DeepSeek Did Not Kill NVIDIA—It Made Powerful AI Cheaper and Expanded the Chip Market
Why did DeepSeek not destroy demand for NVIDIA, but instead make powerful AI cheaper and expand the compute market?
Evaluate DeepSeek and NVIDIA as a physical and industrial system rather than only as software; the next step is to account for chips, energy, cloud capacity, manufacturing, and the cost of scaling.
What to watch for
Key takeaways
The “The market reaction to DeepSeek was immediate” issue should be assessed with one constraint in mind: but that logic is too linear. When computing becomes cheaper, people usually use more of it, not less.
The “DeepSeek: what the model is, pros and cons — part 1/2” scene leads to a working conclusion: the model, reportedly trained on ten or fifty thousand NVIDIA cards, is praised for its visible reasoning zone — reading it is a pleasure in itself — but on the tested cases the final answer still trails o1 from the twenty-dollar tier, and the visible reasoning is partly marketing.
The discussion of “How the DeepSeek release affects the future of AI — part 1/3” yields a practical test: a team that once needed hundreds of millions of dollars now has a chance to run a model locally, fine-tune it, and build it into a product. Even if the public cost figures are incomplete and training used more GPUs than stated, the approach shows that efficiency in architecture and training is nowhere near exhausted.
The decision in “NVIDIA's trouble: a temporary dip or a sunset?” depends on one criterion: a more efficient model does not cancel demand for hardware — NVIDIA will keep selling equipment to new participants, and more distributed demand means less hysterical growth, not the disappearance of the market.
For the “A usage example of OpenAI Operator” scene, the decisive point is this: OpenAI's demo of ordering sushi groceries in Instacart seems aimed at no one in particular, and on a real task — monitoring blocks in a payment system and sending notifications — Operator stopped working when the host switched screens and did the job childishly poorly.
The “OpenAI Operator: pros and cons” issue should be assessed with one constraint in mind: Operator runs in a separate browser and does not always understand the user's real environment — a vivid gap between foundation model and interface: a company can lead in research while releasing an agent that still looks experimental.
The practical meaning of “Conclusion” is that it destroys a more important illusion—that the path to a strong model is known, absurdly expensive, and open to only a few laboratories. Competition will now accelerate and demand for compute will become more distributed. For NVIDIA, that may mean less hysterical growth, but not the disappearance of its market.
What this episode is about
The Chinese model wiped billions from NVIDIA’s market value and forced Silicon Valley to recalculate the cost of training. But a more efficient model does not eliminate demand for compute; it lets more teams launch systems of their own. The real blow was not to hardware, but to the belief that only a handful of companies can build powerful AI.
The market reaction to DeepSeek was immediate: if a Chinese team produced a strong model for less money, then OpenAI’s enormous investments and demand for NVIDIA might be inflated. But that logic is too linear. When computing becomes cheaper, people usually use more of it, not less.
DeepSeek’s main effect is democratization. A team that once needed hundreds of millions of dollars now has a chance to run a model locally, fine-tune it, and build it into a product. Even if the public cost figures are incomplete and training used more GPUs than stated, the approach shows that efficiency in architecture and training is nowhere near exhausted.
For Meta, Mistral, Cohere, Perplexity, and other players, this is both a threat and a gift. It is a threat because a strong open model makes their technology less distinctive. It is a gift because they can build products on a cheaper foundation. NVIDIA still gets to sell equipment to those new participants.
Against the backdrop of DeepSeek, OpenAI is showing Operator and talking about an age of agents. But the product runs in a separate browser and does not always understand the user’s real environment. It is a good example of the gap between a foundation model and an interface: a company can lead in research while releasing an agent that still looks experimental.
DeepSeek does not prove that America has lost or that OpenAI is finished. It destroys a more important illusion—that the path to a strong model is known, absurdly expensive, and open to only a few laboratories. Competition will now accelerate and demand for compute will become more distributed. For NVIDIA, that may mean less hysterical growth, but not the disappearance of its market.
Competition will now accelerate and demand for compute will become more distributed. As a result, for NVIDIA, that may mean less hysterical growth, but not the disappearance of its market.
Episode transcript
The episode is in Russian; below is an English reading guide to the transcript (the full EN transcript is a machine translation). Voice matching applied to 71 segments: 31 identified, 6 mixed, 24 probable, and 10 unresolved.
Read transcript on a separate page
Loading…