Alibaba, ByteDance and DeepSeek are placing huge bets on 10-trillion-parameter models: an unstoppable money-burning game

At the Yunqi Conference on September 22, Alibaba brought the "large parameter" route for large language models back to the forefront.

Alibaba CEO Wu Yongming disclosed on site that Alibaba is preparing to train a new generation of artificial intelligence models with a parameter scale between 5 trillion and 10 trillion.

Subsequently, Liu Daheng, head of the Token Foundry Qwen LLM project of Alibaba's ATH business group, also mentioned on site that the Scaling Law is the key path to Artificial Super Intelligence (ASI), and the parameter scale of Qwen4.5 and Qwen5 in the future is planned to reach 5 trillion to 10 trillion.

On the same day, another piece of news spread across the industry: according to The Information, DeepSeek is currently training a new model with about 2 trillion parameters; Liang Wenfeng, founder of DeepSeek, also stated at a recent investor meeting that the company's next plan is to continue pushing the model to 8 trillion parameters.

Two months earlier, in August, the UK's Financial Times cited people familiar with the matter as saying that ByteDance is training a large language model with a maximum of 10 trillion parameters.

Parameters can be figuratively understood as the "brain capacity" of a large language model, that is, the total amount of empirical weights invoked during model inference and calculation. So what is the concept of 10 trillion? Compared with DeepSeek-R1 that set the industry on fire in 2025, its parameter count at that time was only 671 billion, less than one-fifteenth of 10 trillion.

From Alibaba to DeepSeek, and then to ByteDance, domestic large language model companies have collectively picked up the belief that "the larger the parameter, the better".

But the other side of the scale race: the larger the parameters, the higher the resources and costs invested, while intelligence will not grow at the same rate. After the model expands, the data must expand accordingly, not only in quantity, but also in quality. When everyone is stacking parameters, the real watershed becomes computing power efficiency, that is, how much intelligence can be obtained per unit of computing power.

Alibaba, DeepSeek, ByteDance Place Big Bets on 10 Trillion Parameters

Among the released models, the peak of the domestic camp is Kimi K3 launched by Moonshot AI in July, with a total parameter of 2.8 trillion, which is the world's first open-source model at the 3 trillion level. Alibaba's current flagship Qwen 3.8-Max has about 2.4 trillion parameters, and DeepSeek's active V4-Pro has 1.6 trillion parameters.

If we turn our attention to overseas, the closed-source camp has a higher parameter scale. The industry estimates that Anthropic's Mythos 5 has about 8 trillion parameters, and Fable 5 has about 5 trillion.

In other words, the parameter peak of domestically released models is still 1 to 2 times behind the estimated leading level overseas. And taking 10 trillion as the target, the active flagship of each company generally needs to be expanded by about 5 times more.

Then the question arises: the larger the parameters, is it really the better?

The logic supporting the view that "the larger the better" points to the Scaling Law, which means that when the model, data and computing power are expanded simultaneously, the model's capabilities will be improved accordingly.

In the past few years, from 100 billion to trillion, every jump in parameters has been accompanied by a visible leap in capabilities. Theoretically speaking, larger parameters do mean higher performance.

This is also the confidence for Liu Daheng to call Scaling the key path for Alibaba to achieve ASI.

But the Scaling Law is not infinitely effective. It requires preconditions to ensure that while the parameters become larger, the amount of data and computing power must also keep up in the same proportion, and they must be "high-quality data".

At the trillion-level, simply stacking parameters leads to a significant diminishing marginal return. Therefore, from a practical point of view, a so-called 10-trillion-parameter model cannot achieve a proportional improvement in capabilities if it cannot be fed with an equal amount of high-quality corpora.

There are also existing variables: data, training methods, architecture, and algorithms during inference.

For example, Kimi K3 has a total parameter of 2.8 trillion, but it uses the MoE (Mixture of Experts) architecture. 896 experts only activate 16 at a time, and the actual activated parameters are only about 104 billion. That means, nominally 2.8 trillion, the parameters actually used during operation are far less than that.

In actual use, large parameters do not equal "user-friendly".

On the one hand, the larger the model, the slower the inference, and the user's waiting time is visibly prolonged. On the other hand, the inference cost rises accordingly, which will eventually be passed on to the price. Even for flagship models with only hundreds of billions or 2 to 3 trillion parameters, the actual usage cost for users is already quite high.

So leaving aside "whether it is easy to use" and returning to the competition itself: the parameter scale determines the upper limit of capability, and the upper limit determines the ranking.

This is exactly why Alibaba, ByteDance and DeepSeek rush to 5 trillion to 10 trillion despite knowing the diminishing marginal return and high costs. The competition for large parameters is not for the current experience, but for the admission ticket to the next generation of models.

The 10 Trillion Bill: Large Language Models Burn Money, Who Can Afford It?

Although super models can touch higher intelligence, a larger base means more training and deployment resources. It is a tacit consensus among all major manufacturers that large language models burn money.

Overseas, the capital expenditure guidance for Microsoft, Google, Meta and Amazon in 2026 has soared to about 7200-745 billion US dollars.

The domestic market is equally aggressive: Alibaba's capital expenditure in Q2 2026 reached 67.7 billion yuan, a year-on-year increase of 75%, and its free cash flow plummeted from positive 73.8 billion yuan to negative 46.6 billion yuan; ByteDance's maximum capital expenditure in 2026 can reach up to 700 billion US dollars, about three times that of the previous year, and it may hit 1 trillion US dollars again in 2027.

Where is the money spent? The most direct item is infrastructure.

When OpenAI trained GPT-4, the computing power demand was counted in hundreds of billions of Tokens, and thousands of GPUs needed to operate collaboratively in a high-speed interconnected cluster for weeks or even months. At that time, GPT-4 had only 2 trillion parameters.

And the pre-training of a 10-trillion-parameter model requires the number of GPUs and the duration to rise to dozens of higher levels, which means that tens of thousands of high-end GPUs run continuously for several months. Electricity bills, cluster operation and maintenance, and chip procurement are all astronomical figures.

Judging from the investment in data centers alone, at the Yunqi Conference, Wu Yongming announced that the scale of global data centers operated by Alibaba Cloud will exceed 20GW by 2032; Goldman Sachs previously estimated that Alibaba's currently online capacity is about 3-4GW. This means that about 16-17GW will be added in the next six years, and the scale will expand to more than 5 times the current level.

Referring to the information disclosed by Jensen Huang of NVIDIA on the earnings call, the cost of building 1GW of computing power is about 500-600 billion US dollars. According to this calculation, the total construction investment of Alibaba's new 16-17GW in the next six years may reach the level of 800 billion to 1 trillion US dollars.

The second is the inference cost. The completion of model training is not the end, but the beginning. The larger the parameters, the more computing power is invoked for each generated response; when the model is deployed in the daily calls of tens of millions of users, the electricity cost and computing power overhead on the inference side are often more lasting and more costly than training.

There is also an easily overlooked invisible bill: data. As the model expands, the training data must expand simultaneously, not only in quantity, but also in quality. The current industry consensus is that high-quality, finely cleaned and labeled data is scarcer than raw computing power.

With three bills in front of them, can the players afford it?

Cash flow has already issued a warning. In the past 5 quarters, Alibaba's free cash flow has been negative for 4 quarters. The capital expenditure in the latest quarter is as high as 67.7 billion yuan. The cash and liquid investment of about 474.5 billion yuan on the account can still support the business, and last month, Alibaba raised about 80 billion Hong Kong dollars through share placement in the Hong Kong stock market for AI infrastructure.

But in the face of trillion-dollar level computing power investment, this amount of resources is far from enough, and continuous external financing is almost inevitable.

The situation of ByteDance is similar. According to The Information, ByteDance's net profit in the first half of this year fell by a single-digit percentage year-on-year to about 20 billion US dollars. It is the domestic technology giant with the largest AI capital expenditure, with AI-related investment reaching 150 billion yuan in 2025, and further increasing to 200 billion yuan in 2026.

Against the backdrop of funding gaps, ByteDance recently completed a syndicated loan of about 30 billion US dollars, the largest in history for a technology company, and part of the funds is expected to be invested in the continuously expanding AI infrastructure.

An Unstoppable Arms Race: From Flash to 10 Trillion Parameters

Less than a month before the Yunqi Conference, Alibaba just launched Qwen3.8-Flash-Next: with a total parameter of 125 billion, only 6 billion parameters are activated per inference, the training cost is nearly 90% lower than the previous generation, and the lowest API price is one-third of that of DeepSeek-V4-Flash.

On the same day, Zhipu's GLM-5.3-Flash was also listed on the open source leaderboard, with a total parameter of 320 billion and 18 billion activated parameters.

In those two weeks, the industry was talking about the "cost revolution": the key to winning the model competition is to see who is more cost-effective, who is faster, and who can minimize the unit cost in high-frequency calls.

A month later, the Yunqi Conference pushed the word "large" back to the center of the stage, and the two playing styles began to coexist.

From the perspective of the model itself, models like Flash optimize the "high-frequency, real-time, cost-sensitive" scenarios, where users demand speed and low price, so the activated parameters are minimized.

Large parameter models, on the other hand, optimize the most difficult, most diverse, and most widely covered scenarios, aiming for the most powerful inference, the longest Agent link, and the widest capability coverage.

What allows these two playing styles to coexist and even switch back and forth is the MoE architecture (Mixture of Experts architecture), which splits the model into many "experts" and only activates part of them for each inference. Its core contribution is to completely decouple the total parameters from the actual invoked parameters: the total parameters can continue to expand to support the upper limit of capability, while the single computation amount is controlled and does not increase proportionally with the total scale.

For large language model manufacturers, MoE leverages the cost structure itself: the training cost is controllable, and inference is billed by activated parameters, which means that with the same base, manufacturers can bet on both the small-and-fast route and the large-and-powerful route at the same time.

Both Flash and large parameters are ultimately answering the same question: how to make every cent generate more intelligence.

However, large parameters are still an expectation, and 10 trillion is only a number on the PPT. The peak of active domestic flagship models is still 2.8 trillion. When the models are actually launched, the game table may have been reshuffled. At that time, 10 trillion will no longer be the finish line, but a new starting line.

This article is from the WeChat public account "Tech Planet" (ID: tech618), author: Ren Xueyun, published with authorization from 36Kr.