DeepSeek V4 Shakes AI Market with Low-Cost Power

DeepSeek dropped V4 last week. Open source, open weights, frontier-level, and a fraction of the price of Opus 4.7 or GPT-5.5. This is a bigger deal than R1 was, and R1 dropped the stock market 20% basically overnight.

DeepSeek V4 Pro is a 1.6 trillion parameter mixture of experts (MoE) model with 49 billion active parameters and a 1 million token context length. V4 Flash is the workhorse, 284 billion total with 13 billion active. Both trained on about 33 trillion tokens. On the agentic coding benchmarks, MMLU Pro, GPQA Diamond, SWE-bench Verified, it's right up there with Opus 4.7 and GPT-5.5. A little behind, but just a little.

Most use cases don't require absolute frontier intelligence. The vast majority of companies aren't doing frontier scientific research or trying to crack the hardest coding problems in the world. They're running a business.

So imagine you're a CEO. You look at GPT-5.5 at $30 per million output tokens. Opus 4.7, similarly priced. Then you look at DeepSeek and it's a fraction of that. It does almost everything you actually need. It's open source, so you can fine-tune it, host it how you like, control it precisely. The calculus becomes really obvious. Why would you pay so much more?

That's where the problem comes in. Jensen's been saying China is going to build their own chips and their own models, so they might as well be built on American technology (NVIDIA chips). Fine. But the same argument now works in reverse. If US enterprise companies build their AI strategy on top of Chinese open source models, that's a big geopolitical security risk. If those Chinese AI Labs change their architecture or cut us off, we're suddenly in a really bad spot.

Export Controls - Are they working? Kind of yes, kind of no. Yes, because DeepSeek is openly compute-constrained. Their own paper says Pro service capacity is very limited until their supernodes scale up in the second half of this year. They literally can't serve this model the way they want to. But also no, because that constraint forced them to innovate on the algorithms side. They came up with algorithmic unlocks that let them train and run V4 at a fraction of the price, even with nerfed GPUs.

Distillation Hacking - Anthropic put out a report a few weeks ago, and just yesterday the US Government echoed it, saying foreign entities, primarily China, are running industrial-scale distillation campaigns to steal American AI. This is a real concern. But if you actually look at Anthropic's report, DeepSeek's number was 150,000 exchanges. Moonshot had 3.4 million. Minimax had 13 million. 150K is just not enough to explain the level of quality DeepSeek hit. And they open sourced the whole thing with a thorough whitepaper that explains exactly how they did it. It doesn't mesh.

Now think about the bigger picture. Trillions of dollars are pouring into AI in the United States. The fastest infrastructure buildout in history. That investment requires a return. If global enterprise demand routes around US closed-source models because Chinese open source is good enough and so much cheaper, that return doesn't show up. And we've bet the economy on this.

Then there's the cultural piece. Think about how social media changed the world, and social media came from the US initially. We were able to control the narrative in a lot of places. Now flip that. Imagine we're all built on Chinese models, and they're dictating what the models can and can't say. Yes, they are open source so we can change them, but Chinese cultural subtleties will remain in the models. These are big questions we're going to have to grapple with.

So where do we go from here? Two things. The US needs to go much harder on open source. The frontier labs in the US are not open source friendly, with the exception of Google, and even Google is building small open source models, not at the level of a DeepSeek V4. Second, even if we stay closed source, OpenAI and Anthropic need to get much cheaper much more quickly. Because US enterprise companies are doing the math, and right now the math doesn't favor them.

DeepSeek didn't catch up to America. But they built something good enough, gave it away for free, and a lot of US companies are going to take them up on it.