DeepSeek is on the verge of closing a $7.4 billion funding round that values the Hangzhou AI startup at $74 billion. Its API business posted an 82.9 percent gross margin. Revenue grew roughly tenfold in seven months. And its models, which enterprise teams can run on their own infrastructure under a permissive license, still cost a fraction of what US frontier labs charge. For enterprise AI buyers, this week’s news is not just a funding story. It is a preview of what happens to the open-source cost advantage when the company providing it prepares to become accountable to public markets.
The Numbers: Revenue, Margins, and the IPO Clock
According to The Information and corroborated by Bloomberg data and the South China Morning Post, DeepSeek generated approximately 475 million yuan ($70.7 million) in revenue in the first seven months of 2026. That is roughly ten times the company’s entire 2025 revenue.
What makes the number remarkable is not the size but the structure. DeepSeek’s API business achieved an 82.9 percent gross margin over the same period. Net losses narrowed to 715 million yuan for the seven months, down from 935 million yuan for all of 2025. Infrastructure spending surged to 11 billion yuan, the inevitable cost of keeping pace with US frontier compute. The business is not yet profitable, but the trajectory is clearly toward a unit-economics model that works.
The funding round itself, first reported by the Wall Street Journal on August 27, targets 50 billion yuan ($7.4 billion) at a pre-money valuation of 500 billion yuan ($74 billion). Existing investors Monolith Management, Shixiang Capital, and battery giant CATL are participating. New investors in discussions include CPE, Legend Capital, and Stony Creek Capital, a semiconductor-focused private equity firm linked to the chairman of ChangXin Memory Technologies. State-backed funds from Hefei are also involved.
The fundraise is explicitly a pre-IPO round. DeepSeek has begun preparing to file a prospectus on Shanghai’s Star Market, with a listing target of 2027.
Why the 82.9% Margin Changes the Narrative
The conventional story about open-source AI economics is that it cannot be monetized: a company releases its weights, competitors copy the training insights, and pricing collapses toward marginal inference cost. DeepSeek’s margin data complicates that story significantly.
The 82.9 percent API gross margin is higher than what most enterprise SaaS companies report at scale. It exceeds what OpenAI and Anthropic have publicly disclosed about their API economics. The reason is inference efficiency. DeepSeek’s V4 architecture, with its mixture-of-experts design, activates only a fraction of its parameters per token. The model was designed from first principles around the constraints of Chinese inference hardware, which lacks access to the latest NVIDIA chips. That constraint produced efficiency gains that translate directly into gross margin when the model is served at scale.
The implication for enterprise AI buyers is important: open-source AI models, when operated by their original developer at scale, can be as profitable as proprietary models. The “race to zero” narrative was always partly wrong. What actually happens is that the lowest-cost inference engine wins on margin, and the developer who built the model has a structural advantage over anyone running it on top of commodity cloud infrastructure.
The Price Increase: A Signal for Enterprise Procurement
DeepSeek recently raised prices on its V4-Pro API. Output token pricing went from $0.87 to $3.96 per million tokens, a 4.5x increase. The table below shows where V4-Pro now sits relative to peers.
| Model | Input ($/M tokens) | Output ($/M tokens) | License |
|---|---|---|---|
| DeepSeek V4-Pro (new) | $0.27 | $3.96 | Apache 2.0 |
| DeepSeek V4-Pro (old) | $0.14 | $0.87 | Apache 2.0 |
| Kimi K3 | $2.00 | $15.00 | Proprietary |
| GPT-5.6 Sol | $3.00 | $15.00 | Proprietary |
| Claude Opus 5 | $5.00 | $25.00 | Proprietary |
| Qwen3.8-Max | $0.40 | $1.20 | Commercial |
| Gemini 3.7 Flash | $0.15 | $0.60 | Proprietary |
Sources: DeepSeek API docs, Gate News analysis, vendor pricing pages.
DeepSeek remains cheap by any comparison. But the direction of the move matters. This is the company signaling that its era of below-cost API pricing is ending as it prepares for IPO-grade financial scrutiny. Enterprise procurement teams that have built AI cost models around DeepSeek’s 2025 pricing should revisit those assumptions.
The broader Chinese AI pricing environment is moving in the same direction. Zhipu AI, Alibaba, Tencent, and Baidu have all raised prices in 2026. The discount war that saw some APIs priced at near-zero in 2025 is over. Scale, infrastructure investment, and the demands of a competitive model development cycle have pushed the industry toward sustainable pricing.
What Pre-IPO Status Means for Enterprise Due Diligence
An IPO filing, expected before year-end, will be the first time DeepSeek operates under Chinese mainland disclosure requirements. That is meaningful in both directions for enterprise buyers.
On the positive side, a public company has audited financials, disclosed business relationships, and legal obligations to shareholders. The opacity that characterizes most frontier AI labs, including the company’s current private structure, gives way to something enterprise procurement teams can actually assess.
On the less comfortable side, a Shanghai Star Market listing subjects DeepSeek more formally to Chinese data security and export control regimes. Enterprises in regulated industries, particularly those in financial services, defense supply chains, healthcare, and legal services, should expect that their security and legal teams will have questions about any API that runs on Chinese-hosted infrastructure once the IPO disclosure process begins.
The self-hosted path, which was always the most compliance-friendly option, continues to be the recommended approach for sensitive workloads. DeepSeek’s Apache 2.0 license allows full deployment on your own infrastructure. That option is unaffected by IPO status.
What Enterprise AI Teams Should Do Now
The three practical actions for AI buyers following this week’s news:
Audit your DeepSeek cost assumptions. If your AI cost models were built around 2025 or early 2026 pricing, the 4.5x output price increase on V4-Pro changes your unit economics. Run the updated numbers. For most use cases, DeepSeek API remains cost-effective, but the margin buffer over alternatives has narrowed.
Update your vendor risk register. Pre-IPO status, Chinese regulatory jurisdiction, and large-scale infrastructure expansion are all material changes to the vendor risk profile. Refresh your procurement documentation before the IPO filing period creates more questions than answers.
Evaluate the self-hosted option seriously. DeepSeek V4-Pro is available under Apache 2.0. Running it on your own infrastructure, whether on-premises or in a hyperscaler VPC, sidesteps jurisdiction concerns, locks in pricing, and gives you model-level control. As the enterprise AI community covered in our analysis of open-source AI cost strategies, self-hosting frontier-class models is now viable for teams with a capable ML infrastructure function.
The broader context for these decisions is an AI industry undergoing a structural pricing shift. As explored in our coverage of the AI pricing revolt, enterprise buyers who locked in multi-year API agreements in 2025 at 2025 prices are now sitting on significant structural advantages. Those who did not should expect that the cost landscape of 2027 will look substantially different from 2025.
DeepSeek’s pre-IPO round is not just a China story. It is the moment that open-source AI’s cheapest, highest-quality player starts behaving like a public company, and everything that implies about pricing discipline, regulatory accountability, and long-term vendor relationship management follows from that.
Enterprise teams building AI systems in 2026 should factor in a world where DeepSeek’s marginal cost advantage compresses, its data governance obligations increase, and its financial transparency improves. The model itself is the same. The vendor relationship is changing. That distinction matters for every AI procurement decision you make in the next twelve months.
For a deeper look at how to structure enterprise AI vendor strategy in a multi-model world, book a call with the Enera team or explore our AI-native enterprise transformation framework.