{"id":"1e9bc5ec-550c-4b27-9033-4ea00f8e33aa","name":"DeepSeek","slug":"deepseek","description":"DeepSeek is a Hangzhou-based AI model developer spun out of the Chinese quantitative hedge fund High-Flyer in July 2023 by founder Liang Wenfeng. Unlike the US neocloud pack (CoreWeave, Lambda, Crusoe) that rents GPUs by the hour to third parties, DeepSeek runs its own fleet as a captive research resource: the Fire-Flyer 2 cluster built out from 2021 onward, roughly 10,000 Nvidia A100 GPUs plus a later top-up of H800 accelerators procured before US export controls tightened. The lab does not sell compute; it sells inference on its own frontier models through a low-priced API.\\n\\nThe company vaulted into global view in January 2025 when DeepSeek-R1 briefly displaced ChatGPT as the top free iOS download and Chinese researchers claimed a training run for the underlying V3 base model costing roughly $5.6 million, an order of magnitude below Western frontier labs. That efficiency claim, whether taken at face value or read as a narrow marginal-cost figure, triggered a one-day $600B drop in Nvidia's market cap and reframed the \"how much compute do you need for a frontier model\" debate. Successive releases (V3.1, V3.2, the V4-Flash and V4-Pro tiers in 2026) have kept the cadence.\\n\\nDeepSeek is privately held and until 2026 was funded entirely off High-Flyer's balance sheet. In May 2026 it closed a $7B Series A at a $52B post-money valuation and began IPO preparation targeting a 2027 listing. Its distinctive posture in the neocloud/AI-infra map is that it is an operator of its own inference stack rather than a rentable cloud, and it is doing frontier model work under Chinese export-restricted silicon, which has made it both a proof point for compute-efficient training and a target of US and Australian government use bans.","status":"active","founded":2023,"hq":"Hangzhou, China","hqLat":null,"hqLng":null,"hqGeocodedAt":null,"hqAddress":null,"hqPrecision":null,"website":"https://www.deepseek.com","fundingTotal":"7000000000","type":null,"roles":[],"reviewStatus":"reviewed","lifecycleState":"active","supersededByCompanyId":null,"sources":[{"url":"https://en.wikipedia.org/wiki/DeepSeek","title":"DeepSeek — Wikipedia","publisher":"Wikipedia"},{"url":"https://www.deepseek.com","title":"DeepSeek official site","publisher":"DeepSeek"},{"url":"https://arxiv.org/abs/2412.19437","date":"2024-12","title":"DeepSeek-V3 Technical Report","publisher":"arXiv (DeepSeek)"}],"keyFacts":[{"label":"Parent company","value":"Spun out of High-Flyer, a Chinese quantitative hedge fund, in July 2023; founder Liang Wenfeng also runs High-Flyer","sourceUrl":"https://en.wikipedia.org/wiki/DeepSeek"},{"label":"GPU inventory","value":"Fire-Flyer 2 cluster: ~10,000 Nvidia A100 GPUs (625 nodes of 8) built 2021 onward, later supplemented with Nvidia H800 accelerators for training and deployment; connected over InfiniBand. No public H100/H200/B200 numbers, consistent with US export controls","sourceUrl":"https://en.wikipedia.org/wiki/DeepSeek"},{"label":"Silicon partnerships","value":"Nvidia-only stack (A100, H800). No disclosed AMD, Cerebras, or Chinese-domestic accelerator use","sourceUrl":"https://en.wikipedia.org/wiki/DeepSeek"},{"label":"Positioning","value":"Captive research fleet plus low-priced hosted-inference API, not a rentable GPU cloud. Training-focused frontier lab; inference sold through deepseek.com API and app","sourceUrl":"https://www.deepseek.com"},{"label":"Most recent funding round","value":"$7B Series A closed May 2026 at $52B post-money valuation; earlier April 2026 talks had targeted $70B pre-money","sourceUrl":"https://en.wikipedia.org/wiki/DeepSeek"},{"label":"Public status","value":"Private; began IPO preparation in July 2026 targeting a 2027 listing","sourceUrl":"https://en.wikipedia.org/wiki/DeepSeek"},{"label":"Reported V3 training cost","value":"$5.576M for the DeepSeek-V3 base-model training run, per the company's own December 2024 tech report; widely disputed as a marginal-cost figure that excludes prior R&D and cluster capex","sourceUrl":"https://en.wikipedia.org/wiki/DeepSeek"},{"label":"Latest material release","value":"DeepSeek-V4-Pro (1.6T parameters) released August 13, 2026, following V4-Flash (284B) on July 31, 2026","sourceUrl":"https://en.wikipedia.org/wiki/DeepSeek"},{"label":"R1 market impact","value":"DeepSeek-R1 launched January 20, 2025; briefly displaced ChatGPT as top free iOS app by January 27 and coincided with a one-day ~$600B drop in Nvidia's market cap","sourceUrl":"https://en.wikipedia.org/wiki/DeepSeek"},{"label":"Government restrictions","value":"US FY2026 NDAA bars DeepSeek from Department of Defense systems; Australia issued a government-wide removal directive in 2025","sourceUrl":"https://en.wikipedia.org/wiki/DeepSeek"},{"label":"Employee count","value":"~160 staff as of 2025, per Chinese-press synthesis","sourceUrl":"https://en.wikipedia.org/wiki/DeepSeek"}],"aliases":[],"collisionRisk":"low","reviewNote":null,"atsProvider":null,"atsSlug":null,"factoryLocations":null,"manufacturingCapacity":null,"orgType":null,"burnSignal":null,"litigationEvents":null,"insurancePostureDisclosed":null,"unitEconomicsDisclosed":null,"tickerSymbol":null,"stockExchange":null,"secCik":null,"revenueUsd":null,"revenueBasis":null,"revenueAsOf":null,"employeeCount":160,"employeeCountBasis":"reported headcount, Wikipedia synthesis of Chinese press","employeeCountAsOf":"2025-01-01T00:00:00.000Z","marketCapUsd":null,"marketCapAsOf":null,"marketCapSource":null,"createdAt":"2026-08-28T01:48:32.056Z","updatedAt":"2026-08-28T03:45:17.421Z","jsonLd":{"@context":"https://schema.org","@type":"Organization","@id":"https://registry.deploy.report/companies/deepseek","url":"https://registry.deploy.report/companies/deepseek","name":"DeepSeek","description":"DeepSeek is a Hangzhou-based AI model developer spun out of the Chinese quantitative hedge fund High-Flyer in July 2023 by founder Liang Wenfeng. Unlike the US neocloud pack (CoreWeave, Lambda, Crusoe) that rents GPUs by the hour to third parties, DeepSeek runs its own fleet as a captive research resource: the Fire-Flyer 2 cluster built out from 2021 onward, roughly 10,000 Nvidia A100 GPUs plus a later top-up of H800 accelerators procured before US export controls tightened. The lab does not sell compute; it sells inference on its own frontier models through a low-priced API.\\n\\nThe company vaulted into global view in January 2025 when DeepSeek-R1 briefly displaced ChatGPT as the top free iOS download and Chinese researchers claimed a training run for the underlying V3 base model costing roughly $5.6 million, an order of magnitude below Western frontier labs. That efficiency claim, whether taken at face value or read as a narrow marginal-cost figure, triggered a one-day $600B drop in Nvidia's market cap and reframed the \"how much compute do you need for a frontier model\" debate. Successive releases (V3.1, V3.2, the V4-Flash and V4-Pro tiers in 2026) have kept the cadence.\\n\\nDeepSeek is privately held and until 2026 was funded entirely off High-Flyer's balance sheet. In May 2026 it closed a $7B Series A at a $52B post-money valuation and began IPO preparation targeting a 2027 listing. Its distinctive posture in the neocloud/AI-infra map is that it is an operator of its own inference stack rather than a rentable cloud, and it is doing frontier model work under Chinese export-restricted silicon, which has made it both a proof point for compute-efficient training and a target of US and Australian government use bans.","identifier":"1e9bc5ec-550c-4b27-9033-4ea00f8e33aa","foundingDate":"2023","address":"Hangzhou, China","publisher":{"@id":"https://deploy.report/#organization"}},"framework_metadata":{"framework_schema_version":"0.1.0","verification_status":"verified","maturity_stage":null,"lifecycle_state":null,"architectural_position":{"cohort":null,"sub_cohorts":[]},"within_cohort_verified_vs_claimed_pair":null,"cap_flags":[],"verification_depth":{"sources_count":3,"primary_source_types":["knowledge-base","preprint"]}}}