DeepSeek-V3, the model that turned heads in early 2025, cost roughly $5.6 million to train on H800 GPUs, according to the company's own technical report — a number that helped reframe global assumptions about the compute cost of frontier AI. Its successor, DeepSeek-V4-Pro, released on April 24, 2026, scales to 1.6 trillion total parameters while keeping API input pricing at $1.74 per million tokens, around 13 times cheaper than comparable US models at launch.
How DeepSeek built frontier AI for under $6 million
The headline number from DeepSeek's V3 technical report, published in December 2024 on arXiv, is 2.788 million H800 GPU hours for the full training run, which at roughly $2 per GPU-hour comes to approximately $5.576 million. The model has 671 billion total parameters but uses a Mixture-of-Experts architecture that activates only 37 billion per token, reducing inference and training compute relative to a dense model of the same nominal size. Pre-training covered 14.8 trillion tokens in under two months on a 2,048-GPU cluster.
V4-Pro, released six months later, extends this architecture: 1.6 trillion total parameters, 49 billion active per token, and a default one-million-token context window. At launch it scored 80.6% on SWE-bench Verified, the highest coding benchmark result of any model at the time of release, according to benchmark tracker BenchLM. DeepSeek labels V4-Pro a preview release, meaning final performance figures may shift before general availability.
The scaling pattern across DeepSeek's model generations reflects a deliberate architectural bet. Rather than chase raw parameter count across all weights, the team progressively increases total-model size while keeping the active-parameter footprint small. V4-Flash, the lighter version released alongside V4-Pro, uses just 13 billion active parameters drawn from a 284-billion-parameter pool, making it cheaper to serve than V3 on a per-token basis despite being a newer model.
The $10 billion fundraise and a valuation that tripled in six weeks
Until April 2026, DeepSeek had no external investors. High-Flyer Capital Management funded all research spending directly, without venture-capital dilution or a public markets timeline. That changed when the company opened its first external round, targeting at least 300 million dollars at a valuation above $10 billion, according to reports in April 2026. The China Integrated Circuit Industry Investment Fund — the state-backed vehicle known colloquially as the "Big Fund" — is leading the round, with Tencent Holdings and Alibaba Group in separate talks to participate, per Bloomberg.