DeepSeek, the Hangzhou AI lab founded by hedge-fund billionaire Liang Wenfeng, is finalizing its first external funding round — targeting up to $4 billion at a valuation that has climbed fivefold to $50 billion in under a month — after the lab's April 24 release of V4, a 1.6-trillion-parameter model engineered to cut the memory and compute costs that have made long-context AI prohibitively expensive. The round matters beyond DeepSeek: China's state semiconductor and AI apparatus is effectively co-signing a bet that cheap, efficient open-weight models can displace mid-tier proprietary AI, with real consequences for every business — and every developer — now pricing AI infrastructure.
State Capital Moves In Three Weeks After V4 Ships
According to reporting by Reuters, the Financial Times, and the South China Morning Post, DeepSeek is close to closing its first outside financing round. The lead investor is China's National Artificial Intelligence Industry Investment Fund, an $8.8 billion state-backed vehicle established in early 2025. The China Integrated Circuit Industry Investment Fund — known as Big Fund III — is also involved, alongside Tencent Holdings. Hillhouse Capital was reported in discussions, but a Hillhouse spokesperson told the SCMP the firm is not participating. The final terms remain subject to change.
The speed of the valuation escalation is striking. Early April reports cited a target of $300 million and a $10 billion valuation. By the week of May 8, three sources familiar with the discussions told Reuters the figure had reached $3–$4 billion at up to $50 billion — a fivefold jump in roughly four weeks. The stated trigger was mundane: Liang opened the round to offer employees equity after rivals began poaching DeepSeek researchers, including Luo Fuli, who departed to lead Xiaomi's MiMo model team. The strategic subtext is harder to dismiss.
Big Fund III was created to advance China's semiconductor self-sufficiency. It has previously backed chip manufacturers SMIC and Yangtze Memory Technologies. Backing DeepSeek would be the first known Big Fund investment in a Chinese large language model company — a signal, analysts say, that Beijing now views frontier AI software and domestic silicon as a single strategic problem.
V4's Architecture Is the Investment Thesis
DeepSeek released V4 on April 24, 2026, in the same week as OpenAI's GPT-5.5 and Anthropic's Claude Opus 4.7. It comes in two open-weight, MIT-licensed variants: V4-Pro, a 1.6-trillion-parameter mixture-of-experts model that activates 49 billion parameters per token, and V4-Flash, a leaner 284-billion-parameter version. Both ship with a one-million-token context window as the default — not a premium tier.
That default matters economically. Long-context AI has been technically possible for years, but expensive, because the attention mechanism's key-value cache grows with sequence length and consumes GPU memory. DeepSeek's answer is a hybrid attention design combining Compressed Sparse Attention, which groups tokens and retrieves only the most relevant chunks, and Heavily Compressed Attention, which condenses large spans into single summary representations, alongside an uncompressed sliding window for nearby context. The result, according to DeepSeek's own documentation: at a full one-million-token context, V4-Pro needs only 27% of the per-token inference compute and 10% of the KV-cache memory of its predecessor, V3.2.
On benchmarks, the honest picture is mixed. V4-Pro posts strong coding figures — 80.6% on SWE-Bench Verified, the highest of any open-weight model at launch, according to independent tracking by Morph. But the government's own evaluator puts the overall capability gap wider than DeepSeek acknowledges. NIST's Center for AI Standards and Innovation (CAISI), using nine benchmarks including two non-public held-out tests, concluded in May 2026 that V4-Pro trails leading US models by approximately eight months — not the two-to-three months DeepSeek claims based on public leaderboards. CAISI called V4-Pro "the most capable PRC AI model" it has evaluated to date.
A DeepSeek developer using the pseudonym Ex0bit disputed the finding publicly, writing: "There's no 'gap', and no one's 8 months behind." Independent trackers show the US-China gap on public leaderboards has narrowed to roughly 2.7%, according to a 2026 AI index from an academic institution cited by GNCrypto News. Critics of CAISI's report noted it relied on private datasets and a specific choice of comparators. CAISI said it would publish its full methodology.
The Hardware Question That Defines the Geopolitics
The most politically loaded element of V4 is what it runs on. DeepSeek gave Huawei early access to optimize for V4 while, according to The China Academy, pointedly denying that courtesy to Nvidia and AMD. Huawei confirmed its full Ascend SuperNode product line supports V4 for inference. ByteDance, Tencent, and Alibaba subsequently rushed to place orders for Huawei's Ascend 950 chips, with SMIC shares jumping 10% in Hong Kong trading after the V4 launch.
Whether V4-Pro was trained entirely on Chinese silicon remains disputed. A senior Trump administration official alleged the model used Nvidia Blackwell chips that had been smuggled into China, which would constitute a violation of US export controls. DeepSeek denied this, stating it used H800 GPUs and Huawei Ascend 910C chips. DeepSeek's own paper, as noted by The Register, confirms only that the model was validated on both Nvidia and Huawei platforms. What is confirmed: DeepSeek has acknowledged V4-Pro throughput will remain constrained until Huawei's Ascend 950PR supernodes ship in volume in the second half of 2026.
Nvidia CEO Jensen Huang addressed the stakes directly. Speaking on the Dwarkesh Podcast, Huang warned that if DeepSeek optimized its models for Huawei chips rather than American hardware, it would be "a horrible outcome" for the United States, adding it could lead to a world where "AI models worldwide will be developed and perform best not on American hardware."
He Hui, director of semiconductor research at Omdia, offered a cleaner summary: "Huawei's Ascend chips are the country's best homegrown alternative to Nvidia, and supporting DeepSeek V4 shows that top Chinese AI models can now run on Chinese hardware." The irony, as Capacity Media reported, is that the same US export controls intended to slow Chinese AI are now constraining Huawei's ability to manufacture enough Ascend chips to meet the demand they helped create.
US Congress Called DeepSeek a "Profound Threat." Its Users Face Real Risks.
For any business or developer considering DeepSeek's cheap API, the security and privacy record is not theoretical. In April 2025, the House Select Committee on China called DeepSeek "a profound threat to our nation's security," citing corporate filings showing links between Liang Wenfeng's entities and state-linked hardware distributors and research institutes. The committee alleged DeepSeek was powered by at least 60,000 Nvidia processors and may have circumvented US export controls in acquiring them.
Security researchers have identified specific technical harms to users. Cybersecurity firm Wiz discovered a publicly accessible DeepSeek database containing over one million sensitive records — including user chat histories, API keys, and backend logs — with no authentication or access controls. Mobile security firm NowSecure found the app used hardcoded encryption keys and transmitted some device and user data without encryption. Researchers at Feroot Security found hardcoded links in DeepSeek's web login page connecting directly to China Mobile, a company the Federal Communications Commission has banned from operating in the United States.
Douglas Schmidt, dean of the School of Computing, Data Sciences and Physics at William & Mary, described the data risk as "TikTok on steroids," explaining that unlike a social media app, an AI assistant "gets a much deeper insight into the personalities and interests and hopes and dreams of the users." DeepSeek's privacy policy, cited in a Time investigation, states that user data — including chat history, keystroke patterns, and IP addresses — is stored on servers in China, where national intelligence law requires companies to cooperate with government requests.
Italy's data protection authority, the Garante, imposed an emergency ban on DeepSeek on January 30, 2025 — the first such action against an AI chatbot under GDPR — after DeepSeek provided what regulators called a "completely insufficient" response to questions about its data practices, and claimed EU law did not apply to it. Data protection authorities in France, Ireland, Belgium, the Netherlands, Germany, and the Czech Republic have since launched formal investigations. Bipartisan US legislation, the No DeepSeek on Government Devices Act (H.R. 1121 / S. 765), is pending in both chambers.
Anthropic and OpenAI Accuse DeepSeek of Systematic Model Theft
The state-backed investment arrives after DeepSeek faced a separate front of accusations from US rivals. In February 2026, Anthropic alleged publicly that DeepSeek, Moonshot AI, and MiniMax had created more than 24,000 fraudulent accounts and generated over 16 million interactions with its Claude model in violation of Anthropic's terms of service — a technique called distillation, in which a rival trains its own model on another model's outputs. OpenAI had made similar allegations earlier the same month.
Anthropic framed the alleged extraction as a national security issue, warning that models trained this way could enable "authoritarian governments to deploy frontier AI for offensive cyber operations, disinformation campaigns, and mass surveillance." DeepSeek has not publicly responded to the allegations.
Critics pushed back on the framing. Erik Cambria, professor of artificial intelligence at Singapore's Nanyang Technological University, told CNBC that nuance is needed: "the boundary between legitimate use and adversarial exploitation is often blurry." Multiple commentators noted that Anthropic itself settled a lawsuit for $1.5 billion in September 2025 over bulk extraction of copyrighted books for training data, complicating any clean narrative about intellectual-property norms in AI development. DeepSeek has not announced any legal action in response.
What This Round Means for Businesses Using AI Today
The practical stakes for any organization evaluating AI infrastructure are direct. V4-Pro's list price sits at approximately $3.48 per million output tokens, with a promotional rate near $0.87 running through end of May; V4-Flash costs roughly $0.28 per million output tokens at list, according to DataCamp's pricing review. When a one-million-token context is the default rather than a premium add-on, the cost of running a model continuously — as autonomous AI agents do — drops materially compared with Western frontier APIs. The V4 release already ships with integrations for Claude Code and the OpenClaw agent.
The argument circulating among investors is that capable open-weight models create a pricing ceiling for mid-tier proprietary models. If an open model can complete a large share of enterprise tasks at a fraction of the cost, the pricing power of closed mid-tier competitors diminishes. This is an investment thesis, not a proven outcome: frontier closed models retain leads in the most demanding reasoning and agentic tasks, and enterprise buyers weigh security, regulatory compliance, and vendor support alongside token price — considerations where DeepSeek's privacy record, documented above, becomes directly relevant.
The security concerns are not abstract for enterprise buyers. Any organization in a regulated industry — healthcare, finance, legal — that routes sensitive data through DeepSeek's API is exposing that data to servers in China under Chinese intelligence law. John Scott-Railton, a senior researcher at Citizen Lab, made the point plainly in a January 2025 interview with Time: "It shouldn't take a panic over Chinese AI to remind people that most companies in the business set the terms for how they use your private data." The state investment now funding DeepSeek's infrastructure expansion does not change that legal reality.
A June Model Release and Unresolved Questions
DeepSeek has signaled further near-term activity. An upgraded V4.1 model is expected in June 2026, according to Dataconomy's reporting on the funding round. The lab has also indicated multimodal support is in development for V4, addressing the persistent complaint from developers that V4-Pro currently lacks vision capabilities at launch.
What the state funding does not resolve: the gap CAISI measured, the unresolved distillation allegations, the hardware dependency on Ascend chips that won't ship at volume until late 2026, and the data-security exposure that has already produced government bans across multiple continents. What it does confirm is the structure of the wager. A lab that built near-frontier AI cheaply is taking state money — at a valuation that multiplied fivefold in a month — immediately after releasing an architecture designed to reduce dependence on the hardware US export controls restrict. Enterprise buyers, developers, and policymakers now have several billion state-backed dollars as a signal of how seriously Beijing is taking that bet.
ⓒ 2026 TECHTIMES.com All rights reserved. Do not reproduce without permission.
