Companies / DeepSeek
Frontier AI lab
DeepSeek
State of play as of
China’s most influential AI lab: open-weight models, a hard bet on efficiency, and now its own inference silicon.
DeepSeek spent two years arguing that the way to beat a compute deficit was to need less compute, used better. That bet produced open-weight models that run within a few points of the best Western systems at a fraction of the cost. In 2026 the lab took its first outside money, let a state fund hold the only voting seat, and began designing its own inference chip. We track what it ships and who backs it, with dates and sources.
Fast facts
- Founded
- 2023, Hangzhou. Spun out of the quantitative hedge fund High-Flyer.
- Founder & CEO
- Liang Wenfeng, who also founded High-Flyer.
- Flagship model
- DeepSeek V4, open-weight under the MIT license: 1.6 trillion parameters, 1-million-token context.DeepSeek technical report, Reuters
- Latest valuation
- Roughly $52 to $59 billion, set by its first outside round in June 2026.The Information, SCMP
- Notable investor
- The state-backed National AI Industry Investment Fund, the only investor with a vote.The Information, SCMP, company registry
- Compute
- Trained on Nvidia H800 early, Huawei Ascend since. V4 got day-0 support on Huawei Ascend (via Huawei’s CANN stack), Cambricon and Hygon; Alibaba, ByteDance and Tencent ordered hundreds of thousands of Ascend 950s.SCMP, Tom’s Hardware, Reuters
- Headcount
- About 300 to 500 people. In September 2026 it opened roughly 150 engineering roles at once and none in research; an investor briefed on the plan puts the target for the year at 1,000.QbitAI, Sina Tech, via CAD
- Listing
- Reported to have hired CITIC Securities to prepare a Shanghai STAR Market listing, aiming to start the process in 2026. No formal listing-tutoring agreement has been signed and DeepSeek has not confirmed it publicly.Reuters, TechNode, Sina
- In development
- Its own AI chip, aimed at inference rather than training.Reuters
The state of play
DeepSeek opened about 150 engineering roles in a single round and not one of them is a research seat. The notice went up on 8 September and was read closely by QbitAI and republished by 36Kr: server-side development engineers across six directions, among them the large-model research platform, Agent framework components, the DeepSeek API, online serving and data engineering, plus agent elastic-compute engineers on platform development and low-level systems. QbitAI puts total headcount at 300 to 500 people, so the round is worth between a quarter and a half of the company hired into one function, and an investor at a front-line firm told Sina Tech the goal for the year is 1,000. What they are being hired to build has a name: DSec, for DeepSeek Elastic Compute, disclosed for the first time in the V4 technical report as three Rust components, an API gateway called Apiserver, an Edge agent on every host machine and a cluster monitor called Watcher, wired over a homemade RPC protocol on top of 3FS, the distributed filesystem DeepSeek open-sourced last year. Two reports the same week sit underneath it. Reuters reported that DeepSeek has hired CITIC Securities to prepare a Shanghai STAR Market listing and wants to begin the process this year, with Sina adding that due diligence has started, that no formal listing-tutoring agreement is signed and that the company has said nothing officially. And Bloomberg reported that DeepSeek plans to place at least 160,000 Huawei Ascend 950DT accelerators in an Inner Mongolia data center, earmarked for inference rather than training, forming only part of the site’s eventual capacity and dependent on how fast Huawei can produce the parts.
When DeepSeek’s V4 line launched, its day-0 domestic-chip support ran wider than Huawei alone. Cambricon completed full-stack adaptation on V4’s April release day and open-sourced the deployment code the same day; V4 was also fully adapted to Huawei’s Ascend 950PR, with Hygon named too. Huawei’s Ascend runs V4 through CANN, its answer to Nvidia’s CUDA, a way to build on Chinese hardware without touching an Nvidia part (Cambricon and Hygon reached V4 through their own stacks). By late July, as DeepSeek shipped the finished V4-Flash, Alibaba, ByteDance and Tencent had collectively placed orders for hundreds of thousands of Ascend 950 processors. V4-Flash held its API price at 1 yuan per million input tokens and posted agent-benchmark gains from a fresh round of post-training alone; hours earlier OpenAI cut its fast GPT-5.6 model by about 80 percent.
DeepSeek is secretly developing its own AI chip, aimed at inference rather than training. The project began roughly a year ago, and the company is in talks with chip-design houses, a foundry and memory suppliers, hiring chip engineers through private channels. It still trains on other firms’ hardware and will for a long time, but designing a chip for its own models targets the 80 to 90 percent of a model’s lifetime cost that inference represents.
DeepSeek open-sourced DSpark, a technique that speeds per-user inference on its V4 models by 60 to 85 percent, in the same fortnight a US frontier lab found a comparable efficiency gain and kept it in-house. It is the clearest example yet of DeepSeek’s standards-over-margin strategy: give away the efficiency work so its stack becomes the cheapest to run.
A SemiAnalysis teardown found part of DeepSeek V4’s architecture was co-designed for Huawei’s Ascend 950DT inference accelerator, which carries Huawei’s in-house HiZQ memory at 144GB and 4 TB/s. The model and the domestic silicon are increasingly being built for each other.
DeepSeek closed its first outside round, about 51 billion yuan (roughly $7.4 billion) at a valuation of $52 to $59 billion, the largest single AI raise in Chinese history. Founder Liang Wenfeng wrote the biggest check; the only investor with a vote is the state-backed National AI Industry Investment Fund. The stated uses were domestic-chip compute centers, a self-designed AI chip, and talent.
DeepSeek open-sourced V4 under the permissive MIT license with Huawei Ascend as a primary deployment path: 1.6 trillion parameters, a 1-million-token context, and coding benchmarks a hair behind the best Western models. It was the first time a Chinese chip, model and deployment stack lined up in public with no US dependency.
Our reporting on DeepSeek
- DeepSeek Opened 150 Engineering Jobs and Zero Research Jobs
DeepSeek posted about 150 openings on 8 September and not one of them is a research role, per a WeChat notice read closely by QbitAI and republished by 36Kr. They split into server-side development engineers across six directions, which are the large-model research platform, Agent framework components, R&D efficiency infrastructure, the DeepSeek API, online serving and data engineering, and agent elastic-compute engineers across two, platform development and low-level systems, with a target profile of senior backend people two to ten years in. QbitAI puts the whole company at 300 to 500 people, so a single round is worth somewhere between a quarter and a half of existing headcount, hired into one function; an investor at a front-line firm told Sina Tech the goal for this year is 1,000 people, which would put DeepSeek level with Zhipu, whose 2025 prospectus listed 883 employees. What the 150 are being hired to build has a name: DSec, for DeepSeek Elastic Compute, disclosed for the first time in the V4 technical report as three Rust components, an API gateway called Apiserver, an Edge agent on every host machine and a cluster monitor called Watcher, wired over a homemade RPC protocol on top of 3FS, the distributed filesystem DeepSeek open-sourced last year. Cui Tianyi, who runs the DeepSeek Harness team, gave the reason in one sentence: scale produces complexity, and the complexity turns around and demands more scale. Underneath it is the money. Reuters reported on 9 September that DeepSeek has hired CITIC Securities to prepare a listing on Shanghai’s STAR Market and aims to start the process this year, which TechNode carried the same morning; Sina’s wire adds that CITIC has made contact and entered due diligence, that the two sides have not signed a formal listing-tutoring agreement, and that DeepSeek has said nothing officially.
- The Off-Ramp
DeepSeek finished shipping V4, and its day-0 domestic-chip support runs on Huawei Ascend (via Huawei’s CANN, its answer to Nvidia’s CUDA), Cambricon and Hygon. After the launch Alibaba, ByteDance and Tencent ordered hundreds of thousands of Ascend 950 processors.
- The Tell
DeepSeek is secretly building its own chip, and it targets inference, not training. You do not build an inference chip to out-Nvidia Nvidia; you build one to own the compute bill that recurs.
- The Giveaway
DeepSeek open-sourced an 85% inference speedup. OpenAI found the same trick the same fortnight and kept it secret. One bets on standards, the other on margin.
- The Co-Design
A teardown found part of DeepSeek V4’s architecture was co-designed for Huawei’s Ascend 950DT inference chip. The model and the silicon are now being built for each other.
- The Private Round
DeepSeek closed its first outside round, roughly 7.4 billion dollars at a 50-billion-plus valuation. The founder wrote the biggest check; the only investor with a vote is a state fund.
- The Week the Race Happened
DeepSeek open-sourced V4 under MIT with Huawei Ascend as a primary deployment path: 1.6 trillion parameters, million-token context, coding a hair behind the best Western models. The first sovereign stack.
- Closed in Code
A team completed what it says is the first third-party full-parameter post-training of a 1.6-trillion-parameter model on about 1,000 Huawei Ascend cards, with no instability. The penalty is efficiency, and it is shrinking.
These link to the full issues on the newsletter. New pieces go out in the daily first.
Common questions
Is DeepSeek going public?
Not yet, and nothing has been filed. Reuters reported on 9 September 2026 that DeepSeek has hired CITIC Securities to prepare a listing on Shanghai’s STAR Market and aims to start the process during 2026, which TechNode carried the same morning. Sina’s wire adds that CITIC has entered due diligence but that the two sides have not signed a formal listing-tutoring agreement, the first regulated step in a mainland listing, and DeepSeek has made no official statement. Treat it as preparation reported by others, not a confirmed offering.
How many people work at DeepSeek?
Roughly 300 to 500, per QbitAI. In September 2026 the lab posted about 150 openings in one round, all engineering and none in research, which is worth between a quarter and a half of that headcount. An investor at a front-line firm told Sina Tech the target for the year is 1,000 people, which would put DeepSeek level with Zhipu, whose 2025 prospectus listed 883 employees.
Is DeepSeek building its own chip?
Yes. On 7 July 2026 Reuters reported that DeepSeek is secretly developing its own AI chip, aimed at inference rather than training. Sources briefed on the project said it began roughly a year earlier, and that DeepSeek is in talks with chip-design houses, a foundry and memory suppliers. The company still trains on Nvidia and Huawei hardware and will for a long time; the chip is an early, long-term bet on owning the recurring cost of running its models.
Who owns DeepSeek?
DeepSeek is majority founder-held. In June 2026 it took its first outside money, about $7.4 billion at a $52 to $59 billion valuation. Founder Liang Wenfeng wrote the biggest check, and the only investor with a vote is the state-backed National AI Industry Investment Fund, per The Information, SCMP and the company registry.
What is DeepSeek V4?
DeepSeek V4 is the lab’s flagship model, open-sourced under the MIT license in April 2026 with 1.6 trillion parameters and a 1-million-token context. It launched with Huawei Ascend as a primary deployment path and scores a hair behind the best Western models on coding, at a fraction of the cost to run.
What chips does DeepSeek V4 run on?
DeepSeek V4 launched in April 2026 with Huawei Ascend as a primary deployment path, and its day-0 support spanned a wider set of Chinese accelerators from that launch. Per SCMP and Tom’s Hardware, Cambricon completed full-stack adaptation on the day of release and open-sourced the deployment code the same day, V4 was fully adapted to Huawei’s Ascend 950PR, and Hygon was also named. Huawei’s Ascend runs it through CANN, Huawei’s alternative to Nvidia’s CUDA, so a developer can build on that hardware without an Nvidia part (Cambricon and Hygon reached V4 through their own stacks). It still trains partly on Nvidia hardware.
Does DeepSeek use Nvidia chips?
It has. DeepSeek trained on Nvidia H800 chips early and has shifted to Huawei Ascend since; V4 ships with the Ascend 950 as a primary deployment path. In July 2026 it was reported to be designing its own inference chip to reduce that dependence further.
Who founded DeepSeek?
Liang Wenfeng, who also founded the quantitative hedge fund High-Flyer. DeepSeek spun out of High-Flyer in 2023 and is based in Hangzhou.
Every dated fact above traces to a named source, and where it comes from our reporting, to the dated issue that carried it. Where a claim is self-reported or disputed, we say so.