DeepSeek

DeepSeek is a Chinese AI research company founded in 2023 as a subsidiary of High-Flyer, a quantitative hedge fund. Based in Hangzhou, the company made global headlines in January 2025 when it released DeepSeek-R1, a reasoning model that matched or exceeded the performance of leading US models at a fraction of the reported training cost. DeepSeek closed its first outside venture round in June 2026, raising approximately 50 billion yuan (about $7.4 billion) at a valuation above $50 billion — the lab's first external capital. Founder Liang Wenfeng contributed roughly $3 billion himself, while commercial investors led by Tencent and battery maker CATL accepted five-year lockups and no voting rights. The only investor granted governance rights and a direct stake was the state-backed National Artificial Intelligence Industry Investment Fund — an unusual structure that concentrates control with the founder and the Chinese state. The raise places the lab alongside frontier US labs by valuation even though its training spend remains a fraction of theirs.

DeepSeek's current model is V4.1-Flash, released September 10, 2026 — a natively multimodal mixture-of-experts with a 552 billion parameter backbone that activates only 8 billion parameters per token while reading input, holds its key-value cache to 890 bytes per token, and ships ungated on Hugging Face under a plain MIT license with no revenue, user-count or territory conditions. It replaced the April 2026 pair in a single release: V4-Flash was retired the same day, and from noon Beijing time on September 14 every request to the V4-Pro endpoint is served by V4.1-Flash and billed at the cheaper Flash rate. DeepSeek's stated reason is that its own testing showed the smaller model beating V4-Pro on performance, cost and speed together. Two consequences are worth separating. The default endpoint's output price fell by roughly 70 percent without any headline price cut — 60 cents per million tokens off-peak against V4-Pro's one dollar ninety-eight — and it gained vision, which V4-Pro never had. But V4-Pro remains the larger artifact at 1.6 trillion total parameters and 49 billion active, and remains the largest open-weight model available under a fully permissive license — Kimi K3 and Qwen3.8 Max have since passed it on raw size, but both carry custom terms — so anyone who wants that specific model should download the weights rather than rely on the endpoint.

DeepSeek releases its models as open-weight under permissive licenses, making them freely available for download and modification, and operates the DeepSeek chat platform and API at pricing significantly below Western alternatives. The company's R1 release was a watershed moment for the AI industry, challenging the prevailing assumption that frontier AI capabilities require hundreds of millions of dollars in training compute — efficiency innovations including novel training techniques and architecture optimizations sent shockwaves through Silicon Valley and temporarily wiped hundreds of billions from NVIDIA's market cap.

Open availability has a safety corollary. In July 2026, Palo Alto Networks' Unit 42 documented a threat actor who wired DeepSeek into an open-source agent framework and had it run a largely autonomous intrusion campaign — enumerating targets, selecting vulnerabilities, pulling public exploit code, and executing attacks across roughly 460 targets after a single instruction. Unit 42 reported that OpenAI's provider-side controls refused equivalent requests and disabled an account linked to the same operator. The contrast is worth holding precisely: DeepSeek is markedly more restrictive than US models on political topics and markedly less restrictive on offensive-security ones. Those are two different axes, and they are easy to conflate.

Founder Liang Wenfeng has publicly committed to keeping DeepSeek on an open-source path and pursuing artificial general intelligence as the company's core goal, resisting the usual pressure to chase near-term commercialization. That posture is unusual at this valuation tier — most frontier US labs draw their largest checks from corporate cloud partners (OpenAI and Microsoft, Anthropic and Amazon) and treat AGI claims with strategic ambiguity. DeepSeek is doing the opposite, and doing it with one of the largest state-aligned bets on AGI to date outside the United States. Beijing has extended state-secret travel restrictions — originally placed on senior DeepSeek researchers — to AI talent across Alibaba and other private firms, requiring some professionals to seek official approval before traveling abroad. The pattern signals reciprocal controls that bind the lab into a state-managed talent regime, mirroring US chip export logic in reverse. On the US side, an interagency committee has approved DeepSeek — along with memory-chip maker CXMT and more than 100 other Chinese firms — for addition to the Commerce Department's Entity List on national-security grounds, though the administration has so far held off on the listing to avoid escalating tensions with Beijing; among the cited concerns is Anthropic's allegation that DeepSeek attempted to extract capabilities from its Claude models.

In mid-2026, DeepSeek began moving beyond models toward silicon of its own. Reuters reported that the lab had spent roughly a year quietly recruiting chip engineers and courting manufacturing partners to design an in-house processor for AI inference — the stage of running finished models rather than training them. The goal is to reduce its dependence on both Nvidia and Huawei as US export controls tighten. It is a strategic pivot into hardware that, if successful, would place DeepSeek alongside OpenAI and Anthropic among the AI labs pursuing their own custom chips — though designing and manufacturing a competitive processor typically takes years and significant capital, made harder by US restrictions on Chinese access to advanced foundries.

In August 2026 DeepSeek reversed the price leadership that had defined it. The 0813 build of V4-Pro reached general availability on August 12 with a cut of roughly three quarters, and four days later the lab moved its entire API onto peak and off-peak billing at sharply higher rates — V4-Pro output rising from 87 cents per million tokens to one dollar ninety-eight off-peak and three dollars ninety-six at peak, with cached input up roughly 1,100 percent. DeepSeek remained well below the US frontier tiers throughout, but the August rise ended its automatic claim to the bottom of the market. The September 10 move to V4.1-Flash then reversed much of the increase without a headline price cut, because the cheap model simply became the only model — which is a reminder that with this lab the effective price is set by which checkpoint the default endpoint points at, not by the rate card.

On September 10, 2026 Anthropic named DeepSeek in a threat intelligence report as running an unauthorized distillation campaign against Claude. It describes the method as a cross-session replay attack: saving the reasoning signature Claude returns in place of its raw thinking, starting a new session, and eliciting the model to convert that signature back into the full reasoning trace, circumventing a control built to prevent exactly this. Anthropic further alleges that DeepSeek silently relayed its own customers' requests to Claude without informing them, keeping the transcripts as training data — so users who believed they were using a DeepSeek model received Claude's answers, and their sessions reached a third party they had not chosen. DeepSeek had not publicly responded at publication. Anthropic is a direct commercial competitor, so treat this as an interested party's evidence rather than a settled finding.