Welcome back! This is an article I’ve wanted to write for a while now, exclusively focused on clearing up some of the most prevalent myths in this space, especially regarding China.

With the world suddenly realizing that AI is expensive, the “battle” between the US and China for AI supremacy is hotter than ever. If the tide turns in favor of Chinese open models, the trillions of dollars at stake could be at clear risk.

But as always, the picture is much more nuanced, and several of your beliefs right now that you have been told are simply blatantly false.

Let’s dive in.

Ironically, the first myth that needs to be debunked for good is the idea that China is “catching up.” The truth is that, despite all the hard evidence you may be seeing, it’s simply not true.

But here’s the thing. As you’re about to witness, it might not matter at all.

The first myth is this graph below, claiming that Chinese models (they refer to open-weight models, models that are free to download, but most of the best ones are Chinese) are only 4 months behind.

The data seems hard to dispute; Kimi K2.6 seems to be at the level the US was at with GPT-5.3 Codex in December. However, it’s false.

And the issue is precisely drawing that conclusion based on model benchmarks and not on the product.

And let me tell you that nobody more than I would want to see open-source catching up to proprietary solutions. However, I’m not here to expose you to my desires, but to the truth.

The reason is quite simple. In the absence of fundamental breakthroughs in algorithms that do not exist today, AI progress is driven by two scaling laws:

And the US is enormously ahead in both.

On the former, US Labs have larger models (in the order of ten times the size) and perhaps even two orders of magnitude larger training budgets.

The largest known AI cluster in China is a 10,000-Ascend 910C cluster in Shenzhen with up to 11,000 PetaFLOPs of compute, or 11 ExaFLOPs. That number doesn’t say much by itself, but if we assume it is FP8, which most likely it is, it’s a fourth of the compute of a single Google Ironwood 9,216-TPU pod at 42 ExaFLOPs.

And that isn’t even Google’s most powerful server, and they’ve scaled that number to a whopping 121 PetaFLOPs with the upcoming TPUv8 pods for training using FP4 precision.

In apples to apples, that is 121 vs 22 ExaFLOPs, 6 times more, for approximately the same number of chips, 9,600. That means a cluster with roughly the same number of chips has 11 times the compute potential.

I mean, a single NVIDIA Vera Rubin NVL72 server, recently deployed for the first time, has almost 60% higher compute capacity. One single server.

In the meantime, two US models have already crossed the 10²⁷ FLOP training budget barrier, both Gemini 3.1 Pro and Claude Mythos, which, by the way, just got released minutes ago.

For reference, that is about 100 times as much compute as was used to train GPT-4 (i.e., for that budget, you could have trained 100 GPT-4s).

This is a massive disadvantage for China; it’s just is. On the second scaling law, you need a lot of inference compute to be able to serve long sequences that increase performance, compute that again China doesn’t have (relative to the US).

On the topic of training budgets, how is it that China can “stay close” with budgets 100 times smaller?

The answer is distillation, which massively reduces your training requirements. That is, by training your models on data from larger US models, you artificially close the gap with them very effectively and with much less compute, because you’re basically teaching your model to imitate the larger model.

But distillation and its savings come at a cost that is not appreciable in the benchmarks: generalization.

The fact that you trained a model on 10-100 times less compute means your model is extremely unlikely to be as good as the former outside the prioritized domains, something that benchmarks hide really well.

You can purposefully train models to do well on benchmarks, but if you test those models outside that domain, you’re going to see the cracks pretty fast. But that doesn’t show in the benchmarks, which means you appear to be better than you really are.

The other problem with benchmarks is that they hide reality from us; you aren’t seeing the actual product, just the models.

But what do I mean by that?

When you’re testing for a benchmark, you’re doing so under perfect constraints. Probably batch 1 (meaning you’re serving the model in ideal conditions), with little regard to latency, and with huge compute budgets; you’re allocating maybe thousands of dollars to every task and tens of thousands to the overall benchmark in order to do the best you can.

But that’s not the reality your model will face in the real world.

The real world is not an ideal lab environment. In the real world, models need to fight for scarce compute, run in high batches (slower responses), and, crucially, operate with limited thinking budgets so as not to bankrupt the serving company.

This means that the product, the actual deliverable that customers get, is NOT what the benchmark shows (and to be clear, this applies to US Labs too, just as much, it’s just that they have more compute).

In real-time inference, Labs will quantize models, drop thinking budgets to meet demand, serve you distillations of the real thing, and others.

All this shows that the real experience is not what benchmarks show. And as you can imply from all that I’m saying, for now, this is a compute game all the way.

Therefore, the point here is that it’s not surprising that Chinese Labs are becoming more frugal and thus innovating at the intelligence-per-dollar level; they really have little option.

But if they could compete at the Frontier with equal resources, they would behave exactly like US Labs are behaving (Alibaba planning to 10x its compute, Zhipu co-founder acknowledging the gap could be widening).

But then, how do we compare both?

An actual fair comparison would be product benchmarks that measure model performance within their products and serve them to real users. Because when you do that, you realize that the user experience is like night and day between American and Chinese products.

The only thing that could change this is a fundamental breakthrough that proves that models can be trained and served with less compute. The closest thing we have is Sapient Intelligence’s HRM-Text model.

This is a billion-parameter model that I discussed in detail here recently. You can run it on your smartphone, and it proved to be quite decent and superior to GPT-3.5, despite requiring ~700x fewer training tokens and an estimated 44,000x fewer training FLOPs (operations), as you can see in the graphs below.

Interestingly, GPT-3.5 was the model ChatGPT used when it first launched in late 2022, and it was the state-of-the-art at the time.

Sadly, however, while promising, this is not a particularly useful model, and we need this to scale to more useful capability thresholds. If we saw a model that can run on an iPhone having the performance of the Chinese frontier today, well, that would be a complete revolution. However, that’s simply not a reality today.

But what about Chinese models served on US LLM providers? Are those a threat to US Labs?

Well, that’s a completely different story.

A few days ago, the CEO of LindyAI, a chatbot product that helps users increase productivity with AI models and has millions of users, switched its underlying models from Anthropic’s Claude to DeepSeek.

And not only are they saving millions of dollars, but they claim an actual increase in performance.

And with companies like Uber saying “enough is enough” about uncontrolled AI spending, many are waking up to something that was obvious in hindsight: what really matters is intelligence per dollar, and thinking that everyone would always prefer the frontier model was unequivocally shortsighted.

Currently, the prevailing sentiment is that Chinese models are cheaper and represent the Pareto frontier in terms of intelligence per cost, which suggests they are being widely adopted as we speak.

But are they? Well, like the previous question, yes and no.

The reality is that we already have plenty of evidence that open AI models are being adopted by US and European customers.

Perhaps the best example of this is OpenRouter, a highly popular US LLM router that lets you access most models from a single place.

Another good example of increased open model usage is provided by Ramp, a company expense management company, showing that DeepSeek is becoming extremely popular amongst its enterprise clients.

That looks like decent adoption, but the reality is that private tokens, those generated by Google/OpenAI/Anthropic/SpaceX, are several orders of magnitude larger. Nonetheless, Google’s latest figure is over 3.2 quadrillion tokens per month across its AI surfaces.

That is 3,200,000,000,000,000+ tokens/month.

OpenRouter processes 34 trillion tokens per week, or 147T per month. That means Google alone processes 21 times as many tokens per month as OpenRouter, across all models it serves (including proprietary models), signaling an enormous gap between private and open models.

But Lindy’s transition from Anthropic to DeepSeek could be the start of the “Huge Transition”, where the decision-making algorithm for choosing a model is not “let’s use OpenAI” but “let’s try the open models first, and if they don’t work, then we look at Anthropic”.

All this could smell like a huge opportunity for China, considering the popular belief that Chinese models are unequivocally cheaper and offer better value for your buck. At this point, I believe this is widely accepted.

However, my issue with this entire conversation is that it is discussed from the wrong perspective.

The issue is that the statement “Chinese models offer 80% of performance per 20% the price” is true but incomplete, because the last part is missing: “but so are US open models”.

Examples include the recently announced NVIDIA Nemotron 3 Ultra. Despite being half the size of Kimi K2.6 and three times smaller than DeepSeek v4 Pro (both Chinese models), it offers much better Pareto performance, is much faster, and is also cheaper.

And if we factor in closed models, like SpaceX’s Grok 4.3 model, it’s also clearly on the Pareto curve, offering a better price per token than Kimi K2.6.

Therefore, there’s very little evidence that Chinese models are cheaper. Conversely, I also want to use this moment to debunk a myth that China subsidizes inference costs. It doesn’t, and it’s easily provable in two ways:

Therefore, what investors in OpenAI and Anthropic should be scared of is not Chinese clouds offering competitive models, because inference has to be served from nearby due to latency constraints. The threat is US inference providers, companies like the Hyperscalers or neoclouds like Fireworks, offering Chinese models on US soil while offering equally competitive pricing.

So far, it doesn’t seem like China is ahead on anything. They are pretty good at training cheap models, but so is the US, and the benefits of these models are still being ripped by US companies.

So, where’s the issue then?

Well, there’s a place in this industry where China is ahead of the US, and it’s not even close. An area where China is actually a threat.

It has nothing to do with AI. It has nothing to do with talent. And it has nothing to do with a “secret sauce” nobody knows about.