DeepSeek raises 500 billion yuan funding

DeepSeek said it has completed its first external financing round, raising more than 500 billion yuan, or about $74 billion, in a transaction that would mark a striking reversal for a company that had previously pledged to avoid outside capital and a public listing. The financing briefing put DeepSeek’s pre-money valuation at 3.675 trillion yuan, roughly $543 billion, making the funding itself less notable than the message attached to it: the company now sees compute capacity, not model design alone, as the central constraint in its attempt to close the gap with leading U.S. AI labs.

The company said founder Liang Wenfeng personally contributed 200 billion yuan to the round. Tencent invested 100 billion yuan, CATL invested 50 billion yuan, and NetEase, JD.com and IDG Capital each invested 30 billion yuan, according to the briefing. The National AI Industry Investment Fund added 10 billion yuan. Those figures, if reflected in finalized financing documents, would represent one of the largest private technology fundraising rounds ever disclosed.

The scale of the round also changes the way DeepSeek will be judged. Until now, the company has built much of its reputation on claims of efficiency, open model releases and a research culture presented as deliberately different from the performance-management systems used by many large technology companies. With outside capital now involved, DeepSeek’s ability to convert research speed into durable infrastructure, product revenue and hardware access becomes a more concrete test.

Liang says compute is the main bottleneck

Liang Wenfeng, founder of DeepSeek, told the financing briefing that the company’s technology remains 12 to 18 months behind top U.S. AI developers while operating with about one-twentieth of their compute capacity. He said DeepSeek’s goal is to narrow that lag to three to six months by expanding computing resources.

That framing is central to the transaction. DeepSeek is not presenting the financing as a conventional growth round aimed mainly at sales expansion. It is positioning the capital raise as a resource strategy for training, inference and experimentation, where access to advanced processors increasingly determines how quickly frontier AI systems can be tested and improved.

The company said its V3 model training still depends on Nvidia graphics processing units, but it also uses a self-developed compiler called TileLang to reduce reliance on Nvidia’s CUDA software ecosystem. CUDA is the programming platform that has helped make Nvidia GPUs the default hardware for AI training. Reducing dependence on it would give DeepSeek more flexibility if domestic chips become more available or if access to foreign chips tightens.

DeepSeek said it is working closely with Chinese hardware partners. The briefing stated that Huawei-made AI chips perform at about one-quarter of Nvidia’s speed, while adding that adaptation work has been feasible. The company identified production volume, rather than software compatibility or basic chip capability, as the main constraint.

Liang said DeepSeek expects domestic AI hardware capacity to meet demand within five years. That forecast is ambitious, but it also shows how deeply the company’s roadmap is tied to China’s semiconductor supply chain. A model developer can optimize software aggressively, but repeated frontier-scale training still requires chips, power, networking equipment and data center capacity.

Roadmap moves from reasoning models to embodied AI

DeepSeek’s technical roadmap starts with chain-of-thought reasoning and moves toward agents, continual-learning models, what the company called iterative AI singularity, and eventually embodied AI. Chain-of-thought reasoning refers to models showing intermediate steps while solving complex tasks. Agents are AI systems designed to plan and take actions across software tools with less human intervention.

The company said its current focus is a coding agent that can autonomously improve software and research efficiency. If successful, that would support a common frontier-lab strategy: using AI systems to accelerate the development of later AI systems. DeepSeek did not provide performance benchmarks for the coding agent in the briefing.

The next-generation V4 model series is being developed with integrated multimodal capabilities, according to the company. Multimodal systems process more than one type of input, such as text, code, images, audio or video. DeepSeek said it views sustained learning ability as more important than incremental efficiency gains, suggesting that the company is aiming beyond lower-cost model training toward systems that can adapt over longer operating periods.

DeepSeek also said it plans to open-source its strongest models and help other research teams, including technical competitors, deploy them. The company argued that open releases do not necessarily weaken commercial viability. That position aligns with DeepSeek’s public image as a lab seeking influence through model accessibility, though large-scale open-source releases can increase pressure to monetize through APIs, enterprise services or specialized deployments.

Culture remains part of the pitch

Liang described DeepSeek as a profit-oriented company with selective commercial targets. He said the firm chooses clients and partners whose goals align with its long-term artificial general intelligence strategy, rather than pursuing every available revenue opportunity.

The company said it operates without KPI-based management and leaves roughly half of each employee’s time for self-directed exploration. DeepSeek described team stability and long-term continuity as priorities, saying those conditions are necessary for complex research programs.

Internal decision-making follows consensus rather than top-down direction, according to Liang. Researchers can initiate projects independently, and employees generally keep relaxed schedules with limited after-hours demands. The company presents that structure as a way to preserve research efficiency, not as a workplace branding exercise.

DeepSeek also identified data labeling as a major expense. The company said half of its researchers currently work on data annotation and that it prioritizes datasets with low cost and high expected impact. It described high-quality labeling as limited mainly by time and capital, not by technical ability.

Token-market claims need sharper evidence

The financing is likely to draw attention from crypto traders focused on decentralized compute networks, because DeepSeek’s briefing reinforces a real pressure point: advanced AI development needs more chips, power and data center capacity than many organizations can obtain cheaply. Tokenized compute projects often claim they can aggregate idle GPUs from distributed users and sell that capacity at lower cost than centralized cloud providers.

DeepSeek’s financing materials, as described in the briefing, did not identify tokenized compute networks as part of its hardware strategy. The company instead emphasized Nvidia GPUs, its TileLang compiler, domestic chip adaptation and cooperation with Chinese hardware manufacturers. That distinction matters. A general shortage of AI compute does not automatically translate into demand for blockchain-based compute markets.

Claims that traders should aggressively buy utility tokens tied to shared computing networks require evidence beyond the existence of high GPU prices or long chip delivery times. The relevant questions are specific: whether AI developers are actually using those networks for production workloads, whether the networks can deliver reliable performance, whether revenue comes from real compute buyers, and whether token value is directly linked to usage rather than speculation.

DeepSeek’s disclosure strengthens the case that compute access is a decisive competitive variable in AI. It does not, on its own, validate token-market narratives built around decentralized GPU supply. The more immediate read-through is narrower and more concrete: DeepSeek is raising capital because efficiency gains have limits when frontier model development still depends on scarce, expensive hardware.