The US is winning the AI race where it matters most

avkcode's thesis is that the AI race has a wrong scorecard problem. Papers, engineer counts, benchmark wins are not what's deciding it. What's deciding it is who can finance infrastructure, train and serve models at scale, and apply AI across the economy. By that measure the US is well ahead.

What changed in 2025

Since DeepSeek R1 in January 2025, US companies moved faster on commercialization. OpenAI pushed harder into agents and Codex. Anthropic turned Claude Code into a paying business. China has serious contenders, but the gap shows up in revenue, adoption, tools, and reach.

DeepSeek matters for a different reason. Its strategic value for China is reducing Nvidia dependence and pushing inference toward domestic stacks like Huawei Ascend. That's supply-chain autonomy, not the same thing as profitable AI leadership.

The electricity layer

avk includes a useful price table — household and business electricity in USD/kWh:

The US is cheaper than the major Western European economies. China and Russia are cheaper than the US. Canada is cheaper still. Power matters — turning electrons into compute is the whole game — but power alone doesn't decide it.

The decisive layer is cloud and data

The US owns the global hyperscalers. AWS, Azure, GCP are the channels models reach the world through. The US also owns the data platforms that generate and organize the corpus: YouTube as a video corpus, Google Drive and Microsoft 365 inside daily office work, GitHub inside software development. New models get pushed into products people already use every day. China has much of this domestically. Europe doesn't.

This is why cheap power alone doesn't decide. A country can have cheap electricity and still lose if it lacks cloud scale, platform reach, developer ecosystems, and access to large flows of useful data. The US has all of them at once.

Christian Klein of SAP has argued Europe doesn't need more data centers. avk agrees on the narrow point that LLMs alone aren't enough, but the broader lesson is the opposite of Klein's: data centers matter as part of a larger system. Europe spent ~$58.8B on Indian software services in FY 2023–24 and ~$67.1B the next year. Even if Europe decided today to finance real cloud champions, building infrastructure would only be step one — moving banks, manufacturers, and public agencies onto those platforms is a decade-scale migration during which the hyperscalers keep extending the lead.

The one exception cited: Arkady Volozh trying to build Nebius into a European AI infrastructure company. That confirms the rule — Europe is still at the start.

The weaponized-AI frontier

The next phase, avk argues, may be country-AI vs country-AI in bot networks, cyber campaigns, autonomous weapons. A provider doesn't need magic to do this. Tuning systems to dehumanize rivals, justify violence, or target populations is disturbingly straightforward.

Models like Anthropic's Mythos point to a second-order shift. The old Linux instinct was many eyes on open code. Frontier cyber models may push states and defense firms the other way — closed software, closed tooling, closed firmware, closed chips. If a model can't train on the code and architecture of a target stack, it has less context and less speed. That doesn't make systems safe but it raises the value of proprietary stacks all the way down to hardware.

Where this lands

avk's framing complements the financial-collapse case in subprime-ai-crisis and ai-too-expensive — the same physical infrastructure that makes US commercialization possible is also the bet that has to pay off for the unit economics to work. The two views aren't in opposition. Zitron's case is "the bill comes due"; avk's is "the build is real, and other regions don't have it." Both can be true simultaneously, and if the bubble does pop, the question of who owns the pieces afterward is what avk is foregrounding.

The pieces sit alongside ai-bubble-pale-horses as another reading of the same physical reality from a different angle.