It has been over 150 days since China's large language model DeepSeek R1 was released with reasoning capabilities comparable to OpenAI and a shockingly low price below 90% of the industry. At that time, the market was generally concerned that this move would trigger a cutthroat battle for the commoditization of AI models. However, the data shows a different picture: despite DeepSeek's model being very popular on third-party platforms (like OpenRouter), traffic for its self-operated applications and API services has decreased instead, with market share continuing to decline.
Research institution (SemiAnalysis) points out that the core reason lies in DeepSeek's strategy of sacrificing user experience to cope with computing power limitations under U.S. chip export restrictions. Although it offers low prices, it comes with extremely high latency (Time-to-First-Token) and a very small context window (Context Window), forcing users to wait several seconds for a response.
(SemiAnalysis) analysis indicates that this deliberate trade-off is aimed at concentrating limited high-level computing resources on model R&D in a bid to achieve breakthroughs in the ultimate competition for AGI (General Artificial Intelligence), rather than focusing on current API service profitability.
DeepSeek App Market Share: Dropped from 8% to 4.5%
So, how severe is the decline in DeepSeek's App market share? According to (SemiAnalysis), DeepSeek could be said to have peaked upon debut. After launching the App in January this year, its market share experienced nearly vertical explosive growth, skyrocketing from nearly 0% to a peak of over 8%.
Research emphasizes that due to incomplete user data tracking in the Chinese market and many Western AI services being banned in China, the data likely underestimates DeepSeek's real influence. However, overall, DeepSeek's market share has been declining since peaking in January, having dropped to around 4.5% by the end of May; notably, the overall market's AI App user index has been rising during the same period, indicating a trend of user migration.
Source of the chart: (Digital Age) Comparison trend chart of 'AI Application User Index' and 'DeepSeek Market Share'.
DeepSeek Web Visits: Decline of nearly 30%
On the browser access side, the visits to the DeepSeek web version have also been declining. According to SimilarWeb data, during the three months from February to May this year, DeepSeek's website traffic dropped from 614 million (614 M) to 436 million (436 M), a decline of 29%.
In contrast, ChatGPT, as the market leader, increased its traffic from 3.905 billion (3,905 M) to 5.492 billion (5,492 M), achieving a strong growth of 40.6%. The growth of Gemini under Google is even more astonishing, with traffic rising significantly from 284 million (284 M) to 528 million (528 M), a growth rate of 85.8%.
As for Grok, which joined the fray later, due to its lower traffic base, the traffic exploded from 51 million (51 M) to 179 million (179 M), achieving an astonishing growth rate of 247.1%, the highest among all services. Claude, on the other hand, grew steadily, with traffic increasing from 73 million (73 M) to 100 million (100 M), realizing a growth of 36.5%.
Changes in browser visits of 5 major AI models
Source: (Digital Age)
Sacrificing experience for the future, DeepSeek's computing power is under pressure.
DeepSeek's strategic choices are clearly reflected in various service metrics. Compared to other providers, users using DeepSeek's official services must endure delays of over 25 seconds and only a 64K context window, severely limiting application scenarios that require substantial memory, such as code analysis.
(SemiAnalysis) points out that in stark contrast, the total usage of third-party hosted DeepSeek model instances has grown nearly 20 times since the R1 release, as these service providers utilize better-optimized hardware configurations to offer lower latency and larger context windows.
DeepSeek utilizes high-concurrency batch processing of user requests to maximize the performance of a single GPU, thereby reducing costs per million tokens. This move has brought global model recognition and adoption by the open-source community but has also shifted the burden and opportunities of the service onto third-party cloud platforms.
(SemiAnalysis) points out that this reflects the dilemma faced by Chinese AI companies, which, unable to obtain high-end chips like those from Nvidia on a large scale, can only 'exchange (users') time for (budget) space' to ensure that core R&D does not fall behind.
Three elements of token economy: 'CP value' indicator aside from price.
The DeepSeek phenomenon reveals that the core competition in the current AI market is no longer simply a price war, but a trade-off of 'tokenomics'. The value of a model cannot be defined solely by '$/Mtok' (price per million tokens) but depends on the balance of three key performance indicators (KPIs):
Latency: The time required from sending a request to the model generating the first token.
Interactivity: The speed at which the model generates subsequent tokens, usually measured in tokens per second.
Context Window: The length of conversation or data that the model can 'remember'. Model providers can adjust these variables to determine the final token price.
The battle for AI Tokens is not only about total amount but also about 'output purity'.
Limited computing power is not a unique dilemma for DeepSeek; even well-funded Western AI company Anthropic faces similar challenges. Anthropic's Claude 4 Sonnet has seen its output speed on the API drop by 40% since its launch, reaching about 45 tokens per second. This is similar to the reason for DeepSeek: to cope with massive requests with limited computing power, it had to increase batch processing rates, thus sacrificing interaction speed.
However, Anthropic has also demonstrated another optimization approach, as its model provides more 'refined' answers, generating far fewer total tokens than competitors. This means that although the speed is slower, the total waiting time for users to receive complete answers may be shorter.
This article is authorized to be reproduced from: (Digital Age)
Original title: (The tide has receded? DeepSeek's traffic avalanche near 30%: Why has it taken the 'big bowl' CP value route but failed to retain users?)
Original author: Li Xiantai
'DeepSeek's traffic avalanche near 30%! Why has it taken the big bowl route but failed to retain users?' This article was first published in 'Crypto City'.