I wanted to buy American artificial intelligence. That sentence should not need explaining. My company builds and tests AI systems for a living. A new frontier model comes out, and I want to run it through the same evaluation harness I use for everything else. I don’t need unlimited access. I don’t need another giant monthly plan. I need twenty runs, maybe fifty, maybe a hundred.
Give me fifty dollars of the model. Let me use it. Charge me for what I use. That should be easy. It isn’t.
American AI companies sell their best intelligence through a maze of subscriptions, tiers, rate limits, and product gates. A business can already be paying for capacity it never touches, and still find that the one model it wants to test sits behind another subscription, another tier, another wall. Meanwhile the Chinese competitor is standing there with an API key, ready to go. That’s not a minor inconvenience. It’s becoming an industrial policy problem.
The Price Gap Is Real
Look at two models on the market right now. OpenAI’s GPT-5.6 Luna is a genuinely good model, built for fast, high-volume work with a million-token context window. It runs $0.20 per million input tokens and $1.20 per million output tokens.
Z.ai‘s GLM-5.3 Flash also offers a million-token context window. It’s a multimodal model built for coding and long agentic tasks. Its promotional price: $0.075 per million input tokens, $0.25 per million output.
Do the math. GLM comes in roughly 2.7 times cheaper on input and nearly five times cheaper on output. Other model comparisons show gaps even wider than that. A few cents don’t sound like much when you’re asking a chatbot what to make for dinner. Businesses don’t use AI that way.
An agent can make dozens of calls to finish one job. A coding system can chew through millions of tokens reading a codebase. A retrieval pipeline feeds documents into a model over and over. An evaluation harness runs the same task fifty or a hundred times against competing models. One ordinary interaction with a locally hosted model in my own lab recently burned through more than 6,000 input tokens. Multiply that by thousands of interactions, then millions, then billions. Pennies become dollars. Dollars become budgets. Budgets become architecture. That’s where China’s pricing edge turns strategic.
Good Enough Beats Better
American AI companies seem to believe they’re competing on intelligence. They are competing on intelligence. Businesses, though, don’t buy intelligence in a vacuum. They buy capability, reliability, price, availability, integration, support, security, and convenience, all bundled together. The winning model doesn’t have to be the smartest model. It has to be smart enough.
A Chinese model that delivers 95 percent of the useful capability of an American model at a quarter of the price isn’t an inferior product by default. Sometimes it’s the smarter business decision. That gets even truer when the cheaper model is multimodal, has a long context window, sits behind an ordinary API, shows up on aggregators, and can sometimes be downloaded and run at home.
Today’s Chinese AI ecosystem increasingly combines cheap hosted inference, model aggregators, downloadable weights, self-hosting, and familiar API conventions. All of that gives developers many ways to try a model and deploy it without first signing up for a big recurring commitment. American AI companies may be misreading the market they’re in.
Price Is Distribution
Every technology platform wants developers. Microsoft knew it. Apple knew it. Amazon knew it. Google knew it. Developers build ecosystems. It usually starts small. A developer tries a technology. She writes something around it. She tunes her prompts to it. The benchmarks pile up. The software starts depending on how that particular model behaves. Someone documents the integration. Someone else learns to maintain it. Before long, switching models isn’t a five-minute decision anymore.
AI adoption runs on a sequence: access, experimentation, evaluation, integration, optimization, dependency. Price shapes the first two stages, which makes price more than a revenue decision. Price is distribution. Cheap inference buys experiments. Experiments buy integrations. Integrations buy ecosystems.
China doesn’t need every American business to conclude that GLM, DeepSeek, Qwen, or Kimi beats OpenAI, Anthropic, or Google on merit. China needs those models inside the evaluation set. A model can’t win a benchmark it never enters.
China May Be Playing a Different Game
Here’s the question American policymakers and tech executives should be asking. What if Chinese AI companies aren’t trying to maximize profit on inference right now?
Nobody outside those companies knows for certain. Several explanations are plausible, and probably more than one is true at once. Chinese firms may have genuinely lower costs. They may have found more efficient architectures. Investors may be subsidizing growth in exchange for market share. Chinese industrial policy may treat AI as strategic infrastructure rather than another software line, which makes losses easier to tolerate.
The point is that the objectives might not match. An American AI company answers to investors asking about margins, recurring revenue, and capital efficiency. A strategically backed Chinese competitor can rationally weight developer adoption, global distribution, and technical standards over near-term unit profit.
Those are different games. America may be playing this as a SaaS market. China may be playing it as an industrial contest. That wouldn’t be a new story. Governments and corporations have absorbed years of losses to build railroads, telecommunications, semiconductors, cloud platforms, and solar manufacturing. Artificial intelligence looks like it could be next.
The question isn’t whether a Chinese company profits on today’s million tokens. It’s what becomes valuable once millions of developers have built their tools around those tokens.
Subscriptions Aren’t the Villain or Are they
American AI companies aren’t behaving irrationally. Subscriptions bring in predictable revenue. They make purchasing simple for ordinary users. They help manage scarce compute. They reduce abuse. They fund the interfaces, storage, memory, connectors, and support that make a product usable. Heavy reasoning systems create wildly uneven costs across users, and flat unlimited access would break the math fast.
None of that explains why a paying business customer can’t simply buy more compute directly. That’s the actual mistake. I don’t object to paying more for an American model. I object to being forced to buy the wrong thing.
Give Businesses a Bucket
The fix is almost embarrassingly simple. Keep the subscription. Sell the application, interface, storage, memory, and connectors for twenty, thirty, or fifty dollars a month, same as now.
Add a compute wallet on top of that. Let me put in fifty dollars. Or a hundred. Or five hundred. Let my company decide where it goes. A routine task costs little. A frontier model costs more. Maximum reasoning costs more still. Let me cap an experiment at ten dollars. Let me greenlight a hundred for a benchmarking project. Let an administrator decide which models employees can touch. Let a consultant bill inference straight to a client. Let a research group draw compute against a grant. When the money runs out, it stops.
None of this is technologically radical. Pieces of it already exist. OpenRouter sells dollar-denominated credits across hundreds of models. Mistral mixes plan-based access with pay-as-you-go consumption. Anthropic and OpenAI already run prepaid credit systems in places. What’s missing is mostly packaging.
Sell me the software as a subscription. Sell me the intelligence as a resource.
Washington Should Be Paying Closer Attention
This gets stranger next to Washington’s own worries about Chinese technology. Federal policymakers are increasingly focused on model provenance, data security, and dependence on foreign AI systems, especially in government and defense-adjacent work. Those rules are a patchwork. Chinese-origin models aren’t banned across the board from private-sector use, and restrictions shift by contract, agency, and workload.
The contradiction is still worth naming. Washington worries about long-term dependence on Chinese AI. American developers keep bumping into Chinese models that are dramatically cheaper and far easier to test. American AI companies shouldn’t be the ones helping build that outcome.
The danger isn’t that American businesses wake up one day wanting Chinese AI over American AI. The danger is that we’re making the Chinese version easier to buy.
This Is How Markets Get Captured
Technology leadership was never decided purely by the best benchmark score. Markets get won through distribution. Through price. Through availability. They get won the moment a developer with an idea at eleven at night can enter a card number and start building at 11:05, no permission required.
American AI companies have real advantages: talented researchers, enormous infrastructure, strong brands, and the largest commercial AI market on earth. They can and should charge a premium for that. Businesses already pay premiums for security, reliability, support, and reduced geopolitical risk.
Premium pricing isn’t the problem. Friction is.
I wanted to run the American model. I was ready to pay more for it. I just wanted to buy the amount my business actually needed. China’s companies were more than happy to sell it to me that way. That should worry more people than the executives designing subscription tiers.
America may still have the better model. China doesn’t have to build a better one to gain ground. It just has to make its model cheap enough, capable enough, and easy enough to try that developers keep coming back. Some of those experiments become products. Some of those products become infrastructure. Some of that infrastructure becomes dependence. American AI companies should think hard before they keep helping that happen.
We say we want American businesses buying American AI. Maybe we should start by letting them.