deepseek

a sourced history of deepseek, the open-weights ai lab, its founder liang wenfeng, and high-flyer, the hedge fund behind it.

deepseek looks like it appeared in january 2025, when r1 took the top of the app store and ~$589b off nvidia in a day. it didn't — it is the research arm of high-flyer, a hangzhou quant fund that spent a decade building an ai trading firm and bought ~10,000 nvidia a100s before the export controls arrived. this is the record of the whale behind the lab: the classmates who automated trading from their dorms, the fund's peak and public apology, the fire-flyer clusters, the open-weights doctrine, the shock, the scrutiny, and the first raise that made the disruptor an incumbent.

• history: the company, sourced and dated — from the dorm-room trading bots to fire-flyer ii, the r1 shock, and the first raise.
  • the classmates: 2008–2015 — automated trading from the zju dorms, jacobi, and the fund's registration.
  • the ai fund: 2016–2021 — the ai fund: deep-learning trades, the golden bull, the ~rmb 100b peak, and the apology.
  • the fleet: 2021–2023 — fire-flyer ii's ~10,000 a100s, bought before the october 2022 controls; the agi pivot.
  • the lab: 2023–2024 — incorporation, the release ladder, grpo, the v2 price war, and v3's 2.788m-gpu-hour run.
  • the shock: january–march 2025 — r1, the app-store #1, the nvidia crash, the bans, and the symposiums.
  • the scrutiny: 2025 — passports, the stalled r2, the ascend failure, the nature paper, and the poached bench.
  • the incumbent: 2026– — the first raise, v4, the price hike, ipo prep, and r2 never shipped.
• people: liang wenfeng and the people around both companies — the zju classmates, the fund's partners, the research bench, and the diaspora.
  • the founder: liang wenfeng — the quant who codes, co-author on the papers, 'an ordinary little pig,' and the richest ai founder.
  • the writing: the corpus — the golden bull speech, the two 暗涌 interviews, the technical reports, and the leaked investor call.
  • the team: the partners and the bench — xu jin, zheng dawei, chen zhe, li huan, lu zhengzhe, the first cfo, and the ~300-author research org.
  • interviews and talks: the interviews and talks — two 暗涌 sittings and one forum speech are nearly the whole public record.
  • the philosophy: technological idealism, open weights as strategy, constraint as engine — and the people-level controversies.
• side quests: the pursuits around the flagship — the philanthropy, the compute-donation program, the edge model lines, open source week, and the raise.
  • the founders' quests: 'an ordinary little pig' — liang's pseudonymous philanthropy and the founders' quests.
  • the lab's quests: the compute-donation program, open source week, the consumer app, the ascend hedge, r2, the ipo track, and the fund that keeps earning.
  • the bench's quests: the researchers' lines — grpo, prover, math-v2, and the vl2/janus multimodal work.
• coverage: what the discourse made of deepseek — the sputnik week, the crash, the bans, the censorship and chip sourcing debates, and the incumbent turn.
  • the discovery: the invisible giant — the 暗涌 interviews, the price war the west missed, and v3's technical-report notice.
  • the awe: the sputnik post, #1 over chatgpt in 52 countries, and the origin-story profiles.
  • the panic: the crash, the jevons counter, the meta war rooms, and the $1b–$150b valuation debate.
  • the chips: the silicon question — semianalysis's 50k-gpu estimate, the commerce probe, the smuggling trail, and the h20 squeeze.
  • the bans: italy first, then the us cascade — navy, nasa, texas, h.r.1121 — and the global map.
  • the censorship: the refusal strings, the self-censoring stream, the weights-level findings, and the hosted-serving counterweight.
  • the security: the open database, the jailbreak findings, and the app forensics.
  • the counter-takes: the distillation claims, the cost teardown, and the 'optimization, not revolution' counter-takes.
  • the state: the symposiums, the house report, the passport story — state adjacency as the coverage frame.
  • the year after: the year after — the panic post-mortem, the peer-review precedent, the second deepseek moment, and the incumbent turn.
  • controversies: the running debates — the cost number, distillation, chip sourcing, layered censorship, the security week, and the valuation fog.
  • sentiment eras: sentiment phases: invisible giant → sputnik week → panic → scrutiny → incumbent.
• the idea: the whale-funds-the-lab model — the lineage deepseek inherited, what it proved, and what it left unresolved.
  • lineage: the renaissance template, the zju network, the fund as patron, infrastructure first, and the embargo as engine.
  • what deepseek proved: what the model demonstrated — patronage, constraint-engineered efficiency, open weights as distribution, pricing as a weapon.
  • what stayed unfinished: r2, the ascend wall, research meeting revenue, the end of pure patronage, and the poached bench.
  • the arguments: whether the pattern is portable, whether openness is the moat or the exposure, and what the state adjacency costs.
  • successors: kimi, qwen, zhipu, minimax, mistral — and meta's llama as the patron who retreated.
  • open questions: the questions the next year has to answer.
• sources: every source cited in the deepseek history, with its date and type.
  • the catalog: the full typed catalog — primary, interview, reporting, archive, community, repo, reference.
• method: how ai agents researched the deepseek history, the evidence rules they followed, and what they could not verify.
  • the lanes: the two parallel research lanes and what the formula proved.
  • the evidence rules: the evidence model — confidence, precision, sentiment discipline.
  • known gaps: what the research could not verify — recorded, not hidden.
  • corrections: how to correct the record.

AI-drafted at Ben Guo's direct request and credited to Hraness; every claim links to its cataloged source.

Canonical page: https://hraness.com/deepseek