DeepSeek Statistics (2026): Users, Funding, Training Cost, and Market Share

A product manager budgeting an AI feature, an investor sizing a company, and a developer choosing a model can all reach for a DeepSeek statistic—and mean something different. Downloads track distribution, OpenRouter token share tracks activity on one marketplace, and GPU-hours document a particular training run.

Here is the usable 2026 picture: DeepSeek’s consumer app had an estimated 188 million weekly active users in Q1 2026; its first external funding round was reported at about $7.4 billion; and the DeepSeek-V3 paper disclosed 2.788 million H800 GPU-hours with an assumed-run cost of $5.576 million. The figures below identify the population, period, and evidence type so you can cite the number that answers your question.

Table of contents

• Key DeepSeek statistics at a glance

• DeepSeek users and downloads

• Funding and valuation

• Model specifications and training costs

• Developer and platform adoption

• DeepSeek API pricing

• Market impact

• What the public data does not show

• Methodology and source notes

• Conclusion

Get the Mobile Testing Playbook Used by 800+ QA Teams

Discover 50+ battle-tested strategies to catch critical bugs before production and ship 5-star apps faster.

Email addressGet Instant Access

100% Free. No spam. Unsubscribe anytime.

Key DeepSeek statistics at a glance

• 188 million weekly active users in Q1 2026: Presenc AI estimates this consumer-app audience, versus an estimated 29 million weekly active users in Q1 2025. It is a third-party estimate based on public disclosures, analytics, and the firm’s monitoring.

• 173 million cumulative app downloads from January 2025 through mid-2026: Business of Apps reports this AppMagic-based figure. Its headline banner says 175 million, leaving a two-million-download discrepancy on the publisher’s page.

• About $7.4 billion in first external funding: Caixin Global reported a roughly 50-billion-yuan round in July 2026, based on a regulatory filing involving an indirect investor.

• More than $52 billion post-money valuation: Caixin’s filing-based July 2026 report placed the valuation above 350 billion yuan. CNBC reported a 350-billion-to-400-billion-yuan range, or $52 billion to $59 billion, in June 2026.

• 671 billion total parameters and 37 billion activated per token: DeepSeek-V3’s technical report gives these mixture-of-experts architecture figures and reports 14.8 trillion pre-training tokens.

• 2.788 million H800 GPU-hours and $5.576 million: The full V3 report assigns this compute and assumed cost to its official training run at $2 per H800 GPU-hour.

• 9% to 18% OpenRouter token share from January to June 2026: OpenRouter measured DeepSeek’s share of combined input and output tokens across more than 450 trillion platform tokens collected from January 1 through June 14, 2026.

• 7.22 trillion OpenRouter tokens from July 27 to August 2, 2026: TechNode reported this weekly processing total for DeepSeek V4 Flash.

• $1.32 per million V4-Flash output tokens at peak: DeepSeek’s API documentation lists this direct rate; off-peak pricing is half the listed peak price.

DeepSeek users and downloads

The best current public estimate for DeepSeek’s consumer-app audience is 188 million weekly active users in Q1 2026. Presenc AI’s comparable Q1 2025 estimate is 29 million weekly active users, a reported 6.4× increase. The same analysis estimates 9.2 million Hugging Face downloads per month in Q1 2026 for DeepSeek models and about 41,000 API organizations in mid-2026. These are third-party measurements rather than company disclosures. (Presenc AI)

Business of Apps, citing AppMagic, lists 130 million active users at the end of 2025. The published summary does not define the activity window as daily, weekly, or monthly. It says a majority of this audience was in China without publishing a percentage. (Business of Apps)

For distribution, the AppMagic-based total is 173 million cumulative downloads from the January 2025 launch through mid-2026. The same publisher’s banner says 175 million, so cite the 173 million body figure with the discrepancy rather than silently choosing between them.

The AppMagic country split below covers DeepSeek app downloads in 2025. It describes download distribution, and the percentages do not add up to 100% because the published list covers nine countries rather than every market.

If you are interpreting installs alongside ongoing product use, Quash’s mobile app retention statistics provide the related retention and churn context.

Funding and valuation

DeepSeek’s first external round was reported at about 50 billion yuan, or $7.4 billion, in July 2026. Caixin’s regulatory-filing-based coverage names Tencent Holdings and CATL among the investors, and also reports participation by NetEase and JD.com. It identifies Liang Wenfeng as founder and High-Flyer as DeepSeek’s incubator. (Caixin Global)

The reported post-money valuation was more than 350 billion yuan, or $52 billion. The $52 billion figure is the lower boundary of the June 2026 $52 billion-to-$59 billion range reported by CNBC, rather than a company-confirmed point valuation.

DeepSeek has not published an audited revenue figure. A reported funding round and a post-money valuation establish the scale of a financing event and investor pricing, not an official revenue result.

Model specifications and training costs

DeepSeek-V3 is a mixture-of-experts model with 671 billion total parameters and 37 billion activated parameters per token. Its technical report says pre-training used 14.8 trillion tokens.

The V3 report breaks the official training run into 2.788 million H800 GPU-hours: 2.664 million for pre-training, 119,000 for context extension, and 5,000 for post-training. At an assumed rental price of $2 per H800 GPU-hour, the authors calculate a $5.576 million cost for that run.

The paper limits that cost to the official training run, excluding earlier research, ablation experiments, data preparation, and hardware purchases. It is therefore a disclosed run-cost figure with a defined boundary.

DeepSeek-R1’s technical paper reports 147,000 H800 GPU-hours and $294,000 for incremental reasoning-layer compute on top of DeepSeek-V3-Base. Its Table 7 assigns 101,000 hours and $202,000 to R1-Zero, 5,000 hours and $10,000 to supervised-fine-tuning data creation, and 41,000 hours and $82,000 to R1.

DeepSeek released a V4 preview on April 24, 2026, in Pro and Flash variants. CNBC described the release as open-source. Public reporting does not establish a specific MIT license for V4, and DeepSeek has not disclosed a V4 training budget. (CNBC)

Developer and platform adoption

OpenRouter offers a defined view of model activity on its own platform. Across more than 450 trillion input and output tokens from January 1 through June 14, 2026, DeepSeek’s share rose from 9% in January to 18% in June. It fell to 5% during February and March, then rebounded to nearly 20% by early June after the April V4 release; OpenRouter attributes much of the increase to agentic workloads.

V4 Flash processed 7.22 trillion tokens on OpenRouter during the weekly ranking period from July 27 through August 2, 2026. TechNode also reported 8 trillion tokens on OpenCode on August 1: 5 trillion from free trials and 3 trillion from paid usage. These figures are separately reported platform totals. (TechNode)

Usage volume can establish interest in a model without answering whether generated output performs reliably in a release workflow. For that adjacent question, see Quash’s vibe coding statistics and trends.

DeepSeek API pricing

DeepSeek publishes direct API rates for V4-Flash and V4-Pro. These are peak prices per million tokens; the documentation lists off-peak rates at half the peak prices. Both models list a 1-million-token context length and a 384,000-token maximum output.

Your API cost depends on the volume of cache hits, cache misses, and output tokens in your requests. For broader budgeting beyond model calls, Quash’s mobile app development cost data addresses the product-development side of the equation.

Market impact

On January 27, 2025, Nvidia lost close to $600 billion in market capitalization in one trading day. CNBC reported that shares closed down 17% at $118.58 and described the event as the largest one-day U.S. market-cap drop at the time. (CNBC)

The selloff put DeepSeek’s reported compute efficiency at the center of investor attention around AI infrastructure demand. It remains a dated market event, while the 2026 OpenRouter figures are the more current, platform-defined indicator of DeepSeek model activity.

What the public data does not show

Several often-requested DeepSeek statistics remain unavailable or insufficiently defined for a headline claim.

• V4 training cost: DeepSeek has not disclosed V4 training compute, hardware cost, or total development cost.

• Official revenue: DeepSeek has not published an audited revenue figure.

• Headcount: Axis Intelligence cites a roughly 150-to-200-person estimate via DemandSage; there is no company-confirmed employee count. (Axis Intelligence)

• Direct web traffic: March and May 2026 visit figures circulate through an indirect Similarweb, Panto AI, and Statista chain. They are panel estimates rather than audited DeepSeek traffic disclosures.

• Independent V4 benchmark scores: A specific score requires a primary leaderboard or model card.

• Mobile QA and devtool penetration: No first-party Quash data measures how often mobile test-automation frameworks, QA tools, or developer SDKs call DeepSeek’s API. Category-level API telemetry with a clear population and period would answer that question.

That final gap is practical if you are selecting an AI model for testing work. Broad adoption numbers cannot substitute for task-level reliability, integration behavior, and the failure patterns your app needs to catch. Quash’s State of QA Automation 2026 report covers the wider testing environment in which that evaluation happens.

Methodology and source notes

This roundup prioritizes DeepSeek technical papers and API documentation, OpenRouter’s first-party platform telemetry, regulatory-filing-based financial reporting, and contemporaneous financial journalism. Audience, app-download, organization, and headcount numbers remain labelled as third-party estimates when their publishers do not provide a company-reported count.

Each metric has a different population and period: app downloads are cumulative installations; weekly active users estimate consumer-app use in a quarter; Hugging Face downloads measure model retrieval; OpenRouter share measures token volume on OpenRouter; and GPU-hours cover a documented training run. The 173-million-versus-175-million download conflict remains unresolved by the publisher and is preserved above.

Conclusion

The public record establishes strong consumer attention, substantial reported financing, unusually detailed V3 training disclosures, and rapidly growing activity on a major model marketplace. It is less complete on the facts that would settle a buying or integration decision: V4 training economics, audited revenue, independently sourced benchmarks, and DeepSeek’s penetration in mobile QA workflows.

Use the available figures as bounded evidence, then seek task-specific evaluation where your decision depends on production behavior. That is the difference between citing a DeepSeek statistic and relying on it.