The Geopolitics of AI Infrastructure - Dylan Patel, SemiAnalysis

Nvidia's CFO stated on an earnings call that without export restrictions to China, Nvidia would have sold $50 billion worth of GPUs to China in a single year, while the actual ban prevented the sale of approximately one million GPUs, which the company wrote down as $5 billion in inventory.

“recently Nvidia got banned from selling their H20. This is like a cut down H100 or H200 to China. um they wrote down $5 billion worth of inventory and on the earnings call, Colette, the CFO said, "If we didn't have export restrictions, we would have sold $50 billion worth of GPUs to China this year." Which I thought was like a very interesting comment. Um but uh the ban stopped about a million GPUs and um if you Yeah, it's it's it's it's quite interesting that like you know that amount of compute was also blocked.”

The Middle East (UAE and Saudi Arabia) is building 4-5 gigawatts of power capacity by 2030 to support data center infrastructure, while the Middle East's willingness to invest capital is strategically beneficial to US AI companies regardless of how many times the speaker emphasizes how 'stupid' the investment is.

“Middle East is building, you know, by 2030 like four gigawatts, right?”

The G42 (UAE-based AI company) has secured a deal to purchase 500,000 GPUs annually, retaining 20% for internal use and directing 80% to US hyperscalers and cloud companies, and is constructing a 5 gigawatt data center campus (satellite photos show the first phase is a 1 gigawatt facility), which is approximately 25x larger than XAI's 200 megawatt training infrastructure and comparable to the entire 1.2 gigawatt Stargate project.

“the deal is basically that G42 can buy 500,000 GPUs a year. Um and they get to keep 20% of them to do whatever they want. The other 80% have to go to uh US hyperscalers and cloud companies and AI companies, right? And so uh G42 is building a 5 gawatt data center campus. The one photo pictured is the first parts of a gigawatt data center campus, right? Um that is that is absolutely ridiculously big, right? XAI's data center is like 200 megawatts, right?”

DeepSeek has announced plans to train next-generation models using Huawei chips, representing a concrete deployment of Chinese domestic compute infrastructure for advanced AI model development.

“there's already like big announcements from DeepSeek that they're going to work and use use Huawei chips uh to try and train their next generation models.”

US export restrictions and sanctions enforcement is ineffective at preventing adversary access to critical components because determined actors (China, Huawei) can use shell companies, third-party procurement, and multi-step supply chains to access US, Dutch, Japanese, and Korean manufacturing and components despite official bans.

“it's funny that sanctions are completely useless because they're still able to access this stuff...we know that export restrictions and things like that the enforcement of them sucks so it's not like the US government apparatus is going to actually be able to effectively enforce this right”

China's SMIC (China's TSMC equivalent) can produce 50,000 wafers per month at 7 nanometers, and based on historical precedent where smartphone chips reach advanced nodes before large AI chips, SMIC is likely to begin producing 7-nanometer AI chips in very high volumes this year, contradicting assumptions that China will lack equivalent AI compute.

“Smick, right, their TSMC uh, has enough tools for 50,000 wafers a month. Today, the only 7nmter chip that anyone's found from SMIC is their smartphone chip, right? Smartphone chips are smaller, easier to make. Um, they yield better than large AI chips. Um, so when you look at like, hey, 5nanmter, the first 5nimeter chip was an iPhone chip in 2020, but Nvidia didn't release a 5nanmter GPU until uh, 2022, 2023 with the H100, right? Uh and so likewise, right, Smick is making seven nanometer for phones. Uh they're going to get to the point where they can start making seven nanometer for AI chips likely in very high volumes this year.”

OpenAI will have more compute capacity in 2027-2028 than would have been built domestically, enabled by Middle East financing that reduces US capital constraints on infrastructure buildout.

“you know OpenAI it because of this will have more compute in 2027 or 2028 than they would have had without it, right? It wouldn't have been built in the US.”

Huawei sourced approximately 2.9 million chips from TSMC for their AI systems, with manufacturing arranged through Sofco, a Bitcoin and cryptocurrency mining company that falsely presented itself as unaffiliated with Huawei to evade sanctions, and this supply has allegedly stopped following a US fine to TSMC of $500 million against $500 million in revenue, which the speaker characterizes as a minimal penalty.

“they have roughly 3 million chips worth 2.9 million chips worth uh from TSMC. That's allegedly stopped now, right? the US gave TSMC like a billion dollar fine uh for $500 million of revenue that they made uh which which doesn't seems like a slap on the wrist honestly...the chip is manufactured through uh at TSMC through Sofco uh which is a Bitcoin cryptocurrency mining company uh that pretended to not be attached to Huawei and just bought these chips.”

The United States faces a massive power supply deficit of 63 gigawatts to meet the electricity demands of data center construction projects already under construction or in advanced planning, based on satellite photography, construction permits, and development tracking of every data center site by site, while only 44 gigawatts of new power generation capacity is being added even with aggressive assumptions about renewables, batteries, and combined-cycle gas and nuclear reactors that are planned or being installed.

“there is a 63 gawatt uh shortfall of power in the US right based on the data centers that we see under construction and we're tracking every single one site by sight etc building by building um with satellite photos permits regular providings all this stuff right...you have 44 gigawatts being added but you have a um on on the power side...whereas you're adding 100 gawatts of data center capacity so the US simply just doesn't have enough power unless we do something”

US export restrictions stopped approximately one million GPUs from reaching China, representing significant but incomplete enforcement of policy.

“the ban stopped about a million GPUs and um if you Yeah, it's it's it's it's quite interesting that like you know that amount of compute was also blocked.”

Huawei absolutely destroyed competitors in 5G and telecom through engineering superiority rather than cost advantages, demonstrating that Huawei operates at the technological frontier in communications technology.

“Huawei is is really really cracked. Um, they absolutely destroyed everyone in 5G and telecom um by just engineering better.”

US power constraints are caused by multiple structural failures including skilled labor shortages, regulatory delays at federal, state, and local levels, and utility monopolies that lack incentives to expand generation capacity and allow utilities to set rates without competitive pressure.

“the argument sort of is like well geopolitically the US can't build um enough data centers can't build enough power whether it's for the lack of skilled labor whether it's for the lack of you know regulatory regulators holding stuff up. Um, federal, not as much anymore, but local, state, and also in America, we have this beautiful thing called utility companies that are regulated monopolies that get to do whatever they want, right? If anyone has a power bill in California, they understand that the utilities suck.”

Stargate, the first six components of which total 1.2 gigawatts, is significantly larger than current AI training infrastructure and represents an order-of-magnitude increase in data center scale.

“Stargate, the first uh first six parts of it in total is 1.2 gigawatts, right? So like this is a this is a massive um data center that they're building.”

SMIC (China's TSMC equivalent) currently has manufacturing capacity for 50,000 wafers per month at 7 nanometers, but to date has only demonstrated 7nm production for smartphone chips, which are smaller and easier to manufacture with better yields than large AI chips, yet based on smartphone-to-GPU production timelines (iPhones reached 5nm in 2020 while Nvidia's first 5nm GPU was the H100 in 2022-23), SMIC will likely be able to produce large 7nm AI chips in high volumes this year.

“China's uh, Smick, right, their TSMC uh, has enough tools for 50,000 wafers a month. Today, the only 7nmter chip that anyone's found from SMIC is their smartphone chip, right? Smartphone chips are smaller, easier to make. Um, they yield better than large AI chips. Um, so when you look at like, hey, 5nanmter, the first 5nimeter chip was an iPhone chip in 2020, but Nvidia didn't release a 5nanmter GPU until uh, 2022, 2023 with the H100, right? Uh and so likewise, right, Smick is making seven nanometer for phones. Uh they're going to get to the point where they can start making seven nanometer for AI chips likely in very high volumes this year.”

Despite US sanctions on Huawei and China, Huawei accessed TSMC manufacturing through Sofco, a cryptocurrency mining company that posed as unaffiliated, enabling production of approximately 2.9 million Ascend chips, demonstrating the ineffectiveness of export controls.

“despite the fact that Huawei is a sanctioned entity, China is a sanctioned country, they're still a they were still able to actually access uh TSMC to manufacture these chips, right? Um the the chip is manufactured through uh at TSMC through Sofco uh which is a Bitcoin cryptocurrency mining company uh that pretended to not be attached to Huawei and just bought these chips. Um, and then there's, you know, HBM, which is high bandwidth memory from Samsung and Highex in Korea. Um, and then all the equip all the equipment to package it together is from the US, Netherlands, and Japan. So, it's funny that sanctions are completely useless because they're still able to access this stuff.”

Huawei's Ascend 910B and C chips, integrated into the Cloud Matrix 384 system architecture with 384 chips connected across 12 racks using optics and power infrastructure, represents an architecture that Nvidia attempted and failed to deploy called DGX H100 Ranger, which had 256 Nvidia GPUs connected in NVLink but proved too expensive, power-hungry, and unreliable to bring to production.

“Nvidia talks a lot about Blackwell NVL 72. It's one rack with 72 GPUs...Huawei's done something similar. It's called the Cloud Matrix 384. 384 of their chips, right? Connected together in not one rack, but like 12...Their architecture is actually one that Nvidia tried to deploy and they failed at um, and that was called uh, DGX H100 Ranger, right? Which was 256 Nvidia GPUs...They tried and it did not work. They could not bring it to production because it was expensive, power hungry and unreliable.”

Ramco Digital (also known as Humane) is the dominant revenue source for xAI's Grok, with over 90% of Grok's revenue or funding coming from Ramco Digital, indicating that xAI's commercial viability is dependent on corporate cloud infrastructure revenue.

“Humane is actually the Ramco uh folks the Ramco Ramco digital folks which is the vast majority of Grock's revenue as well so there's uh likely Grock stuff in there um I think like 90 plus% of Grock's revenue or or they've got fun money from them is coming coming from uh from from the Ramco digital folks.”

The Microsoft-OpenAI partnership partially dissolved because OpenAI demanded GPU quantities that Microsoft considered untenable, forcing OpenAI to seek alternative funding and infrastructure partners to meet its computational requirements.

“Microsoft OpenAI deal sort of split up partially because OpenAI wanted so many GPUs and Microsoft was like, "No, you're crazy. Like that's way too much. like you don't have revenue. Um, and OpenAI is like but we can convince someone to build it for us, right? Because OpenAI is really nice, right? Like they get to rent the GPUs.”

China's advantage in the geopolitical AI race rests partly on having no power problems; even if China's chips are less efficient, the abundant, cheap power eliminates this disadvantage.

“Uh China's got no pro power problems. Even if their chips are less efficient, doesn't really matter, right? Like I think the AI race is a very geopolitical and interesting one.”

Stargate project phases are adding approximately 220 megawatts each with two already near completion and six more in development, totaling 1.2 gigawatts, while facing significant project failures compared to completed projects in the broader portfolio.

“So for every Stargate that's out there, right, um you know, 220 megawatts for these four building like things and there's two of them already almost completed and then there's another six going up. Uh that's 1.2 2 gawatts total for Stargate.”

Middle East investors and SoftBank are willing to take larger financial risks on data center infrastructure because they are less constrained by traditional financial discipline than US investors, effectively absorbing the capital cost burden that US companies cannot justify.

“the nice thing is like you know a fool is easily parted with their money um and SoftBank and the Middle East are the most stupid investors in the world potentially um and so you know I actually I think they're fine I think They're fine. I think these are good deals, but like you know um you know OpenAI gets to have someone pay for the cluster um buy all these GPUs, build all the data centers and then like with the promise they're going to rent them, right?”

Inference clusters in the US are increasingly being used for training during off-peak hours through reinforcement learning with verifiable rewards, blurring the distinction between inference and training infrastructure.

“people today now uh in their inference clusters at night in the US are just running reinforcement learning with verifiable rewards generating you know trajectories and then you know keeping the good tokens uh when there's low utilization right so like inference clusters are now also training clusters right so I think I think that there's like the lines are blurring between what is an inference and training cluster”

OpenAI's desire for massive GPU capacity (repeatedly stated as 'seven trillion' or similar ambitious figures) was partly the cause of a partial split in the OpenAI-Microsoft partnership, as Microsoft declined to build clusters at the scale OpenAI demanded and at the pace needed, forcing OpenAI to seek external partnerships (such as Middle East deals) where third-party investors build, purchase, and own the data centers while OpenAI rents the compute.

“Microsoft OpenAI deal sort of split up partially because OpenAI wanted so many GPUs and Microsoft was like, "No, you're crazy. Like that's way too much. like you don't have revenue. Um, and OpenAI is like but we can convince someone to build it for us, right? Because OpenAI is really nice, right? Like they get to rent the GPUs."”

Huawei has accumulated approximately 13 million HBM stacks through the Samsung-Coasia-Faraday supply chain workaround and continues to receive additional shipments.

“Um, and so Huawei has stackpiled roughly 13 million HBM stacks. Um, you know, already. Um, and they're continuing to receive more shipments.”

There are legitimate geopolitical security risks to offshore GPU placement in the Middle East, including the possibility that GPUs could be smuggled to China, rented to China in violation of deal terms, or used to support authoritarian regimes, but these risks are offset by the alternative scenario where the US cannot build sufficient domestic compute and thus cedes AI leadership to China by default.

“there are real criticisms, right? Like, what if these GPUs get smuggled, right, to China?...there's all these security requirements like okay what if okay they're not maybe the GPUs don't get smuggled there maybe they get rented to China right...there are real risks here um and and also like giving power to authoritarian countries and that have like you kingdoms, right?”

Inference clusters operated by US AI companies at night (during low utilization periods) are increasingly being used to run reinforcement learning with verifiable rewards to generate training trajectories and token sequences, effectively converting idle inference infrastructure into training infrastructure, blurring the distinction between inference and training clusters.

“people today now uh in their inference clusters at night in the US are just running reinforcement learning with verifiable rewards generating you know trajectories and then you know keeping the good tokens uh when there's low utilization right so like inference clusters are now also training clusters right”

US companies have pitched coal stocks to clients because of power shortages, demonstrating that market participants are responding to the perception of power constraints in the US energy system.

“Like I've literally pitched a coal stock to my clients in Ohio and the stocks 3x because there's power issues. Like like it's it's crazy how um how power constrained America is, right?”

Humane (the cloud/data center company part of the deal structure) is substantially funded by or connected to Ramco Digital, which also generates ~90% of Grok's revenue, suggesting that Elon Musk's Grok AI model training may be supported by Middle East infrastructure and investment.

“Humane is actually the Ramco uh folks the Ramco Ramco digital folks which is the vast majority of Grock's revenue as well so there's uh likely Grock stuff in there um I think like 90 plus% of Grock's revenue or or they've got fun money from them is coming coming from uh from from the Ramco digital folks.”

SMIC relies entirely on Western tools (US, Japanese, Dutch equipment including ASML) for chip fabrication, but is developing domestic alternatives by purchasing billions of dollars of Western equipment, running wafers through both Western and domestic tools in parallel, and reverse-engineering Western equipment to improve Chinese tool capabilities.

“SMIC, China's TSMC is uh dependent entirely on Western tools today. Uh, but there are many uh Chinese tool companies...they'll buy billions of dollars of uh American, Japanese, and Dutch equipment, including ASML, but then they'll put it next to their domestic tool, and then they'll run uh wafers through both, and then they'll just like learn how to improve it. They'll also tear them down and reverse engineer them.”

Stargate's first six phases total 1.2 gigawatts of compute capacity, each phase comprising 220 megawatts with two phases nearly completed and six more under construction.

“Stargate, the first uh first six parts of it in total is 1.2 gigawatts, right? So like this is a this is a massive um data center that they're building.”

UAE is providing matching investments to the US for AI infrastructure: for every dollar spent in UAE on AI, UAE will spend an equivalent dollar in the US on AI infrastructure; G42 has sites in Kentucky and New York.

“UAE is uh providing matching investments to uh, US, right? For any dollars they spend in UAE, the UAE spends in UAE on AI infrastructure, they're also going to spend in the US. Um, and so that's already started, right? G42 has sites in like Kentucky and New York, um, that they're spending in.”

The total announced investment in Middle East + US data center infrastructure is approximately $80 billion across multiple Saudi, UAE, and American companies.

“Um likewise um you know and then there's also an AWS thing, right? So there's all all these American companies are also like coming together with the Middle East. So the question is like you know why the heck are we sending all these GPUs to the Middle East is what like the EAS would say and then the like capitalists will be like yeah GPUs and money, right? So like it depends on what side of the fence you're on. Um, and I like to think, you know, I'm uh I'm on one side of the fence, so I'll tell you which side.”

Many AI data center projects in the US are failing despite high announced investment figures, illustrating that power constraints are blocking execution independent of capital availability.

“Um for every one of these that's happening, there's so many projects that are failing, right?”

Despite irrational economics, the Middle East data center investments benefit US AI companies because foreign investors absorb capital expenditure and payback risk, allowing US companies like OpenAI to access compute that would not have been financed domestically, and the speaker characterizes this as a favorable outcome even if Middle East investors are economically devastated.

“the nice thing is like you know a fool is easily parted with their money um and SoftBank and the Middle East are the most stupid investors in the world potentially um and so you know I actually I think they're fine I think They're fine. I think these are good deals”

Coal stocks in Ohio have appreciated 3x due to power shortages and energy constraints, with the speaker having pitched coal stock investments to clients based on power-scarcity driven valuations.

“Um for every one of these that's happening, there's so many projects that are failing, right? Like I've literally pitched a coal stock to my clients in Ohio and the stocks 3x because there's power issues. Like like it's it's crazy how um how power constrained America is, right?”