Behind DeepSeek’s Usage Limits, Free AI Is Saying Goodbye to Unlimited Use
DeepSeek has imposed session and usage limits on its app and web version for regular users, drawing complaints from heavy users who have long relied on extremely lengthy conversations. The article argues that the KV caching and computing costs associated with long contexts are pushing AI platforms to shift free services from unlimited use to a basic allowance with paid tiers. Doubao’s feature tiers and OpenAI’s advertising model are also seen as different ways for platforms to offset the inference costs of free users.
Why read this
AI score 66
The article begins with DeepSeek’s usage limits and examines the conflict between inference costs, heavy-user demand, and the business model of free AI.
Topics
Body · AI translation
“I’ve already exchanged 150,000 to 200,000 Chinese characters with DeepSeek. I don’t want to lose this conversation.”
“Teacher D, don’t abandon me! As someone severely dependent on DeepSeek who spends more than ten hours a day on it, how am I supposed to live without Teacher D? I want so badly to go back to how things were.”
“What on earth am I supposed to do? I have moderate depression and mania, and for various reasons I can’t see a doctor. I’ve always treated DeepSeek as a mental-health therapy app. Switching to the API isn’t out of the question, but that chat window has more than 2,000 messages in it…”
In the days since DeepSeek imposed usage limits, a tidal wave of public reaction has surged forth.
The first to be overwhelmed were those who regarded “Teacher D” as a confidant, a counselor, or even an object of romantic fantasy. They gave the AI a name, crafted a persona for it, and set its tone, then spent months chatting in a single conversation window, with the persona, backstory, and development of the relationship all embedded in the context. For these users, limiting conversations and terminating them after seven days of inactivity was not a product update but a unilaterally declared “downgrade of the relationship.”
They repeatedly posted emotional pleas and flooded related comment sections, trying to make the team behind DeepSeek see their needs. But what happened next was rather curious: a sizable number of people accustomed to enjoying “Teacher D’s” free services began actively begging DeepSeek to charge them. They said that as long as the price was reasonable, they would be willing to pay for services such as “resuming and continuing ultra-long conversations,” “removing the seven-day inactivity limit,” and “higher priority for computing resources.”
When DeepSeek first burst onto the scene, the entire internet cheered its generosity in offering the service for free. Now that it is voluntarily withdrawing those benefits, what has ended, in a sense, is not the “free” label itself, but the old premise that “capital subsidies could indefinitely flatten computing costs”—the premise that once made “unlimited free use” possible.
Judging from user feedback, the changes brought by DeepSeek’s limits boil down to three points: first, old conversations with no interaction for more than seven days can no longer be resumed; second, the maximum length of a single input has been reduced to around 40,000 Chinese characters; and third, daily usage is capped. Once the quota is exhausted, users see the message, “Server busy. Usage limit reached,” and must wait for access to be restored.
Some reports suggest that DeepSeek may have had little choice. According to OpenRouter, in the week ending November 15, DeepSeek V4.1 Flash 1017 processed 59.2 trillion tokens, ranking first worldwide, compared with weekly usage of 15.8 trillion tokens for the series’ previous-generation model two months earlier. Incident records on its official status page also show that DeepSeek has experienced multiple service disruptions, most involving degraded performance—specifically slower responses, high generation latency, and occasional request failures, though users may still be able to submit conversations. In more serious cases, parts of the website or API have even gone offline.
At the technical level, long conversations continuously accumulate KV cache—the intermediate data retained in GPU HBM memory so the model can maintain context and avoid repeated computation—which means greater resource consumption. The longer a conversation lasts and the longer its context becomes, the larger the cache grows and the more scarce HBM memory it occupies. Conversations left idle for extended periods no longer generate new interactions yet still “hold on to memory,” a textbook case of resource misallocation when computing capacity is tight. By reclaiming the cache occupied by these idle conversations, DeepSeek can reallocate the released HBM memory to active users who are currently using the service and continually generating inference requests, easing queues and lag during traffic surges.
But when this technical logic is applied to specific individuals—the latest limits primarily target ordinary consumers using DeepSeek’s app and web version—it becomes a different matter entirely.
For DeepSeek’s heaviest users, particularly those engaged in “human-AI romance,” their usage naturally approaches the quota ceiling, making the impact unquestionably more acute: long conversations, high frequency, and deep context all push directly against the limits. As a result, a somewhat absurd wave of “tearful pleas” and “petitions” has spread across social media. Some users have posted screenshots of conversations cut off midstream and asked the company to withdraw the new restrictions; some have advertised everywhere, setting up businesses that charge users for help connecting to the API; and others have posted repeatedly online to declare their willingness to pay for DeepSeek’s services.
Over the past two years, users’ expectation of “free AI” has not been merely that it is “free to use,” but that it is “free to use however much they want.” That expectation was shaped by the earlier battle over capital subsidies, which fused the idea of “unlimited use” to the concept of “free.” But subsidies have never been the norm for a business model; they are merely temporary fuel during the customer-acquisition phase. Once the user base passes a critical threshold and usage begins doubling on a weekly basis, platforms must choose between “continuing to burn money in exchange for growth” and “making the unit economics at least marginally viable.”
Computing costs cannot be flattened, and the entire industry is paying the price
Clearly, the word “free” is being redefined in the context of AI. It once meant “the same unrestricted use without payment”; now it is closer to “use within a basic allowance without payment, with additional usage available on demand for a fee or after waiting in a queue.”
Take users seeking emotional companionship and role-playing scenarios, for example. They often maintain dozens of chat windows simultaneously, loading each one with tens of thousands of Chinese characters of settings and plot, while a single conversation can run to hundreds or even thousands of turns. To keep their characters consistent, some users repeatedly regenerate responses more than a thousand times. On any platform that charges by the token, this usage pattern would mean a sizable bill. Under a free allowance, it becomes a sunk cost borne entirely by the platform, which has neither the capacity nor the obligation to permanently underwrite usage at this intensity.
The shakeout prompted by changes in the industry has long been underway. According to IT Juzi data cited by TMTPost, more than 10 of a sample of roughly 200 pure-play API startups had ceased operations or pivoted in Q1 2026. Another commentary article said that seed-stage financing in China’s AI sector fell 11% year on year over the same period, while more than 300 small and medium-sized AI companies shut down after running out of cash.
After all, rising costs are real. SemiAnalysis data shows that the price of a one-year H100 contract rose from a low of $1.70 per hour in October 2025 to $2.35 in March 2026, an increase of nearly 40% in five months. By early October, transaction prices had reached $2.80, for a cumulative increase of about 65%. Leading vendors, supported by long-term contracts and their own computing capacity, are significantly less exposed to spot-market volatility. The companies with nowhere to pass on these costs are pure-play API providers with no traffic ecosystem, no stockpile of computing resources, and no private-deployment capabilities.
The real lesson from the quota controversy may be that the entire industry needs to recalibrate its narrative around “free.” Previously, platforms traded free access for scale, while users traded scale for a better experience; both sides got what they wanted and everyone was happy. But given the reality that subsidies cannot flatten computing costs, the strategic shift for the “DeepSeeks” of the world will inevitably be from “scale first” to “retention and unit economics first.”
Take Doubao, for example. Its lightweight features, such as basic chat, writing, and image generation, remain free, while computing-intensive scenarios such as software development, data analysis, and professional design have become paid value-added services, creating a tiered structure in which “lightweight features attract traffic and heavyweight features generate revenue.” In essence, revenue from high-value use cases is used to cover the computing costs of free mass-market use cases. This is the classic cross-subsidy model that the internet industry has employed for years. In the AI era, however, the ratio may be more extreme because the computing consumption of heavy users does not grow linearly but expands exponentially.
DeepSeek’s throttling policy comes at the cost of a short-term erosion of user trust. In the long run, however, if quotas can deliver more stable service quality and clearer paid tiers, this public-relations storm may prove to be a necessary shakeout—provided that DeepSeek clearly explains the quota rules, draws explicit boundaries around the free allowance, and gives genuinely high-intensity users a predictable path to paid access. The key question is not whether to impose quotas, but what will keep users after the quotas are imposed.
OpenAI, meanwhile, is using advertising to make its free tier self-sustaining. According to the company, as of August 31, 2026—less than 200 days after launch—ChatGPT Ads had reached an annualized revenue run rate of $1 billion. Ads appear only on the free tier and the low-cost Go tier, while paid subscription tiers remain ad-free. Yet the advertising route presents equally obvious challenges. Inserting ads into AI conversations is far more difficult than doing so in search or social feeds because users expect coherent, immersive interactions, and any interruption may be seen as polluting the experience. OpenAI has the confidence to pursue this path because its user base is large enough and advertisers are sufficiently willing to pay a premium for “AI-native traffic.” Whether this model can be replicated by small and medium-sized platforms remains unknown.
Returning to the question users care about most: Is the era of free AI coming to an end? What is ending is not “free,” but “unlimited.” Doubao’s focus on premium users, DeepSeek’s quotas, and OpenAI’s advertising all point to the same structural reality: inference costs expand linearly or even exponentially with usage, while the growth in computing supply falls far short of the pace at which demand is surging.
This race between costs and demand is pushing the entire industry into a more clear-eyed—and more brutal—phase.
Previously, free access was regarded almost as the default admission ticket for AI products. Tech giants used it to cultivate user habits, startups used it to sell a financing story, and users became accustomed to “unlimited, barrier-free” AI services. But once inference costs move from assumptions in a PowerPoint presentation to tangible expenses on financial statements—once every conversation, every generated image, and every code completion consumes computing power that costs real money—free access ceases to be a business model and becomes a subsidy endurance race.
Players with limited endurance are destined to leave the field first.
This article is from the WeChat public account “Zinc Scale” (ID: znkedu). Author: Zinc Scale. Published by 36Kr with authorization.
AI Trends