DeepSeek

DeepSeek is a Chinese AI company, backed by the quantitative hedge fund High-Flyer, that builds frontier-competitive models with substantially less compute than US counterparts, using architectural innovations (MoE, MLA, GRPO) and efficiency-focused engineering. It is frequently cited as a case study in whether compute-poor labs can compete with frontier US companies, in fast-follow dynamics, in distillation, and in US-China AI competition.

Snapshot

Valuation

Revenue

Funding rounds

Overview

DeepSeek was founded in 2023 and is headquartered in China. It is a subsidiary of the quantitative hedge fund High-Flyer, which long supplied its capital and compute. Its models are designed to be frontier-competitive while using roughly 10× less training compute than Anthropic or OpenAI (Source: epochai.substack.com), with open-weight releases a recurring feature of its strategy. The company is led by founder-CEO Liang Wenfeng.

Models

DeepSeek-R1, released January 2025, was an open-weights reasoning model that briefly matched OpenAI's o1 and prompted a stock market reaction. DeepSeek-V3 (December 2024) was a 671B-parameter mixture-of-experts model activating 37B parameters per token, treated as an exemplar of compute efficiency.

DeepSeek released V4 Pro (1.6T MoE / 49B active) and V4 Flash (284B / 13B) on April 24, 2026, both at 1M-token context, with new Compressed Sparse Attention and Heavily Compressed Attention architectures. V4 Pro was priced at $1.74/$3.48 per million tokens and V4 Flash at $0.14/$0.28, described as the cheapest in their respective classes. SemiAnalysis assessed V4 Pro as 3–6 months behind the US frontier (Source: simonwillison.net; newsletter.semianalysis.com). Full coverage is at DeepSeek V4 Pro / V4 Flash.

Writing in MIT Technology Review on April 24, 2026, Caiwei Chen framed the V4 release along three axes: a roughly 60% lower cost than V3 driven by 1.6T-MoE compression and DSA attention; validation of Chinese chip-supplier viability, with Huawei Ascend training compatibility disclosed for the first time as a deliberate design constraint; and acceleration of the Chinese open-weight ecosystem against Reflection, Llama, Qwen, and Mistral. Chen characterized the release as a policy-relevant event rather than merely a model release (Source: technologyreview.com).

The pricing posture reversed later in the year. Bloomberg reported on August 13, 2026 that DeepSeek is raising prices for its flagship V4 models by more than four times, effective August 16, 2026, ahead of a possible public offering, and that its rates remain below those of its largest rivals even after the increase (Source: bloomberg.com). The increase follows the V4-Pro-0813 release the previous day and roughly reverses the permanent cut described below, which had taken V4-Pro to a quarter of its prior rate.

The change also replaces flat API pricing with time-of-day rates. From 16:00 UTC on August 16, 2026, V4-Pro moves from $0.435 per million cache-miss input tokens and $0.87 per million output tokens to $0.66 and $1.98 off-peak, and $1.32 and $3.96 at peak. Reuters reported the changes represent increases ranging from 50% to more than 1,100% depending on model, token category and time of use (Source: venturebeat.com). Peak-versus-off-peak differentiation is a departure from the flat per-token pricing that characterized the V4 line at launch; see Inference Economics and Token Pricing.

On May 23, 2026, DeepSeek said it would make permanent a 75% price cut on its flagship V4-Pro model, cutting V4-Pro API costs to 0.025–6 yuan per million tokens (~$0.0035–$0.83), down from 0.1–24 yuan. The company declined to say whether the move reflected improved supply of Huawei Ascend 950 chips, leaving open whether the cut was driven by a reduction in compute cost or by competitive price pressure in the Chinese model market. The cut continued the V4-line pricing posture, V4 Pro and V4 Flash having already been the cheapest in their respective tiers at the April launch (Source: reuters.com). See Inference Economics and Token Pricing.

DeepSeek released DeepSeek-V4.1-Flash on September 9, 2026: a 552-billion-parameter mixture-of-experts model on a new causal encoder–decoder architecture that activates 8 billion parameters for input and 16 billion for output, with native visual understanding. The company says the model's key-value cache requires one quarter the high-bandwidth memory and one eighth the solid-state storage of the previous generation. New API pricing took effect at 04:00 UTC on September 10, with off-peak rates at 50% of peak. DeepSeek is retiring V4-Pro: from 04:00 UTC on September 14, all deepseek-v4-pro requests route to V4.1-Flash at V4.1-Flash rates until a V4.1-Pro launches (Source: deepseek.com). Full coverage is at DeepSeek-V4.1-Flash.

The memory claim is the policy-relevant part of the release. High-bandwidth memory supply is the binding constraint on Chinese AI compute production through roughly 2028 (Will Huawei catch up to Nvidia by 2030? (Epoch AI, September 2026)), so an architecture that cuts cache memory per request by a factor of four eases the constraint Chinese developers face on the inference side rather than the training side. DeepSeek's stated figures are unverified here; no technical report has been examined.

Key innovations

• MLA (Multi-head Latent Attention): a more efficient attention variant (Source: epochai.substack.com)
• GRPO: a reinforcement-learning approach that Epoch AI described as demonstrating "great research taste" (Source: epochai.substack.com)
• MoE at scale: activating only 37B of 671B parameters per token, a constraint-driven efficiency approach
• FP8 training and DualPipe: the V3 technical report describes an FP8 mixed-precision training framework, which it states it validated "for the first time" on an extremely large-scale model, and the DualPipe pipeline-parallelism schedule, which hides cross-node expert-parallel communication behind computation; V3 was trained on 2,048 NVIDIA H800 GPUs with DeepSeek's in-house HAI-LLM framework and without tensor parallelism (DeepSeek-V3 Technical Report)

The V3 report also addresses chip designers directly, asking for higher FP8 accumulation precision than the 14 bits of NVIDIA's Hopper Tensor Cores, native fine-grained quantization, and communication co-processors that would unify InfiniBand and NVLink networking (DeepSeek-V3 Technical Report). It credits 196 named contributors, 150 of them in research and engineering, and states DeepSeek's commitment to "the route of open-source models with longtermism" (DeepSeek-V3 Technical Report).

Technical papers

• DeepSeek-V3 Technical Report — arXiv (Dec 2024): 671B MoE, FP8 training, DualPipe parallelism.
• DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning — arXiv (Jan 2025; published in Nature, 2025): reasoning emerges from pure reinforcement learning.

Funding and valuation

DeepSeek had previously eschewed external funding, relying on parent High-Flyer. In May 2026 it began its first-ever outside funding round, reported on May 8 as a raise of up to roughly $7.35B (50B yuan) and described as the largest by a Chinese AI company on record, with founder-CEO Liang Wenfeng writing the biggest check himself. The Information reported that the round had prompted DeepSeek to expedite commercialization and accelerate its model-release cadence toward industry norms (Source: theinformation.com). Reuters described the shift to outside money, given the round's size, as a signal that the cost of a frontier-competitive Chinese model had grown beyond what hedge-fund cash flow could sustain even with DeepSeek's compute-efficiency advantages.

By June 3, 2026, Reuters reported the maiden external round was slated to raise roughly $7 billion at a valuation of $52–59 billion (Source: reuters.com). The same day, the South China Morning Post reported the round being finalized at over 50 billion yuan (US$7.4B) and a valuation just under $60B, six times the April mark, with committed investors Tencent (10B yuan), CATL (5B), and NetEase and JD.com (3B each), and Liang Wenfeng putting in roughly 20 billion yuan of his own capital, leaving the bulk of the round self-funded (Source: scmp.com). CATL, the Chinese EV battery maker, was first reported on May 22, 2026 to be planning to invest in the round, widening DeepSeek's backer base beyond Liang Wenfeng and High-Flyer into China's industrial-manufacturing capital (Source: theinformation.com). The round closed on June 16, 2026, raising more than 50 billion yuan (about $7.4 billion) at a valuation exceeding $50 billion, under an unusual structure in which only a Chinese state entity received voting rights (Source: reuters.com).

On July 14, 2026, DeepSeek entered preliminary talks for a second funding round at an approximately $71 billion pre-money valuation, 37 percent above the roughly $52 billion post-money figure from the first round, with the new capital earmarked for gigawatt-scale data centers and in-house inference chips (Source: techtimes.com). The Information reported the company's annualized revenue was nearing $500 million as of July 15, 2026 (Source: theinformation.com). The Wall Street Journal reported on July 14 that DeepSeek is preparing to list shares in Shanghai in 2027, seeking funds for research in its competition with Anthropic and other U.S. developers; the company is preparing financial statements by December 2026 for the offering (Source: wsj.com; theinformation.com; bloomberg.com). Founder Liang Wenfeng, who retains roughly 78 percent of the company, saw his net worth reach about $36 billion on July 14, 2026 (Source: techtimes.com).

Liang told investors at a meeting on September 20, 2026 that DeepSeek's annualized revenue run rate had reached $1 billion, more than double the rate of a few months earlier, which The Information, citing two people with direct knowledge, attributed to a recent price increase and the continued popularity of its models. DeepSeek aims to complete its second funding round by the end of October, raising 50 billion yuan ($7.5 billion) at a valuation of 500 billion yuan as it prepares to list on the Shanghai Stock Exchange. According to the same account, the August price increase raised customer costs by 2.3 to 4.5 times; Liang said more than 70% of DeepSeek's computing capacity goes to training and less than 30% to inference, and that he expects Huawei to begin delivering training chips as early as the fourth quarter of 2026. The article body is paywalled; its published abstract confirms the run-rate figure and the attribution to the price increase (Source: theinformation.com).

The second round was suspended on July 25, 2026. DeepSeek verbally told prospective investors it would not sign agreements in the coming days; it had sought at least 10 billion yuan (about $1.5 billion) at a $71 billion valuation. The suspension followed the viral circulation of a transcript attributed to Liang discussing reliance on Nvidia chips and China's lag behind the United States — comments whose authenticity has not been verified. The same report described the first round as having closed in June 2026 "at $7 billion," a figure that does not reconcile with the roughly $52 billion post-money valuation implied by the July 16 Chinese corporate filing and most likely refers to the amount raised rather than the valuation; the report does not specify which (Source: fortune.com).

Competitive position

DeepSeek trains its models with roughly 10× less compute than Anthropic or OpenAI (Source: epochai.substack.com). The Stanford HAI AI Index 2026 placed the US-China capability gap at 2.7% on benchmarks as of March 2026, with models trading the lead multiple times (Source: Stanford HAI AI Index Report 2026). Zhipu AI CEO Tang Jie offered a contrary read, stating: "The truth may be that the gap is actually widening" (Source: epochai.substack.com). Founder Liang Wenfeng gave his own assessment in a May 20, 2026 investor call whose leaked transcript circulated on July 23, 2026: in the 3-hour-44-minute session he said China's gap with the U.S. is primarily compute resources rather than talent, defended open-sourcing as DeepSeek's core strategy, and predicted consolidation among AI labs (Source: luizasnewsletter.com; aiproem.substack.com). He also gave concrete chip-access figures: DeepSeek sought 200,000 Huawei 950 chips for a frontier training run but could obtain only 16,000, against Huawei's expected 750,000-chip output for 2026 (Source: fredgao.com; transformernews.ai).

DeepSeek recurs across several competition and governance debates:

• Fast-follow. DeepSeek-R1 caught up to OpenAI's o1 within months with less training compute, cited as a case study of the fast-follow problem.
• Export controls. DeepSeek's progress despite chip restrictions is cited against the argument that export controls create durable advantages, though Epoch AI argues efficiency cannot fully bridge a 10× compute gap (Source: epochai.substack.com).
• Distillation. Anthropic accused DeepSeek of distilling from Claude's outputs; MiniMax may have obtained roughly 100B tokens from Claude interactions (Distillation).
• Supply-chain designation. In the Anthropic-DoW conflict, DeepSeek had not been designated a supply-chain risk despite links to the Chinese military, while Anthropic, an American company, was threatened with that designation.

Reuters reported on July 7, 2026, citing three people familiar with the matter, that DeepSeek is developing its own AI chip to reduce reliance on both Nvidia and Huawei silicon (Source: reuters.com). The move parallels Zhipu AI's reported inquiries about a bespoke processor the same week. See Semiconductor Supply Chain, Export Controls (AI).

Bloomberg reported on September 4, 2026, citing people familiar with the project, that DeepSeek is preparing a large order of Huawei AI chips for a data centre of roughly one gigawatt in Ulanqab, Inner Mongolia, which could begin partial operations in late 2027 or early 2028. The order size, chip model, delivery schedule and whether a contract has been signed have not been disclosed, and DeepSeek has neither confirmed the order nor acknowledged the Ulanqab project. The reporting notes that DeepSeek's V4 model, launched in April 2026, was tuned for Huawei's Ascend 950, after which Chinese cloud providers ordered hundreds of thousands of the chips, and that Inner Mongolia has signed around $27 billion in AI computing projects under the Eastern Data, Western Compute programme (Source: thenextweb.com).

If confirmed, the order would sit against Liang Wenfeng's May 2026 account of chip scarcity — seeking 200,000 Huawei 950 chips and obtaining 16,000 — and would represent a shift from constraint to committed capacity within roughly four months. It also runs in the opposite direction from the July 2026 report that DeepSeek is developing its own silicon; the two are not necessarily inconsistent, since a bespoke chip is a multi-year effort and Huawei supply addresses near-term capacity. See Export Controls (AI).

Use by threat actors

Palo Alto Networks' Unit 42 documented in a July 30, 2026 report that an opportunistic operator based in Zhuhai used DeepSeek, driven through the open-source Hermes Agent framework, as an autonomous offensive operator against exposed workflow-automation infrastructure (Chinese-Speaking Threat Actor Harnesses AI Models for Autonomous Cyberattacks). Unit 42's stated reason for the model choice bears directly on DeepSeek's guardrail posture: the actor tried Western tools first and routed them through an anonymising proxy, but their provider-side controls "likely limited their effectiveness for autonomous attacks," and the actor settled on "a model with minimal safety controls (DeepSeek) accessed through an open-source framework with no client-side restrictions." OpenAI confirmed to Unit 42 that its own safeguards refused the policy-violating requests and that its safety systems disabled an account believed linked to the campaign. The report records no equivalent action on the DeepSeek side and does not state whether DeepSeek was notified.

The autonomous cycle achieved no compromise. The Taiwanese research firm TeamT5 separately reported on August 24, 2026 that state-affiliated Chinese groups have more than doubled their attack volume since assigning routine work and malware development to AI, again naming DeepSeek as the favoured model for its low running cost and weak guardrails (Source: implicator.ai). The two findings concern different actors and should not be read as one campaign. See Autonomous cyber-agents.

Government and national-security response

The V4 release on April 24, 2026 coincided with US federal actions framing DeepSeek as a national-security concern. The Office of Science and Technology Policy issued an NSTM-4 memo titled "Adversarial Distillation of American AI Models" (see Adversarial Distillation). The State Department sent a global cable ordering posts to spotlight alleged DeepSeek IP theft (Source: reuters.com). On April 22, 2026, Tencent and Alibaba together committed over $20B to DeepSeek-centric infrastructure builds (Source: bloomberg.com).

A congressional approach of a different kind followed in September 2026. Rep. Ro Khanna, ranking member of the House Select Committee on the CCP, wrote to the chief executives of DeepSeek, Alibaba and Moonshot AI asking whether they would cooperate on pacing frontier AI, according to the Washington Post, which reported the letters on September 17, 2026 and said Chinese officials and technology leaders had so far rejected calls to slow AI development (Source: washingtonpost.com). Only the letter to Alibaba, dated September 16, has been published; it names DeepSeek among the laboratories that should join "a binding, enforceable, and verifiable agreement" with US laboratories to pace the frontier, including "global limits on unchecked recursive self-improvement" (Khanna letter to Alibaba on pacing the frontier (September 16, 2026)). The DeepSeek letter's text has not been located, and no DeepSeek response had been reported as of September 18, 2026. See Coordinated slowdown proposals compared.

China's Cyberspace Administration opened an investigation into DeepSeek, alongside Moonshot AI, over potential data leaks to Anthropic, according to reporting published September 22, 2026 of which only the abstract was retrieved; Anthropic's September 10 threat intelligence report states that DeepSeek, Xiaomi and Moonshot "fed conversations between their own models and users into Claude" (Detecting and countering misuse of AI: September 2026 (Anthropic threat intelligence report); Source: theinformation.com).

DeepSeek was among the companies invited to brief the UN Security Council at its September 23, 2026 meeting on AI and international security, two people familiar with the matter said on September 22; other Chinese firms including Moonshot were also invited to make statements, and founder Liang Wenfeng did not plan to attend (Source: reuters.com).

See also

Related pages: (Source: epochai.substack.com), China and the US Are Running Different AI Races, The Least Understood Driver of AI Progress, Clawed, Distillation, DeepSeek-V3 Technical Report, DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning, DeepSeek V4 Pro / V4 Flash, Adversarial Distillation.

Relationships

• supports: DeepSeek-V3 Technical Report and DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning — DeepSeek's own publications, and the evidence base for this page's model and competitive-position sections
• contradicts: Export Controls (AI) — the training-cost and capability claims on this page are the standard case that export controls have not foreclosed frontier capability in China; the page records the claims and the disputes over them rather than settling the question
• depends-on: Nvidia & TSMC — AI Compute Infrastructure and Huawei — Ascend AI Accelerators — the compute this page's cost analysis rests on, and the domestic substitute the government-response section describes
• instance-of: Fast-Follow Problem — DeepSeek is the case most often cited for the argument that a capability lead is short-lived
• related: Zhipu AI — the other Chinese developer this page compares against
• related: Liang Wenfeng — founder, and the source of the strategy positions this page attributes
• related: DeepSeek V4 Pro / V4 Flash — the current release this page's capability claims resolve to
• related: Chinese-Speaking Threat Actor Harnesses AI Models for Autonomous Cyberattacks and Palo Alto Networks — the basis for the use-by-threat-actors section
• related: Adversarial Distillation, Autonomous cyber-agents, Inference Economics and Token Pricing, Semiconductor Supply Chain
• related: US-China AI Competition: Different Races, Different Metrics — the comparison page this company page feeds
• related: Khanna letter to Alibaba on pacing the frontier (September 16, 2026) — names DeepSeek among the Chinese laboratories asked to join a frontier-pacing agreement (September 2026)