Why DeepSeek Matters
DeepSeek is the company that broke the assumption that only US labs with tens of billions of dollars of compute could build frontier AI. When it released the R1 reasoning model in January 2025 under a permissive MIT license, claiming a training cost far below what OpenAI and Google were reported to have spent on comparable systems, the model briefly became the top free app on the US iOS App Store and triggered a historic sell-off in AI infrastructure stocks, with Nvidia alone losing roughly $600 billion in market value in a single trading day, the largest one-day loss for any US company on record. Commentators labeled it AI's 'Sputnik moment': proof that compute scale was not the only path to frontier capability, and that a comparatively small, then-eighteen-month-old lab operating under US export restrictions on advanced chips could still compete with the best-funded labs in the world.
The disruption reshaped how the industry talks about AI economics. DeepSeek's models still routinely undercut US frontier labs' API pricing by an order of magnitude or more, and its choice to publish open weights rather than keep its models proprietary forced a rethink of what a competitive moat in foundation models even means once the cost of training a capable model keeps falling.
From a Hedge Fund Side Project to a Frontier AI Lab
DeepSeek's roots are in quantitative finance, not a university AI lab or a Big Tech spinout. Founder Liang Wenfeng, a Zhejiang University-trained engineer born in 1985, co-founded the quant hedge fund High-Flyer in 2015 with two college classmates, building trading systems around machine learning that grew High-Flyer into one of China's largest quant funds, reportedly peaking near $14 billion in assets under management around 2021. Regulatory pressure on China's quant trading industry between 2019 and 2023 pushed the team to lean harder into AI research as a side project, including building out High-Flyer's own GPU clusters, before Liang spun the effort out as DeepSeek in Hangzhou on July 17, 2023, funded entirely by the hedge fund.
The lab spent its first eighteen months mostly out of the spotlight, but each release escalated the story. V2, in May 2024, undercut domestic rivals so aggressively on API pricing that it was nicknamed 'the Pinduoduo of AI,' forcing Alibaba, Baidu and Tencent to cut their own prices. V3, in December 2024, was trained for a claimed ~$5.6 million (2.79 million H800 GPU-hours) by DeepSeek's own technical report, a fraction of estimates for comparable Western models. Then came R1 in January 2025, which turned DeepSeek from a regional price disruptor into a name recognized across the global AI industry within a single week.
Efficiency as Architecture: How DeepSeek Builds Cheaper Models
DeepSeek's central technical bet is that clever architecture and training method can substitute for raw compute. Its models use a Mixture-of-Experts design, DeepSeekMoE, that splits each layer into many small 'expert' subnetworks and activates only a fraction of them per token. R1 carried 671 billion total parameters but activated only about 37 billion per token, giving it the knowledge capacity of a much larger dense model at a fraction of the inference cost; the V4 family that followed in April 2026 pushed the same idea further, with V4-Pro reaching 1.6 trillion total parameters against 49 billion active. Paired with Multi-Head Latent Attention, which compresses the key-value cache that ordinarily balloons memory use at long context lengths, the architecture is a big part of why DeepSeek can offer a 1-million-token context window at prices far below its Western peers.
On the training side, DeepSeek showed that a model could learn to reason, including self-verification and error correction, largely through large-scale reinforcement learning (using its own Group Relative Policy Optimization algorithm) rather than expensive human-annotated chain-of-thought data, cutting a cost center other labs treated as unavoidable. By 2026 the company pushed the efficiency story further still: it says the V4 family was trained on Huawei's domestically produced Ascend chips rather than Nvidia GPUs, a claim that, if accurate, is among the highest-profile demonstrations yet that a near-frontier model does not require the Nvidia hardware US export controls were designed to keep out of China.
A Business Model Built on Giving the Model Away
DeepSeek monetizes almost entirely through its API rather than its free consumer chatbot, pricing V4-Flash at $0.14 per million input tokens and $0.28 per million output tokens, a small fraction of comparable US frontier offerings, while publishing its model weights under an MIT license that lets any cloud provider, Chinese or otherwise, host and resell the models without paying DeepSeek anything. That combination has been good for adoption: annualized revenue reportedly approached $400-500 million by mid-2026, roughly double the prior year. But DeepSeek has slipped to third place among China's own AI-native apps by monthly active users (about 130 million), trailing ByteDance's Doubao and Alibaba's Qwen, even as it remains one of the most closely watched model families among global developers for its reasoning and coding performance.
The open-weight strategy is a genuine tension in the business model, not just a marketing choice. Analysts describe it as a 'moat inversion,' where publishing weights maximizes adoption and positions DeepSeek as infrastructure for a non-US AI stack, at the cost of being easy for others to commoditize. DeepSeek's own August 2026 warning that it would raise prices significantly across its services, after V4-Flash usage reportedly hit 8 trillion tokens processed in a single day, is the clearest signal yet that the free-and-cheap era of its pricing was straining the compute behind it.
From Self-Funded to a Record Raise, Under One Man's Control
For nearly three years, DeepSeek operated on High-Flyer's balance sheet alone, with Liang Wenfeng holding an estimated 84% founder stake as of mid-2024 and no outside investors. That changed in June 2026, when the company closed roughly $7 billion in its first external round, the largest AI funding round in Chinese history, at a roughly $52 billion valuation, with Tencent, CATL, JD.com, NetEase and a state-backed AI investment fund all participating largely through non-voting instruments that leave Liang's control intact.
A planned follow-on raise wobbled in July 2026, pausing after leaked remarks from Liang frustrated some prospective investors, before reopening in early August 2026 targeting close to $8 billion more at a valuation near $74 billion, with Monolith Management, an early backer of rival Moonshot AI, reportedly among the parties in talks. DeepSeek is simultaneously preparing paperwork for a mainland China IPO on Shanghai's STAR Market, targeted for a 2026 filing toward a possible 2027 listing.
The Geopolitics: Chips, Bans, and a Distillation Fight
DeepSeek has become a proxy in the US-China AI rivalry almost as much as a company in its own right. Its early breakthroughs ran on Nvidia H20 chips that were legal to export to China at the time but were reclassified as restricted in April 2025; by 2026 a senior US official was alleging DeepSeek had trained a newer model using smuggled Nvidia Blackwell chips explicitly banned from China, while DeepSeek's public position moved toward Huawei's Ascend 910C chips, independently estimated at roughly 60% of an Nvidia H100's inference performance, as its declared training hardware for V4.
That dependence on Chinese-controlled infrastructure, combined with a privacy policy that collects extensive user data subject to Chinese data-access law, has driven a wave of government restrictions: the US has barred DeepSeek from federal devices and, under the FY2026 NDAA, from defense and intelligence systems outright, and Italy, South Korea, Taiwan, India and multiple US states have imposed their own bans or restrictions over data-sovereignty concerns. Separately, Anthropic publicly accused DeepSeek, alongside Moonshot AI and MiniMax, in February 2026 of large-scale distillation of Claude's outputs through roughly 24,000 fraudulent accounts, an allegation that sharpened the broader dispute over whether Chinese labs are catching up to the frontier organically or by harvesting it.
What to Watch Next
DeepSeek's near-term test is whether it can convert engineering efficiency into a durable, profitable business rather than a one-time shock. The official general-availability release of the full V4 family, including V4-Pro and new peak/off-peak API pricing, has already slipped past its original mid-2026 target, and the significant price increase DeepSeek flagged in August 2026 will be an early signal of how much pricing power it actually has now that free-tier-scale usage is straining its compute. In parallel, the company is racing to stand up roughly a gigawatt of self-built and leased data center capacity in Ulanqab, Inner Mongolia, on Huawei silicon, targeting partial capacity online in late 2027 or early 2028, a bet that domestic chips can support frontier-scale training and inference at the pace Nvidia GPUs once did.
Further out, the mainland China IPO filing expected later in 2026 will be a major test of whether outside investors, even non-voting ones, are comfortable underwriting a company built largely around one founder's control and one geopolitically exposed technology stack. A newly disclosed move into embodied AI, a strategic stake and model-development pact with humanoid-robot maker Unitree Robotics announced alongside Unitree's own IPO in August 2026, suggests DeepSeek intends to extend its efficiency playbook beyond chatbots and APIs into robotics, a business line worth watching for its first shipped results.