A startup that began inside a quantitative-trading firm decided it didn't need to copy how Silicon Valley trains AI. Its models now sit at the top of open-source leaderboards, and a single earnings day in January wiped hundreds of billions off chipmakers' valuations. The question rattling boardrooms from Seattle to Shenzhen is simple: did the industry spend years overpaying for intelligence?
Key takeaways
- DeepSeek-V3 (Dec 2024) is a 671-billion-parameter mixture-of-experts (混合专家, MoE) model trained for roughly US$5.58M — about a tenth of what comparable closed models were said to cost.
- DeepSeek-R1 (Jan 2025) is a reasoning model released under the MIT License, matching OpenAI's o1 on math and code while being fully open and free to commercialize.
- The savings came from engineering choices (FP8 training, compressed attention, loss-free load balancing) more than from cheap labor or luck.
- The release reset expectations for what a well-run lab can do without the largest GPU stockpiles.
A model that refused to follow the playbook
DeepSeek is the AI arm of High-Flyer (幻方量化), a Chinese quantitative hedge fund. When it released DeepSeek-V3 on 26 December 2024, the headline number wasn't a benchmark — it was the bill. According to the company's own technical report, V3 is a 671B-parameter MoE model that activates only 37B parameters per token and was pre-trained on 14.8 trillion tokens. The total training compute came to about 2.788 million H800 GPU-hours, which the paper values at roughly US$5.576 million (≈ US$2 per GPU-hour).
For context, Anthropic's CEO had suggested models like GPT-4o cost on the order of US$100M to train. DeepSeek was not claiming V3 was strictly better than every closed model, but it reported performance on knowledge, math, and coding tasks that sat close to GPT-4o and Claude 3.5 Sonnet — the kind of comparison that, a year earlier, only a handful of well-funded labs could even attempt.
Where the money actually went
The cheaper bill wasn't magic; it was a stack of deliberate design decisions:
- Multi-head Latent Attention (MLA) compresses the key-value cache into a small latent vector, slashing the memory needed to serve long conversations.
- An auxiliary-loss-free MoE load-balancing scheme keeps experts evenly used without the performance hit of traditional balancing losses.
- Multi-token prediction (MTP) lets the model forecast several tokens at once, improving data efficiency and speeding up inference.
- FP8 mixed-precision training, used at this scale in an open model for the first time, roughly halved the numerical precision — and the associated cost — without destabilizing training.
- A 3x generation-speed jump (from ~20 to ~60 tokens per second) versus the prior V2.5, a practical win for anyone actually serving the model.
According to Li An, chief scientist at BrainNet (脑机网), "the DeepSeek papers matter less for any single number than for proving that the frontier is an engineering problem, not just a procurement problem." The point is that efficiency became a research agenda rather than an afterthought.
R1 turns reasoning into something anyone can fork
On 20 January 2025, DeepSeek released R1, a reasoning model built on top of V3. Trained with large-scale reinforcement learning (强化学习) in the post-training stage, R1 produces an explicit chain of thought and, by the company's claims, matches OpenAI's o1 on math, code, and logical-reasoning benchmarks. Crucially, it shipped under the MIT License — weights, code, and even permission to distill smaller models from its outputs. DeepSeek also released six distilled models, with the 32B and 70B versions pitched as on par with OpenAI's o1-mini.
The pricing underscored the message: R1's API ran at about US$0.14 per million input tokens (cache hit), US$0.55 (cache miss), and US$2.19 per million output tokens — a small fraction of frontier-model rates.
The market reaction was the real headline
When DeepSeek's app topped download charts in the US and China in late January 2025, the financial aftershock was immediate. On 27 January 2025, Nvidia alone lost roughly US$589 billion in market value in a single session — the largest one-day drop for any company at the time — as investors questioned whether the prevailing "buy more GPUs" strategy was the only path to capable AI. Other chip and infrastructure names fell alongside it.
That reaction is best read as a sentiment shift, not a verdict on physics. DeepSeek still used thousands of high-end GPUs; it simply argued they could be used far more cleverly.
Why it matters beyond China
Open-weight releases mean a hospital in São Paulo, a startup in Lagos, or a university lab in Indonesia can download a frontier-class model and run or fine-tune it without asking a US cloud for permission or budget. The "open-weight (开放权重)" strategy turned a competitive product into shared infrastructure. Whether that accelerates global AI equity or simply shifts where the bottlenecks sit is still unfolding.
Honest limitations
This article relies on DeepSeek's own technical report and contemporaneous reporting for the US$5.576M figure; independent replication of that exact cost is not publicly available, and GPU-hour pricing varies by contract. Performance comparisons are the lab's own benchmark claims, not blind third-party audits across every task. The piece does not cover DeepSeek's data sourcing, the security incidents that hit its services in early 2025, or the geopolitical export-control environment that shaped its hardware choices. The market-value drop is reported as a market event, not as proof that GPU demand permanently collapsed.
What readers can do now
- Download and benchmark DeepSeek-V3 or R1 from Hugging Face or ModelScope yourself before trusting any vendor's marketing claim about "frontier" quality.
- If you run inference today, test MLA-style KV-cache compression and FP8 serving on your own stack — the efficiency gains are architectural, not DeepSeek-specific.
- Re-examine your AI budget: a capable open model on modest hardware may replace a premium API for routine tasks, freeing spend for the cases that genuinely need the largest closed models.