For a year, DeepSeek has been the proof that a frontier chatbot can be free. On November 16, the free part started to crack. Users of the company’s app and web chat began hitting limits on three fronts at once: how much they can type, how long a conversation survives, and how much they can use in a day. The notices inside the product give the same explanation each time: insufficient computing power.

What changed on November 16#

The limits, as reported by users and carried by Chinese financial media on Monday, cover three dimensions of the free service:

  • Shorter prompts. A single input is now capped at roughly 40,000 characters. Routine questions are unaffected, but pasting in a full report, a long codebase, or a batch of documents is not.
  • Shorter-lived conversations. Long chats left inactive for more than seven days can no longer be resumed. The cutoff is the one detail reports disagree on — some cite seven days, others 72 hours — and DeepSeek has published nothing to settle it.
  • A daily quota. Once daily usage hits the ceiling, the service shows a message saying the quota has been reached because of server congestion, and users have to wait for it to reset. The exact daily number has not been disclosed.

What makes the episode unusual is the silence around it: no formal announcement, no unified quota rules, no timeline for lifting the limits. Users are learning them from product-page prompts and from hitting them. The topic trended on Weibo the same day.

A long queue of glowing chat bubbles winding toward a single dimming server rack
AI-generated illustration: a queue of free-tier conversations meeting finite capacity.

Demand outran the hardware#

The numbers explain why the company is reaching for the brakes. DeepSeek’s V4.1-Flash-1017 model completed 59.2 trillion tokens of calls in the week of November 9–15, ranking first globally by token volume. That is the model doing the heavy lifting in the free chat service, and the pace at which capacity is being added has plainly failed to keep up with the demand curve.

Analysts quoted in the coverage point to a specific technical pressure: idle sessions. A long conversation that nobody has touched in days still occupies memory — the key-value cache that lets a model pick up where it left off. Cutting off week-old threads and capping input length reclaims that memory for active users. It is a capacity decision dressed up as a product rule, and it is the kind of trade-off a lab makes when it cannot simply buy more servers.

That constraint is not abstract. DeepSeek sits at the sharp end of the US export-control regime, and on November 16 Washington published a rule licensing remote access to advanced AI chips — the cloud route Chinese firms had used to rent compute abroad. The free tier is where the squeeze shows first.

The API stays open — for a price#

Notably, none of this touches the developer API. It keeps separate call limits and strict pay-as-you-go billing, and three days before the quotas appeared, DeepSeek opened its full API suite to developers: web search, URL reading, file parsing, OCR, speech recognition and text-to-speech, anchored by V4.1-Flash-1017. Search costs $0.75 per 1,000 requests and OCR $0.225 per 1,000 pages, according to the company’s pricing.

The split is deliberate: the free chat is a loss-leader for adoption; the API is a business. And the next step may already be visible in the code — developers inspecting DeepSeek’s frontend found a quota-management system that syncs remaining API quotas every five minutes, plus a subscription field explicitly named “Coding Plan.” No pricing or launch date has been announced, but the plumbing for a paid tier appears to be in place. There was a warning shot in May, when DeepSeek quietly limited the regenerate and edit functions in its free chat.

A computer chip under a magnifying glass with a small padlock beside it
AI-generated illustration: export controls and chip scarcity are the backdrop to DeepSeek’s capacity squeeze.

Why it matters beyond DeepSeek#

DeepSeek is the stress test for a whole model of AI distribution: frontier capability plus free, unlimited access. The first promise still holds — its models top the usage charts. The second just broke, in public, with a quota message. Free tiers have always been subsidised, and when compute gets scarce the free tier is the first thing rationed, because it is the only part of the business with no price signal to throttle demand. The practical lesson for users: a free chatbot is not infrastructure. Conversations can be cut off, long threads can expire, and the terms can change without an announcement — so export what matters.

What to watch#

Three things will tell us whether this is a blip or a turn: whether DeepSeek publishes actual quota rules, whether the “Coding Plan” in the frontend code becomes a real product with a price and a date, and whether the daily quota tightens further. A cap that starts loose and quietly shrinks is how a free tier becomes a funnel. Until then, the most-used chatbot in the world is also, for the first time, a rationed one.

Sources: BigGo, “DeepSeek Imposes Compute Quotas on Free Users as Global Token Usage Ranks No. 1”; InvestingLive, “Daily quotas come to DeepSeek, a visible sign of China’s AI chip crunch”; KuCoin News, “DeepSeek Limits Free API Usage, Directs Heavy Users to Paid Plans”; XTechPioneer, “DeepSeek users report usage limits on app and web versions”.