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
- 2023-05 – 2024-07the 暗涌 interviewsthe two 36kr/暗涌 interviews are the discovery coverage — 'a hidden ai giant' (2023) then 'a more extreme chinese techno-idealism story' (2024). after the shock, chinatalk, crossing river, and jiexu translate them into english and they become the most-cited chinese ai texts in the world.english translations circulate after r1
- 2024-05the price war the west missedv2's ~rmb-1-per-million-token pricing ignites china's llm price war — bytedance's doubao answers at ¥0.0008/1k tokens (a 99.3% cut) and alibaba, baidu, and tencent follow. the coverage is chinese and trade-scale; the west hasn't noticed the lab yet.80–97% price cuts across the chinese majors
- 2025-01crippled hardware, frontier modelv3's technical-report coverage — 'how deepseek trained a frontier model on crippled hardware' — is the trade-press tell that something moved in december. the $5.576m figure enters circulation here, scoped correctly at first.
the awe
- 2025-01-24the sputnik postandreessen's 'sputnik moment' post becomes the frame for everything after — 'a profound gift to the world.' within days the analogy is being critiqued for its premises.the defining metaphor of january
- 2025-01-26#1 over chatgptthe app-store coverage — deepseek passes chatgpt at #1 in the us and tops 51 other countries, with launch-day western coverage of the model itself framing it as an open-weights event.#1 in 52 countries
- 2025-01-28who built thiswired's origin-story feature — the lab 'made' by a hedge fund — and the first wave of profiles that explain to western readers who built this.
the panic
- 2025-01-27the crashthe crash coverage — nvidia −17%, ~$589–600b in a day, 'single-biggest one-day loss,' nasdaq −3.1%, ~$1t wiped. trump calls it 'a wake-up call' and 'a positive'; the market story and the policy story split from here.~$1t market cap wiped
- 2025-01-27the jevons counternadella posts the jevons-paradox counter — 'as ai gets more efficient and accessible, we will see its use skyrocket' — and the energy-economics debate follows: efficiency gains consumed by demand.
- 2025-01-29the war roomsthe meta war-room coverage — reports of scramble teams reverse-engineering r1, zuckerberg's leaked all-hands, and the 'hundreds of billions' capex pledge. the disruption story becomes an incumbent-defense story.
- 2025-01-29on deepseek and export controlamodei's 'on deepseek and export control' essay — the 7–10-months-older, on-trend argument that r1 does not obviate export controls. simon willison's notes flag what the essay omits: the open-weights point.the incumbent counter-argument
- 2025-02-10what is deepseek worththe valuation-range debate — bloomberg surveys guesses from ~$1b to $150b for a company that had never taken outside money. nobody can price the whale.estimates ranged ~$1b to $150b
the chips
- 2025-01-31the 50k-gpu estimatesemianalysis estimates ~50,000 hopper-class gpus accessible to deepseek — 10k h800s, 10k h100s, plus h20 orders — far beyond the 10k-a100 documented fleet. techspot amplifies the ~$1.6b-scale infrastructure reading.estimate, not registry
- 2025-01-31the commerce probereuters reports the us commerce department is probing whether deepseek used restricted chips; organized smuggling routes through malaysia, singapore, and the uae enter the coverage.
- 2025-02the smuggling trailthe smuggling trail hardens — singapore charges three men with fraud over server shipments containing nvidia gpus; singapore's share of nvidia revenue is reported climbing 9%→22%; customs-documented 'lobster' cases show chips mislabeled as other goods.three charged in singapore
- 2025-04 – 2025-06the h20 squeezethe h20 squeeze and the huang–liang meeting — the last compliant channel requires licenses, nvidia's ceo flies to beijing, and a state-department official's allegation that deepseek aided china's military adds the national-security layer.
the bans
- 2025-01-30italy blocks firstitaly's garante orders deepseek blocked — the first national action — calling the company's response 'totally insufficient.'first national block
- 2025-01-28 – 2025-02the us cascadethe us cascade — the navy's all-hands email, nasa's memo, texas's first state ban, then h.r.1121 (the 'no deepseek on government devices act') with gottheimer's 'five-alarm national-security fire' framing; state actions follow in new york, virginia, oregon, north carolina.a federal bill and state bans
- 2025-02the ban mapthe global map — australia, taiwan, and korean ministries restrict the app; korea's pipc halts new downloads over data transfers to volcano engine, resuming in april after ~two months.a dozen jurisdictions
the censorship
- 2025-01-28the refusal stringsthe tiananmen refusal coverage — the guardian publishes the verbatim refusal string; the independent reports the app elicits pro-ccp-biased outputs on sensitive topics. app-layer serving behavior, not the weights.
- 2025-01-28the self-censoring streamnbc's real-time self-censoring observations — answers appearing, then deleting themselves mid-generation on sensitive prompts.
- 2025-01 – 2025-09censorship in the weightsthe layered-censorship coverage — wired separates app-layer serving from in-weights tendencies; the r1dacted paper argues censorship sits in the base model; nist's caisi evaluation and secondary summaries keep the model-level findings in the record.r1dacted + caisi findings
- 2025-01-27uncensored by hostingthe hosted-serving counterweight — perplexity's srinivas announces us/eu-hosted r1 as uncensored, demonstrating that the serving layer, not the weights alone, determines refusal behavior.
the security
- 2025-01-29 – 2025-01-30the open databasewiz discloses the open clickhouse database — chat history, api secrets, backend details, unauthenticated — secured after disclosure. the first security finding lands inside the hype week.
- 2025-01-31 – 2025-02the jailbreak findingsthe jailbreak pile-on — cisco reports 100% harmbench attack-success on r1, qualys documents distilled-r1-8b failures, unit42 catalogs working bypasses, and wired's testers 'didn't stop a single one' of their attempts.100% harmbench asr reported
- 2025-02the app forensicsthe privacy forensics — krebs documents device fingerprinting; feroot finds a direct path toward china mobile infrastructure; nowsecure finds hardcoded 3des keys in the ios app.
the counter-takes
- 2025-01-29the distillation claimsthe distillation claims — openai tells the ft it has evidence deepseek used its models 'inappropriately'; microsoft reportedly flagged suspicious api exfiltration; axios and the verge amplify; scmp notes distillation is industry-common. no public evidence drop ever follows.asserted, not evidenced publicly
- 2025 – 2026what $5.576m actually boughtthe cost teardown — the $5.576m figure is real but narrow: one final pre-training run at $2/gpu-hour, per allen's house testimony and the postmortem writeups. the teardown literature separates the number from 'r1 cost $6m to build.'
- 2025optimization, not revolutionthe counter-takes — ifri's 'optimization, not revolution' reading and the sputnik-analogy critique argue the shock was a pricing and narrative event more than a capability leap.
the state
- 2025-01-21r1 day at the symposiumthe premier's symposium — liang briefs li qiang on the draft government work report as one of nine speakers, the same day r1 ships. the state-adjacency coverage starts here.
- 2025-02-17the xi symposiumthe xi symposium coverage — liang seated with ma, ma, ren, wang, and lei at xi's private-enterprise meeting, confirmed through state media's yuyuantantian account; the domestic 'deepseek fever' deployment wave follows through march.state-media confirmed
- 2025-04-16deepseek unmaskedthe house select committee's 'deepseek unmasked' report and the $420m high-flyer funding claim push the patron into the national-security frame — the whale itself becomes a coverage object.
- 2025-02 – 2025-03the passport storythe travel-curbs coverage — the information's passport reporting, the wsj's us-travel warning to founders, zhejiang officials screening investors. the '~130 lab staff / ~200 fund staff' headcount enters circulation.
the year after
- 2025-05the panic post-mortemthe post-mortem on the january panic — r1-0528 'did not have a moment'; the coverage reads the shock's narrative factors rather than a capability gap that persisted.
- 2025-09-17the peer-review precedentthe nature coverage — the first major llm through full peer review; the referee exchange surfaces the $294k rl-cost disclosure and deepseek's on-record denial of training on openai-generated reasoning examples.nature 645:633–638
- 2025-07 – 2025-11the second deepseek momentthe 'second deepseek moment' frame — kimi k2 thinking beats gpt-5 on hle and browsecomp; the chinese open-weights landscape coverage treats the lab's ecosystem strategy as the new normal.kimi k2 thinking
- 2026-02-23the anthropic disclosurethe anthropic disclosure coverage — ~24k fraudulent accounts, >16m claude exchanges, deepseek's ~150k+ called 'most technically sophisticated,' including censorship-safe prompt generation. the distillation story returns with numbers attached — and an accused-party response still unpublished.150k+ exchanges attributed
- 2026the incumbent turnthe incumbent coverage — v4's ascend-tuned delay, the $52b-ish close and star-market ipo prep, the august api price hike (up to ~12x on cache hits) that reads like the price-war catfish raising prices.ipo prep reported
- 2026a year onthe retrospective layer — scmp's year-on look-back, the nature study of chinese-vs-us media framing, and the departures coverage ('~10 of ~300 v4 authors now former') close the arc from shock to institution.
controversies
- 2025-01 – 2026what the $5.576m boughtthe $5.576m misreading — the v3 report's figure is one final pre-training run (2.788m h800 gpu-hours at an assumed $2/hour), not the v3 or r1 program cost. 'r1 cost $6m to train' propagated anyway; the teardown literature had to keep re-anchoring the number.allen's house testimony and the postmortems re-scope the figure; the $294k rl-stage disclosure in the nature paper adds a second, separately-scoped number.
- 2025-01 – 2026-02the distillation questionthe distillation accusation — openai told the ft it had evidence; microsoft reportedly flagged suspicious api exfiltration in fall 2024; sacks claimed 'substantial evidence.' no public technical evidence was published. in february 2026 anthropic released its own figures — ~24k fraudulent accounts, ~150k+ claude exchanges attributed to deepseek — the most detailed account to date, still one-sided.deepseek's on-record position, in the nature referee exchange: r1 was not trained on openai-generated reasoning examples, while conceding the base model ingests web ai-content like every llm.
- 2025-01 – 2025-08where did the gpus come fromthe chip-sourcing question — ~10,000 a100s are documented (bought pre-controls); the 2,048 h800s in the v3 run are self-reported; semianalysis estimates ~50,000 hopper-class gpus accessible; singapore charged three men over smuggled servers; a state-department official alleged military linkage. nvidia's position: its customers acquired chips lawfully.each layer of the inventory claim carries a different evidence class — registry/paper, self-report, analyst estimate, and prosecutor allegation — and the dossier keeps them separate.
- 2025-01 – 2025-09three kinds of censorshipthe censorship layering — app-level refusal strings, real-time self-censoring streams, in-weights tendencies (r1dacted, caisi), and hosted-serving behavior are different claims; perplexity's uncensored us-hosted r1 demonstrated the serving layer matters.the dossier treats each layer separately — serving behavior is not a weights property, and a weights tendency is not identical to the app's refusals.
- 2025-01 – 2025-02the security pile-onthe security week — an open clickhouse database with chat history and api secrets (wiz), 100% harmbench attack-success (cisco), hardcoded 3des keys (nowsecure), device fingerprinting (krebs), and a reported path to china mobile infrastructure (feroot).the database was secured after disclosure; the findings stand as the fastest security scrutiny any lab has received.
- 2025-02 – 2026-09the valuation fogthe valuation fog — february 2025 guesses ran ~$1b to $150b; the 2026 round reports conflict: ~rmb 300b (~$45b) at open, ~$52b per forbes, and higher headline figures elsewhere; htx reported a much smaller ≥$300m / ≥$10b sizing. liang's own injection (up to rmb 20b) and the state-capital lead are reported, not filed.no filing exists; every figure in the round coverage is reported. the dossier keeps the conflicting numbers rather than picking one.
sentiment eras
- 2023–2024the invisible giant · positivethe invisible-giant phase — the lab exists in chinese trade coverage and two 暗涌 interviews; the v2 price war and v3's release register locally while western coverage is essentially zero.
- 2025-01-20 – 2025-01-27the sputnik week · positivethe sputnik week — 'a profound gift to the world,' #1 in 52 countries, janus shipped into the chaos. coverage is awestruck and adoption-led: r1 lands in azure ai foundry within days.
- 2025-01-27 – 2025-02the panic · criticalthe panic — a ~$589–600b nvidia session, war rooms, the jevons counter, a $1b–$150b valuation debate, and the first bans inside ten days.
- 2025-02 – 2025-12the scrutiny · criticalthe scrutiny year — distillation claims, the commerce probe and smuggling cases, layered-censorship findings, passports and investor screening, the failed ascend run, and the stalled r2.
- 2026–the incumbent · mixedthe normalization — peer review in nature, the 'second deepseek moment' frame applied to rivals, the first external raise, v4, and ipo-prep coverage. the disruptor is covered as an incumbent.
AI-drafted at Ben Guo's direct request and credited to Hraness; every claim links to its cataloged source.