In January 2025, a Chinese lab’s chatbot app briefly overtook the incumbents at the top of app store charts, and by the time markets opened that week, chip stocks had lost a meaningful chunk of their value in a single session. The trigger wasn’t a benchmark chart. It was a number: DeepSeek’s team put its V3 training run at somewhere around 5 to 6 million dollars, at a time when comparable frontier training runs at closed labs were widely understood to cost tens of millions, and in some cases well over a hundred million. That gap, more than any leaderboard position, is why DeepSeek matters.

The engineering that made the price possible

The cost story didn’t start in January 2025, and it didn’t come from a single trick. DeepSeek V2, released in May 2024, introduced Multi-Head Latent Attention, a technique built specifically to shrink the memory footprint of the attention mechanism, the part of a transformer that usually scales expensively as context grows. That change alone made larger models cheaper to run and cheaper to train at scale.

DeepSeek V3, in December 2024, is where the pieces came together. It used a mixture-of-experts design with 671 billion total parameters, but only 37 billion of those are active for any given token, so a huge model was trainable and runnable more efficiently than its total size suggests. It trained on 14.8 trillion tokens, and it did so largely in FP8, an unusually low-precision numerical format for training rather than just inference. Lower precision means less memory and faster computation throughout the run, not only at the end when the model is deployed. Together, the architecture and the precision choice are most of the answer to how a 671-billion-parameter model ended up with a training bill in the single-digit millions instead of the nine figures its raw size might suggest.

Reasoning, distillation, and the open license

A month later, DeepSeek R1 added reasoning behavior on top of that foundation, trained through reinforcement learning rather than only supervised fine-tuning on human-written examples. What made R1 spread as fast as it did wasn’t just the flagship model. DeepSeek released it alongside a family of distilled versions ranging from 1.5 billion up to 70 billion parameters, all under the MIT license, meaning developers could run a capable reasoning model on a laptop or a single GPU and build commercial products on top of it without asking permission. By April 2026, DeepSeek V4 had scaled the approach further still, up to 1.6 trillion total parameters with a context window of 1 million tokens, showing the cost-conscious architecture wasn’t a one-time trick but a direction the whole family kept pushing.

Why the reframing mattered more than the score

Think of a new airline entering a route dominated by a few legacy carriers. It doesn’t need to fly a nicer plane or offer a better seat. It only needs to fly the same route for a fraction of the cost, and suddenly every passenger who used to accept the old fare starts asking why they were paying so much in the first place. DeepSeek did the equivalent to frontier AI. Its models were good, but plenty of closed models were also good. What DeepSeek did that others hadn’t was force the entire industry to answer a question it had mostly avoided: how much should any of this actually cost.

That question is why the story lands better as an efficiency story than a capability story. Understanding why a 5 to 6 million dollar figure was so startling requires knowing what training compute typically costs and why, which is exactly the ground covered in Compute and GPUs. And V3’s split between 671 billion total and 37 billion active parameters is a concrete, real-world instance of the total-versus-active parameter distinction that makes this kind of efficiency possible, explored in more depth in Dense vs Mixture of Experts. DeepSeek didn’t win by topping a leaderboard. It won by making the cost of frontier capability a question nobody in the room could keep dodging.