anyone wants. S&P Global's forecast has U.S. data center capacity growing from about 62 GW in March 2026 to 152 GW by 2030, which implicitly assumes the queue and equipment problems get worked out over roughly that window. \[8\]

## China

China's position is close to the inverse of the U.S. one. Power is the advantage, not the constraint. State-directed capital is effectively unlimited for this priority. What China cannot currently buy at will is the most advanced AI-training silicon.

As of late 2025, U.S. firms controlled about 96 percent of global AI compute capacity, according to Stanford HAI's AI Index Report as cited by the American Action Forum, with Nvidia alone accounting for 67 percent. \[9\] Export controls on advanced chips and the equipment used to manufacture them, coordinated with the Netherlands and Japan since 2022, remain the single largest lever the U.S. has over Chinese frontier AI development. China's response has been to lean harder on domestic alternatives (SMIC fabrication, Huawei-designed accelerators) and to shift toward more compute-efficient model architectures, of which DeepSeek is the most visible example. \[10\]

Meanwhile the energy side of China's buildout looks genuinely strong, just not in the specific way Zuckerberg described it. China's total data center capacity is projected to reach roughly 60 GW by 2030, nearly double the current level, according to Rystad Energy. \[11\] China's own nuclear fleet, while much smaller in absolute growth than Zuckerberg's number implied, is real and accelerating: 60 operational reactors and 36 more under construction as of mid-2026, representing more than 49 percent of all nuclear construction happening anywhere in the world right now, with a 15th Five-Year Plan target of 110 GW of nuclear capacity by 2030\. \[12\] State Grid Corporation of China invested roughly 650 billion yuan in the grid in 2025 alone and plans about 4 trillion yuan over the 2026 to 2030 plan period. \[13\] None of that is "1 GW every other week," but it is a genuinely large, state-coordinated buildout that most Western coverage understates relative to its chip-export-control coverage.

So China's constraint is capital-solvable in theory (money can eventually buy or build around any chip shortfall through domestic fabrication) but on a much longer and more uncertain timeline than the U.S. transformer problem, because semiconductor fabrication at the leading edge is a much harder manufacturing problem than transformer manufacturing.

## European Union

The EU case surprised me most, because it's the one region where the binding constraint is neither a physical shortage nor a geopolitical one. It's regulatory fragmentation.

Grid connection approval in major EU markets currently runs 24 to 36 months, similar in magnitude to the U.S. queue problem, and Gartner projects 40 percent of AI data centers will be power-constrained by 2027\. \[14\] But the EU also layered new rules on top of that baseline in 2026: mandatory energy reporting under the Energy Efficiency Directive, Germany's hard PUE ceiling for new builds, and new legislation (referred to in industry coverage as CADA) that explicitly ties priority grid connections and lower network fees to data centers built with European-manufactured chips, published alongside a "Chips Act 2.0" in June 2026\. \[15\] That's a genuinely different kind of constraint than the U.S. or China face: it's policy design choosing to slow down or redirect capacity in service of semiconductor sovereignty, not a physical inability to deliver power.

The result is a fractured map within the EU itself rather than a single national number. Dublin is reportedly no longer viable for new AI campuses above 50 MW without on-site generation. \[16\] Meanwhile France is attracting massive investment specifically because it can offer nuclear baseload and a simplified process: SoftBank committed €45 billion in 2026 to build three data centers in northern France drawing 3.1 GW directly from the grid. \[17\] One trade report cited €5.8 billion in projects across the bloc holding permits but no grid capacity to actually connect. \[18\]

The EU's constraint is the most solvable of the four in principle: a single harmonization effort or a France-style national energy advantage replicated elsewhere could unlock capacity relatively quickly, because the underlying physical resources (nuclear baseload, industrial land, capital) exist. What's missing is political and regulatory alignment across 27 member states, which is a genuinely different problem than manufacturing more transformers or fabricating more advanced chips.

## Gulf states

This is where the story I expected to write broke completely.

Before 2026, the standard framing of Gulf AI infrastructure was accurate as far as it went: sovereign capital is effectively unlimited (the UAE has more than $30 billion in announced AI and data center investment, Saudi Arabia has committed $18 billion under Vision 2030, and the HUMAIN initiative alone represents a reported $77 billion investment roadmap toward roughly 6 GW of capacity over the next decade) \[19\], and the real constraint was widely described, as recently as July 2026, as energy rather than chips or capital, with water flagged as the more acute long-term issue given that eleven of the world's seventeen most water-stressed countries sit in the Middle East and North Africa. \[20\] The UAE's AI sector alone could require an estimated 61 billion liters of water a year by 2030, and heavy reliance on desalinated drinking water (up to 90 percent of supply in Kuwait, 70 percent in Saudi Arabia) makes any cooling strategy that competes with drinking water politically dangerous even before it becomes physically binding. \[21\]

That analysis is no longer the whole picture, because since February 28, 2026, the U.S. and Israel have been at war with Iran, centered on the Strait of Hormuz, with a truce that broke down in July. Iranian drones have struck data center facilities in Bahrain and the UAE. \[22\] The OECD warned in June 2026 that continued conflict could delay or halt Gulf AI infrastructure projects outright, specifically naming Saudi Arabia's $2.7 billion Hexagon facility as an at-risk asset and citing the March 2026 strikes on cloud facilities in the UAE and Bahrain. \[23\] Iran has also threatened to sever undersea cables and mine the strait itself. \[24\]

No amount of sovereign capital fixes a region where facilities are being struck by drones and shipping lanes are being closed. Money solves the water problem eventually, through desalination and closed-loop liquid cooling that can cut direct water use by 90 to 98 percent compared to conventional cooling. \[25\] Money does not solve an active war. This is the one region of the four where the constraint is neither capital-solvable nor policy-solvable on any near-term timeline. It's contingent on something entirely outside the infrastructure conversation: whether the conflict ends.

## Revisiting Zuckerberg

Putting the four cases together changes how I read his argument.

He is right that infrastructure delivery speed, not model capability, is the variable most likely to separate the U.S. and China over the next few years, and right that this is underappreciated relative to how much attention goes to model benchmarks. He is also right, separately, that maintaining export controls on advanced chips to China remains one of the few levers actually constraining a well-capitalized, energy-rich competitor. \[26\] Both of those are defensible positions supported by the evidence above.

He is wrong about the specific number he used to make the case, by close to an order of magnitude, and the number he chose happens to be the most dramatic-sounding one available: "1 GW every other week" is vivid in a way "2.2 GW in five months" is not. I don't think that means he fabricated the figure maliciously. It's more likely a case of a compelling but unverified stat making it into a widely-read note because it supported an argument he already believed, which is a failure mode worth naming because it's common, not because it's unusual.

It's also worth naming who is making the argument. Zuckerberg is not a neutral analyst. Meta is mid-buildout on its own multi-gigawatt "superintelligence" data centers, including a cluster the company has said could rival a significant portion of Manhattan's footprint, funded by "hundreds of billions of dollars" the company says it can cover from its advertising business. \[27\] A public argument that national policy should prioritize faster energy and data center permitting is also, directly, an argument for policies that benefit Meta's own balance sheet. That doesn't make the argument wrong. It does mean it shouldn't be read as disinterested infrastructure analysis, and most of the coverage I found treated it that way anyway.

## Why this matters

If the framing above is right, the practical question "where does AI inference get cheaper first" doesn't have a single answer, because the four leading buildouts are gated by different kinds of obstacles with different resolution timelines.

The U.S. constraint should ease first among the physically-bound ones, because transformer manufacturing capacity, while slow, is a known industrial problem that responds to sustained capital investment over a three-to-five-year horizon, and FERC's 2026 order at least starts addressing the process side of the queue. China's constraint resolves on a longer and more uncertain timeline, gated by how fast domestic chip fabrication can close the gap with the export-controlled frontier, which is a harder manufacturing problem than transformers and has resisted rapid solutions for years. The EU's constraint is, on paper, the fastest to resolve, because it doesn't require inventing new physical capacity, just harmonizing rules across member states, but the EU's track record on cross-border regulatory speed doesn't inspire confidence that "on paper" translates to "in practice" quickly. The Gulf's timeline is not really a timeline at all right now. It's a contingency on whether a war ends.

## What this doesn't establish

This is a synthesis of public reporting and a handful of primary sources (EIA, IAEA's PRIS, World Nuclear Association, PJM's own forecast), not a first-hand audit of grid filings or company disclosures. Several of the specific figures here, the U.S. interconnection queue total, the Wood Mackenzie transformer shortfall percentages, and the Gulf water-demand projections, come from secondary reporting on primary sources I did not independently pull and reconcile myself. Where sources disagreed materially, I said so rather than picking the more convenient number. The Gulf war's trajectory could change the entire regional picture within weeks of this piece publishing, in either direction. And I did not attempt to verify every figure to EIA-level rigor, only the one Zuckerberg used as his headline claim, because that was the specific, falsifiable statement driving the piece.

## How to further validate

A version of this worth trusting more would pull PJM, ERCOT, and CAISO interconnection queue data directly rather than through trade press aggregation, reconcile the transformer shortfall numbers against the actual Wood Mackenzie report rather than press coverage of it, and track the Gulf war's effect on the specific projects named at risk (Hexagon, Stargate UAE, HUMAIN's Riyadh and Dammam sites) against their original delivery timelines over the next two quarters. I'd also want China's grid investment figures translated into the same units and time windows used for the U.S. and EU numbers, since right now they're not directly comparable across the sources I used.

---

## Sources

1. Zuckerberg's infrastructure note, as reported: Tribune India, "Zuckerberg warns China could gain AI edge, calls for faster US AI infrastructure buildout" (Aug 2026); ANI News, same story (Aug 11, 2026); Detroit News, "Five takeaways from Zuckerberg's AI manifesto" (Aug 10, 2026)
2. U.S. Energy Information Administration, "China's nuclear power capacity nearly doubled since 2016," citing IAEA Power Reactor Information System (June 2026)
3. Techplustrends, "AI Data Center Power Requirements 2026," citing Lawrence Berkeley National Lab; EnkiAI, "AI & Data Center Energy 2026, 2,600 GW Queue & PJM Plan"; Hanwha Data Centers, "Data Center Grid Limitations: The Power Bottleneck"
4. Quartz, "America's power grid can't keep up with AI demand," citing RMI analysis and PJM's 2025 long-term load forecast (May 2026)
5. Quartz, same article, citing Wood Mackenzie transformer shortfall and lead-time data
6. CNBC, "Hidden China risks are emerging in America's multibillion-dollar AI data center boom" (Sept 2026)
7. Tech Insider, "US Grids Get 60 Days to Fix AI Data Center Power" (2026); Data Center Knowledge, "Gridlocked: Power Constraints Shape the Future of Data Centers" (Mar 2026)
8. CNBC, "Hidden China risks..." citing S&P Global data center capacity forecast (June 2026)
9. American Action Forum, "Beyond Chips: Can Expanding Export Controls Slow China's AI Progress?" citing Stanford HAI 2026 AI Index Report (June 2026)
10. IEEE ComSoc Technology Blog, "China vs U.S.: Race to Generate Power for AI Data Centers" (Feb 2026); Al Jazeera, "China's secret weapon in AI race with US? Lots of cheap energy" (May 2026)
11. Al Jazeera, same article, citing Rystad Energy data center capacity projection
12. EIA / IAEA PRIS data (as above); World Nuclear Association, "Nuclear Power in China" country profile
13. World Nuclear Association, same profile, citing State Grid Corporation of China investment figures
14. Spheron Network, "Power-Bound, Not GPU-Bound: AI Data Center Power Constraints Are the Real 2026 Bottleneck," citing Gartner projection (June 2026)
15. Jones Day, "EU Data Center Rules Combine Expansion Incentives with New Energy Obligations" (June 2026); Data Center Knowledge, "Data Center Compliance in 2026" (May 2026)
16. Techplustrends, "AI Data Center Energy Cost Europe: 2026 Market & Cost Guide" (June 2026)
17. CRV Science, "AI vs. Net-Zero Emissions: The Physical Energy Limits of Europe's Digital Expansion" (June 2026)
18. Moduledge, "EU Data Center Regulations 2026: PUE, Reporting & Compliance Guide" (Apr 2026)
19. ValuStrat, "UAE Data Centre Investment 2026"; PodTech Data Center, "AI Data Center Infrastructure in the Middle East" (Aug 2026)
20. Vision2030.ai, "Saudi AI's Real Constraint Is Power, Not Chips," citing IDCA CEO Mehdi Paryavi interview with Asharq Al-Awsat (July 2026); Computer Weekly, "Designing water-smart AI datacentres in GCC and MENA"
21. Middle East Council on Global Affairs, "Desert Bytes: Why Gulf AI Ambitions Must Align with Energy and Water Realities" (May 2026)
22. Stimson Center, "Beneath the Strait: Iran Could Threaten Gulf Data Centers, Undersea Cables" (Apr 2026)
23. Vision2030.ai, "Saudi AI's Real Constraint Is Power, Not Chips," citing OECD warning (June 7, 2026)
24. Stimson Center, same article
25. Computer Weekly, "Designing water-smart AI datacentres in GCC and MENA," on liquid cooling water reduction
26. Tribune India, "Zuckerberg warns China could gain AI edge..." on export control position
27. Reuters, "Meta's Zuckerberg pledges hundreds of billions for AI data centers in superintelligence push" (2025); TechXplore, "Zuckerberg, Musk make plea at G20 for more AI data centers" (Sept 2026)

\*Several of these sources are industry trade press and SEO-oriented aggregator sites rather than primary sources. Where a claim traces back to a named primary source (EIA, IAEA PRIS, RMI, Wood Mackenzie, Gartner, S&P Global, Stanford HAI, the OECD), I've named that primary source in the text alongside the outlet that reported it. Where I could not trace a figure to a primary source, I've flagged the range or disagreement explicitly.