Huawei has built a 384-chip AI system that uses an optical interconnect design Nvidia previously failed to bring into production.
2
Export controls have not stopped China from obtaining advanced chips, high-bandwidth memory, and semiconductor equipment through indirect supply chains.
3
The US may need Middle Eastern investment and overseas data centers because it cannot build power and data center capacity fast enough at home.
Summary
Dylan Patel examines AI infrastructure as a geopolitical contest involving chips, memory, data centers, power, and supply chains. He describes Huawei's Cloud Matrix 384, which connects 384 Ascend chips across 12 racks with optical links, and compares it with Nvidia's failed DGX H100 Ranger design. Patel argues that sanctions have been circumvented through intermediaries, allowing Huawei to obtain TSMC-made chips, Korean HBM, and Western packaging equipment. He then turns to large data center deals in the UAE and Saudi Arabia, where most of the purchased GPUs are intended for US companies. Patel accepts the risks of smuggling, re-renting compute to China, and relying on authoritarian states, but argues that these projects may give Western AI labs more compute than they could build in the US. The central constraint is power. He says the US has a large gap between planned data center demand and available generation, while China and the Middle East are adding capacity much faster.
Huawei has built a large optical AI system around its Ascend chips
Patel describes Huawei's Ascend 910B and 910C chips and the Cloud Matrix 384 system. It connects 384 chips across 12 racks, with optics and power infrastructure linking them at high bandwidth. Nvidia's comparable Blackwell NVL72 system uses 72 GPUs in one rack connected through NVLink. Patel says Huawei's design is similar to Nvidia's earlier DGX H100 Ranger project, which attempted to connect 256 GPUs with optics but never reached production because it was expensive, power hungry, and unreliable.
China obtained sanctioned components through a supply chain spanning several countries
Patel says Huawei-accessible chips were manufactured at TSMC through Sophgo, a cryptocurrency mining company that bought them while presenting itself as separate from Huawei. The systems also used HBM from Samsung and SK hynix in Korea, while packaging equipment came from the US, the Netherlands, and Japan. He says Huawei had accumulated roughly 2.9 million TSMC chips and about 13 million HBM stacks. His example of an HBM workaround involves Samsung selling to CoAsia, which sells to Faraday for packaging before the memory is moved onto an Ascend chip.
Domestic Chinese manufacturing could produce substantial AI compute despite lower yields
SMIC has enough equipment for about 50,000 wafers per month, according to Patel. The 7-nanometer chip publicly identified from SMIC is a smartphone processor, which is easier to make and has better yields than a large AI chip. Patel expects SMIC to begin producing 7-nanometer AI chips in high volumes. He rejects the idea that China will lack equivalent compute, especially as DeepSeek has announced plans to work with Huawei chips for future model training.
Export controls also removed a large Nvidia sales opportunity in China
Patel discusses Nvidia's H20 ban, describing the product as a cut-down H100 or H200. Nvidia wrote down $5 billion in inventory. He quotes CFO Colette Kress saying that without export restrictions, Nvidia would have sold $50 billion of GPUs to China that year. Patel says the ban blocked about one million GPUs, making the measure significant for both Nvidia's business and China's access to compute.
Middle Eastern data centers are being built partly for American AI companies
Patel explains a UAE deal in which G42 can buy 500,000 GPUs per year and keep 20 percent, while 80 percent must go to US hyperscalers, cloud companies, and AI companies. G42 is building a 5-gigawatt data center campus. The UAE is also matching AI infrastructure spending in the US, with G42 sites already operating or planned in Kentucky and New York. Patel says OpenAI is expected to have a cluster in the Middle East, although that arrangement was not publicly stated.
Saudi projects show how much larger new AI campuses are than current training sites
Patel describes DataVolt's project connected to Saudi Arabia's NEOM development. The company has broken ground on a 2-gigawatt data center and plans to invest $20 billion in US data centers. The total investment associated with the project is about $80 billion. He compares the scale with xAI's roughly 200-megawatt training infrastructure, saying the Saudi project is about ten times larger. He also connects Humain, Qualcomm, AMD, Nvidia, and Aramco Digital to the wider set of Middle Eastern AI infrastructure deals.
The Middle East deals carry real security and financial risks
Patel lists concerns that GPUs could be smuggled to China or rented to Chinese customers. He also questions whether US enforcement can reliably police the arrangements and whether the US should give infrastructure power to authoritarian monarchies. The financial risk is also substantial. A provider may spend billions on GPUs and a data center, then wait two or three years for rental payments. If an AI lab cannot raise enough money, the provider could be left with an oversized facility. Patel still thinks the deals are worthwhile because they could give OpenAI and other Western labs more compute in 2027 or 2028 than they would otherwise have.
US power shortages may force AI infrastructure outside the country
Patel says the US has a 63-gigawatt power shortfall based on the data centers under construction that SemiAnalysis tracks. He describes a model in which the US adds 44 gigawatts of power while data center capacity rises by 100 gigawatts. He attributes the problem to construction delays, a shortage of skilled labor, regulation, and utilities that operate as regulated monopolies. By contrast, he says China added the equivalent of an entire US power grid in seven years. His conclusion is that the US cannot build enough power and data centers without major changes, while the Middle East is planning roughly four gigawatts by 2030.
"The thought that China will not have equivalent compute is kind of wrong. They will have a lot of compute."05:05
Who should watch
You are planning AI clusters and need to understand how chip supply, memory, power, and export rules affect where they can be built.
You are evaluating Middle Eastern data center investments or cloud capacity and want the commercial risks alongside the geopolitical ones.
You work on AI policy or semiconductor strategy and need concrete examples of how sanctions are being bypassed and why power may be the limiting resource.
