American Enterprise Institute essay framed computing power as the new geopolitical high ground, emphasizing AI infrastructure dominance in US-China competition.

Raw computing power—"compute" in industry parlance—is rapidly becoming the critical bottleneck in global AI development, and China has taken note.

Since February, I've operated an AI analytics platform called "Digital Embassy," a hybrid intelligence service that synthesizes daily news from 22 world capitals in the style of the CIA's Open Source Enterprise. Maintaining it has grown prohibitively expensive. I process approximately 10 million tokens daily—8 million words—costing more than $1,000 monthly to use Anthropic's AI models for analysis and reporting that previously required a hundred government analysts.

Seeking to reduce costs, I reached the same conclusion as many software developers: the freedom to code requires the freedom to compute. Earlier this week, I set out to purchase a computer capable of running open-weight AI models locally. I wanted to run mid-sized models from US companies like Google and Meta, or Chinese labs like DeepSeek and Qwen—not be confined to small systems with low throughput.

A sophisticated machine was necessary. I saved $8,000 for a top-of-the-line Mac Studio with a latest-generation M3 Ultra chip and 256 gigabytes of RAM—sufficient to run mid-sized 80-billion-parameter models. But when I attempted to purchase this morning, Apple's website blocked the order. As of this weekend, they no longer manufacture them. The South Korean factories producing the Mac Studio's high-bandwidth memory are sold out. No Mac Studios with M3 Ultra chips remain available for purchase anywhere globally.

This episode illustrates the harsh reality of compute supply chains in 2026: America's third-largest company no longer finds it worthwhile to produce an $8,000 personal computer for a "low-end" customer like myself.

The global compute shortage's implications extend far beyond personal computing. A durable moat is being erected around computational access—one that affects China's AI developers as well.

Last week, Anthropic's release of Claude Mythos startled the world by discovering novel cybersecurity vulnerabilities in nearly every major web browser and operating system. This "cyber wonder weapon" triggered panic among senior Trump administration officials, who convened emergency meetings with technology and financial institution CEOs to coordinate response.

The question dominating Washington and San Francisco conversations—"What happens when Chinese open-source models catch up?"—is fundamentally misguided. There is compelling reason to believe superintelligence will remain gated behind vast computational resources.

The more probable trajectory is a marked bifurcation in the AI value chain. Chinese labs will capture increasing global market share at the low end—edge devices people already own or can afford—while the most powerful systems remain locked behind an increasingly formidable compute moat at the top.

Chinese AI labs currently lead at the low end. Sparse-attention models from DeepSeek, Qwen, and Moonshot run locally on edge devices and gain increasing popularity with global developers, eroding American AI labs' margins. China's 15th Five-Year Plan explicitly focuses on AI adoption, encouraging citizens to use personal agents like OpenClaw to create "one-person companies" that will drive economic development and address youth unemployment.

However, today's commercially available AI systems—powerful though they are—have not threatened the Chinese Communist Party's legitimacy or its monopoly on violence. Mythos-class models could. A system autonomously generating thousands of zero-day exploits in legacy software represents a qualitatively different capability than the chatbots driving China's "AI Plus" economy, first articulated by Premier Li Qiang in 2024.

The capability gap between these two tiers is defined almost entirely by computational access. Vulnerability-hunting operations like Anthropic's Project Glasswing cannot be performed by individuals or most organizations. It demands massive parallelization across enormous clusters spanning hundreds or thousands of NVIDIA's latest B200 GPUs—a capital investment in the billions, deployed over years.

Beyond GPU sticker prices, Anthropic's achievement required the equivalent of months of industrial capacity from Taiwanese logic-chip factories, South Korean high-bandwidth memory factories, and Japanese silicon-wafer factories.

If Mythos contains 5 to 10 trillion parameters, running a single instance requires 25 to 50 of NVIDIA's latest B200 GPUs (each with 192 gigabytes of high-bandwidth memory) to store model weights. For Project Glasswing's vulnerability work, Anthropic reportedly runs 10,000 model copies in parallel. This likely consumes 250,000 to 500,000 GPUs—potentially 10 percent of all Blackwell chips NVIDIA produced in 2025, worth $5 to $20 billion in aggregate.

Compute production at this scale is verifiably scarce—a substantial problem for China, whose domestic AI chip production represents only 1 to 2 percent of total US output. Huawei was projected to produce just 800,000 AI chips in 2025, none approaching NVIDIA's quality. The UAE is permitted to import up to 500,000 B200-class chips annually. A Chinese equivalent to Project Glasswing launched in spring 2026 would consume nearly all Huawei's production indefinitely.

US export controls on advanced semiconductors matter more now than at any point since their 2022 inception. These controls constrain the quantity and quality of compute available to Chinese military and intelligence services while reinforcing structural conditions that impede Beijing's distribution of frontier AI capabilities widely.

Loosening these restrictions—as the Trump administration attempted with its December 2025 decision permitting H200 exports to China—narrows the compute moat currently separating consumer-grade AI from Anthropic's capability to generate exploits in every known operating system and web browser.

2026 is the year the compute moat becomes visible to all—and when China will almost certainly reconsider its pure-play open-source strategy. Washington should not ease Beijing's path to close this gap at the precise moment it matters most.

The United States must compete where possible, regulate where necessary, and move with sufficient speed to maintain its advantage.