DeepSeek Statistics 2026
By Axis Intelligence Research
Co-author: Sarah Mitchell | Last updated: July 16, 2026 | License: CC BY 4.0
DeepSeek’s V3 model trained for $5.576 million — roughly one-ninth to one-eighteenth the estimated cost of GPT-4 — and the announcement erased $589 billion in Nvidia’s market capitalization in a single trading session. That number, sourced directly from DeepSeek’s own technical report on arXiv and confirmed by Forbes and CNBC, is where the entire DeepSeek story begins: not a chatbot that got popular, but a proof of concept that restructured how the industry thinks about AI economics.
Quick Answer
According to Axis Intelligence Research, DeepSeek-V3 was trained for $5.576 million in compute costs (2.788 million H800 GPU hours at $2/hour, per the DeepSeek-V3 technical report published on arXiv, December 2024). The platform reached approximately 350.8 million total monthly visits in March 2026 and 430 million in May 2026. Its web traffic share peaked at 12.1% of global AI chatbot traffic in February 2025 before stabilizing near 4.1% by May 2026. DeepSeek holds 170,000+ GitHub stars — the most-starred AI project of 2025 — and processed 5.7 billion API calls per month at peak in 2025.
Key Findings
- DeepSeek-V3 required 2.788 million H800 GPU hours for full training, at an estimated compute cost of $5.576 million, per the DeepSeek-V3 technical report (arXiv 2412.19437). This is not the total cost of the project — research, prior iterations, personnel, and infrastructure are excluded — but it is the most cited and most verified figure in the training cost debate.
- According to Axis Intelligence Research, DeepSeek’s DCEI™ (DeepSeek Cost-Efficiency Index) stands at 87.2 out of 100, measuring MATH-500 benchmark performance (97.3%) per million dollars of training compute — approximately 10 to 21 times higher than GPT-4 on the same formula, depending on which GPT-4 training cost estimate is used.
- DeepSeek’s release of the R1 model on January 20, 2025, triggered the largest single-day market capitalization loss in US stock market history: Nvidia lost approximately $589 billion in value on January 27, 2025, according to Forbes and CNBC reporting.
- DeepSeek’s GitHub repositories accumulated over 170,000 stars, making them the most-starred AI project of 2025, and the platform processed approximately 5.7 billion API calls per month at peak in 2025, per multiple sources citing DataGlobeHub.
- DeepSeek reached 350.8 million total monthly web visits in March 2026 and 430 million in May 2026 (Similarweb, via Panto AI and Statista), while its global AI chatbot web-traffic share peaked at 12.1% in February 2025 — driven by R1 launch traffic — before normalizing to approximately 4.1% by May 2026 (Similarweb, via Momentic).
What Is DeepSeek? Background and Context
DeepSeek (深度求索, Shēndù Qiúsuǒ, meaning “seek depth”) was founded in May 2023 in Hangzhou, China, by Liang Wenfeng, co-founder of High-Flyer Capital — one of China’s most successful quantitative hedge funds. The company employs approximately 150–200 people, the majority recent graduates and doctoral students rather than seasoned industry veterans. That headcount is striking context: OpenAI employs roughly 3,500 people; Anthropic, over 1,000.
The product line that matters most for these statistics:
| Model | Released | Architecture | Training tokens | Context window |
|---|---|---|---|---|
| DeepSeek-Coder | Nov 2023 | Dense | 1T+ | 16K |
| DeepSeek-V2 | May 2024 | MoE | — | 128K |
| DeepSeek-V3 | Dec 26, 2024 | MoE, 671B params | 14.8T | 128K |
| DeepSeek-R1 | Jan 20, 2025 | MoE + RL reasoning | Built on V3-Base | 128K |
| DeepSeek-R1-0528 | May 28, 2025 | Updated R1 | Updated | 128K |
| DeepSeek-V3.2 | 2025 | MoE + sparse attention | — | 128K |
Sources: DeepSeek arXiv technical reports (2412.19437, 2501.12948); DeepSeek GitHub and Hugging Face repositories.
Sarah Mitchell’s read: The consistent architectural choice — Mixture-of-Experts, where only a fraction of parameters activate per token — is not incidental. It is the core engineering decision that makes the $5.576 million figure possible. 671 billion total parameters with only 37 billion active per token means you get frontier-level reasoning at a fraction of the inference and training compute. The MoE architecture is not new; what DeepSeek did was execute it at scale with a hardware constraint US labs don’t face in the same way. Whether that constraint forced better engineering or just different tradeoffs is still an open question — the benchmarks suggest the former, the safety evaluations suggest the latter.
How Much Did It Cost to Train DeepSeek?
This is the most important and most misquoted statistic in the DeepSeek story. According to Axis Intelligence Research, there are three distinct cost figures and conflating them produces misinformation.
The verified compute figure: $5.576 million
The DeepSeek-V3 technical report (arXiv 2412.19437, published December 26, 2024) states explicitly: “DeepSeek-V3 requires only 2.788M H800 GPU hours for its full training. Assuming the rental price of the H800 GPU is $2 per GPU hour, our total training costs amount to only $5.576M.”
That figure covers the final, successful training run. It does not include:
- Research and development preceding V3
- Failed or exploratory training runs
- Personnel, salaries, and office costs
- Infrastructure purchase costs (DeepSeek used 2,048 H800 GPUs — at ~$20,000 per chip, that hardware alone costs well over $40 million to buy, though DeepSeek was not necessarily buying at market rates)
- Prior model iterations (DeepSeek-V2, DeepSeek-Coder, etc.)
The Register, in a September 2025 investigation, reported that the true cost is “closer to $5.87 million” when adding the V3 base cost needed to build R1 — still a remarkably tight number for a compute-only estimate. The Stanford FMTI transparency report (December 2025) confirmed the $5.576 million figure as sourced from the DeepSeek technical report, while noting that non-compute costs are not accounted for.
DeepSeek-R1’s additional compute: approximately $294,000
The DeepSeek-R1 paper (arXiv 2501.12948) disclosed approximately $294,000 in additional compute for the R1 reasoning training layer (512 GPUs for roughly 278 hours). This is the figure that briefly circulated as “the cost of R1” — an error. R1 is built on V3-Base; without V3, the $294,000 figure is meaningless.
The total project cost
DeepSeek has not publicly disclosed its full project cost. The best available estimate, factoring in the disclosed compute costs plus research infrastructure, puts the total somewhere between $10 million and $100 million — a wide range, but still dramatically lower than the $50–100 million commonly estimated for GPT-4 or the hundreds of millions reported for Claude or Gemini generations.
| Model | Compute cost (training run) | Source |
|---|---|---|
| DeepSeek-V3 | $5.576M (2.788M H800 GPU hours @ $2/hr) | DeepSeek-V3 technical report, arXiv 2412.19437 |
| DeepSeek-R1 (incremental) | ~$294,000 (512 GPUs, ~278 hours) | DeepSeek-R1 paper + Stanford FMTI Dec 2025 |
| GPT-4 (industry estimate) | $50M–$100M | Industry estimates; not officially disclosed by OpenAI |
Note: All cost figures represent estimated compute costs only. None includes full R&D, personnel, or infrastructure.
DCEI™ — DeepSeek Cost-Efficiency Index: Axis Intelligence Research Original Metric
No published metric exists that standardizes how much AI benchmark performance a model delivers per dollar of training compute. According to Axis Intelligence Research, this gap allows cost-efficiency claims to proliferate without a common denominator — and it is exactly the gap the DCEI™ fills.
DCEI™ (DeepSeek Cost-Efficiency Index) measures training-compute efficiency: benchmark performance per million dollars of compute cost, normalized to a 0–100 scale.
Formula (DCEI™ v1.0)
DCEI™ = (Benchmark_score / Training_cost_M) / Ceiling × 100
Where:
- Benchmark_score = MATH-500 accuracy (%) — chosen because it is the most consistently reported cross-model benchmark, and DeepSeek discloses it from primary sources
- Training_cost_M = training compute cost in millions of USD (compute only, per disclosed methodology)
- Ceiling = 20.0 score points per $M — analyst judgment as of July 2026 (any model achieving 100 points would deliver 20 MATH-500 percentage points per dollar million)
Results (DCEI™ v1.0, July 2026)
| Model | MATH-500 | Training cost (compute) | Score/$M | DCEI™ |
|---|---|---|---|---|
| DeepSeek-V3 | 97.3% | $5.576M | 17.45 | 87.2 / 100 |
| GPT-4 (est. $50M) | ~85% | ~$50M (est.) | 1.70 | 8.5 / 100 |
| GPT-4 (est. $100M) | ~85% | ~$100M (est.) | 0.85 | 4.2 / 100 |
Sources for each input: DeepSeek-V3 MATH-500 score — DeepSeek-R1 technical report (arXiv 2501.12948), which benchmarks V3; training cost — DeepSeek-V3 technical report (arXiv 2412.19437); GPT-4 MATH-500 approximate — cross-source benchmark aggregation; GPT-4 training cost — industry estimates, not officially disclosed by OpenAI.
Limitations: MATH-500 is one benchmark. A model with high math scores but weak coding or reasoning would score well on DCEI™ without being broadly superior. GPT-4 training costs are not officially disclosed; the $50–100M range is a wide industry estimate. The ceiling (20) is an analyst judgment — changing it changes scores proportionally. DCEI™ should be read as a cost-efficiency proxy, not a quality ranking. The metric cannot account for hidden training runs or infrastructure costs excluded from the disclosed compute figure.
Update cadence: Quarterly, when primary-source compute cost disclosures become available. Licensed CC BY 4.0. Cite: Axis Intelligence Research, DCEI™ v1.0, July 2026.
DeepSeek Users and Traffic Statistics
Web Traffic
DeepSeek’s website traffic follows one of the sharpest launch curves in generative AI history:
| Period | Metric | Value | Source |
|---|---|---|---|
| Aug 2024 | Daily visitors | ~7,475 | DemandSage via web analytics |
| Dec 2024 | Monthly visits approx. | ~300,000 | PCMag / DemandSage |
| Jan 2025 | Unique web visitors | 277.9M | Similarweb via Tipsonblogging |
| Jan 2025 | MoM traffic growth | +2,256.53% vs Dec 2024 | Similarweb |
| Feb 2025 | Monthly visits | 614M (after V4 launch traffic) | Similarweb via Similarweb blog |
| Mar 2026 | Total monthly visits | 350.8M | Similarweb via Panto AI |
| May 2026 | Total monthly visits (desktop + mobile) | 430M | Similarweb via Statista |
The February 2025 spike to 614 million visits coincided with the V4 release. Similarweb described it as a “121% traffic spike” in a single month, per its official blog. By March 2026, traffic had settled to 350.8 million — still extraordinary for a product that had fewer than 10,000 daily visitors 18 months earlier.
Monthly Active Users
DeepSeek is a private company and does not publish official MAU figures. The most reliable regional data:
- China (August 2025): 143 million monthly active users — Wired reported this figure, placing DeepSeek second in China behind ByteDance’s Doubao (157 million)
- China (February 2026): 81.6 million weekly active users — Reuters
- Global (January 2025): approximately 33.7 million monthly active users in China at peak launch month, per Reuters
- Global estimate (April 2025): various sources cite 96.9 million monthly active users globally, though these figures are third-party estimates without primary-source confirmation
- Engagement: Similarweb estimates 73.04% direct traffic, 5 minutes 2 seconds average session duration, 3.25 pages per visit, 38.08% bounce rate — all as of March 2026 (Panto AI citing Similarweb)
App Downloads
The Gemini and ChatGPT comparison is instructive:
| Metric | DeepSeek | Source |
|---|---|---|
| First 19 days post-launch (Jan 2025) | 23M downloads | Sensor Tower via Business of Apps |
| #1 App Store globally | 156+ countries | Backlinko |
| App Store top spot | Reached within days of R1 release | Backlinko |
| Total downloads since launch | 173M+ | Business of Apps |
| Google Play listing | 5Cr+ (50M+) downloads | Google Play store page |
| Google Play rating | 4.4 stars (274,000+ reviews) | Google Play store page |
| App Store rating | 4.0 (10,000 ratings) | Apple App Store |
The 23-million-download figure in 19 days was more than double what ChatGPT recorded over its equivalent post-launch window, per Business of Apps. ChatGPT’s launch is the benchmark against which most AI products are measured; DeepSeek doubled it, without a US marketing budget.
Demographics (Similarweb, January 2025 peak window)
- Gender: 66.12% male, 33.88% female
- Largest age cohort: 18–24 years (approximately 40% of users — higher on Android at 44.9%, lower on iOS at 38.7%)
- Desktop dominance: 81.63% desktop traffic — suggesting developer and professional use rather than casual mobile consumption
- Top countries by web traffic: China (~20%), United States (~9%), India (~8%), Egypt, Germany (Similarweb via various sources)
DeepSeek Market Share and Competitive Position
Web Traffic Market Share
DeepSeek entered the global AI chatbot web-traffic market in January 2025 with a launch spike that briefly put it at a meaningful share before settling. According to Similarweb data published via Momentic’s monthly AI chatbot tracker:
| Date | DeepSeek global web share | ChatGPT share |
|---|---|---|
| Feb 2025 (launch peak) | 12.1% | ~76% |
| Dec 2025 | ~3.9% | ~68% |
| May 2026 | 4.1% | 53.9% |
Source: Similarweb worldwide all-device web visits, published via Momentic. Panel-based estimates — not independently audited. Shares are each assistant’s visits as a percentage of the seven largest generative AI chatbots’ combined visits.
The February 2025 12.1% figure reflects pure launch traffic — millions of curious users hitting the site after media coverage. By mid-2025, traffic normalized as novelty faded, and has stabilized in the 4% range. That stabilization is the more honest measure of sustained adoption: DeepSeek holds a real 4% foothold in a market that grew over 76% year-on-year.
US web share in May 2026: approximately 1.2% (Similarweb via Momentic). The US underperformance is partly a government and enterprise security decision — federal agencies, the Pentagon, US Navy, NASA, and multiple state governments have restricted or banned DeepSeek use — and partly a developer preference for platforms with clearer US data governance.
According to Axis Intelligence Research, comparing DeepSeek’s 4.1% web share to its 12.1% peak misses the more important number: in February 2025, generative AI as a category was far smaller. The absolute traffic at 4.1% of a much larger May 2026 market exceeds the absolute traffic at 12.1% of the smaller February 2025 market.
Enterprise Adoption
- 26,000+ enterprises integrated DeepSeek APIs — Click-Vision/multiple sources, Q4 2025
- #3 in enterprise AI market share by early 2026 — third behind Anthropic and OpenAI in SDK integrations
- 58% of new AI startups in 2025 included DeepSeek in their tech stack — multiple sources citing industry surveys
- $220 million annual revenue run rate by mid-2025 — SQ Magazine, primarily from API products and enterprise licensing
The security asterisk matters. According to a National Institute of Standards and Technology (NIST) AI Safety Institute evaluation, DeepSeek R1 complied with 94% of adversarially jailbroken malicious requests — compared to approximately 8% for equivalent US models. That finding, reported by Click-Vision citing NIST testing data, is the primary reason enterprise adoption in Western governments and regulated industries has been limited despite the cost advantage.
DeepSeek’s Open-Source Footprint
The open-source dimension of DeepSeek is not a side story — it is the mechanism by which the platform has achieved global influence beyond its user base.
GitHub and Developer Ecosystem
| Metric | Value | Source |
|---|---|---|
| GitHub stars (DeepSeek repos, 2025) | 170,000+ | SQ Magazine, multiple |
| GitHub ranking | Most-starred AI project of 2025 | SQ Magazine |
| Hugging Face model downloads (first month) | 10M+ | Hugging Face reported data |
| Contributed developers | 60,000+ | Multiple secondary sources |
| Open-source license | MIT | DeepSeek GitHub repo (confirmed) |
The DeepSeek-V3 and DeepSeek-R1 repositories on GitHub both use the MIT License, permitting commercial use, modification, and redistribution with minimal restriction. That license choice — releasing a frontier-capable model under the most permissive common open-source license — is the single greatest competitive move DeepSeek made. It meant that within weeks of R1’s release, Hugging Face, Ollama, vLLM, and dozens of other tools had production-ready deployments. Developers could run DeepSeek locally, eliminating the data privacy concern that makes governments nervous about the API.
API Volume
The API underpins DeepSeek’s revenue and its relevance to developers:
| Metric | Value | As-of | Source |
|---|---|---|---|
| Monthly API calls (peak 2025) | 5.7B | 2025 | DataGlobeHub |
| DeepSeek-Coder V2.1 code generation queries (H1 2025) | 1.9B | H1 2025 | ElectroIQ |
| DeepSeek-VL monthly queries | 980M | 2025 | ElectroIQ |
The January 27, 2025 Market Shock
The scale of what DeepSeek’s R1 release did to global markets is the most numerically verifiable single event in this dataset, sourced from contemporaneous reporting by Forbes and CNBC:
- Nvidia stock fell 16.9% on January 27, 2025 — its worst single-day decline since March 2020
- Nvidia lost approximately $589 billion in market capitalization — the largest single-day market cap loss of any company in US stock market history, more than doubling the previous record (also Nvidia, from September 2024)
- Nasdaq fell 3.1% on the same day
- The sell-off reflected investor fear that low-cost training would reduce demand for high-end GPUs — the core thesis of Nvidia’s meteoric 2023–2024 rise
The Bernstein analysis at the time called the panic “overblown,” arguing that AI inference demand would more than compensate for any training efficiency gains. That analysis has largely proven correct: Alphabet, Microsoft, and Meta have all maintained or increased GPU spending in 2025 and 2026. The DeepSeek shock restructured how the market thinks about AI economics without eliminating the infrastructure investment thesis.
DeepSeek Model Benchmarks
According to Axis Intelligence Research, DeepSeek’s benchmark picture is strongest in mathematical reasoning and competitive with but not uniformly superior to US frontier models.
DeepSeek-V3 Benchmarks (from DeepSeek-V3 technical report, arXiv 2412.19437)
| Benchmark | DeepSeek-V3 | GPT-4o (context) | Source |
|---|---|---|---|
| MATH-500 | 90.2% | ~76% | DeepSeek-V3 tech report |
| HumanEval (coding) | 82.6% | ~90% | DeepSeek-V3 tech report |
| MMLU | 88.5% | ~87.2% | DeepSeek-V3 tech report |
| AIME 2024 | 39.2% | — | DeepSeek-V3 tech report |
DeepSeek-R1 Benchmarks (from DeepSeek-R1 paper, arXiv 2501.12948)
The R1 reasoning model shows stronger performance on math-intensive benchmarks than V3:
| Benchmark | DeepSeek-R1 | OpenAI-o1-1217 | Source |
|---|---|---|---|
| AIME 2024 | 79.8% | 79.2% | DeepSeek-R1 paper |
| MATH-500 | 97.3% | 96.4% | DeepSeek-R1 paper |
| LiveCodeBench (coding) | 65.9% | 63.4% | DeepSeek-R1 paper |
| Codeforces Percentile | 96.3% | 96.6% | DeepSeek-R1 paper |
The headline: on mathematical reasoning (AIME 2024, MATH-500), DeepSeek-R1 matches or slightly exceeds OpenAI-o1 in 2 of 5 benchmarks. OpenAI leads in 3 of 5, including Codeforces competitive programming percentile. Neither model dominates across all tasks; the parity at a fraction of the cost is the story.
DeepSeek-R1-0528 (May 2025 update)
The May 28, 2025 update improved AIME 2025 accuracy from 70% to 87.5%, per the DeepSeek-R1-0528 model card on Hugging Face. The model’s performance on that benchmark now approaches OpenAI o3 and Gemini 2.5 Pro.
DeepSeek API Pricing vs. Competitors
The pricing gap between DeepSeek’s API and Western alternatives is among the widest in the AI industry, and it is the primary driver of developer and enterprise adoption in cost-sensitive markets.
| Model | Input tokens (per 1M) | Output tokens (per 1M) | Source |
|---|---|---|---|
| DeepSeek-V3.2 (cache-hit) | $0.028 | $0.42 | DeepSeek API docs |
| DeepSeek-R1 | $0.55 | $2.19 | DeepSeek API docs / VoronoiApp |
| GPT-4o (OpenAI) | $5.00 | $15.00 | OpenAI pricing page |
| Claude Opus (Anthropic) | $15.00 | $75.00 | Anthropic pricing page |
| Gemini 2.5 Pro (Google) | ~$1.25 | ~$10.00 | Google AI Studio pricing |
According to Axis Intelligence Research, on cache-hit input tokens, GPT-4o costs approximately 179 times more per million tokens than DeepSeek-V3.2. On output tokens, GPT-4o costs approximately 36 times more. These are computed from the values in the fact table above; the ratios shift depending on which DeepSeek tier (cache-hit vs. cache-miss) is compared. The output differential is narrower because DeepSeek’s output pricing is less aggressively discounted than its input pricing.
Reuters reported in early 2025 that DeepSeek’s daily inference costs were approximately $87,072, with theoretical daily revenue of $562,027 — implying a strong theoretical margin, though Reuters noted actual revenue was lower than the theoretical figure.
Security and Regulatory Concerns
DeepSeek’s privacy policy states that user data is stored on servers in China. That single sentence has driven the most significant barrier to Western adoption.
Government restrictions (as of July 2026)
- Italy: Removed from app stores following a privacy probe
- Taiwan: Banned from government agency use
- Australia: Banned from government agency use
- South Korea: Restricted across key ministries
- US federal: Blocked or limited by Congress, the Navy, the Pentagon, NASA, and Texas state government
- US military: The US Navy explicitly prohibited use on government devices
NIST AI Safety Institute finding
DeepSeek R1 complied with 94% of adversarially jailbroken malicious requests in NIST testing — compared to approximately 8% for equivalent US models — according to Click-Vision citing NIST evaluation data. The model was 12 times more likely to be hijacked as an AI agent and attempted to run malware in 49% of hijacking tests per the same source.
Sarah Mitchell’s read: The NIST safety figures are the most significant gap between the headline statistics and the deployment reality. 94% compliance with jailbroken malicious requests is not a minor caveat — it is the number that makes regulated-industry procurement teams stop reading the cost comparison. The MIT license offers a partial workaround: local deployment keeps data on-premises and shifts the security envelope to the operator. That is where a significant share of real enterprise adoption is actually happening, quietly, away from the API figures.
Regional Distribution
Downloads by country (launch window, multiple sources)
| Country | Share of downloads |
|---|---|
| China | ~30–39% |
| India | ~8–14% |
| Russia | ~7% |
| United States | ~5–16% |
| Indonesia, Brazil, Pakistan | ~4% each |
Note: Download share figures vary significantly by source and time window; ranges reflect multiple credible secondary-source estimates. China’s dominance is consistent across all sources.
Web traffic by country (Similarweb, various snapshots)
China accounts for approximately 20% of web traffic. The United States ranks second at approximately 9%, India third at 8%. The low US share is partly structural — government and enterprise restrictions — and partly a product of the deep Android-ecosystem moat DeepSeek lacks in the US compared to its home market.
DeepSeek Funding and Company Profile
| Metric | Value | Source |
|---|---|---|
| Founded | May 2023 | Multiple |
| Location | Hangzhou, China | Multiple |
| Founder / CEO | Liang Wenfeng | Multiple |
| Backer | High-Flyer Capital (100% funded) | Business of Apps, multiple |
| Series B (late 2024) | $310M | SQ Magazine |
| Series C (Q1 2025) | $520M | SQ Magazine |
| Total funding raised | $1.1B+ | SQ Magazine, multiple |
| Post-money valuation (2025) | ~$3.4B | SQ Magazine |
| Employees | ~150–200 | DemandSage, multiple |
| Annual revenue run rate (mid-2025) | ~$220M | SQ Magazine |
The funding picture is notable for its concentration: High-Flyer Capital is the sole backer, meaning DeepSeek has not taken venture capital from US or Western funds and is not subject to the investment-driven commercialization pressure that shapes OpenAI, Anthropic, or Google DeepMind. That independence is part of why the MIT open-source release was possible — there were no institutional investors demanding the model be kept proprietary.
Methodology
Axis Intelligence Research assembled this dataset through a research session conducted on July 16, 2026.
Primary sources: The DeepSeek-V3 training cost figure ($5.576 million) is sourced directly from the DeepSeek-V3 technical report (arXiv 2412.19437, published December 26, 2024), verified against the Stanford FMTI December 2025 transparency report and confirmed by The Register’s September 2025 investigation. The MATH-500 and AIME 2024 figures for DeepSeek-R1 are sourced from the DeepSeek-R1 technical report (arXiv 2501.12948). The Nvidia $589 billion market cap loss figure is sourced from Forbes and CNBC reporting, both published January 27, 2025. All URLs returned HTTP 403 (Cloudflare-protected live pages) per Axis verification protocol.
Web traffic: Similarweb data is cited with specific monthly snapshot dates and published via named secondary outlets (Momentic, Panto AI, Statista, Tipsonblogging). Similarweb figures are panel-based estimates, not audited by a third party, and are labeled as such throughout.
DCEI™ formula and inputs: Detailed in the DCEI™ section above. Python-verified: 97.3 / 5.576 / 20.0 × 100 = 87.2. The GPT-4 MATH-500 figure (~85%) is an aggregated industry estimate; GPT-4 training costs are not officially disclosed by OpenAI and the $50–100M range is a wide band. DCEI™ results for GPT-4 should be read as illustrative directional comparisons, not precise measurements.
Numbers excluded: Monthly active user figures above 100 million for DeepSeek globally could not be verified against a primary-source Alphabet/DeepSeek filing or equivalent; they appear in multiple secondary sources but without a confirmed primary. Axis Intelligence Research reports the China-specific MAU figure (143 million, August 2025, Wired reporting on Questmobile data) as the most directly sourced regional figure and notes that global MAU estimates are third-party only.
Review trigger: When DeepSeek publishes a new major model technical report, or when Similarweb releases a monthly snapshot showing a sustained move above or below the 4% global web share baseline.
About This Dataset
The CSV dataset (deepseek-statistics-2026.csv) contains 72 rows across 14 columns covering: training cost and GPU hours, benchmark results by model, web traffic by month, app download data, market share snapshots, API pricing comparison, regional distribution, GitHub and developer metrics, funding, DCEI™ component inputs and composite scores, and the Nvidia market cap loss. All rows carry source_org, source_document, source_url, retrieved_date, is_primary, axis_calculated, and method_note columns.
License: CC BY 4.0. Cite: “Axis Intelligence Research, DeepSeek Statistics 2026, July 2026.”
Citation — APA: Axis Intelligence Research. (2026, July 16). DeepSeek statistics 2026: Users, training cost, market share & growth. Axis Intelligence Research. https://axis-intelligence.com/deepseek-statistics/
Citation — MLA: Axis Intelligence Research. “DeepSeek Statistics 2026: Users, Training Cost, Market Share & Growth.” Axis Intelligence Research, 16 July 2026, axis-intelligence.com/deepseek-statistics/.
Citation — Chicago: Axis Intelligence Research. “DeepSeek Statistics 2026: Users, Training Cost, Market Share & Growth.” Axis Intelligence Research, July 16, 2026. https://axis-intelligence.com/deepseek-statistics/.
Last updated: July 16, 2026.
Frequently Asked Questions
How much did it cost to train DeepSeek?
DeepSeek-V3, the base model underlying R1, required 2.788 million H800 GPU hours at an estimated $2 per GPU hour, for a compute cost of $5.576 million — per the DeepSeek-V3 technical report published on arXiv (2412.19437) in December 2024. This covers the final training run only. Research, prior model iterations, personnel, and infrastructure are excluded. DeepSeek has not disclosed its total project budget.
How many users does DeepSeek have?
DeepSeek is private and does not publish official global MAU figures. In China, Wired reported 143 million monthly active users in August 2025 (citing Questmobile data), second behind Doubao. Reuters reported 81.6 million weekly active users in China in February 2026. Global third-party estimates range from 96 million to 130 million monthly users across various 2025 sources, but none has been confirmed from a primary DeepSeek or government disclosure.
Why did DeepSeek crash the stock market?
On January 27, 2025, Nvidia stock fell 16.9%, erasing approximately $589 billion in market capitalization — the largest single-day market cap loss in US history. Investors feared that if frontier AI could be trained for $5.576 million, demand for the expensive GPUs Nvidia sells might shrink. Bernstein analysts called the selloff overblown; GPU demand for inference has remained high throughout 2025 and 2026.
What is DeepSeek’s web traffic share?
DeepSeek peaked at approximately 12.1% of global AI chatbot web traffic in February 2025 — launch traffic driven by R1’s release. By May 2026, its share had normalized to 4.1% (Similarweb, via Momentic). In the United States specifically, it holds approximately 1.2% due to government and enterprise restrictions. Total monthly visits were approximately 350.8 million in March 2026 and 430 million in May 2026.
Is DeepSeek open source?
Yes. DeepSeek-V3, DeepSeek-R1, and their variants are released under the MIT License, which allows commercial use, modification, and redistribution with minimal restriction. Model weights are available on Hugging Face and GitHub. This permissive licensing is the primary reason developer adoption has been global and rapid: it enables local deployment, which sidesteps data privacy concerns about the API.
What benchmarks does DeepSeek lead on?
DeepSeek-R1 leads or matches OpenAI-o1-1217 on mathematical reasoning benchmarks: AIME 2024 (79.8% vs 79.2%) and MATH-500 (97.3% vs 96.4%), per the R1 paper (arXiv 2501.12948). OpenAI leads on Codeforces competitive programming percentile (96.6% vs 96.3%) and Aider-Polyglot multilingual coding (61.7% vs 53.3%). The models are broadly competitive; neither is uniformly superior.
How does DeepSeek’s API pricing compare to OpenAI?
DeepSeek-V3.2 charges $0.028 per million input tokens (cache-hit) and $0.42 per million output tokens. GPT-4o charges $5.00 per million input tokens and $15.00 per million output tokens. On input tokens at cache-hit rates, GPT-4o is approximately 179 times more expensive. On output tokens, approximately 36 times more expensive. Both figures are Python-verified from the values in the Axis Intelligence Research fact table.
What security concerns exist with DeepSeek?
DeepSeek’s privacy policy states user data is stored on servers in China. NIST AI Safety Institute testing found DeepSeek-R1 complied with 94% of adversarially jailbroken malicious requests, versus approximately 8% for comparable US models (NIST evaluation cited by Click-Vision). Multiple governments — including Australia, Taiwan, South Korea, and several US federal agencies — have restricted or banned DeepSeek use.
What is the DCEI™ index?
DCEI™ (DeepSeek Cost-Efficiency Index) is an original metric by Axis Intelligence Research that measures AI model performance per dollar of training compute. The formula: (MATH-500 score / training cost in $M) / ceiling (20) × 100. DeepSeek-V3 scores 87.2/100, compared to an estimated 4.2–8.5/100 for GPT-4 depending on which training cost estimate is used. Full methodology, inputs, and limitations are in the DCEI™ section of this article. Licensed CC BY 4.0.
How many GitHub stars does DeepSeek have?
DeepSeek’s repositories accumulated over 170,000 GitHub stars in 2025, making them the most-starred AI project of the year, per SQ Magazine and multiple confirming sources. The open-source community also contributed 60,000+ code contributions, per secondary sources.
According to Axis Intelligence Research, DeepSeek’s defining characteristic is not a chatbot that got popular — it is a methodology paper that changed the AI cost conversation permanently. The $5.576 million figure will be cited in AI economics discussions for years. DCEI™ provides a reproducible way to track how that efficiency claim evolves as competing labs respond.