In the Battle for AI Dominance, Computing Power Is the New High Ground

Ryan Fedasiuk

Raw computer processing power—known within the industry as “compute”—is quickly becoming the bottleneck in global AI development, and China has taken note.

Since February, I’ve run an AI analytics platform called “Digital Embassy.” Half-private intelligence service, half-public good, it synthesizes daily news wires from 22 world capitals in the style of the CIA’s Open Source Enterprise. It has also become very expensive to maintain. I’m processing around 10 million “tokens” each day—8 million words, amounting to more than $1,000 a month—for Anthropic’s AI models to do the thinking and writing that used to command the attention of a hundred government analysts.

Rather than continue spending through the nose, I’ve reached a similar conclusion as many other software developers: the freedom to code requires the freedom to compute. So, earlier this week, I went to go buy a new computer capable of running “open-weight” AI models myself. I didn’t want to be confined to running a “small” AI system with low-throughput rate—I wanted to run some of the best mid-sized models offered by US companies like Google and Meta, or Chinese labs like DeepSeek and Qwen.

I needed a sophisticated machine, not just any old laptop. So I scrimped and saved $8,000 to buy a top-of-the-line computer: a Mac Studio with a latest-generation M3 Ultra logic chip, with 256 gigabytes of RAM—plenty to hold and run a mid-sized model’s 80 billion parameters. I spent the past week getting my finances in order.

But this morning, when I finally went to click “purchase,” Apple’s website wouldn’t let me place the order. That’s because, as of this weekend, they aren’t making them anymore. The South Korean factories that produce the Mac Studio’s high-bandwidth memory are sold out. I couldn’t go to pick one up from an Apple store in Baltimore, Chevy Chase, or Arlington—or even San Francisco, New York, or Chicago. In fact, as best I can tell, there are no more Mac Studios with M3 Ultra chips available for purchase anywhere in the world.

This episode is a snapshot of the bleak reality of compute supply chains in 2026. It is simply no longer worthwhile for America’s third-largest company to produce an $8,000 personal computer for a “low-end” customer like me.

The implications of the global compute shortage extend far beyond my home computer setup. There is a durable moat being erected around access to computational power. China’s AI developers are feeling it, too.

Frontier AI Is Gated by Compute

Last week, the release of Anthropic’s Claude Mythos stunned the world with its capacity to discover novel cyber vulnerabilities in nearly every web browser and operating system. The development of such a “cyber wonder weapon” sparked a panic among senior officials in the Trump administration, who convened an emergency meeting with CEOs of tech companies and financial institutions to plan around its release.

But the question now dominating conversations in Washington and San Francisco—What happens when Chinese open-source models catch up?—is the wrong one. There is good reason to believe superintelligence will remain gated behind vast quantities of computational power.

The more likely trajectory is a marked split between the low and high ends of the AI value chain, with Chinese labs gaining global market share at the bottom—on edge devices people already own, or can still easily buy—while the most powerful systems are locked behind an increasingly formidable compute moat at the top.

At the low end, Chinese AI labs are winning over the world’s AI developers. Sparse-attention models engineered by DeepSeek, Qwen, and Moonshot can be run locally on edge devices. They are increasingly popular with global publics, which is already eating into American AI labs’ bottom lines. China’s 15th Five-Year Plan is squarely focused on bolstering AI adoption—encouraging netizens to use personal agents like OpenClaw to birth “one-person companies” (OPCs) that will accelerate China’s broader economic development and absorb some of its legions of unemployed youth.

But today’s commercially available AI systems—powerful as they are—have not threatened the legitimacy of the Chinese Communist Party or its monopoly on the instruments of violence. Mythos-class models could. A system capable of autonomously generating thousands of zero-day exploits in legacy software is a qualitatively different instrument from the kinds of chatbots that have motivated China’s AI strategy since Premier Li Qiang first articulated the concept of an “AI Plus” economy in 2024.

The gap between these two classes of capability is defined almost entirely by access to computational power. Running the kind of vulnerability-hunting that Anthropic’s Project Glasswing requires is not something any individual—or most organizations—can do. It demands massive parallelization across enormous compute clusters, spanning hundreds or thousands of NVIDIA’s latest B200 GPUs—a capital investment measured in billions of dollars, spent over several years.

The Billion-Dollar Math Behind Mythos

More than the sticker price of the GPUs used to train and run Mythos, it is important to understand that what Anthropic has achieved requires the equivalent of months of industrial capacity at the Taiwanese factories producing logic chips, the South Korean factories producing high-bandwidth memory, and the Japanese factories producing silicon wafers.

If Mythos is a 5 or 10 trillion-parameter AI system, then running just a single instance of the model would require somewhere between 25 and 50 of NVIDIA’s latest B200 GPUs (each sporting 192 GB of high-bandwidth memory) to store its model weights. To power its vulnerability-sleuthing “Project Glasswing,” Anthropic is reportedly running 10,000 copies of the model in parallel. That means the project is likely using somewhere between 250,000-500,000 GPUs—perhaps as many as 10 percent of all the Blackwell chips NVIDIA made in 2025, worth in aggregate of $5–20 billion.

Compute production at this scale is verifiably scarce, and it is a big problem for China—whose domestic AI chip production amounts to only 1–2 percent of total US output. For context, Huawei was slated to make just 800,000 AI chips in all of 2025—and they do not nearly approach NVIDIA’s quality. The UAE is reportedly permitted to import up to 500,000 B200-class chips per year. If it were launched in spring 2026, a Chinese “Project Glasswing” would occupy nearly all of Huawei’s production for the foreseeable future.

Custody of Compute Is More Important than Ever Before

This is precisely why US export controls on advanced semiconductors are more important now than at any point since their inception in 2022. Such controls constrain the quantity and quality of compute available to Chinese military and intelligence services, and they simultaneously reinforce the structural conditions that make it harder for Beijing to distribute frontier AI capabilities widely.

Loosening these restrictions—as the Trump administration attempted with its December 2025 decision to permit H200 exports to China—only narrows the compute moat that currently separates consumer-grade AI from the kind of capability that Anthropic has used to generate exploits in every known operating system and web browser.

2026 is the year that the compute moat will become visible to everyone—and the year China will almost certainly change its mind on pure-play open-source. Washington should not make it easier for Beijing to close the gap at the precise moment that gap matters most.

About the Author: Ryan Fedasiuk

Ryan Fedasiuk is a fellow at the American Enterprise Institute, where he focuses on US-China relations, technology, and national power. Concurrently, Mr. Fedasiuk is an adjunct assistant professor at Georgetown University. He was previously an advisor for US-China Bilateral Affairs at the State Department, where he served as the US government’s main point of contact with the Chinese Embassy in Washington.