After Nvidia chips became hard to buy, what keeps China's large models running?

[[**As U.S. chip controls tighten, domestic compute can already keep large models running, but the efficiency gap is clear; whether it can turn from "backup compute" into low-cost production capacity depends on three metrics—capacity, efficiency, and cost—and each leading company is taking a different path.**

**Key points**

**1. Domestic compute can already support large-model operation, but efficiency is significantly lower than Nvidia's**

Zhipu's domestic-chip inference cluster has reached 100,000 cards and carries real token production; Meituan's LongCat-2.0 also completed training on domestic compute. But Ascend 910C achieves only 30%-35% MFU in trillion-parameter post-training, while H100 can reach 40%-50%, meaning the same investment cannot train the same model.

**2. Compute efficiency is determined by chips, software, and models; software and models are the current room for active optimization**

Chips set the upper limit of compute supply and are hard to make up for in the short term; software determines how much hardware performance is released, and models determine how much compute is needed to complete the same task. When chips are constrained, large-model companies improve token output efficiency through compilers, inference frameworks, MoE architectures, low-precision computing, and other means.

**3. The paths of four leading companies diverge: Zhipu has the highest short-term certainty, DeepSeek the greatest medium-term elasticity**

Zhipu pursues an asset-heavy full-stack closed loop; based on a theoretical inference gross margin of 80%, its 30 billion yuan compute investment could yield annual revenue of up to 40 billion yuan, but it needs to prove continued improvement in per-token cost and gross margin. DeepSeek offsets chip shortcomings with model efficiency; the V4-Pro price increase exposes its constraints from high-end compute, and its 160,000-card Ulanqab cluster is expected to go into operation in 2027-2028, with on-schedule capacity delivery being key. MiniMax is shifting from leasing to building its own domestic compute, still in the validation stage from construction to production. Kimi lowers adaptation costs through open-source models, but suspended new user registrations after the K3 release due to compute bottlenecks.

**4. Investors need to track three metrics: capacity, efficiency, and cost**

For capacity, watch data-center commissioning progress and domestic chip procurement scale; for efficiency, watch growth in model calls and token usage; for cost, watch declines in per-token cost, improvement in inference gross margin, and whether profits can be maintained after API price cuts. If all three improve simultaneously, domestic compute will shift from a backup option to a cost advantage.Original2026-09-26 10:02After Nvidia's cards can no longer be bought, what do Chinese large models rely on to keep running?Zhang BeibeiQuick overview

But what is new is that on September 12, when Dario Amodei, CEO of U.S. frontier AI company Anthropic, talked about chips, he placed chip controls within a larger proposition: the United States can slow the pace of AI development, provided that China is not allowed to catch up.

In a long post published on X titled "We Must Pace the Frontier," Amodei also proposed that the United States should continue restricting sales of advanced AI chips and semiconductor manufacturing equipment to China, while cracking down on chip smuggling, remote access to advanced compute from overseas data centers, and model distillation. He believes these measures can significantly widen the U.S. AI lead over China in the next 3-5 years.

What the capital market is really trading, however, is precisely the expectation chain: compute constrained → can domestic chips catch it → what is the cost of catching it.

The question has therefore become urgent.

And the next 3-5 years Amodei refers to are precisely the period when domestic large models will continue to iterate and inference demand will keep growing.

If high-end chips really cannot be bought, can Chinese large models continue to run?

More importantly, who can, under compute constraints, squeeze more value out of every domestic chip and produce more, cheaper tokens?

This is what truly deserves a fresh look from the capital market in this round of domestic compute substitution.

In recent years, against the backdrop of constrained high-end chip supply and sharply rising uncertainty in obtaining compute, Chinese large-model companies have begun advancing domestic compute substitution.

Judging from cases that have already emerged, domestic compute already has the capability to support large-model operation.

That is to say, even if high-end chips cannot be bought, domestic compute can at least provide a layer of backup for Chinese large models.

And the evidence for that backup is increasing.

First, on the inference side. Zhipu's domestic-chip inference cluster has reached 100,000 cards; all online traffic for GLM-5.3-Flash runs on domestic chips, and its anonymous model Ox-Alpha has cumulatively processed 62 trillion tokens on overseas platforms, already entering the stage of real token production.

Second, on the training side. Meituan's LongCat-2.0 has completed training on large-scale domestic compute; projects such as the Guangdong-Hong Kong-Macao "domestic chips train domestic models" initiative are also advancing 10,000-card-level domestic training.

But being able to run and running well are two different things.

For example, in full-parameter post-training tasks for trillion-parameter models, Ascend 910C's MFU—a core metric for measuring training efficiency—is about 30%-35%, while H100 clusters typically reach 40%-50% on tasks of the same level. This means that with the same money and time, domestic compute still cannot train the same model for now.

Therefore, for large models, compute supply is only the first hurdle.

What truly determines competitiveness is the second hurdle: compute efficiency. It answers how much effective output limited compute can generate.

The compute efficiency of large models is jointly determined by at least three variables: chips (per-card performance, interconnect bandwidth, memory, energy efficiency), software (compilers, operators, inference frameworks, communication libraries, scheduling), and models (MoE architecture, sparse attention, low-precision computing, KV Cache optimization).

Roughly speaking, cost per unit Token ≈ chip cost ÷ (chip performance × software utilization × model efficiency).

The current state of the three variables is uneven. Chips determine the upper limit of compute supply and are also the variable that is currently hardest to make up for quickly; software determines how much of the hardware's performance can be released, while models determine how much compute is needed to complete the same task.

When chip supply is difficult to change in the short term, software and models become the areas where large model companies have more room to optimize proactively.

This also explains why domestic compute substitution is not just the business of chipmakers; large model companies themselves have also begun to participate in software adaptation, system optimization, and model efficiency improvement.

At present, the practices of several leading large model companies are precisely different paths unfolding around this logic.

Once domestic compute can run, what do large models compete on?

The four companies' answers are all different.

Among them, Zhipu wants to prove that heavy assets can bring cost advantages, MiniMax wants to prove that self-built compute can handle real traffic, DeepSeek wants to prove that model efficiency can replace chip performance, and Kimi wants to prove that the ecosystem can make up for its compute.

They are solving different problems, but all ultimately point to converting limited compute into more and cheaper Tokens as quickly as possible. The specific analysis is as follows:

From renting GPUs, to purchasing Token factory services, to going All-in on Infra, domestic chips are already the main inference compute, and it has also acquired Zhongke Jiahe to complete the compiler and inference engine software stack, putting more compute infrastructure in its own hands.

More noteworthy is that after completing a $5 billion financing round in September, Zhipu obtained the funds to continue expanding training and inference infrastructure. And management explicitly stated at an exchange meeting that compute supply is no longer the main constraint on revenue growth in the short term.

But self-built compute does not mean costs will necessarily fall.

Data centers, chips, networks, and power all require capital investment, and depreciation and O&M costs will follow.

So what Zhipu really needs to prove is whether, after turning compute from external leasing into owning it itself, it can improve utilization and ultimately lower the cost per unit Token.

Although Zhipu's management, at the exchange, for the first time gave a quantitative estimate of compute investment. A 30 billion yuan compute investment corresponds to about 100,000P of compute, of which 40% is used for training R&D and 60% for business inference. Relying on the inference cost-performance of GLM-5.3, the theoretical inference gross margin can reach as high as 80%; with compute utilization maxed out at full price, annual revenue could reach 40 billion yuan.

That is, in a single year it could not only cover all compute investment but also bear model training R&D expenses, theoretically turning compute from a cost center into an asset with calculable returns.

Whether this set of estimates can truly hold still needs to be verified against operating data. So what is most worth watching in the next stage is whether cost per unit Token and gross margin can continue to improve.

Only if these indicators all improve next year can it prove that heavy-asset compute investment has truly turned from a cost center into production capacity. Otherwise, the capital market may adjust downward.

(2) MiniMax: moving from renting compute to owning controllable compute, with implementation first.

MiniMax faces another problem.

In the first three quarters of 2025, the company's cloud computing service costs accounted for more than 90% of cost of sales. This means that as model call volume grows, compute costs will also rise rapidly.

So MiniMax began to gradually shift from renting compute to building its own domestic compute clusters, and models such as M3 and H3 have also begun to advance adaptation to domestic chips.

Its verification sequence has one more step than Zhipu's.

Because the compute cluster has just begun construction, the first step to verify is actually whether this compute can truly go from being built to running. Only when MiniMax, like Zhipu, has already begun using domestic chips to carry real inference traffic will the shift from renting compute to owning controllable compute have truly completed its first step.

By then, the market will further ask whether, after owning compute, MiniMax can enable more Tokens to be produced at lower cost.

So MiniMax is still in the verification stage of compute moving from construction to production. What can be tracked going forward is the implementation progress of domestic clusters and the progress of model adaptation. Only when the cluster is truly online and carrying real traffic can model adaptation possibly translate into actual cost improvement. Before that, MiniMax's compute autonomy remains a promise yet to be fulfilled.

(3) DeepSeek (preparing for a STAR Market IPO): turning model efficiency into part of competition.

If Zhipu and MiniMax are solving the questions of whether domestic compute can run and whether it can be held in one's own hands, then DeepSeek's highlight is that it stands on the efficiency side, not fixating on how much compute it owns, but insisting on how many Tokens the same compute can produce.

The most distinctive thing about this route is that it naturally hedges against the logic of controls. Controls on chips aim to make compute more expensive, while model efficiency aims to make compute cheaper, and the two offset each other. This is precisely the strategic undertone that distinguishes DeepSeek from other companies.

It should be noted that DeepSeek does not uniquely possess any single technology; MoE, low precision, KV Cache optimization, sparse attention and others have become common directions of continuous optimization across the entire large-model industry.

What is truly worth noting is that DeepSeek has long placed model efficiency at the core of architecture design, rather than treating it as a nice-to-have engineering measure.

But on the front of domestic compute substitution, whether this route can continue to work remains to be verified.

When V4 was released in April 2026, DeepSeek explicitly stated that, constrained by high-end compute, the Pro version's throughput was limited, and it expected prices to drop substantially after Ascend 950 super nodes go on sale in volume; the price increase for V4-Pro in August precisely exposed the real-world constraints. Before domestic compute supply is stable, the Pro version remains constrained by chips.

The real capacity bet is the 160,000-card cluster DeepSeek has deployed in the Ulanqab data center, which, according to Bloomberg, will only gradually enter production between the end of 2027 and early 2028.

In other words, DeepSeek's current pricing embeds capacity expectations for 2027–2028. If the 160,000 cards enter production on schedule and the 950 ramps up on schedule, unit Token cost is expected to fall to the lowest in the industry, giving it the greatest elasticity among the four; but the words "on schedule" are its biggest uncertainty. For every quarter delivery is delayed, the elasticity narrative shifts back by a quarter as a whole.

Therefore, what is most worth watching in the next stage is price. After domestic compute forms stable capacity, whether the Pro version's price can be substantially lowered. That is the acceptance point for whether the three lines—self-developed chips, domestic adaptation, and self-built data centers—can converge into a real cost advantage.

(4) Kimi (Moonshot AI, filing for IPO): using the ecosystem to reduce adaptation cost and time.

Kimi represents another path: through open models and ecosystem cooperation, shortening the adaptation cycle for domestic chips from the moment they are obtained to the moment they are truly usable. For example, Huawei Ascend, Alibaba T-Head, Hygon, Moore Threads, Biren and others all announced adaptation immediately after Kimi's models were open-sourced.

(Chart production: Huxiu Miaotou)

The advantage of this route is that it does not require investing a large amount of its own resources, and the adaptation cost is relatively low.

But it also exposes another problem: successful adaptation does not equal sufficient compute supply.

After the release of K3, user requests surged, approaching the carrying limit of the existing cluster, and it was later forced to suspend new consumer registrations. Compute remains a hard constraint, and when product demand surges, it is easy to be bottlenecked by compute.

So what Kimi's route solves is how to put domestic chips to use faster, but whether it can solve the compute bottleneck still needs verification. Before that, the company's performance growth may be constrained by compute.

Putting the four companies together, one finds that the adaptation route solves "can run," while compute autonomy solves "can run affordably and can keep running." The observation point is cluster expansion, and the verification point is whether they will still be bottlenecked the next time demand surges.

Final thoughts

In summary, when chips are increasingly hard to buy, replacing Nvidia cards with domestic cards is not the goal. What truly needs to be observed is whether domestic compute can transform from "backup compute" into a genuinely efficient production capacity that can generate commercial returns.

How the progress is going can be tracked with three major indicators:

First, capacity (can run).

For us investors, it is hard to know exactly how many cards a cluster has, but the production progress of data centers and compute clusters, the scale of domestic chip procurement and deployment, and whether large-model companies have begun using this compute to carry real business can all be tracked.

For example, the progress of MiniMax's domestic compute cluster deployment; the ramp-up of Ascend 950, and the production progress of DeepSeek's 160,000-card domestic compute in Ulanqab, etc.

Second, efficiency (whether it runs well).

At this layer, there is no need to insist that companies disclose internal indicators such as Token output per card; what matters more is whether the same compute investment can support more model calls and a larger business volume. Observation signals include: growth in model call volume, growth in Token usage, the scale of business carried by domestic compute, and whether the company continues to release model-side optimizations that reduce compute and memory usage.

Third, cost (whether it can make money).

Ultimately it still comes back to financial data. Look at whether unit Token cost has declined, whether the gross margin of the inference business has improved, and whether profit can keep growing after API prices fall. For listed large-model companies or those preparing to list, these data will ultimately be reflected in revenue, gross margin and cash flow.

If the above indicators improve simultaneously, domestic compute will have transformed from a backup option into a cost advantage; but if there is only adaptation and no efficiency, the narrative of domestic compute substitution will be hard to sustain.

And if one must rank the above four paths, the one with the highest short-term certainty is Zhipu, the one with the greatest medium-term elasticity is DeepSeek, and MiniMax and Kimi are still in the verification period.

Returning to the question at the beginning: if high-end chips can no longer be bought, China's large models will not necessarily stop. But by then, whoever can continue running at a sufficiently low cost may, on top of the "domestic substitution narrative," gain a higher valuation premium.

Article title: After Nvidia's cards can no longer be bought, what will China's large models rely on to keep running?

244 articles published on Huxiu. This content may not be reproduced without permission. For authorization, please contact hezuo@huxiu.com. Everyone who is changing and wants to change the world is on the Huxiu APP[ ]