How Strong is U.S. AI Really?

The last few weeks saw another upheaval in the tech markets. The release of Kimi K3 seemed to shake confidence in the idea that Open AI and Anthropic had a lock on the best models for the foreseeable future. Kimi K3 and before that GLM 5.2 come close on some AI benchmarks, and you can get access to them at lower prices. Furthermore, they release the weights so large scale users possessing powerful hardware can host them locally, meaning the users pay nothing to any AI developer. Or users can pay third-party inference providers who host free models at low prices.

All those facts raise the question of how Claude or Chat GPT can justify their massive near-trillion dollar valuations. Can they pull off an IPO now? Even if the U.S. were to ban use of Chinese models would that protect the leading companies? There’s a whole world out there outside the USA after all.

What if the U.S. labs take part in a safety slowdown? Does that put them at risk of getting overtaken?

Free,(open weight), innovative Chinese AI appears to be catching up to U.S. AI model capability, which seems like a bad sign for prospective Anthropic and Open AI shareholders. It’s also a bad sign for current investors in most parts of the AI economy. Huge investments go into developing and refining AI models, but the Chinese makers do it cheaper. That would seem to show the huge investments represent overspending. The Chinese models typically cost less to use too. It would seem like the U.S. models are doomed to fall behind.

The actual situation isn’t so simple. The question of whether Chinese AI constitutes a threat to U.S. AI dominance hinges on how AI model development works (something we’re all still figuring out.) The idea that Chinese models (Qwen, GLM, Kimi K, Deepseek) could over-take Claude, Chat-GPT and Gemini soon in terms of capability is probably incorrect. Probably. It could be a trillion-dollar question.

Really there are two questions:

  1. How likely are the Chinese models to overtake U.S. leading models in capability? I’ll explore this more in the next part.
  2. How likely are Chinese models to take away significant amounts of business? I don’t know, but have some thoughts.

If who has the leading models matters the most , U.S. companies should be safe for a while. But if access to the next best model costs 25% as much and 90% of customers don’t actually need the leading models most of the time, U.S. AI is in trouble financially. It’s not that they can’t figure out how to serve up AI at Chinese prices (outside of subsidies,) it’s that selling tokens (inference) at cost squeezes their ability to demonstrate they have a viable business.

Competing for Users

Businesswise, competitive pricing is certainly a vulnerability. Once six-month old models perform well enough for 99% of users it may not matter to them that they don’t use the latest and greatest when good enough costs 25% as much. A lot more happens in six months of AI development than it used to, and maybe the leading models will stay attractive to the most valuable customers. It’s too early to say.

Still, if Chinese models capture 75% of the useage by what is essentially dumping they can cut the legs out from under Open AI and Anthropic. Google is pretty safe since most of their value lies elsewhere.

The shift away from U.S. AI in favor of Chinese models is well underway. The share of tokens on Open Router from 2025 to mid-2026 went from 4.5% to 58%

Tokens by country

That’s not a full picture. It only shows what Open Router customers do, it doesn’t necessarily mirror the whole market, though it’s a dramatic trend. The situation is changing rapidly.

On the other hand, not all tokens are equal. Some models (including Kimi K3) are notably less efficient token users, meaning it costs more tokens to accomplish a task on Kimi K3 than equivalent tasks done by Claude Fable 5.

On the third hand, every provider charges very different amounts for different model tokens. Claude’s Fable 5 is simultaneously the most expensive per token, but also the most efficient on tokens per task. To get a good measure you’d want to see tasks completed by model and token broker, which may never be available.

Inference and Model Training

Let’s clarify that there are two distinct costs involved: First, model development (training), and second, serving the AI model to users (called inference.) Chinese labs offer their models for low rates (inference) and seem to have developed their models more efficiently than the U.S. labs.

Anthropic says the inference part of their business is profitable (if you could somehow separate it financially from future model development.) U.S. companies ought to be able to stay competitive with others when competing on inference – it’s just a matter of keeping up withthe the latest techniques, and special purpose inference hardware will probably be available roughly equally. [1] If competing on model development costs, it’s a more complicated question which I go into next.

To understand if Chinese AI really poses an existential threat to U.S. AI business you have to know how they have developed their models so far, and how model development actually works generally. Just looking at Kimi K3 current capability next to Anthropic’s Fable 5 and viewing it as a race is misleading – how they got to where they are is apples and oranges.

How to Build an AI Model

Quick outline:

  • “Pre-training”:

Consists of ingesting vast amounts of high quality language. Communication like messaging, books, articles, FAQs, scientific papers and everything else. The technical challenge here is to consume the material efficiently at scale distributing the work.

This is where language gets encoded into tokens and relationships between words get formed and the models learn to predict what likely comes next. At this stage the model will complete both sides of a conversation or whatever you feed it, the model will try to continue.

  • Post-training
  1. Supervised fine-tuning

The model is taught to distinguish between two sides of a conversation. New tokens are added to represent sides of the conversation and the model can learn to complete only its side. This takes less compute than the pre-training phase.

You feed the model curated high quality examples either synthetic or human generated. At this stage you’re tuning the model to learn desirable behavior patterns for reasoning and conversation and problem-solving.

Success at this stage is known as achieving a “cold start.” Now you can throw as much re-enforcement learning at the model and it can actually make effective use of the additional training.

  1. Re-enforcement learning

This is where the model is trained to reach goals. It can learn from example scenarios or from human feedback. To start with you feed the unformed model thousands or millions of examples and problem solving scenarios.

In scenarios like coding challenges where the solution can be automatically checked (the solution compiles and produces a correct output,) the training process records many attempts by the model to solve the challenges. These are called “trajectories.” This is incredibly useful for teaching coding and general computer use.

You grade the results either by another model or a human, or pass / fail if there’s a specifically correct answer. The model learns by scoring well, but the scoring system has to be designed carefully so the models don’t cheat to obtain a good score. It’s possible to score the training with another LLM.

Tradjectories now are pretty complex as the models get more and more capable. Running enough challenging ones to advance the model abilities costs a lot. Thousands will produce no successes. To improve a model in a reasonable timeframe, thousands of instances have to run at once, which takes a lot of the best hardware. Rather than simple questions, some tradjectories now resemble full software projects.

On top of all this, you want to make sure it interacts well with human users, so you do RLHF (re-enforcement learning with human feedback.) This will take a while and thousands of human trainers.

  • Distillation:

Borrow successful examples of problem solving from already trained models. You don’t need to spend time calibrating the rewards and running hundreds of thousands of failed tradjectories.

Instead of most supervised fine-tuning you can use an already trained model to provide example reactions to your model too. In other words, you don’t need as much human feedback; it’s already embedded in the fully trained model that required human feedback.

The re-enforcement learning phase could be shortened if you can work from the most effective tradjectories.

Distilling vs original training

There’s strong evidence most of the Chinese models were distilled from U.S. models, mainly Claude and Chat-GPT. An overwhelming portion of online public discussion around that claim centers on moral or legal angles, but those don’t really matter. Debating if Anthropic or Open AI or Moonshot AI (Kimi K) are more in the wrong is a distraction – deliberately, I think. [2] What matters is (1) if it’s true (it is,) and (2) once you have an accurate understanding of the situation, what are the technical and financial implications?

During distillation you’re extracting information created during the laborious post-training phase. Think of the outputs of an AI like a function: If you get enough examples of inputs and outputs you can work out an approximation of that function, at least for inputs you care about. You capture hundreds of thousands of x = f_chatgpt(y) for a good variety of y.

By skipping the expensive post-training, especially the human feedback part which takes a good deal of calendar time you can release models much faster and cheaper than a developer that has to do all the original work. The RLHF stage takes a lot of people a lot of time. It’s harder to just staff up super fast, compared to provisioning a bunch of servers for a week to distill that same RLHF from another model. The human feedback is a crucial step that can’t be rushed too much by the original developer. A developer distilling a model downstream of the original can purchase temporary compute to quickly extract the value the post-training added to the model.

Part of the motivation for this approach had to be that the most advanced Nvidia GPUs were banned for export to China and they had to make do with less powerful processors. That put even more pressure on Deepseek to not only skip RLHF but also use less compute overall. Lacking the most powerful GPUs, Chinese developers had to innovate hard to squeeze out as much training and inference as they could with what they had.

Just how much time and money distillation saves is an impirical question, but we could roughtly estimate it could be up to 50% of the cost and more than that of the total time to finish building their model. Getting access to vast amounts of interactions with the leading models is interesting but could fairly be called cheating.

The legal or moral high ground doesn’t matter much. If Chinese AI companies have “cheated” to catch up, there’s not much U.S. companies or the government can do to stop them; if the U.S. “unfairly” blocks use of Chinese models domestically over national security concerns rather than IP or trade secret theft claims, it won’t stop Chinese AI companies. Since China already blocks Gemini, ChatGPT and Claude, the U.S. labs don’t stand to lose that business in retaliation – they’re already blocked. What China can do is to keep releasing open weight models. If the U.S. closed model AI gets too expensive, users in the U.S. and around the world can use their own hardware or go through third party hosts of the open-weight models.

How they do it

While Claude is banned in China, there are re-sellers or “relays” Chinese users can pay that have collected ssubscriptions and pool them for their users. To make even more money the relays can copy the transcripts and sell them to the Chinese labs to use in post-training. Here’s a recent example. Operators at an AI lab can of course also use these relays to intentionally submit training scripts.

The scheme is alleged to be fueled with credit card theft and gift card theft to save money and make the number of users appear higher and geographically random. Anthropic claimed Alibaba (Qwen) alone created around 25,000 fraudulant accounts to make 28 million interactions to train from.

The other major Chinese labs Moonshot, Minimax, and Deepseek are also using this approach.

Leading AI Labs set the Pace

Although distilling allows you to skip a lot of slow and expensive process, you’re never adding your own original post-training ideas, only distilling an approximation of the source AI. Someone actually has to pay for the original post-training to get fundamentally new capabilities on top of a base model.

So, the fact that several Chinese models approach the newest Claude or Chatgpt capability may not present much of a threat to a monopoly on the SOTA (state of the art.) If skipping original post-training is a major contributor to the Chinese AI development speed, they aren’t going to surpass U.S. SOTA models.

The question remains if most customers require that extra edge, but in national security terms the U.S. could still expect to stay in the lead. Furthermore if the Chinese labs mainly “draft” off the U.S. leading labs, in principle a deliberate slowdown on development would slow down the Chinese AI lab progress too.

Cold Start to Recursive Self-Improvement

Even when distillation provides a cold-start, it’s not enough to sustain future development. What actually matters is if the Chinese models can achieve recursive self-improvement or not.

Recursive self improvement simply means the ability for an AI system (it would have to be more than a mere model) to get powerful enough to improve itself without the bootstrap phase provided by humans. This would probably look like AI providing substantial help in developing a successor model, and that successor becoming even more capable assisting in the development of its successor, and so on.

To go from close behind to ahead of the frontier models would require either an independent full scale post-training operation, or effective recursive self improvement.

If recursive self improvement really is a thing in the near-tterm, even blocking (somehow) Chinese AI makers from U.S. model output won’t do much for long.

Summing up

Not only could cheaper models take a lot of the market, if they can reach a critical threshold where they can do recursive self improvement, they can achieve parity with SOTA U.S. models.

  1. There’s alternative specialized hardware being developed for LLM inference. It’s expensive and limited in quantity currently but no reason to think it will only be available in one country Cerebras has very fast (up to 15x faster) processors. This hardware favors high throughput and low latency. It’s currently expensive compared to generic GPUs, but has a quality all its own. GPUs were general purpose LLM chips useful for training and inference. However, training can tolerate large latency and favors large batch jobs, where inference is just the opposite. Google has developed their TPU (Tensor Processing Unit) for inference, and Amazon has their own custom hardware as well. It’s a competitive space. More AI Inference Hardware Guide.

  2. There are hundreds upon hundreds of comments in social media and forums of the form “Anthropic stole thousands of books and information in general to make their model, so what’s wrong with Chinese AI stealing from them?”. It’s a deceptive nihilistic tactic since obviously two things can be wrong at once. Plus, Anthropic and Open AI have actually settled and compensated authors a lot and now have policies that they at least plausibly can claim sticking to “fair use.” It’s certainly not all settled law yet though. Still, the attacks count on readers being uninformed.

The relentless attacks on Anthropic (now that they have eclipsed Open AI) are an obvious campaign to demoralize Anthropic users and investors. That’s it. I don’t know if it’s orchestrated by Open AI, or Chinese companies or the Chinese government or what, but you can see the aim clearly enough. Spend a little time looking at who posts the comments and their similarity and it’s clear.

The main things the campaign do are (1) change the subject of if the Chinese labs deserve sanctions / consequences (“What about Anthropic’s stealing?”) and (2) give the impression everyone is fleeing U.S. AI and demoralize their users.

By flooding social media with this sentiment they steer discussions away from people caring about the actually interesting questions – if Chinese AI developers are taking distillation shortcuts and why that matters economically and technically. There’s plenty of ill-will against the big AI companies and it’s easy to harness it to derail discussion.

The question of if it’s okay to read the whole internet and most books and create a art / music / text generation machine that overwhelms all the output of actual creative humans is something else and to be clear I think it’s bad.

There are also a lot of comments praising the services that provide access to Chinese models, and open weight policies. They say the price for GLM 5.2 or Kimi K3 are way more affordable than Claude or Chat GPT. That may be (it’s complicated) but what they don’t point out is that these prices are probably even more heavily subsidized than Chat GPT or Claude pricing. It can’t go on forever. That’s fine, but the tone of the comments implies the U.S. providers are ripping off their users and you’re a chump to pay full U.S. prices.

Whatever else you think of the U.S. AI companies, clearly it isn’t the case that they are extracting huge profits from their users. The prices are barely enough to pay for inference directly, and not enough to cover all its externalities. They’re burning cash. Certainly not enough to underwrite future model development – that’s mostly done with private investment still. For those Chinese models that are open weight the services selling inference don’t have to pay for any of the model development. And for the developers selling their service directly, the original development didn’t cost nearly as much if they distilled the U.S. models.