The $5.6M AI Training Run That Shook $1 Trillion of AI Stocks

For years, the AI industry followed a fairly simple story: bigger models required more GPUs, more GPUs required larger data centers, and larger data centers required more electricity, more capital and ever more expensive infrastructure. If you wanted to compete at the frontier of artificial intelligence, the assumption was that you needed access to enormous financial and computational resources.

Then DeepSeek appeared.

A Chinese AI company that relatively few people outside the industry had heard of suddenly released a model competitive with some of the strongest systems coming out of Silicon Valley. Attached to that release was one number that immediately caught everyone’s attention: $5.576 million.

A few weeks later, Nvidia lost around $593 billion in market value in a single trading session. Semiconductor companies, infrastructure providers and other businesses tied to the AI boom dropped with it, and what initially looked like another technical model release suddenly became a Wall Street story as well. (reuters.com)

The popular version of the story quickly became that a $6 million Chinese startup had somehow destroyed a trillion dollars of market value. That makes for an excellent headline, but the real story is both more complicated and, in many ways, more interesting. DeepSeek did not build an entire AI company for $6 million. What it demonstrated was how much efficiency might still be hidden inside modern AI systems.

Inhaltsverzeichnis

The $6 Million Number Everyone Misunderstood
Then Came R1
Then Developers Looked at the Price
The Bigger Problem for Silicon Valley
Wall Street Noticed
Open Models Changed the Equation Too
No, DeepSeek Did Not Make Compute Irrelevant
The Real DeepSeek Story
Further Reading

The $6 Million Number Everyone Misunderstood

The famous number comes from DeepSeek’s own technical report. The company reported that training DeepSeek-V3 required approximately 2.788 million Nvidia H800 GPU hours. Using an assumed cost of $2 per GPU hour, DeepSeek calculated the cost of the official training run at $5.576 million. (arxiv.org)

That number is real, but it needs context.

It does not mean somebody started an AI company with $6 million and walked out a few months later with a competitor to OpenAI. The figure covers the official training run of DeepSeek-V3. It does not include all the research that came before it, failed experiments, salaries, data preparation, hardware ownership, infrastructure, earlier model iterations or the many other costs involved in building an AI lab.

Even with that qualification, the result remains striking. DeepSeek-V3 contains 671 billion parameters, although its Mixture-of-Experts architecture activates only around 37 billion parameters for each token. The model was trained on 14.8 trillion tokens, which makes the reported training cost interesting not because the entire company somehow cost $6 million, but because a highly capable model of that scale could apparently be trained far more efficiently than many people expected.

This was not a small experimental model running on a workstation under someone’s desk. It was a serious frontier-scale system, and that is exactly why the number attracted so much attention.

Then Came R1

DeepSeek-V3 attracted interest among researchers, but DeepSeek-R1 was the release that turned the company into a global story.

R1 focused heavily on reasoning tasks such as mathematics, coding and multi-step problem solving, which makes OpenAI’s o1 family a more useful comparison than GPT-4. DeepSeek built R1 on top of the V3 foundation and used reinforcement learning to push the model further in tasks where simply generating a plausible next word is not enough.

One particularly interesting part of that work was R1-Zero. Instead of first teaching the model through large amounts of carefully curated supervised examples, DeepSeek used reinforcement learning and allowed useful reasoning behaviours to emerge through training. The result was not particularly polished, so DeepSeek later added additional training stages to produce the more usable R1 model, but the experiment itself was important.

It suggested that some capabilities we had become used to teaching directly might instead emerge through reinforcement learning when the right incentives and training setup are provided. That did not make supervised fine-tuning obsolete, but it showed that the standard recipe was not necessarily the only one available.

Then Developers Looked at the Price

The next surprise was not hidden inside a technical paper. It was right there in DeepSeek’s API pricing.

When R1 launched, DeepSeek charged $0.55 per million uncached input tokens, $0.14 per million cached input tokens and $2.19 per million output tokens. OpenAI’s o1 pricing at the time was $15 per million input tokens and $60 per million output tokens.

On the published token prices alone, DeepSeek R1 was therefore roughly 27 times cheaper than OpenAI o1.

That does not automatically mean every task became 27 times cheaper. Reasoning models may use very different numbers of tokens to complete the same piece of work, and the real cost of AI is ultimately the cost of completing the task rather than the cost of one million tokens in isolation. Still, a price gap that large is difficult to ignore, especially for companies discovering that their largest AI expense may not be training a model once, but running it millions of times afterward.

The Bigger Problem for Silicon Valley

This is where DeepSeek became more than another model release.

For several years, the AI infrastructure story had been built around scarcity. The industry needed more GPUs, more data centers, more power capacity and more capital. Companies were committing tens of billions of dollars to infrastructure based on the expectation that increasingly capable AI systems would require ever larger amounts of compute.

DeepSeek introduced another possibility. Perhaps the industry did not only need more compute; perhaps it also needed to become much better at using the compute it already had.

The engineers behind DeepSeek-V3 used a collection of architectural and systems techniques to improve efficiency. Its Mixture-of-Experts design means that only part of the full model is active for each token. Multi-head Latent Attention reduces some of the memory demands associated with attention, FP8 mixed-precision training lowers computational and memory requirements, and multi-token prediction is intended to improve training and inference efficiency.

None of those techniques mean that GPUs suddenly became irrelevant. DeepSeek still used millions of GPU hours and thousands of accelerators. What changed was the assumption that better AI necessarily required nothing more sophisticated than buying more hardware.

The uncomfortable question became whether part of the industry had begun to confuse access to enormous amounts of compute with efficient use of it.

Wall Street Noticed

Investors understood the implication very quickly.

If better architectures, training methods and engineering could allow companies to produce competitive models with less compute than previously expected, then some of the assumptions behind the enormous valuations of AI infrastructure companies might deserve another look.

On January 27, 2025, Nvidia shares fell almost 17%, wiping approximately $593 billion from the company’s market capitalization in a single trading session. Other semiconductor and AI-related companies also declined as investors reconsidered how much hardware future AI systems might actually require. (reuters.com)

That does not mean DeepSeek literally destroyed hundreds of billions of dollars of economic value. Market capitalization reflects investor expectations and can change very quickly in both directions. Some of those losses were recovered, and a one-day market reaction should not be confused with a permanent change in the economics of AI.

What DeepSeek did was introduce uncertainty into a story investors had become very comfortable with. The equation had been straightforward: more AI required more compute, more compute required more GPUs, and more GPUs meant more revenue for the companies supplying them. DeepSeek added another variable to that equation, and that variable was efficiency.

Open Models Changed the Equation Too

There was another part of the DeepSeek story that mattered beyond training costs and API pricing.

DeepSeek released R1’s model weights and code under the MIT License, allowing commercial use, modification and derivative work. (github.com)

For researchers and developers, that changed the nature of the model. They did not have to interact with R1 exclusively through DeepSeek’s API. They could download the model, experiment with it, study it, run it on their own infrastructure and build their own systems around it.

DeepSeek also released smaller distilled versions based on Qwen and Llama models, making some of the reasoning capabilities accessible to teams that could not realistically run the full 671-billion-parameter system.

That openness matters because progress spreads faster when other teams can inspect and build on what has already been done. A proprietary API can be impressive, but an accessible model becomes part of a much larger engineering ecosystem.

No, DeepSeek Did Not Make Compute Irrelevant

The easiest conclusion to draw from all of this would also be the wrong one: that AI no longer needs expensive hardware.

It absolutely does.

DeepSeek used thousands of GPUs and millions of GPU hours. Training models at this level remains an enormous engineering exercise, and the demand for compute may even increase if improvements in efficiency make AI cheap enough to use in far more applications.

We have seen this pattern repeatedly in computing. Making computation cheaper does not necessarily reduce total consumption. Very often, it creates new uses that were previously uneconomical and increases demand overall.

The real lesson from DeepSeek is therefore not that infrastructure stops mattering. It is that efficiency may matter just as much as scale, and that the industry can no longer assume that more spending automatically translates into a proportional advantage.

The Real DeepSeek Story

The „$6 million startup destroys $2 trillion“ version is memorable, but it is not accurate.

DeepSeek was not built for $6 million. The $5.576 million figure refers specifically to the reported official training cost of DeepSeek-V3, not the total cost of the company or the full development of R1. And DeepSeek did not literally erase $2 trillion from the stock market.

Once the exaggeration is removed, however, the underlying story becomes even more interesting.

A relatively small Chinese AI lab built a highly capable model using hardware that was not considered the absolute frontier, trained it at an unexpectedly low reported compute cost, developed a strong reasoning system on top of it, released the model openly and offered API access at a fraction of the prevailing price.

That was enough to make investors question some of the assumptions sitting underneath one of the largest investment cycles in the history of technology.

Marc Andreessen described DeepSeek as an “AI Sputnik moment.” Whether that phrase ultimately proves justified is less important than the question DeepSeek forced the industry to confront.

For years, the AI race had largely been framed around who could acquire the most compute. DeepSeek reminded everyone that another race is happening at the same time: who can extract the most intelligence from the compute they already have?

That may turn out to be the more interesting contest.