DeepSeek's Pricing Isn't Sustainable. Here's the Math.
At $0.14 per million tokens, DeepSeek is undercuting the market by 10x. But the unit economics don't add up — even with their efficiency claims.
DeepSeek shocked the AI industry with pricing that’s roughly 10x cheaper than OpenAI and Anthropic. $0.14 per million input tokens. $0.28 per million output tokens. That’s cheaper than running your own infrastructure in most cases.
But pricing below cost isn’t a business model. It’s a customer acquisition strategy with an expiration date.
The cost floor
Training and inference costs for a GPT-4 class model are well-documented:
- GPU rental (H100): ~$2.50/hour per GPU at scale
- Inference throughput: ~1000 tokens/second per H100 for a 70B model
- Per-million-token GPU cost: ~$0.69 at theoretical maximum efficiency
Even at maximum utilization with zero overhead, the GPU cost alone exceeds DeepSeek’s input price. Add bandwidth, storage, research, staffing, and the gap widens.
How they might be doing it
Model architecture efficiency. DeepSeek’s MoE (Mixture of Experts) design is genuinely innovative. By activating only a subset of parameters per token, they reduce inference compute significantly. This could plausibly cut costs by 2-3x.
Cheaper hardware. Chinese GPU alternatives and bulk deals with domestic providers could reduce per-GPU costs below Western rates.
Government subsidies. China’s strategic investment in AI means DeepSeek may be receiving indirect support that artificially lowers their operating costs.
Loss-leading. The simplest explanation: they’re pricing below cost to capture market share and mindshare. It’s working.
Why it matters
DeepSeek’s pricing is forcing the entire industry to reconsider their margins:
- OpenAI cut GPT-4o pricing by 50% in response
- Google reduced Gemini pricing for high-volume users
- Anthropic introduced batch processing at a 50% discount
This is great for developers in the short term. But if DeepSeek’s pricing is unsustainable, the industry is racing toward a price floor that will eventually snap back.
The VC vs. state question
Western AI companies are funded by venture capital — investors who expect 10x returns within 7-10 years. They can’t price below cost indefinitely.
DeepSeek’s funding structure is murkier. If they’re backed by state resources with strategic rather than financial objectives, they can sustain losses far longer than any VC-backed competitor.
This creates an asymmetric price war. One side needs to eventually make money. The other might not.
Our take
DeepSeek’s pricing is real, their models are competent, and developers should absolutely use them where appropriate. But don’t build your entire product around $0.14/M tokens. Build cost buffers into your architecture. Assume prices will eventually rise.
The current AI pricing environment is an aberration, not the new normal. Plan accordingly.
