Behind the Kimi K3.1 Pro quota controversy: the stronger the model, the tighter the compute

[[**Kimi K3.1 Pro's performance approaches the global top tier, yet a subscription quota controversy exposes its compute bottleneck: the stronger the model, the more tokens Agents consume, and the commercialization and adoption of Chinese large-model companies remain constrained by compute.**

**Key points**

K3.1 Pro scores 65.7% on Terminal-Bench 4.0, close to Claude Opus 5.5; but users of the 199-yuan and 699-yuan plans complain of insufficient quotas, and a developer estimates the API value of the old plan at about 1.94 times that of the new plan.

**2. API prices fall, but Agents still push compute demand higher**

K3.1 Pro's per-million-token input/output prices drop to 13 yuan and 68 yuan, below K3's 20 yuan and 100 yuan; yet a complex Agent can consume millions of tokens. In June 2026, domestic daily average tokens were about 500 trillion, nearly 4 times the level of December 2025.

**3. Compute shortages will constrain independent model companies over the long term, and domestic supply is still hard-pressed to fully resolve them**

This article is from the WeChat public account: 01Founder, editor: Max, author: Max, always on the road

In the early hours of October 24, Moonshot AI released its new-generation model Kimi K3.1 Pro.

This is a large model with 3-trillion-level parameters and support for a 1-million-token context.

In the officially published Terminal-Bench 4.0 test, it achieved a score of 65.7%, already very close to global top-tier models such as Claude Opus 5.5.

For a Chinese startup founded only a few years ago, such a result is already astonishing enough.

Two years ago, Chinese large-model companies were still struggling to catch up with GPT-4; today, they are able to compete head-on with those American companies that possess enormous compute and capital reserves on the world's most cutting-edge technology stage.

Yet this launch, which should have drawn applause, produced a rather ironic scene.

The models are getting smarter, but users find they can afford them less and less.

After the release of K3.1 Pro, complaints about membership quotas quickly appeared on social media.

Some users bought the 199-yuan-per-month Pro plan, ready to experience the new model's coding and Agent capabilities, only to find themselves nearing the usage limit after running just a few complex tasks.

At the same time, discussion began to emerge about Kimi adjusting membership quotas around the release of the new model.

One developer used two 699-yuan subscription accounts to run tests through 49 API requests and, converting by the same token structure, found that the API value obtainable under the old plan was about 1.94 times that of the new plan.

Although this result cannot directly represent the actual quota changes for all users, it still sparked considerable discussion.

What is more interesting is that the API prices of the new model Kimi released this time have actually gone down.

According to official pricing, the previous-generation K3's per-million-token input and output prices were 20 yuan and 100 yuan respectively, while K3.1 Pro's dropped to 13 yuan and 68 yuan.

Looking at API prices alone, Kimi is still working hard to maintain the cost advantage that Chinese models have consistently had.

But for users who buy subscription plans, model capabilities have improved and the per-token price has fallen, yet in actual use they need to calculate their quotas even more carefully.

The model company has produced a handsome report card, but users are starting to complain about their bills.

When I saw these discussions, I suddenly recalled a dinner in 2024.

Kimi in 2024 was probably China's most glamorous large-model startup.

At that time, ChatGPT had already shown the world the potential of large language models, but there were still few AI products in China that ordinary users could remember by name and actively use.

Kimi quickly broke out with its long-text capabilities and at one point became one of the most closely watched products in China's large-model industry.

Its user growth, product experience, and the financing story of Moonshot AI behind it all put the company in the spotlight.

For entrepreneurs of that period, Kimi almost represented one possible path for the commercialization of Chinese large models.

It happened that I had a friend working in enterprise services at the time, with some B2B clients in hand, hoping to integrate Kimi's model capabilities to provide customized AI applications for enterprises.

So, through an acquaintance's introduction, he arranged a dinner with a co-founder of Moonshot AI, hoping to discuss the possibility of cooperation.

Because he did not know much about model training and the technical side, he specifically asked me to accompany him.

Before the meeting, I thought this should have been a natural thing.

After all, at that time the entire large-model industry was discussing commercialization; how to find clients truly willing to pay, and how to get models out of the lab and into enterprises' production processes, were questions almost every startup could not avoid.

And right in front of us happened to be exactly such clients, who had real business needs and were also willing to pay for model services.

English text is being prepared.

Of course, I can understand the choice the other party made at the time.

In 2024, Kimi was growing rapidly, and limited computing resources needed to be prioritized to ensure the stable operation of its own products.

For a startup still fumbling between technology and commercialization, how to allocate these resources was inherently a difficult multiple-choice question.

After that meal, we did not continue to pursue the cooperation, and this matter gradually faded from my memory.

Until today, two years later, Kimi has grown from a chat assistant that focused on long-text capabilities back then into a frontier model capable of executing complex programming and long-horizon Agent tasks, and Moonshot AI's technical strength and model scale are no longer what they were.

But when I saw users complaining about insufficient quota for K3.1 Pro, that phrase from back then—"not enough compute"—surfaced in my mind once again.

Two years ago, an enterprise client was willing to pay to use Kimi, but Moonshot AI ultimately did not take on this business because of tight compute.

Two years later, Kimi has trained a world-class large model, and ordinary users are willing to spend 199 yuan or even 699 yuan per month to purchase the service, yet they find that they still need to carefully allocate each call.

Today's membership quota controversy involves specific product pricing, inference costs, and capacity management, and we cannot yet judge, based solely on external user experience, just how much compute pressure exists inside Kimi.

But these two events, two years apart, made me start thinking about a bigger question.

Chinese large-model companies are already capable of building world-class models, but how much compute do we actually still need for these models to be fully used by the world?

To understand this question, we first need to look at how much large models have changed over the past two years.

In 2024, our relationship with large models was still simple.

Most of the time, we sat in front of a computer, asked ChatGPT or Kimi a question, and waited for it to give an answer.

Whether translating an article, organizing materials, or writing a piece of code, the model's work usually came to a pause when a conversation ended.

But by 2026, large models are changing in two directions.

First, the models themselves are becoming increasingly massive.

If we go back two years, parameter scales in the hundreds of billions, such as 100B or 120B, were already enough for a model company to make a big deal of.

But today, hundreds of billions of parameters is no longer anything rare, and the scale of leading models is advancing toward the trillions.

Kimi K3's total parameter count reaches 2.8T, equivalent to 2.8 trillion;

Zhipu GLM-5.3 reaches 753B, or 753 billion;

DeepSeek V4 Pro further reaches 1.6T, equivalent to 1.6 trillion.

Of course, these models generally adopt the MoE architecture, and each inference only needs to activate part of the parameters, so the amount of computation per run is far smaller than the number implied by the total parameter scale.

But for the companies responsible for deploying these models, larger model weights, more complex parallel computing, and context windows that often run to millions of tokens still mean higher requirements for VRAM, interconnect bandwidth, and infrastructure.

And the expansion of model scale is only half the story.

Second, the spread of Agents is completely changing how people use large models.

Today's users have begun to get used to handing complete tasks to models.

After a Coding Agent receives a task, it may need to read hundreds of source code files, understand the structure of the entire project, then modify code, call tools, run tests, and repeatedly adjust its approach based on error messages.

If the task is complex enough, it may also launch multiple Subagents, letting different model instances handle different problems simultaneously.

From the user's perspective, the whole process may involve only a simple instruction, but behind that sentence, the model may have already worked continuously for dozens of minutes, performed hundreds of inferences and tool calls, and consumed millions or even more tokens.

As a result, a more thorny problem has emerged.

The models themselves are becoming increasingly massive, and the work we ask them to complete is also becoming increasingly complex.

English text is being prepared.

This reminds me of the era when electricity was just becoming widespread.

At first, people simply wanted to replace kerosene lamps with electric lights to obtain a cheaper and more convenient form of lighting.

But as electric power technology developed, electric motors began to enter factories, elevators, rail transit, and various electrical appliances appeared one after another, and people's demand for electricity continued to expand.

A single electric light consumes less and less electricity, but humanity has already begun to use electricity to drive entire cities.

Technological progress has lowered the cost of each use, but it has also opened up a demand market far larger than before.

Today's large models are undergoing a similar change.

In the past, we only occasionally asked a model a question; now we hope it can work continuously for several hours like a real employee, or even run dozens of Agents at the same time.

Every time a model becomes a little smarter, we discover more work worth handing over to it, and this new work in turn requires more compute to support.

On October 23, the China Academy of Information and Communications Technology released a "Large Model Token Service Performance Monitoring Report (2026)," and two sets of data in it happen to provide a footnote to this change.

In June 2026, domestic average daily Token consumption had already reached about 500 trillion, nearly 4 times that of December 2025.

At the same time, the normalized average price of domestic large model services rose from 6.8 yuan in the second half of 2025 to 13.46 yuan in 2026, an increase of nearly 100%.

The normalized average price here is a statistical result after uniformly processing different models and billing methods. It reflects changes in the market price structure and cannot be directly understood as all models' APIs having doubled in price.

But looking at these two sets of data together, the changes taking place across the entire industry are already quite obvious.

Over the past two years, one of the most important stories in China's large model industry has been the price war.

From several leading vendors constantly refreshing API prices, to DeepSeek impacting the global market with highly competitive pricing, Chinese models have gradually established a distinct industry impression: capabilities are getting stronger, and prices are getting cheaper.

However, as models gradually transform from chat tools into production tools capable of continuous work, the economic logic of this competition has also begun to change.

In the past, people mainly focused on how much it cost for a model to generate one Token.

Today, as model scale continues to expand and Agent tasks become increasingly complex, how much computing resources are needed to complete a piece of work has begun to become a more important question.

This also explains why, even though Kimi lowered its per-Token price, some subscribers still feel it is getting more and more expensive.

For model companies, Tokens are a cost that can be precisely measured; for users, what they really care about is how much work output a subscription can actually bring.

When models begin to work on people's behalf, computing resources become one of the most important means of production in this era.

And China's large model startups have grown up precisely in an environment where this means of production is relatively scarce.

If we turn the clock back to 2023, the large model startup booms in China and the United States arose at almost the same time, but the starting points of these young companies were quite different.

In January 2023, Microsoft announced a new multi-billion-dollar investment in OpenAI, and their cooperation extended from funding to Azure cloud computing and supercomputing infrastructure.

In the same year, Anthropic established deep cooperative relationships with Google and Amazon successively, obtaining huge funding and compute support.

And tech giants like Google and Meta already possessed massive AI computing infrastructure themselves.

The situation in China was much more complicated.

In October 2022, the United States began imposing export controls on advanced computing chips to China, and further tightened restrictions in 2023.

It was precisely during this period that the domestic large model startup boom began to rise. A group of companies including Moonshot AI and Baichuan Intelligence entered the competition in frontier models, and existing teams such as Zhipu also accelerated their R&D pace.

Although these companies were likewise backed by capital and large technology enterprises, and DeepSeek also had computing resources accumulated over many years by High-Flyer Quant, Chinese model companies have always faced more restrictions in their channels for obtaining advanced compute, costs, and supply certainty.

This is also why I have always felt that compute is the original-family wound of China's large model startups.

A company's origins often influence its development path for a very long time.

For China's large model startups, constraints on computing resources have gradually permeated every link of technology R&D, product supply, and commercialization.

DeepSeek is a very representative example.

From the MoE architecture and MLA attention mechanism to a series of engineering innovations centered on training efficiency and inference costs, DeepSeek has consistently shown a high degree of emphasis on computational efficiency.

These technologies are also directions jointly explored by frontier models worldwide, but in an environment where advanced computing resources are relatively limited, how to make each compute card deliver greater value through architecture and engineering optimization clearly has more urgent practical significance.

This pursuit of efficiency has gradually become a distinct engineering culture among Chinese model companies, and has also helped them continuously approach the global technological frontier under limited computing budgets.

English text is being prepared.

For an independent model company, limited compute must simultaneously satisfy R&D for the next-generation model and inference services for existing products.

Competition in frontier models is becoming increasingly fierce, and it is hard to slow R&D investment. Once a new model succeeds, it will attract more users and further push up inference demand.

For tech giants with large cloud businesses, such as Alibaba and ByteDance, part of the pressure can be alleviated through existing infrastructure and diversified revenue.

But for independent model startups, every resource allocation may mean giving up another important opportunity.

What is crueler is that for these companies, technological success sometimes instead becomes an expensive burden.

If one day Kimi really builds the world's best Coding model and attracts millions of developers to use it every day, then how much computing resources would be needed to let these users run their Agents stably and continuously, and how enormous an inference cost would have to be borne?

This real-world pressure can also help us understand why open source has become such an important strategic choice for China's large model industry.

Once model weights are open, cloud providers, inference service providers, and enterprise users around the world can deploy the models using their own infrastructure, allowing the models' influence to break through the compute-scale limits of the original developer and simultaneously build a broader developer ecosystem.

Model companies are responsible for creating intelligence, while the entire ecosystem shares the cost of keeping that intelligence running.

However, as more and more enterprises are able to deploy models and provide inference services on their own, how can the companies that initially bore the enormous R&D costs earn enough revenue from the technological value they created to sustain development of the next generation of models?

This may also be the hardest part for China's independent model startups.

Over the past few years, they have gradually proven that even in an environment with limited resources, China can still produce world-class large models.

But between training a world-class model and building a world-class model company capable of sustained operation, there still lies massive computing infrastructure and a business model that is still being explored.

Fortunately, over the past two years, domestic compute has progressed faster than many people imagined.

I myself used Huawei's Ascend 910 series chips a few years ago, and the experience at the time was really not pleasant.

Hardware performance was one aspect; what was more troublesome was that the entire software ecosystem was far from mature enough. From framework adaptation and operator support to various low-level tools, engineers had to invest a great deal of effort in debugging.

Back then, using domestic cards was like driving on a road that had not yet been repaired—after going a stretch, you had to stop and clear obstacles.

But the changes over the past few years have indeed been substantial.

Huawei's Ascend chips have evolved from the early 910 series to the 950 series. According to officially published technical specifications, Ascend 950 reaches 1 PFLOPS of FP8 compute, while the 950DT, to be launched in the second half of this year, has 144GB of memory capacity and 4TB/s memory access bandwidth. Some hardware metrics are already comparable to NVIDIA's H200 compute cards.

Although actual training and inference efficiency still need to be tested under real workloads, such progress is already quite remarkable.

At the same time, Cambricon continues to iterate its SiYuan series accelerator cards, and T-Head released the new-generation Zhenwu V900 in September this year, claiming computing capability 3 times that of the previous generation M890. Domestic AI chips are gradually forming a more complete product system.

Huawei has also put Ascend 960, 970, and even 980 on its product roadmap for the coming years, and has begun organizing large numbers of domestic AI chips into larger-scale computing systems through supernode interconnection.

Chip companies are busy increasing compute supply, and model companies are also stepping up construction of their own computing infrastructure.

In September this year, Bloomberg reported that DeepSeek plans to deploy at least 160,000 Huawei Ascend 950DT chips in a data center under construction in Inner Mongolia to support large-scale inference services.

Although this plan has not yet been fully and publicly confirmed by both parties, and actual delivery also faces capacity constraints, its planned scale is already enough to reflect leading model companies' expectations for future computing demand.

Zhipu provides another very interesting case.

On September 17, Zhipu founder Tang Jie disclosed that they used an Infra Agent driven by GLM-5.3 to participate in building and optimizing the inference infrastructure for GLM-5.3-Flash.

According to officially published information, this system runs on a cluster composed of more than 100,000 domestic AI accelerator cards, and the Infra Agent raised end-to-end throughput to about 3 times the initial baseline in less than two weeks.

This case sparked a lot of discussion about RSI at the time, but I think it has another, more practical significance: Chinese large model companies have begun trying to use the intelligence they created to in turn improve the efficiency of computing infrastructure.

After all, when the supply of a compute card is constrained, enabling it to handle more computing tasks through compilers, inference frameworks, and resource scheduling itself has very direct commercial value.

These advances are indeed encouraging, but I still believe that the compute shortage in China's large model industry will be difficult to thoroughly resolve for a considerable period of time to come.

Because what truly constrains compute supply now involves the entire semiconductor industry chain.

Progress in chip design capability is certainly important, but turning those designs into sufficiently large-scale, stable, and cost-controllable production capacity still requires joint progress in many areas, including advanced processes, lithography equipment, high-bandwidth memory, and advanced packaging.

In September this year, Huawei publicly stated that domestic demand for its AI computing equipment exceeds existing production capacity.

I had even previously heard a claim that Huawei's order schedule has extended to around 2030.

There is still a lack of publicly verifiable evidence for this specific timeline, but judging from the supply and demand situation that has been disclosed, the pressure on expanding domestic compute capacity is indeed very great.

In my view, to truly alleviate this structural compute shortage, China ultimately still needs breakthroughs in more foundational areas such as advanced manufacturing equipment, processes, storage, and packaging, and this will clearly take longer than developing a new generation of chips.

Moreover, even if domestic compute supply capacity continues to improve, the growth of model scale and the spread of Agent applications may still continuously create new computing demand.

It is like a race with no finish line.

Chip manufacturers keep expanding capacity, model companies are rushing to build clusters, and Infra teams are working to improve the utilization efficiency of every card—and each cost reduction may make more previously expensive AI applications viable, which in turn brings a new round of demand growth.

We are creating compute at an unprecedented speed, and we are also consuming compute at an unprecedented speed.

For tech giants with massive capital reserves and mature cloud computing infrastructure, this may be a competition they can invest in over the long term.

But for independent model companies that need to rely on continuous financing and limited commercial revenue to sustain R&D, every increase in demand may mean new infrastructure investment and greater financial pressure.

Over the past few years, China's model startups have already achieved remarkable technological progress under limited resource conditions, but to truly serve more users with these capabilities at a sustainable cost, a sufficiently powerful computing infrastructure is still needed.

At that time, Kimi was standing in the spotlight, and with its long-text capabilities it became one of the brightest stars in China's large-model industry.

We went to discuss cooperation full of expectations, hoping to leverage its capabilities to build products for enterprise customers that had been difficult to achieve in the past, but in the end, because of insufficient compute, we could only regretfully give up.

Looking back now, I have somewhat come to a deeper understanding of this matter.

In two years, China's large-model industry has already traveled a quite difficult road.

Technological goals that were once out of reach are now being achieved one by one by a group of young Chinese researchers and engineers, and we finally have the opportunity to stand on the world's most cutting-edge stage and compete head-on with those tech giants that possess enormous resources.

I am very happy about these advances, and because of that I hope even more that they can truly be transformed into widely used products and services.

After all, when a group of people have gone through hardships and finally brought a model to world-class level, yet still need to repeatedly weigh how to let more people use it, this is somewhat regrettable.

And for Kimi, as well as this generation of Chinese large-model startups, I hope that one day in the future they will finally be able to break free from this constraint, invest in R&D with greater ease, and welcome every customer willing to use their products with greater confidence.

I still remember those four words that Moonshot AI's co-founder told us at that dinner in 2024:

Not enough compute.

Two years later today, Chinese large-model companies have come a long way.

And I hope that the next time someone goes to them to discuss cooperation, they will finally no longer have to hear those four words. AI Venture Capital Daily Channel: Frontier Technology This content comes from the internet, original link; the views represent only the author and do not represent Huxiu's position. If there are copyright issues, please contact hezuo@huxiu.com, and we will verify and handle them promptly. Everyone who is changing the world and wants to change the world is on the Huxiu APP [ ]