DeepSeek, the Hangzhou-based artificial intelligence lab that stunned Silicon Valley last year with an open-weight reasoning model trained for a fraction of frontier costs, is in advanced talks to raise at least $300 million in its first-ever external funding round at a valuation exceeding $10 billion, according to reports published on April 17, 2026 by The Information and corroborated by Reuters, TechCrunch, and Investing.com. The round, confirmed by people familiar with the discussions on April 23, 2026, marks a sharp philosophical pivot for a company that spent its first two-and-a-half years refusing every venture check offered to it by Chinese tech giants and state-linked funds.
The $10 billion target represents a nearly threefold jump from the $3.4 billion valuation DeepSeek carried on secondary markets in 2025, and it lands at a moment when global AI capital has concentrated in American labs. OpenAI is marked at $500 billion in recent secondary trades and Anthropic sits near $350 billion following its rejected tender offer. Yet the more revealing number is not the price tag. It is the fact that founder Liang Wenfeng, who has personally funded DeepSeek through his quantitative hedge fund High-Flyer Capital Management since 2023, has finally concluded that reasoning models, agent systems, and the compute behind them cost more than any single hedge fund can sustain.
This analysis unpacks the DeepSeek funding round’s structure, the technical and financial pressures that forced the shift, how the $10 billion price compares against global AI peers, and what the deal means for the US-China AI race, export controls, and the open-weight movement that DeepSeek has come to symbolize.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
DeepSeek’s $300M First Funding Round: What We Know
According to The Information’s April 17, 2026 report, DeepSeek is seeking a minimum of $300 million in primary capital, with the total round size potentially expanding depending on investor demand. The $10 billion pre-money valuation places DeepSeek firmly in the decacorn category, though still an order of magnitude below the American frontier labs. No investors have been officially named, but people briefed on the talks have pointed to a mix of Chinese strategic backers and offshore sovereign capital. Reports from Investing.com and TechStartups on April 17 confirmed that High-Flyer Capital Management will remain the controlling shareholder after the round closes.
Liang Wenfeng, the 40-year-old founder, told Chinese state media earlier in 2026 that “China’s AI industry needs confidence as much as capital,” a framing now reflected in the decision to open the cap table. The round is expected to close in the second quarter of 2026, though the company has not confirmed an official timeline. DeepSeek declined multiple requests for comment.
The $10 Billion Valuation in Context: How DeepSeek Stacks Up
A $10 billion valuation is an awkward number in 2026’s AI capital landscape. It is large enough to place DeepSeek among the most valuable privately held AI companies in the world, yet small enough that the gap to OpenAI and Anthropic is now measured in multiples rather than percentages. The table below compares the leading frontier AI labs by their most recent public or reported valuation.
| Company | Valuation (2025-2026) | Country | Last Round Type | Model Licensing |
|---|---|---|---|---|
| OpenAI | ~$500 billion | United States | Secondary tender | Proprietary |
| Anthropic | ~$350 billion (rejected) | United States | Tender offer | Proprietary |
| xAI | ~$200 billion (reported) | United States | Acquisition by SpaceX | Proprietary + open weights |
| Mistral AI | ~$15 billion | France | Debt + equity | Open weights + API |
| DeepSeek | ~$10 billion (target) | China | First primary round | Open source (MIT) |
| Moonshot AI | ~$3.3 billion | China | Series C | Proprietary |
| Zhipu AI | ~$2.7 billion | China | Strategic round | Hybrid |
The comparison illustrates DeepSeek’s unique position. It commands the highest valuation of any Chinese AI lab but trails American peers by roughly 50x. More telling is the licensing column. DeepSeek is the only company in the top tier releasing its flagship model under the permissive MIT license, a stance that has driven enterprise adoption across Asia, Europe, and increasingly the United States despite Washington’s policy skepticism.
How High-Flyer’s 56.6% Return Funded DeepSeek for 30 Months
To understand why DeepSeek could refuse outside capital for so long, it helps to understand High-Flyer Capital Management. Founded by Liang in 2015, High-Flyer is a quantitative hedge fund that manages assets across Chinese equity and futures markets using deep-learning models. In 2025 the fund posted a reported 56.6% return, generating enough operating cash flow to bankroll DeepSeek’s training runs, salaries, and data center leases without dilution.
That arrangement was rare but not unprecedented. Renaissance Technologies famously funded internal research for years without external investors. What is different with DeepSeek is the capital intensity of the product. Training a frontier reasoning model costs tens to hundreds of millions of dollars per run, and serving it to consumers at sub-cent pricing requires continuous GPU purchases and data center expansion. Even a $1 billion hedge fund profit stream cannot absorb that indefinitely while also running a quantitative trading book.
Liang has publicly emphasized that DeepSeek rejected prior offers because it wanted to protect a research-first culture unencumbered by quarterly revenue pressure. The pivot to external capital in April 2026 reflects a judgment that the next model generation cannot be built within the cash flow of a hedge fund alone.
DeepSeek R1 and V3: The Models That Shook the Market
The reason investors are now willing to price DeepSeek at $10 billion comes down to two model releases: DeepSeek V3 in December 2024 and DeepSeek R1 in January 2025. V3 was a 671-billion-parameter mixture-of-experts model that DeepSeek claimed was trained for roughly $5.6 million in GPU rental cost, a figure that triggered intense debate but was not convincingly refuted. R1 followed as a reasoning-tuned variant that applied reinforcement learning on top of V3 to produce chain-of-thought outputs comparable to OpenAI’s o1.
When R1 launched on January 20, 2025, its weights were released under the MIT license and accompanied by a technical paper published on arXiv. Within a week the DeepSeek mobile app topped the Apple App Store free rankings in the United States, surpassing ChatGPT for the first time any Chinese AI product had done so. On January 27, 2025, the market reaction was severe. Nvidia shares fell roughly 17% in a single session, erasing nearly $600 billion in market capitalization, as traders questioned whether American hyperscalers had overbuilt GPU capacity for a training paradigm that could be matched with orders of magnitude less compute.
The panic was short-lived. Within weeks Nvidia recovered much of the loss as buyers recognized that even efficient training still requires enormous inference capacity. But R1 left a durable mark. It proved that a non-American lab with restricted access to H100 chips could reach frontier performance, and it reset expectations about the floor on training cost.
Benchmark Performance: DeepSeek R1 vs Frontier Rivals
Published benchmark results from the DeepSeek R1 technical report, later verified by independent evaluators, put the model within a few percentage points of GPT-4 and o1-preview on reasoning-heavy tests. The table below compares R1 and its successor R1-Zero against contemporary American and Chinese frontier models on the most widely cited benchmarks.
| Benchmark | DeepSeek R1 | OpenAI o1 | Claude 3.5 Sonnet | Qwen 2.5-Max | Meaning |
|---|---|---|---|---|---|
| MATH-500 | 97.3% | 96.4% | 78.3% | 94.5% | Graduate math reasoning |
| AIME 2024 | 79.8% | 79.2% | 16.0% | 23.3% | Competition math |
| MMLU | 90.8% | 91.8% | 88.3% | 87.9% | General knowledge |
| SWE-bench Verified | 49.2% | 48.9% | 50.8% | 40.0% | Real-world coding |
| GPQA Diamond | 71.5% | 75.7% | 65.0% | 60.1% | Graduate science |
| Codeforces Elo | 2029 | 2061 | 717 | 1327 | Competitive programming |
The numbers matter for the funding conversation because they answer the obvious investor question: is DeepSeek actually competitive, or is it a clever replication? The data says competitive. R1 edges o1 on MATH-500 and AIME while trailing slightly on GPQA Diamond and Codeforces, a spread that is within the error bars of most evaluation pipelines. For deeper side-by-side analysis, see our DeepSeek vs ChatGPT 2026 comparison.
The Cost Advantage: $0.55 vs $15 Per Million Tokens
Perhaps the most disruptive element of the DeepSeek story is pricing. Running R1 through the DeepSeek API costs roughly $0.55 per million input tokens and $2.19 per million output tokens, compared with $15 per million input and $60 per million output for OpenAI’s o1 at launch. That is a 27x spread on input and a 27x spread on output, sustained across a model with broadly comparable benchmark performance.
The cost gap reflects three structural factors. First, DeepSeek’s mixture-of-experts architecture activates only 37 billion of its 671 billion parameters per token, reducing GPU memory pressure during inference. Second, DeepSeek trained V3 on a reported cluster of 2,048 H800 GPUs, a cut-down version of the H100 sold legally into China before the October 2023 export control tightening. Third, Liang’s team optimized the training stack aggressively, using FP8 mixed-precision training and custom communication kernels to extract more tokens per GPU-hour than most Western labs have reported.
None of these optimizations are secret. DeepSeek published the architecture details in open technical reports, which is part of why Meta, Mistral, and others have since adopted similar MoE designs. The cost advantage may narrow over time, but for 2026 it remains a defining competitive edge.
Why DeepSeek Is Raising Capital Now
Three forces converged in early 2026 to push DeepSeek toward external capital. The first is the arms race in reasoning models. OpenAI, Anthropic, Google DeepMind, and xAI each released next-generation systems in the first quarter of 2026, and the training runs behind them reportedly consumed tens of thousands of GPUs over months. DeepSeek’s next model, widely expected to be designated V4 or R2, will require compute at a scale that stretches even High-Flyer’s balance sheet.
The second is agent infrastructure. The frontier has shifted from chat models to autonomous agents that plan, call tools, and execute tasks across sessions. Building production-grade agent systems requires specialized training data, evaluation pipelines, and inference infrastructure that DeepSeek has so far deprioritized. Closing that gap means hiring aggressively and purchasing significant new GPU capacity.
The third is talent retention. With Anthropic paying staff researchers reported salaries above $1 million and OpenAI matching or exceeding that tier, Chinese labs face compensation pressure that equity upside can help address. A priced external round establishes a clearer market value for DeepSeek equity, which makes employee stock options meaningful in negotiations.
Export Controls and the Geopolitical Backdrop
DeepSeek operates under the shadow of US export controls that have progressively restricted Chinese access to leading-edge AI chips since October 2022. The October 2023 update banned H800 sales into China, the January 2025 AI Diffusion rule imposed country-tier caps on advanced GPU exports, and the May 2025 update further tightened licensing for Chinese data centers. CSIS analysis published in early 2026 argued that DeepSeek’s efficiency gains have been interpreted in Washington as evidence that the controls are either insufficient or being circumvented.
The practical effect on DeepSeek is twofold. On one hand the company has credibly demonstrated that frontier performance is reachable with cut-down hardware, which strengthens its strategic position inside China. On the other hand any new round raises questions about whether US-linked capital can participate and whether DeepSeek models can continue to be hosted on American clouds. Several US hyperscalers quietly removed DeepSeek R1 from their managed model catalogs in the second half of 2025 after political pressure, while leaving it available on bring-your-own-weights configurations.
The funding round is also a signal to Beijing. By accepting outside investors for the first time, Liang is effectively opening DeepSeek to strategic scrutiny from Chinese regulators and industrial policy planners, a shift that mirrors what happened to Alibaba and ByteDance as they matured. For the broader context on how capital is reshaping the AI chip stack, see our coverage of Huawei’s Ascend 950PR chip.
Expert Reactions: Analysts and Researchers Weigh In
The response to the DeepSeek fundraise has been mixed across the analyst community. Dan Wang, a China tech scholar at the Hoover Institution, commented in an April 2026 note that “the DeepSeek round is the end of the self-funded founder-purist era in Chinese AI. From here on every serious lab in China will operate on external capital, which means external governance.” Wang’s view aligns with what Liang himself has acknowledged in public remarks this year.
Gil Luria, head of technology research at D.A. Davidson, told TechStartups on April 17 that “$10 billion is a conservative mark against what DeepSeek could have commanded. The founder is clearly prioritizing alignment of investors over maximizing the price.” Luria noted that a more aggressive round could have cleared $20 billion given current secondary-market demand for AI exposure.
Clement Delangue, CEO of Hugging Face, posted on X on April 18 that “DeepSeek raising external capital does not change the open-weight commitment. The team has been consistent that releasing under MIT is core to their strategy, and investors coming in now understand that.” Delangue’s platform has hosted more than 10 million downloads of DeepSeek R1 derivatives since the January 2025 release.
Paul Triolo, partner at DGA-Albright Stonebridge Group, cautioned in remarks to Bloomberg that “the capital unlock for DeepSeek will trigger a wave of follow-on rounds at Moonshot, Zhipu, and Baichuan that Beijing may choose to coordinate through the state-owned fund network. The strategic question is whether this becomes organic or steered.” Triolo tracks US-China technology policy and has previously flagged DeepSeek as a pivotal test case for export controls.
Gary Marcus, AI researcher and NYU professor emeritus, offered a more skeptical take on his Substack on April 20: “A $10 billion valuation for any pure-play AI lab assumes a path to durable revenue that does not yet exist for DeepSeek. The company charges almost nothing per token and releases weights for free. That is defensible as a research mission but not as a commercial thesis at this price.”
The Competitive Landscape: China’s AI Market in 2026
DeepSeek is one of roughly half a dozen Chinese labs pursuing frontier AI in 2026. The competitive set includes Moonshot AI (backed by Alibaba and Tencent), Zhipu AI (state-adjacent with strategic investment from Saudi Arabia’s Prosperity7), Baichuan, 01.AI, and the internal labs at Alibaba (Qwen team), ByteDance (Doubao), Baidu (Ernie), and Tencent (Hunyuan).
Alibaba’s Qwen series is arguably the most capable Chinese model family on published benchmarks, with the Qwen 2.5-Max variant landing within a few points of Claude and GPT-4 on MMLU and SWE-bench. Qwen has the advantage of Alibaba Cloud distribution and enterprise contracts across Southeast Asia. DeepSeek counters with reasoning performance, open weights, and a significantly lower API price.
Moonshot’s Kimi K1.5 matched R1 on several benchmarks in early 2025 but has not been released open-weights, limiting its adoption. Zhipu’s GLM-4 series is strong in Chinese-language tasks but trails on English-language reasoning. The net result is that DeepSeek holds the most defensible position among the pure-play labs going into the fundraise, which is part of why $10 billion is readable as a reasonable rather than aggressive mark.
Open Source Strategy and the MIT License Commitment
Every DeepSeek flagship model to date has shipped with weights available under the MIT license, an unusually permissive stance even within the open-source AI community. Meta’s Llama models use a custom license with use-case restrictions, Mistral ships its most capable models under commercial license, and Google’s Gemma requires acceptance of a custom acceptable-use policy. DeepSeek’s MIT choice means organizations can fine-tune, redistribute, and commercialize the models without royalty or restriction.
The open-weight posture drives three commercial effects. It accelerates enterprise adoption because legal reviews are frictionless. It generates a distributed talent funnel as researchers worldwide build on DeepSeek weights and often cite the team in publications. And it creates defensibility against API price wars because customers who run the model on their own hardware are insulated from future price changes.
The question for new investors is whether the MIT stance survives dilution. Liang has publicly committed to maintaining it, and reports suggest the term sheet will include covenants protecting the open-weight release policy for flagship models. Enforcement of such covenants is unusual but not unprecedented in AI-focused rounds.
Five Predictions for DeepSeek and Chinese AI After This Round
Based on the structure of the round, the competitive environment, and the historical arc of Chinese AI companies, five outcomes look probable over the next 12 to 24 months.
- Follow-on round within 12 months. A $300 million first raise at $10 billion is a starter, not a war chest. Expect a Series B at $20 to $25 billion by early-to-mid 2027, coinciding with the launch of DeepSeek’s next frontier model.
- V4 or R2 ships in Q3 2026. DeepSeek’s cadence has been one major release every 9 to 12 months. The next generation will likely combine R1’s reasoning techniques with larger context, agentic tool use, and multimodal inputs.
- State-linked capital participates. While the first $300 million may come from private investors, a follow-on round at scale will almost certainly include China’s national AI fund or a provincial equivalent, mirroring how SMIC and YMTC were backstopped.
- US cloud availability narrows further. Political pressure will push remaining US hyperscalers to delist DeepSeek models from managed catalogs, even as bring-your-own-weights and self-hosted deployments continue to grow.
- Open-weight consolidation. DeepSeek, Meta Llama, and Mistral will emerge as the three durable open-weight providers, with other entrants either consolidating or shifting to proprietary strategies.
Market Impact: Ripple Effects Across AI Infrastructure
The DeepSeek round arrives during an AI capital cycle that is already distorted. US venture capital deployed $297 billion in the first quarter of 2026, with 81% of the dollars going to AI, according to data covered in our Q1 2026 venture capital analysis. Against that backdrop a $300 million Chinese round is a rounding error, but the reputational and strategic signal is outsized.
For Nvidia the round is mildly bearish. If DeepSeek can extend its cost leadership with new capital, the argument that frontier AI requires tens of billions in GPU spend becomes harder to sustain. Nvidia’s response has been to accelerate the rollout of its Vera Rubin platform, which targets inference efficiency rather than raw training throughput.
For hyperscalers the effect is mixed. AWS, Azure, and Google Cloud all benefit when enterprises run DeepSeek on their infrastructure because it drives compute consumption, but they lose margin versus hosting proprietary models with premium pricing. Expect a quiet shift toward tiered pricing that rewards customers bringing their own weights.
Historical Context: From Sputnik Moment to Strategic Asset
When R1 launched in January 2025 it was described as a “Sputnik moment” for American AI, language that originated with commentators in Silicon Valley and spread to Washington within days. The comparison was imprecise. Sputnik represented a surprise capability leap by a strategic adversary. R1 represented a cost-and-openness leap by a commercial lab that happened to be headquartered in China.
What has changed in the fifteen months since is the integration of DeepSeek’s techniques into the broader AI ecosystem. Meta adopted mixture-of-experts patterns similar to DeepSeek’s in its Llama 4 training. OpenAI open-weight releases in 2025 were in part a response to pressure from DeepSeek’s R1. Anthropic’s pricing on Claude 3.5 and Claude Opus 4 moved to compete more directly with the DeepSeek API tier. The market absorbed the shock and incorporated the techniques.
The funding round formalizes what was already clear. DeepSeek is a strategic asset, for China, for the open-weight community, and for any enterprise that wants use in vendor negotiations with American frontier labs. Pricing that asset at $10 billion is a starting position, not a ceiling.
Risks and Open Questions
Several risks could undermine the round’s trajectory. Regulatory risk is the most immediate. If the Trump administration or Congress moves to restrict US persons from investing in or purchasing services from Chinese AI companies, DeepSeek’s international revenue and cloud availability could contract sharply. The January 2025 AI Diffusion rule already created a precedent for country-tier restrictions that could be expanded.
Technical risk matters too. The cost advantage that defines DeepSeek depends on continued algorithmic innovation. If American labs match the efficiency gains, which several are publicly attempting, the API price spread will compress. That would weaken the economic case for the $10 billion valuation without a clear path to premium monetization.
Talent risk is underappreciated. DeepSeek’s core research team numbers fewer than 200 people. Aggressive recruitment by Alibaba, Tencent, and American labs with local presence could erode the team’s coherence. A priced round helps with retention but does not eliminate the exposure.
What This Means for Enterprise AI Buyers
For technology leaders at enterprises evaluating AI platforms, the DeepSeek round reinforces three strategic considerations. First, multi-model strategies are now table-stakes. Relying on a single frontier API exposes organizations to pricing and availability shocks that the DeepSeek effect has proven are possible. Second, open-weight options need to be evaluated in parallel with proprietary APIs. Running DeepSeek R1 on self-managed infrastructure can cut inference costs by an order of magnitude for high-volume workloads.
Third, geopolitical exposure must be factored into AI vendor selection. Hosting DeepSeek models in US production environments carries regulatory risk that did not exist with earlier open-weight models from Meta or Mistral. Enterprises with US federal contracts, or in regulated industries, should anticipate restrictions.
Our Anthropic vs OpenAI 2026 analysis and Claude vs ChatGPT 2026 comparison provide deeper coverage on how the American frontier labs are positioning against these dynamics.
Frequently Asked Questions
How much is DeepSeek raising in its 2026 funding round?
DeepSeek is seeking at least $300 million at a valuation exceeding $10 billion, according to reports from The Information and Reuters published on April 17, 2026. The round is in advanced discussions but has not officially closed as of April 24, 2026.
Who founded DeepSeek and who owns it?
DeepSeek was founded in 2023 by Liang Wenfeng in Hangzhou, China. It is owned by High-Flyer Capital Management, the quantitative hedge fund Liang founded in 2015. High-Flyer is expected to remain the controlling shareholder after the 2026 round closes.
How does DeepSeek’s $10 billion valuation compare with OpenAI and Anthropic?
DeepSeek’s $10 billion target trails OpenAI at roughly $500 billion and Anthropic at roughly $350 billion by about 50x and 35x respectively. DeepSeek is still the most valuable privately held Chinese AI lab and the highest-valued open-source frontier AI company globally.
Is DeepSeek R1 really as good as OpenAI o1?
On published benchmarks including MATH-500, AIME 2024, and MMLU, DeepSeek R1 scores within a few points of OpenAI o1. R1 edges o1 on some math benchmarks and trails slightly on science and coding tests. Real-world performance depends on use case, latency requirements, and deployment constraints.
Will DeepSeek keep releasing models under the MIT license?
Liang Wenfeng has publicly committed to maintaining the MIT open-weight release strategy, and reports suggest the 2026 term sheet includes covenants protecting that policy. Investors participating in the round are understood to accept the open-weight strategy as part of the deal.
Can US companies invest in DeepSeek?
US person participation in the DeepSeek round is subject to Treasury Department outbound investment restrictions that took effect in January 2025. The restrictions cover certain AI investments by US persons into Chinese entities. Most of the round is expected to be closed with Chinese and non-US international investors to avoid these complications.
How did DeepSeek train its models so cheaply?
DeepSeek reports using a mixture-of-experts architecture with 671 billion total parameters but only 37 billion activated per token, FP8 mixed-precision training, and a cluster of approximately 2,048 H800 GPUs. The V3 training run was reported at $5.6 million in GPU rental cost, a figure that has been debated but not disproven.
What happened to Nvidia stock when DeepSeek R1 launched?
On January 27, 2025, Nvidia shares fell approximately 17% in a single session, erasing nearly $600 billion in market value, as traders questioned whether hyperscaler GPU buildouts had been overbuilt relative to DeepSeek’s demonstrated training efficiency. The stock recovered much of the loss within weeks.
Where can I try DeepSeek’s models?
DeepSeek’s chatbot is available at the DeepSeek website and through the iOS and Android apps. Model weights are hosted on Hugging Face under the deepseek-ai organization, and the API is accessible through DeepSeek’s platform with competitive per-token pricing.
External Sources
- DeepSeek official website
- DeepSeek on Hugging Face
- DeepSeek R1 GitHub repository
- DeepSeek R1 technical paper on arXiv
- CSIS analysis: DeepSeek, Huawei, and US-China AI race
- The Information
- TechCrunch
Last updated: April 24, 2026
