DeepSeek Statistics 2026
Users, Growth & Market Impact
From a $6M training cost that shook Silicon Valley to a $45B valuation and 130M+ monthly active users — the complete data story behind AI's biggest disruptor.
On January 20, 2025, a Chinese AI startup called DeepSeek released a reasoning model called R1 — and within 72 hours, it had erased an estimated $1 trillion in US market value from Nvidia, Microsoft, and other AI-adjacent stocks. The model matched OpenAI's o1 in key benchmarks, but had reportedly been trained for just $5.6–6 million — a fraction of the hundreds of millions competitors were spending.
What followed was one of the fastest user adoption curves in AI history. This page consolidates 100+ verified statistics on DeepSeek's users, growth, benchmarks, funding, and global impact — updated through May 2026.
About This Data
DeepSeek is a private company and does not publish official user figures. Statistics on this page are sourced from Backlinko, Business of Apps, SimilarWeb, Reuters, Bloomberg, TechCrunch, Sensor Tower, and Panto AI research. Where sources differ, we note the range. See our broader AI statistics 2026 and most popular AI tools 2026 reports for context.
What is DeepSeek?
A fast primer on the company, its founders, and what makes it genuinely different from OpenAI and other Western AI labs.
DeepSeek (深度求索, meaning "Seek the Deep Truth") is an AI research lab founded in 2023 in Hangzhou, China. It is wholly funded by High-Flyer Capital Management, a quantitative hedge fund that uses AI for trading algorithms. The fund's founder, Liang Wenfeng, bootstrapped DeepSeek without external investment for its first two-and-a-half years — a philosophical choice rooted in protecting a research-first culture.
The company has only 101–200 employees — making it one of the smallest AI labs operating at frontier scale anywhere in the world. Its models are open-source under the MIT license, meaning anyone can download, modify, and deploy them commercially for free. This single decision has been central to its explosive global adoption.
DeepSeek User Statistics 2026
Monthly active users, daily active users, and how DeepSeek grew from a niche research tool to one of the world's top AI apps in under six months.
Context: How It Compares
DeepSeek's 96.88M MAU in April 2025 made it the #4 most popular AI app globally — behind ChatGPT, Google Gemini, and Microsoft Copilot. It surpassed this in just 3 months after a global launch. By August 2025, DeepSeek had 143 million monthly active users in China alone (Wired), making it second only to ByteDance's Doubao (157M) in the Chinese market. Read more in our ChatGPT users report for comparison.
The growth rate itself is historically unprecedented for a non-gaming app: DeepSeek grew from 33.7M to 96.88M monthly active users in just three months — a 187% increase in 90 days. The March-to-April 2025 growth alone was +25.81%. To contextualize this, it took ChatGPT approximately 5 months to reach 100M users (a record at the time); DeepSeek did it in roughly the same window but with a fraction of the marketing budget.
Website Traffic Statistics
How DeepSeek's web traffic exploded from under 8,000 daily visitors to hundreds of millions in fourteen months.
DeepSeek's traffic pattern is unlike any other AI platform. 81.63% of its traffic comes from desktop, suggesting it's primarily used for professional and development work rather than casual mobile consumption — a sharp contrast to ChatGPT and Gemini, which skew more mobile-heavy. Average visit duration is 5 minutes and 2 seconds with 3.25 pages per visit — indicating high engagement, not just curiosity clicks.
Before the R1 launch, DeepSeek had about 7,475 daily visitors in August 2024. The value of their traffic (if they had to pay for it through ads) grew from about $2,800/day to over $100,000/day in January 2025 — a 35x increase in traffic value overnight.
Downloads & Geographic Distribution
DeepSeek became the #1 most downloaded app in 156 countries within days of launch. Here's where it landed.
Monthly Active Users by Country
Countries That Banned DeepSeek
Several countries removed DeepSeek from app stores over data security and privacy concerns. Italy launched a privacy probe and removed it from stores. South Korea, Taiwan, Australia (government devices), and parts of the EU raised formal concerns. The US Congress introduced legislation restricting its use on government devices. DeepSeek censors politically sensitive topics by default due to its Chinese origins — a limitation for unrestricted enterprise use outside China.
User Demographics
Who uses DeepSeek? Age, platform, and device breakdown based on available app store and analytics data.
DeepSeek's user base skews younger than ChatGPT: the 18–24 cohort dominates across both platforms, with Android showing even stronger youth adoption. This demographic is typically dominated by students, developers, and early adopters who are drawn to the model's open-source nature and coding capabilities. The 50–64 age group is surprisingly strong (23.3% iOS, 26.1% Android) — suggesting broader appeal than its "developer tool" positioning might suggest.
AI Benchmark Performance
How DeepSeek's models score against OpenAI, Claude, and other frontier models on standardized tests.
DeepSeek R1's January 2025 launch sent shockwaves precisely because its benchmarks matched or exceeded OpenAI's o1 on reasoning tasks — at a reported 214x lower inference cost than Claude 3.5 Sonnet. The models use a Mixture of Experts (MoE) architecture: despite having 671B total parameters, only 37B activate per query, making it dramatically cheaper to run than dense models.
AIME 2024 (Math Reasoning)
MMLU (General Knowledge)
HumanEval (Coding)
GSM8K (Grade School Math)
Model Architecture: Why It's Efficient
DeepSeek uses a Mixture-of-Experts (MoE) design — the model has 671B total parameters but only activates 37B per query. This is like having a huge team of specialists but only calling on the relevant experts for each task, rather than running every specialist's brain for every request. The result: frontier-level performance at a fraction of the compute cost. DeepSeek was also trained using reinforcement learning rather than expensive supervised fine-tuning — trained on just 2,048 Nvidia H800 GPUs over 55 days.
Training Cost & Efficiency
The numbers that shook the AI industry: how DeepSeek built frontier-level AI at a fraction of what Western labs were spending.
The $5.6 million training cost figure was contested by some AI researchers who argued it didn't include R&D amortization, failed experiments, or infrastructure costs. Even accounting for these caveats, DeepSeek's efficiency is genuinely and dramatically better than Western frontier labs. It was trained on 14.8 trillion tokens using Nvidia H800 chips — older, export-restricted GPU models that are less powerful than the H100s used by US labs. This constraint forced innovations in training efficiency that ended up being competitive advantages.
The cost of DeepSeek's API is also a fraction of rivals: at $0.14 per million input tokens for V3, it represents a 214x cost advantage over Claude 3.5 Sonnet ($30/million tokens). This pricing is reshaping enterprise AI budgets globally. See our AI market size statistics 2026 for the broader market context.
Funding & Valuation
DeepSeek went from zero external funding to a reported $45B valuation in the space of a few weeks in mid-2026 — here's how the numbers evolved.
For two and a half years, DeepSeek refused every venture check that came its way. Founder Liang Wenfeng publicly cited the desire to protect a research-first culture unencumbered by quarterly revenue pressure. The pivot to external capital in April 2026 reflects a judgment that the next model generation — more complex reasoning, agent systems, and infrastructure — cannot be built within the cash flow of a hedge fund alone.
The round is reportedly led by China Integrated Circuit Industry Investment Fund (a state investment vehicle), with Tencent offering to buy up to 20% and Alibaba also in negotiations. China is strategically funding homegrown AI development to sidestep US chip export restrictions — and DeepSeek, which has already been optimized to run on Huawei chips, is central to that strategy. See our AI startups 2026 report for context on the broader landscape.
Revenue Reality: Open Source Challenge
DeepSeek has generated very little direct revenue — its models are open-source and its chatbot is free for consumers. This is why early valuation discussions ranged from $10B to $45B: investors disagree on how to price a company with frontier AI capabilities, massive user scale, and near-zero revenue. In April 2026, DeepSeek began offering developers a temporary 75% discount on API usage amid intensifying competition in China's AI market.
Market Impact — The $1 Trillion Day
What happened to global markets, US tech stocks, and the AI industry the day DeepSeek R1 launched.
The January 27, 2025 "DeepSeek Day" triggered the largest single-day tech sector decline since September 2020. Nvidia alone lost approximately $600 billion in market cap — the largest single-day loss for any company in stock market history. The reasoning: if frontier AI could be trained for $6 million rather than hundreds of millions, the massive datacenter GPU buildout driving Nvidia's valuation might be overstated.
The broader AI industry consequences were structural: ByteDance, Tencent, Baidu, and Alibaba all cut their AI API prices within weeks to compete. OpenAI accelerated its release of reasoning-optimized models. The "compute is everything" consensus in AI was permanently disrupted. Read our AI market size statistics 2026 for the full macro picture.
DeepSeek Timeline — From Lab to Global Force
The key milestones in DeepSeek's rapid rise from a hedge fund side project to one of the most consequential AI labs in the world.
Enterprise Adoption Statistics
Beyond consumer chatbot users — how DeepSeek is being adopted by enterprise, developers, and the open-source community.
DeepSeek's open-source model has accelerated enterprise adoption because companies can self-host the model without sending data to a third-party cloud — a critical requirement for regulated industries. Banks, healthcare systems, and government agencies globally have deployed private DeepSeek instances. The model has also been optimized to run on Huawei Ascend chips, making it accessible to Chinese enterprises that cannot use Nvidia hardware due to export restrictions.
DeepSeek vs ChatGPT — Side by Side
The two most-discussed AI platforms compared across every dimension that matters for users and enterprises.
| Factor | DeepSeek | ChatGPT (GPT-4o) |
|---|---|---|
| Founding | 2023, Hangzhou China | 2022, San Francisco USA |
| Model Architecture | Mixture of Experts (671B params, 37B active) | Dense (estimated 1.8T params) |
| Training Cost (reported) | ~$5.6M (R1) | $100M+ (GPT-4) |
| API Cost (input tokens) | $0.14/1M tokens | $10–30/1M tokens |
| Open Source | ✓ MIT License | ✗ Closed source |
| Monthly Active Users | 130M+ (Dec 2025) | 400M+ (Feb 2025) |
| AIME 2024 Score | 79.8% | 79.2% (o1) |
| MMLU Score | 90.8% | 91.8% (o1) |
| Multimodal | Partial (V3.2) | ✓ Full |
| Content Censorship | Political topics restricted | Partial (safety policies) |
| Self-hosting possible | ✓ Yes | ✗ Cloud only |
| Free Consumer Tier | ✓ Unlimited | Limited (GPT-3.5) |
| Data Privacy | China jurisdiction concern | US jurisdiction |
| Valuation (2026) | $20–45B (in talks) | $300B+ (OpenAI) |
When to Use DeepSeek vs ChatGPT
Choose DeepSeek for: math and reasoning tasks, coding (especially complex algorithms), cost-sensitive API applications, self-hosted enterprise deployments, and open-source integration projects. Choose ChatGPT for: multimodal tasks (images, audio), enterprise data environments where Chinese data jurisdiction is a concern, consumer applications requiring broader language support, and when GPT-4's ecosystem integrations matter. For a broader comparison, see our most popular AI tools 2026 guide.
Frequently Asked Questions
How many users does DeepSeek have in 2026? ▼
DeepSeek reached approximately 130 million monthly active users globally by the end of 2025, according to Business of Apps. In China alone, it had 143 million MAUs by August 2025 (Wired). Daily active users peaked at 22.15 million as of January 2025. The platform has been downloaded 173 million times since its global launch in January 2025. As of March 2026, its website received 350.8 million monthly visits.
How much did it cost to train DeepSeek? ▼
DeepSeek reported training costs of approximately $5.6–6 million for its R1 model, using 2,048 Nvidia H800 GPUs over 55 days. This compares to estimated training costs of $100M+ for GPT-4. Some AI researchers disputed whether the figure included all R&D costs, but even critics acknowledge DeepSeek's training efficiency is dramatically better than Western frontier labs. The V3 model was trained on 14.8 trillion tokens using reinforcement learning rather than expensive supervised fine-tuning. See our AI statistics 2026 for more training cost comparisons.
What is DeepSeek's valuation in 2026? ▼
DeepSeek is raising its first-ever external funding round in April–May 2026. Initial reports (The Information, April 17) cited a $10B+ target; within 48 hours investor interest pushed this above $20B (Reuters). By May 2026, the Financial Times and Bloomberg reported the valuation had reached $45 billion. The round is reportedly led by China's state investment fund, with Tencent and Alibaba also in negotiations. No deal has been finalized and valuation could change. See our AI startups 2026 report for context.
Is DeepSeek safe to use? What are the privacy concerns? ▼
DeepSeek raises legitimate data privacy concerns for users outside China. It is subject to Chinese jurisdiction and Chinese laws that can require companies to share data with the government. A cloud security firm identified a significant data exposure incident involving DeepSeek's systems. Several countries — including Italy, South Korea, Taiwan, and Australia (for government devices) — have restricted its use. For enterprise use outside China, self-hosting the open-source model (rather than using DeepSeek's cloud API) mitigates data transfer concerns. The model also applies censorship to politically sensitive topics by default, limiting its use for certain research applications.
How does DeepSeek compare to ChatGPT on benchmarks? ▼
DeepSeek R1 matches OpenAI o1 on math reasoning (79.8% vs 79.2% on AIME 2024) and scores 90.8% on MMLU general knowledge (vs o1's 91.8%). DeepSeek-Coder V2.1 achieved 85.6% on HumanEval — the highest for any open-source coding model in 2025. It's the second most preferred coding assistant per StackOverflow's 2025 survey, behind only GitHub Copilot. Where ChatGPT has a meaningful advantage is multimodal capabilities (images, audio) and its broader ecosystem. For detailed benchmarks, see our most popular AI tools 2026 guide.
Why did DeepSeek wipe $1 trillion in US market value? ▼
When DeepSeek R1 launched on January 20, 2025, its reported $6 million training cost challenged the core thesis that frontier AI required massive, expensive compute infrastructure — which had been driving record GPU sales for Nvidia and datacenter buildouts for Microsoft, Amazon, and Google. If AI could be trained at a fraction of assumed costs, the billions being invested in AI infrastructure seemed potentially overvalued. Nvidia's stock fell ~17% on January 27, 2025 — the largest single-day market cap loss in stock market history (~$600B). The broader tech selloff totaled approximately $1 trillion. Chip stocks, AI infrastructure names, and power/cooling companies all declined sharply.