On November 15, users opening DeepSeek saw a line of notice.
"Insufficient compute. Long conversations inactive for more than 7 days are no longer supported."
Two changes came with it: a single-input ceiling of roughly 40,000 characters, and a daily quota that ends with "usage limit reached due to server load."
By November 16, no announcement, no timetable.
Eight outages and the four hours
Jiemian News counted the official status page: 8 disruptions since October — October 2, 8, 12 (twice), 15, 20, 27, and November 3. Eight events across 48 days, one every 6 days.
The longest, October 15, ran over 4 hours: consumer web and app chat plus the V4 Pro and V4.1 Flash APIs, all limited at once. DeepSeek later sent credits to API users. The rest ran minutes; some were degraded performance, not downtime.
Four hours against a quarter's 2,208 is 0.18%. Unavailability for the quarter was 0.54%. One incident accounts for a third of it.
Seven of the 8 last minutes; the 8th drags the average.
What 40,000 characters holds
At 1,500 characters a page, 40,000 is 27 pages.
A 100-page due diligence report at the same density runs to 150,000 characters, so it takes 4 pastes. Once split, the model cannot see the first half, and cross-chapter checking is gone — a bigger loss than a few pages. The 7-day rule hits old material; this hits today's work.
Seven days governs the stock; 40,000 characters the flow.
Turning a percentage into hours
From August to October, the status page shows V4.1 Flash at 99.46% and V4 Pro at 99.74%.
99.46% sounds fine. August to October is 92 days, 2,208 hours; 0.54% of that is 11.92 hours. Annualizing by 365÷92, or 3.97, gives 47.3 hours — 1.97 days. A Flash user then spends two full days a year unable to reach the model.
V4 Pro: 2,208×0.26% is 5.74 hours, 22.8 annualized. Flash is cheaper, serves more calls and fails more often.
On November 17 the same page read 99.55%, 99.73% and 99.83% for Flash, Pro and chat, over a window marked August–November. A reporter's screenshot the day before showed August–October. The first half of November moved the denominator from 92 days to 122: 2,928×0.45% is 13.18 hours, annualized to 39.4 hours. A window 30 days longer produced an annual figure 8 hours lower.
November was clean, and 47.3 is not a constant; it rolls forward with the window.
Why seven days, why 40,000 characters
All three limits point at one thing: KV cache.
Every turn, the model keeps all previous tokens' key-value cache in VRAM so the next turn can attach. Longer conversation, bigger cache; more users, tighter memory. A session untouched for 7 days holds hundreds of thousands of tokens of cache and produces nothing new. So: clear idle sessions and reclaim; cap input to bound any single cache; impose a daily quota, which is a ceiling on concurrency.
DeepSeek did not get dumber; its VRAM ran out.
The API was spared
Jiemian's report has the key sentence: the API keeps an independent concurrency quota, isolated from the consumer pool. That is why developers felt it less. The 40,000-character and 7-day rules apply to the app and web product; the API context window is unchanged.
Free traffic is the buffer, trimmed first.
Demand doubled, supply is waiting
On November 13, DeepSeek opened its full-suite application API: parameters in the existing interface unlock deep reasoning, search and file input, and OCR, ASR and TTS can be called separately. Two days later the consumer side started rationing. API bills by usage; free consumer quota is pure cost, so it gets squeezed first.
One more lever: off-peak pricing at half the peak rate. Batch jobs can move; online products cannot.
OpenRouter's weekly board shows the pressure. October 19–25: V4.1-Flash-1017 served 33.6 trillion tokens, up 801% week on week and first globally; the older V4.1 Flash took second at 18 trillion. In late July the week ran 7.22 trillion. 33.6÷7.22 is 4.65 times across roughly 12 weeks, a weekly compound rate near 13.7%.
At that pace, capacity doubles every 5 weeks. The top 5 were all Chinese models.
33.6 trillion a week is 55.56 million tokens a second. A week is 604,800 seconds; the division is easy, sustaining the rate is not.
For API users: 2,500 concurrent accounts for deepseek-flash, 500 for v4-pro.
Science and Technology Daily reported in October that DeepSeek plans a GW-scale data center in Ulanqab, with at least 160,000 domestic AI chips, producing in stages from late 2027 to early 2028. 160,000 sounds like a lot until you see the date. Another figure: a second funding round of 105.5 billion yuan in October, at a post-money valuation of 615.5 billion yuan, per China Newsweek. DeepSeek has confirmed neither the amount nor the use.
A building goes up in a year; power and chips do not. DeepSeek has never published its compute footprint, so the gap can only be inferred from how hard the limits bite.
A queueing rule
The limits are a queueing rule, not a product decision. When resources fall short, someone decides who gets compute first, and DeepSeek chose "active first": use it and you are protected; stay away 7 days and you yield your seat. More efficient than first-come-first-served, and colder: it splits "used to use" from "using now."
For individuals, old threads stop connecting. For developers, the concurrency cap. For product teams, storing long conversations costs money the platform used to absorb.
No longer.
Fold 99.46% into 47.3 hours and then 2 days: that is what the number means, and DeepSeek is moving those days to places active users will not feel. Export the conversations that matter, and stop treating the chat box as cloud storage.
*Disclaimer: This article is based on publicly available information and is for reference only. It does not constitute investment advice.*
