From Nvidia’s Leadership Change to DeepSeek’s Limits, the AI Industry Is Shifting from “Competing on GPUs” to “Doing the Math”
DeepSeek’s tightening of its free usage allowance and Nvidia’s increasingly cautious investment stance are presented as two sides of the AI industry’s shift from an “expansion mindset” to a “budget mindset.” The article says DeepSeek no longer supports long conversations that have been inactive for more than seven days and has imposed limits on daily usage and per-request input. As computing budgets tighten, the industry’s focus is also shifting toward closed-loop business models and the efficient use of limited resources.
Why read this
AI score 57
The article juxtaposes DeepSeek’s limits with Nvidia’s shifting investment strategy, offering a framework for observing the industry’s transition from expanding computing power to operating under budget constraints.
Topics
Body · AI translation
From Nvidia’s Leadership Change to DeepSeek’s Usage Limits: The AI Industry Stops “Competing for GPUs” and Starts “Balancing the Books”
Tech New Knowledge · Download the app for exclusive early access
November 20, 2026, 19:35:44 · Beijing Updated November 20, 2026, 21:10:31
DeepSeek has imposed a “seven days of inactivity” threshold on long conversations, while Nvidia has adopted a more cautious investment stance following the departure of its founder. The two developments happened in quick succession, but one did not cause the other. Rather, they represent opposite ends of the same shift: suppliers are beginning to install speed limiters, service providers are clearing out sunk resource usage, and application developers are starting to demand closed-loop business models. Computing power has not become cheaper; it is simply no longer assumed that capacity can be expanded without limit. Once “there will always be more GPUs” is no longer taken for granted, the entire industry must answer a basic question anew: Whose problems, exactly, are all these GPUs solving?
Original content © Tech New Knowledge, Digital Hardware Desk
By Pineapple · Edited by Jiuli
A few days ago, DeepSeek—long known for offering generous capacity—suddenly tightened its free usage allowance. It no longer supports long conversations that have been inactive for more than seven days, has also introduced daily usage limits, and has capped a single input at roughly 40,000 Chinese characters. At almost the same time, across the Atlantic, Jensen Huang, who founded Nvidia, stepped down as CEO, and CFO Colette Kress became interim CEO. Reuters described her as “the only person at the company who challenged Huang on AI investment.” The reason was specific: She had reduced a proposed $250 billion plan to support OpenAI’s data centers to $120 billion, earning her the label of a “pragmatist.”
Placed side by side, the two developments most readily invite a causal interpretation—and causality is precisely what should not be inferred. DeepSeek’s limits are not a direct consequence of Huang’s departure. They are not links in a chain reaction, but opposite ends of the same shift: The AI industry is moving from an “expansion logic” to a “budget logic.”
01
Expansion Logic: The Years When Computing Power Needed No Brakes
Jensen Huang was the most emphatic proponent of this set of assumptions. He said that as long as humanity’s demand for intelligence was limitless, Nvidia’s business would have no ceiling. Nvidia itself was a token factory: Every inference performed by a large model and every thought generated by an agent paid it a token fee, with no end to demand in sight. At GTC in March 2026, he offered a more concrete figure: Demand for Blackwell and Rubin through 2027 would total at least $1 trillion, adding that “actual computing demand will be far higher than that figure.”
Model companies therefore designed products on the assumption that “there will be more GPUs,” data centers expanded on the assumption that “demand will keep rising,” and investors priced assets on the assumption that “expansion will not stop.” Three roles and three sets of decisions, all sharing the same unproven premise. Nvidia has never merely dominated chip supply; it has also shaped an entire industry’s expectations of limitless expansion—expectations that never prove demand, but simply presume it.
The supply side moved first. On data-center construction, the new CEO will inevitably be less aggressive than Huang, because financial risk has moved to center stage. The earlier reduction of a $250 billion commitment to $120 billion was already a preview. Reuters quoted a venture capitalist as saying, “No one but Jensen Huang would invest in his own customers as aggressively as he did.” That observation gets to the heart of the matter: Part of the previous wave of computing-power expansion was supported not by end-user demand, but by Nvidia’s large-scale financing and commitments to customers. When financial prudence replaces the impulse to expand, what disappears is not merely a single investment, but the illusion that demand can be manufactured from within the industry itself.
Capital markets confirmed this shift almost simultaneously. On November 13, SemiAnalysis sharply lowered its forecast for capital expenditure on AI accelerators. The reason was not a sudden disappearance of supply, but slowing AI revenue growth, tighter financing conditions, and mounting constraints involving electricity and grid connections. The return calculus behind “adding computing power regardless of cost” was being questioned anew. One distinction is important here: What was revised downward was the forecast curve, not spending that had already occurred. Under the new model, global capital expenditure on AI accelerators will still rise from approximately $392 billion in 2026 to $923 billion in 2030. The direction has not reversed; the slope has merely flattened. Once budget logic takes hold, it lowers expectations for the marginal growth of global computing capacity. Those expectations will then pass through chip deliveries, rental prices, and capital-spending schedules, layer by layer, until they reach every model company dependent on computing power.
The usage side followed by clearing its books. Returning to DeepSeek, the technical motivation is no mystery: Long conversations continuously accumulate KV cache, while idle sessions occupy scarce GPU memory. Usage limits reclaim resources and preserve capacity for active users. But this technical explanation answers only “how to cut,” not “why cut now, what to cut, and what to preserve.” Viewed through the lens of budget logic, the trade-offs become very clear: Cut idle long conversations and the sunk context of inactive users, while preserving active inference and the core pursuit of AGI. DeepSeek would rather inconvenience free users than rush to erect a paywall; it would rather sacrifice consistency of experience than allow computing power to be used inefficiently. This is not an experiment in monetization, but an exercise in resource allocation.
In the Chinese market, there is a third dimension to this calculation. As global supply expectations narrow and access to overseas computing power becomes more difficult, tying model architectures more closely to domestic computing hardware is another proactive way of balancing the books. DeepSeek V4 completed Day 0 adaptation across everything from low-level operators and communication protocols to GPU memory scheduling and framework compatibility. For the first time, its official technical report listed “Huawei Ascend NPUs” alongside “Nvidia GPUs” in its hardware validation checklist. IDC data show that Nvidia shipped 678,000 AI chips for Chinese data centers in the first half of 2026, down 44% year on year; Ascend shipments climbed from just over 300,000 in the same period a year earlier to 572,000, narrowing the market-share gap to six percentage points and bringing the two nearly neck and neck. On one side, others can no longer supply as much; on the other, domestic alternatives have been prepared in advance. This so-called binding is not a gesture—it is a way to factor supply uncertainty into costs ahead of time.
03
The Value Logic: Narrative Takes a Back Seat, Applications Take the Stage
When both the supply side and the user side begin balancing the books, assessments of the application layer are reordered as well.
At the APEC CEO Summit over the past few days, Lei Jun put it bluntly: Investment in AI is enormous, but the scenarios in which it can actually be deployed are extremely limited, and relatively few companies have truly closed the commercial loop and made money. Robin Li offered a similar assessment: The most urgent task is to identify the application scenarios for which AI is genuinely suited; if companies adopt AI but their profits do not rise, the bubble could become a reality.
Under the logic of expansion, these two statements were correct but went unheard as empty truisms—chips were still soaring, demand was still chasing supply, so who had time to talk about profitability? Under the logic of budgeting, they have become the main theme. When chips were abundant, no one talked about profitability; while demand was still chasing supply, no one talked about closing the loop; once allowances become quotas, everyone has to talk about it. Precisely when computing power is most constrained is when the computing-power narrative most needs to be given less weight.
Ultimately, the value of AI must be realized at the application layer. When Nvidia’s supply shifts from “as much as needed” to “cautious,” and DeepSeek’s conversations shift from “unlimited” to “quota-based,” the industry’s center of gravity is forced to move away from “who can get more chips” toward “who can create real value with a limited number of chips.” This does not mean computing power is no longer important; it remains one of the hardest constraints. But once computing power is no longer assumed to be infinitely expandable, model companies, chip companies, and application companies must all answer a more basic question: Whose problems, exactly, are these chips solving, and what problems are they solving for them?
In the age of expansion, the competition was over who could get the most; in the age of budgeting, it is over who can use it most effectively.
“Special Notice: The above content—including any videos, images, or audio—was uploaded and published by users of Dafeng Hao, a self-media platform under Phoenix New Media. This platform merely provides information-storage services.
Notice: The content above (including the videos, pictures and audios if any) is uploaded and posted by the user of Dafeng Hao, which is a social media platform and merely provides information storage space services.”
AI Trends