A Full Breakdown of DeepSeek's "Cost-Cutting, Efficiency-Boosting" Shockwave

I'm finally writing about DeepSeek. Of everyone covering AI, I'm probably the last to weigh in on DeepSeek — which, as it turns out, gives me a small but genuinely useful advantage.

Industry Buzz》

DeepSeek's Three Real Contributions — and How They're Reshaping the AI Market

First, I think a lot of the details and facts about DeepSeek only came out quite late. Second, given how brutally competitive the AI space is, there was never going to be a scenario where DeepSeek released something and rivals just sat there without responding — a response was inevitable, and that response, in turn, forces more disclosure. So basically, all the details about DeepSeek, and how competitors have reacted to it, have really only come together as a full picture recently — as in, right now — including things very much worth discussing, like OpenAI's Deep Research feature (built on GPT, the generative pre-trained transformer), which only got pushed out the door because of this pressure.

That's roughly the situation, and it's the upside of writing about this a bit late.

There's already a mountain of material out there on DeepSeek's technical details, essays, and reasoning, but I don't think the pieces that came out earliest were especially reliable, for a few recurring reasons.

First, there's a lack of clarity or understanding out there about the fact that the underlying technical integration isn't actually that unusual — DeepSeek's real achievement is executing that same logic at much higher performance and dramatically lower cost. These techniques weren't invented by DeepSeek; what they do have is genuinely strong execution and optimization. Neither model distillation nor MoE (mixture of experts) fell out of the sky. In fact, MoE has long been a core strength of France's Mistral AI — which is exactly why Mistral has been one of the companies hit hardest by this.

Then there's the cost question. DeepSeek claims it only spent $6 million, but that figure doesn't include a lot of the upfront training costs — it only counts the final training run.

That's understandable as a piece of marketing; it's just the logic of a pitch. Sure, they say it only cost $6 million, but actual deployment costs likely run somewhere between $500 million and $1 billion, and they used something like 50,000 GPUs — computing power roughly equivalent to at least 20,000 higher-end chips. The "$6 million" framing is a bit like that old joke: a young saver buys a house after ten years of saving, except what nobody mentions is that those ten years of savings only covered a 5% down payment — mom and dad quietly covered the other 95% mortgage.

Setting aside the overhyped parts, DeepSeek's biggest contributions really come down to three things.

First, because it's open-source (which, sure, plenty of people in the market still dispute), it can be deployed independently, and DeepSeek has released the corresponding development logic alongside it. The associated papers disclose the "recipe" to a meaningful degree, which is a genuine boon for the AI field as a whole.

Second, it proves there's still room to build something new outside the handful of giant labs. Plenty of smaller tech companies had already given up on building foundation models because of GPU constraints, and now they have a real shot at getting back into the game. There's suddenly a lot more perceived room for engineering-based solutions, which is good news for investment across the whole market.

Third, it has created a ripple effect that's pushing the big labs to compete harder and release more of what they've been sitting on. If it weren't for DeepSeek, it probably would have taken a lot longer before we got access to Google's dramatically cheaper Gemini API and GPU-powered deep research tools.

DeepSeek's biggest gift to the rest of us is showing that there's still enormous room left in how the entire engineering stack gets deconstructed — there's a lot of interesting work still possible, and we don't necessarily have to depend solely on the giant labs. I think that's a genuinely good dynamic. Sure, $500 million is still a lot of money, but plenty of players can write that check.

The approach itself is fundamentally engineering optimization, so costs came down by one to two orders of magnitude. Once DeepSeek went public, plenty of people used similar engineering tricks to build comparable models — AI data godmother Fei-Fei Li's team, for instance, distilled a similar model using only about $20 worth of compute.

Business Opportunities》

Cost-Cutting Potential Rises — Long-Term Outlook for Semiconductors Stays Bullish

Does this change the current landscape of Big Tech dominance in AI? I think that's still a somewhat optimistic take right now, though in plenty of domains that dominance is genuinely under serious pressure — especially in edge computing and "good enough" use cases.

Judging by recent moves, the big labs' lead is still intact, and they can quickly cut API (application programming interface) pricing below DeepSeek's rates (at least for now) — Google's Gemini 2 being one example. So on that front, I think the optimism is a bit premature; I don't think the big labs' technical level or cost-efficiency is actually that far behind.

Because DeepSeek's release happened to land right in the gap between the big labs' update cycles, it created an immediate sense of catching up — but the speed of the big labs' response shows they were already prepared for the next generation of products.

Coming back to the core question: now that a cost-cutting, efficiency-boosting path has been found, what happens next for companies like Nvidia and TSMC? For one thing, it was never realistic to expect Nvidia to keep monopolizing the market with margins this high forever. But rising potential for cost-cutting and efficiency gains should, if anything, increase long-term semiconductor demand — that much seems like a fair consensus — even as short-term volatility is unavoidable.