On Monday, January 20, 2025, DeepSeek, a Chinese AI developer, launched two new open-source “reasoning” large language models (LLMs), DeepSeek-R1-zero and DeepSeek-R1, which performed equivalently to or even outperformed OpenAI’s reasoning model OpenAI-o1 — at a fraction of the cost. By Friday of that week, DeepSeek’s app had climbed to the top of the Apple App Store, and on Monday, January 27, 2025, a huge selloff in U.S. tech and AI stocks (like Nvidia) removed a collective USD 1 trillion from their market capitalization. This prompted quick America-first rhetoric from Silicon Valley figures like Marc Andreesen, who called it a “Sputnik moment”; Donald Trump was more sanguine, describing it as a “wake-up call” for American Innovators. But how big of a deal is it, really?
DeepSeek’s launch is the firmest evidence yet of the commodification of large AI models, something that we predicted last year. Pure AI companies like OpenAI will be hit the hardest, and this could be the event that bursts the bubble in AI valuations. Paradoxically, it’s also a net positive for the long-term use of AI, as it highlights there is a clear path to profitably running an AI business. We’ll break down the improvements DeepSeek made, the implications for companies, and what this means from a policy perspective below.
Is DeepSeek really a fraction of the cost of competitive models? The short answer is yes. DeepSeek V3 — its latest LLM — cost around USD 5.7 million to train, an order of magnitude less than the estimated hundred million dollars spent on training similar models like Meta’s Llama 3. The DeepSeek team piloted a number of impressive improvements to train its models in a highly data- and memory-efficient way, including by compressing much of the data on the tokens used to train the model, using 8-bit rather than 32-bit storage for the training data, and using “experts,” that is, only activating part of the model at a time for specific tasks rather than using all the model all the time. Much of this was done due to hardware limitations — DeepSeek was trained on memory-limited Nvidia H800 GPUs, as more powerful Nvidia H100 GPUs are banned from China by the U.S. government’s export controls. The net result is that DeepSeek claims that it trained the model in just 2.78 million H800 GPU hours, compared to Llama 3.1, which was trained in roughly 30 million H100 hours. This is roughly one tenth the number of hours, and less than one tenth the cost due to the cost differences between the H800 and H100 GPUs. You can’t just go and make your own AI for USD 5.7 million, however; as that cost doesn’t include spending on DeepSeek’s R&D, headcount, and other essential services. It’s also worth noting that DeepSeek almost certainly got much of the training data for V3/R1 by “distillation” from other models (a process of getting tokens to train on from completed models), so this breakthrough would not be possible without previous open-source AI developments.
A tenfold reduction in cost does not necessarily mean a tenfold growth in markets. AI has been priced at a loss since the explosion of ChatGPT two years ago. This has encouraged its growth — companies like OpenAI wouldn’t have experienced the astronomical expansion in users if the average consumer had to pay, or the average business user had to support the profitable operation of AI. Like Uber, venture capital has been footing the bill for hypergrowth. The cost reductions from DeepSeek shouldn’t have a significant impact on demand growth, though, because the prices really can’t get much lower. Rather, they should have the impact of making the AI business more commercially viable within existing markets by reducing both the cost of operating AI models and (especially) capital costs (capex) of chips. Starting with a blank slate, companies will have a better line of sight to profitability by operating AI at scale. This doesn’t apply, however, to companies that have already made major capital investments in GPUs or data centers.
The near-term implications are grim for competitive AI developers, particularly OpenAI. Prior to this announcement, OpenAI was the only company with a “reasoning” model, a type of LLM that breaks prompts into smaller segments and makes multiple attempts to answer a single prompt to produce the highest quality answer. This isn’t reasoning as humans do (or even as other algorithms) but is a way to juice performance by spending more compute. OpenAI’s reasoning models were one of the reasons it was considered a market leader, but now that gap has been erased. More concerning, however, is the issue of money. OpenAI lost around USD 5 billion last year and spent nearly USD 8.5 billion on training models and staffing. OpenAI has (optimistically) projected revenues of USD 11.6 billion in 2025 and USD 100 billion in 2029, the year it hopes to become profitable. But these revenue figures depend on price increases to its ChatGPT service — doubling the price by 2029. Such price increases seem totally unsustainable in the face of open-source competition causing a price war. Already, DeepSeek is pricing it’s R1 model at USD 2.19 per one million tokens — a fraction of OpenAI’s o1 model at USD 60 per million tokens or USD 12 per million tokens for o1 mini. OpenAI will have to cut pricing to remain competitive, but it’s hard to imagine it achieving profitability under those circumstances. It will, of course, eventually develop cheaper models, but much of its cost structure is in labor, which will be hard to save on. It’s also hard to see how it hits its lofty USD 100 billion revenue goal with price cuts instead of increases; this could cause investors to balk at giving OpenAI the hundreds of billions of dollars in extra funding it says it needs.
Data center energy demand expectations should come back down to earth now. The AI hype has fueled a corresponding hype for businesses selling energy to data centers — the metaphorical picks and shovels for the AI gold rush. This excitement led to a huge surge in plans to build data centers, along with a huge surge in forecast energy demand — but as Lux has consistently pointed out, much of this demand won’t materialize. Even before DeepSeek, the physical constraints such as the production of chips, the speed of permitting, and the sourcing of electricity meant that the most optimistic estimates weren’t likely. The widely cited U.S. Department of Energy report from last month that forecast data centers reaching 6.7% to 12% of energy demand by 2028 now looks very out of date; breathless estimates as high as 16% of energy demand by 2030 now look foolish. This doesn’t mean that the data center energy business is dead, however. Data center energy demand will grow, just at much more realistic levels. Data centers are attractive because they are premium energy consumers, with a track record of paying extra for clean, reliable power. There’s still money to be made selling picks and shovels — but megaprojects like the proposed USD 100 billion to USD 500 billion for OpenAI’s Stargate now seem fanciful and unnecessary.
The longer-term implications for the AI sector are mostly positive. If your thesis is that AI is core to your company’s future success and profitability (as tech company CEOs uniformly believe) then cheap, powerful, energy-efficient AI is a good thing. Obviously, there are going to be some hits in the near term: Mark Zuckerberg’s decision to stockpile 600,000 Nvidia H100 GPUs by the end of 2024 and desire to double that number this year now look a little silly; that’s roughly USD 24 billion in capex in 2024, only to get your clock cleaned by a startup with far fewer resources. But the open-source nature of DeepSeek means that just as it built on previous models, so too will groups like Meta leverage DeepSeek’s advancements to make its own AIs better. The commodification of AI cuts both ways: DeepSeek is unlikely to be a leader for very long. Unlike OpenAI, companies like Meta have the cash cushion to blow billions on GPUs and are more interested in the long-term potential of AI. Cheap AI could mean more profits in such a scenario. The users of AI are also winners here: A price and performance war benefits early adopting sectors like banks, coding, and customer support. Startups developing AI wrappers are also somewhat vindicated, as the commodification of AI models means that the differentiation in AI will come elsewhere, and a user interface is one likely source. Of course, it may mean that the AI developers begin to aggressively compete with and buy out the wrapper companies to access that differentiation, but the decision to stay out of the cutthroat model race seems smart.
The biggest winners are the platform owners — like Apple and Microsoft. Apple’s AI strategy has been roundly criticized by many (including the author) for missing the mark on developing useful applications. Despite that, it owns the most valuable difference: a captive audience. The commodification of AI means that there will be a race to access end-users, not a race to have the best models. Apple’s walled garden ecosystem is a valuable prize: Google famously pays Apple more than USD 20 billion per year to be the default search engine on Safari. It’s easily worth it for Google, as paying to maintain its monopoly means it can make the money back (and more) in ad revenue. The same dynamics may apply to AI companies, assuming a valuable consumer application for AI can be found. The biggest single winner of all this may be Microsoft, which is still the biggest player in corporate operating systems. Its access to businesses — where the clearest use-cases for AI reside — puts it in the best position to sell AI access and reap the rewards from major cost reductions. The fact that these business use-cases represent a smaller (potential) market than consumer applications is also much less of an issue: If AI capex going forward is an order of magnitude less than before, those markets have a much more appealing payback.
The U.S.’ chip controls appear to have backfired spectacularly. Necessity is the mother of invention, and the DeepSeek team cites the lack of access to H100 GPUs as a key driver of the memory innovations that powered the breakthroughs in cost. It’s tricky to say exactly what this means for chip restrictions, but it’s certainly a major blow to those who think that U.S. supremacy in AI is crucial for the future. The most likely outcome of all this is an even stronger push by U.S. AI companies (including OpenAI) to tie AI to the security state to ward off foreign competition. Expect more efforts by tech CEOs to lobby Trump to support their efforts, either indirectly (by not regulating them) or directly (though government spending). Trump, for his part, has remained relatively positive toward China, making further chip controls seem unlikely — but he could pivot at any moment.
Lux Take
This is a moment that rivals the launch of ChatGPT — a pivot point for the development of AI as an industry. It makes it clear that AI will be cheap enough for business use-cases. It also makes it clear that AI models are commodified. Moving forward, AI integration will be a key focus — especially platform owners like Apple and Microsoft, which are in a prime position to capitalize on low-cost, high-quality AI. Successful businesses are still possible — even likely in data center power, but pure AI demand won’t reach the lofty heights forecast as recently as a month ago. Executives that had already developed a robust AI strategy don’t need to change direction; if anything, they should accelerate their deployment of AI and look forward to a generation of cheaper, more powerful models.
