The Geopolitics of AI Infrastructure

How do you assemble a competitive AI system when your individual accelerators are not the most efficient available? Huawei’s answer is to connect more of them. Its Ascend 910B and 910C chips introduce the hardware story; CloudMatrix 384 turns that story into a system architecture. NVIDIA’s Blackwell GB200 NVL72 connects 72 GPUs through NVLink within one rack. Huawei connects 384 accelerators across multiple racks using optical links. Patel describes roughly twelve racks; the SemiAnalysis companion analysis clarifies that these are twelve compute racks, supplemented by four switch racks.

The surprising comparison is with an earlier NVIDIA design. In Patel’s account, DGX H100 Ranger would have connected 256 GPUs in one optical NVLink network, but cost, power consumption and reliability prevented it from reaching production. Blackwell took a different physical approach: concentrate the system into a rack so its high-bandwidth connections can remain copper. Instead of distributing the system across servers drawing roughly 10–20 kW each, the rack becomes an approximately 120 kW machine.

A larger interconnect domain can compensate for weaker individual chips, but it carries its own costs. Huawei accepts greater power consumption and extensive optical connectivity. Patel credits Chinese manufacturing economics with making this more affordable, while explicitly saying that reliability data is unavailable. Building the architecture and establishing its reliability at scale are separate achievements.

The next constraint is obtaining the components. Huawei faces entity restrictions, and China faces targeted semiconductor export controls—not a comprehensive embargo. Patel alleges that Huawei obtained TSMC fabrication through SOPHGO, a cryptocurrency-mining company acting as an intermediary. His account then follows the rest of the supply chain: high-bandwidth memory, or HBM, from Samsung and SK Hynix in Korea, and packaging equipment from the United States, the Netherlands and Japan. A nominally domestic accelerator can still depend on several foreign industrial systems.

Patel treats these routes as evidence of weak enforcement. SemiAnalysis estimates that Huawei obtained 2.9 million TSMC dies, rather than 2.9 million finished accelerators; Patel says that supply route had allegedly stopped. He also claims that TSMC received a roughly $1 billion fine against approximately $500 million of revenue, figures that should remain his account rather than an established enforcement outcome. Alongside the imported supply, he introduces SMIC as a prospective domestic fabrication route.

HBM creates another potential bottleneck because an accelerator needs memory as well as a compute die. The December 2024 US controls cover HBM under specified jurisdictional and authorization conditions. Patel describes an alleged route that changes the form in which the memory crosses the border:

1. Samsung sells HBM to CoAsia in Taiwan.
2. CoAsia passes it to Faraday, which packages it alongside a chip Patel describes as nonfunctional.
3. The package ships to China.
4. The HBM is removed and repackaged onto an Ascend accelerator.

The mechanism is physical repackaging: the desired memory arrives attached to another device, then becomes an input to the final accelerator.

Patel characterizes this route as legal circumvention and estimates that Huawei had stockpiled roughly 13 million HBM stacks, with shipments continuing. Both the transaction-specific legality and the stockpile remain claims in his account. The larger manufacturing question then changes: how long must imported inventories bridge the gap before domestic production can supply the required components?

Patel estimates that SMIC has tools sufficient for 50,000 wafers per month, while describing its observed seven-nanometer output at the time as smartphone chips. That distinction matters because die size and yield determine how much useful output a manufacturing process produces. Small phone chips are easier to manufacture successfully than large AI accelerators; demonstrating a process on one does not immediately establish economical production of the other.

His analogy is the gap between the iPhone’s 2020 move to a newer process generation and NVIDIA’s later H100 generation. The useful comparison is the sequence—small consumer dies first, large accelerator dies later—rather than treating the rounded process-node labels as exact specifications. Patel expects SMIC to follow that sequence into high-volume seven-nanometer AI-chip production during the year of the talk. His expectation of millions of chips depends on the yields it achieves.

The resulting compute need not consist of individually identical chips to become substantial in aggregate. Patel points to reported DeepSeek plans to use Huawei hardware for next-generation training as a possible demand source. This remains an anticipated training deployment, not evidence that the next model had already been trained on those systems.

Circumvention does not mean that restrictions have no effect. NVIDIA’s H20, described here as a reduced-capability H100/H200 offering for China, provides the counterexample. Patel rounds the financial impact to a $5 billion inventory write-down. NVIDIA’s reported Q1 FY2026 result gives the more precise accounting: a $4.5 billion charge covering excess H20 inventory and purchase obligations following the April 2025 licensing requirement.

Patel also paraphrases CFO Colette Kress as saying NVIDIA could have sold $50 billion of GPUs to China that year without restrictions. The earnings-call transcript instead describes nearly $50 billion as a future Chinese AI-accelerator market opportunity, not assured NVIDIA annual sales. Patel separately estimates that the restriction blocked about one million GPUs. His discussion therefore contains both effects: alternative supply routes can survive controls, while a particular restriction can still withhold a large amount of compute.

The Middle East agreements introduce a different policy instrument: allowing access in exchange for allocation and investment commitments. Following Trump’s visit to Saudi Arabia and the UAE, Patel turns to G42’s UAE construction. He describes an annual allowance of 500,000 GPUs, with 20% retained by G42 and 80% assigned to US hyperscalers, cloud providers and AI companies. These are the deal terms as he presents them, not operating capacity already delivered.

Patel describes a five-gigawatt campus ambition, with the satellite photograph showing initial portions of a one-gigawatt development. His contemporaneous comparison puts xAI at roughly 200 MW, infrastructure behind released OpenAI models at 200 MW or less, and the initial Stargate portions at 1.2 GW. The comparison is about the enormous scale of the proposed build; planned campus capacity should not be confused with a cluster already drawing that power.

Microsoft appears as both an investor in G42 and a likely large customer. Patel connects OpenAI’s anticipated Middle East cluster to Sam Altman’s earlier multitrillion-dollar infrastructure ambitions. The public record includes OpenAI’s May 22, 2025 Stargate UAE announcement, so the cluster should not be treated as permanently undisclosed simply because the talk calls it not publicly stated. It is a planned deployment.

The other side of the bargain is investment in America. Patel describes dollar-for-dollar matching: UAE spending on domestic AI infrastructure would be accompanied by equivalent investment in the United States. He cites G42 sites in Kentucky and New York as early examples, while leaving fulfillment of the full commitments conditional. Together with the proposed allocation to US companies, this is the political case for permitting the exports: overseas construction could also finance and serve the American AI ecosystem.

An audience question asks whether the UAE infrastructure is intended for inference or training. The distinction is less stable than those labels suggest. Patel describes US inference clusters using spare nighttime capacity for reinforcement learning with verifiable rewards: generate trajectories, evaluate their results and retain useful tokens for the training workflow.

A cluster serving users during busy periods can therefore contribute training data when demand falls. This does not mean every inference deployment becomes a full training system; it means that generating model outputs is itself part of some training workflows. A campus at the proposed UAE scale could support training, but its power rating alone does not determine which workloads it will run.

Saudi Arabia supplies the next example. Patel introduces DataVolt through the wider development story around The Line, describing a two-gigawatt data center as having broken ground. The NEOM project announcement identifies a different precise scope: a planned 1.5 GW campus in Oxagon, rather than The Line, with phased development rather than verified groundbreaking. The distinction matters when judging how much capacity is committed, under construction or operating.

Using his two-gigawatt description, Patel compares the project with xAI’s then roughly 200 MW training infrastructure: ten times the power scale of a build he says had already cost more than $10 billion. He also reports a $20 billion DataVolt commitment to US data centers and an approximately $80 billion broader investment total including American companies. These are separate project and investment claims in his account, not interchangeable measures of delivered compute.

Patel describes DataVolt as providing infrastructure for HUMAIN, the cloud-facing part of this ecosystem, and connects HUMAIN with Aramco Digital personnel. The supplier mix he outlines includes custom Qualcomm CPUs and both AMD and NVIDIA GPUs. He infers likely Groq participation from its Aramco Digital relationship; his estimate of more than 90% shifts between revenue and money received, so it does not establish a precise revenue concentration. AWS participation adds another route for American technology companies into the Saudi buildout.

The objections to these deals concern more than where a data center stands. Patel distinguishes several risks:

• Physical diversion: GPUs could be smuggled onward to China.
• Remote access: hardware could remain in the Middle East while Chinese customers rent its compute.
• Domestic opportunity cost: critics argue that the same GPUs could have served US deployments.
• Political power: greater AI capability strengthens authoritarian monarchies as well as their commercial partners.

Patel sees Saudi Arabia and the UAE as balancing relationships with multiple powers, and he doubts that US enforcement can reliably police every condition. Preventing a shipment from moving again does not, by itself, control who ultimately uses the machines.

The supporting case begins with demand from American AI companies. Patel links part of the Microsoft–OpenAI tension to OpenAI wanting more GPUs than Microsoft was willing to finance. A lab can promise to rent capacity, but somebody else must first build the data center and purchase the hardware. The provider carries that initial cash requirement.

Patel’s illustrative economics put recovery of the cluster investment two to three years into a roughly five-year rental contract. His hypothetical Microsoft example makes the exposure concrete: build a $100 billion data center for OpenAI, then depend on OpenAI raising $150 billion to pay for its commitments. Those are teaching figures, not a disclosed contract. Expected profit over the contract’s life does not eliminate the risk that the customer cannot fund the intervening payments. If demand disappoints, finding another customer for such a large specialized cluster may also be difficult.

External capital can increase compute access by accepting construction and customer-payment risk. After joking about investor judgment, Patel says he considers the deals worthwhile: SoftBank and Middle Eastern investors can fund the buildings and GPUs against promised rental demand. OpenAI obtains access without supplying all the construction capital itself, and the same arrangement can benefit other Western labs. His forecast is that these deals leave OpenAI with more compute in 2027–2028 than it would otherwise have had.

Financing is only one reason overseas capacity might be additional rather than a relocation of US construction. The other is electricity. Patel expects roughly four gigawatts of Middle East capacity by 2030, below the five-gigawatt promise he discussed earlier. The US faces a much larger mismatch between proposed data centers and available power.

Patel reports a projected US power shortfall of 63 GW, based on his team’s tracking of data centers under construction. The method follows individual sites and buildings through satellite imagery, permits and regulatory filings. It attempts to ground demand in physical development rather than relying solely on company announcements.

On the supply side, his model assumes strong additions from renewables and batteries, new single-cycle and combined-cycle gas generation, and retention of almost all coal capacity otherwise scheduled to retire. He separately cites 44 GW of power additions against 100 GW of data-center capacity additions. These figures are not a simple subtraction that reproduces the 63 GW estimate; their common horizon and accounting basis are not established in the discussion. The substantive point is that even favorable supply assumptions leave a large gap.

Closing that gap requires more than ordering generators. Skilled labor constrains construction; state and local approvals can delay projects; regulated utilities control critical parts of power delivery. Patel is especially critical of the utility structure, but the practical consequence is broader: an investor may have the money and the GPUs without being able to secure an energized site on the required schedule.

Patel contrasts America’s difficulty expanding electricity supply with China’s recent construction pace, which he characterizes as adding the equivalent of an entire US grid in seven years. That is his scale comparison, not a conversion between the chart’s energy quantities and data-center power ratings. His underlying concern is that the United States has not sustained comparable expansion for decades.

Stargate provides the visible construction example. Patel describes roughly 220 MW building groupings, two nearly complete, with additional construction underway, and gives 1.2 GW as the overall total. The groupings and total are his descriptions, not enough to reconstruct an exact building-by-building capacity ledger. For each successful headline project, he says, many others fail to obtain the power they need. His anecdote about an Ohio coal-stock recommendation tripling illustrates how he sees electricity scarcity affecting investment opportunities; it is not proof of the cause of that return.

The deployment comparison returns to Huawei’s opening tradeoff. Less efficient chips require more electricity, but abundant electricity can make that disadvantage manageable. More efficient chips still cannot run without power. The policy question is therefore whether Middle East construction enables useful compute that US infrastructure cannot deliver on the same schedule, despite the security and political risks of building it overseas.

The final audience question asks whether SMIC remains dependent on ASML or is developing an entirely domestic alternative. Patel describes SMIC as wholly dependent on foreign tools at the time, while emphasizing that numerous Chinese equipment companies are working on substitutes. Purchases of American, Japanese and Dutch equipment—including ASML systems—therefore coexist with domestic tool development.

The learning mechanism is concrete. Place an imported tool beside a domestic one, run wafers through both, compare the results and use the differences to improve the domestic process. Patel adds teardown and reverse engineering as another route. In his account, imported manufacturing equipment is both an immediate production dependency and a reference against which future substitutes can be developed.