In the past two years, during the craziest time in the AI industry, everyone was focused on the models. How powerful is GPT, has Claude caught up, why can DeepSeek achieve amazing results at a lower cost, and can domestic large models continue to progress.

Written by: Non-Small Account

In the past two years, during the craziest time in the AI industry, everyone was focused on the models. How powerful is GPT, has Claude caught up, why can DeepSeek achieve amazing results at a lower cost, and can domestic large models continue to progress.

But by 2026, things started to change.

On April 24, DeepSeek released the preview version of DeepSeek-V4, the model scale and context capabilities once again refreshed market expectations. V4-Pro reached a total parameter of 1.6 trillion, with 49B active parameters, supporting millions of contexts. By August, DeepSeek further updated the official version of V4-Pro and began to enhance the capabilities of Agent, Responses API, and coding scenarios. In the past, people would interpret these developments as "the large model has been upgraded again," but looking at it now, what is truly noteworthy may no longer be the model itself.

The significance of this development is much greater than "DeepSeek has released an even stronger model."

Because it signifies that the core of competition in the AI industry is shifting from "whose model is stronger" to "who can obtain computing power at a lower cost and with more stability." The model is just a layer of the product presented to users, while chips, data centers, electricity, networks, and computing power scheduling are the underlying foundations that support these models in continuous operation.

Ironically, just as DeepSeek is moving into the chip field, Xiaomi is also doing the same thing.

On August 24, Xiaomi released the Xuanjie O3 and further launched the Xuanjie O100 aimed at AI computing and the Xuanjie D100 aimed at smart driving. Prior to this, Xiaomi had already invested heavily in establishing its own chip research and development system, and this layout also means that Xiaomi's chip strategy is further extending from mobile phone SoC to AI computing and smart vehicles. Reuters reported that Xiaomi has currently invested over 20 billion RMB in the development of Xuanjie chips, with related team sizes exceeding 3,000 people.

A company that builds large models begins to research chips, while a company that makes phones and cars starts to layout AI acceleration chips. On the surface, the two companies seem unrelated, but they are actually doing the same thing behind the scenes: minimizing dependence on external computing power systems.

This is the real change worth paying attention to in today's AI world.

1. The most expensive thing in AI may never be the model, but the computing power behind the model

Many people in the past understood AI by treating the model as a core asset. A team spends years training a large model, so the model itself becomes the company's moat. But after AI began to commercialize, everyone gradually discovered a more realistic problem: training the model is just the first step; what is truly expensive is keeping it running constantly every day.

After a model is trained, each user question, each task executed by the Agent, each code generation, each image and video generation requires computing resources. The more users there are, the larger the call volume, and the more GPUs are needed. The more complex the model, the higher the inference cost.

This is also why DeepSeek has repeatedly emphasized inference efficiency in the past. Its advantage is not just "the model is cheap," but through algorithm, architecture, and engineering optimizations, striving to maximize the value extracted from each GPU.

However, no matter how much one optimizes, one cannot escape a reality: if AI continues to grow, the demand for computing power will continue to grow.

This leads to a very interesting change. Early internet companies needed servers, so they handed over servers to IDC; later, in the cloud computing era, companies no longer needed to buy machines themselves and could rent resources directly from cloud vendors. But the AI era may have returned to another extreme; the demand for computing resources from large models is so huge that cloud computing itself is also beginning to become a new scarce resource.

So you will see more and more large companies starting to develop their own chips. NVIDIA is certainly still wildly expanding its GPU ecosystem, but Google has TPU, Meta is advancing its own AI accelerator, Amazon has Trainium and Inferentia, and other large AI companies are also looking for ways to reduce computing power costs.

DeepSeek's entry into the chip field is essentially the same.

What it really wants to solve is not the simple question of "can I produce a GPU stronger than NVIDIA," but rather "can I make my models run more efficiently, cheaply, and controllably on specific tasks."

This distinction is very important.

Because after the AI industry enters commercial competition, model capability only determines whether you can win users, while computing power cost determines whether you can survive.

2. Why is Xiaomi also making AI chips? Because phones are becoming a computing node

If DeepSeek is making chips because models need computing power, then Xiaomi's entry into the AI chip field represents another change: AI is moving from data centers to endpoints.

Past phones had the most important tasks of communication, photography, and running apps. Now more and more functions are directly handled by AI, and voice assistants, image processing, translation, search, content generation and various Agent capabilities all require increasingly stronger local computing capabilities.

Phones are not the only endpoints.

Cars, robots, AI glasses, home devices will all need computing power in the future. A large number of AI tasks may not all be uploaded to the cloud but may be executed directly on the device locally. Doing so not only reduces latency but also minimizes network transmission and some cloud costs.

So what Xiaomi really wants to do is not just a smartphone chip.

It is attempting to establish its own edge AI computing system.

This is why the combination of the Xuanjie O3, O100, and D100 is worth more attention than simply releasing a mobile processor. It reflects that Xiaomi is simultaneously occupying significant computational accesses in phones, AI computing, and smart driving.

This brings us back to the question we just mentioned: those who control computing power also control part of the infrastructure during the AI era.

If a company’s entire AI business relies on other people's chips, other people's clouds, and other people's data centers, then its AI operations are essentially built on the infrastructure provided by others.

Once costs rise, supply issues arise, or ecosystems change, the company itself will be affected.

That is why there is now a very obvious trend: more and more enterprises are starting to find ways to possess their own computing capabilities.

But this also raises another question.

If in the future, global AI truly needs such massive computing power, do we really have to rely on a few companies to build larger and larger super data centers?

The answer may not be singular.

This is where Web3 has a genuine opportunity to re-enter this story.

3. The truly interesting aspect of DePIN is not "renting GPUs for profit," but re-organizing idle computing power

A few years ago, the Web3 industry loved to talk about DePIN.

Various projects told users that as long as they contributed GPUs, storage, bandwidth, or other hardware, they could earn Token rewards. This story was very lively at the time, but the problems were also very obvious: if the network itself has no real demand, then having more nodes makes no sense.

Many projects ended up becoming, "issue Tokens to attract machines, machines join the network for Tokens, Token prices rise to attract more machines, and once prices fall, nodes quickly exit."

This is also why the market gradually started to reassess DePIN.

But AI is providing it with a completely different opportunity.

Because AI really needs massive computing power.

This is not a false demand created by Tokens, but a demand in the real world that is rapidly increasing.

Currently, many AI companies need GPUs, but not all GPUs worldwide are running at full capacity. Many enterprises have idle computing power, many data centers have temporarily idle resources, and many high-performance personal computers are in low-utilization states for most of the time.

The question becomes: can this dispersed computing power be reorganized?

The traditional cloud computing method of solving this problem is to build large data centers, gather resources, and manage them uniformly. What DePIN aims to try is another path: making the hardware scattered in different places into an open network.

At this point, the truly meaningful part of Web3 is not "blockchain + GPU," but rather it provides a new coordination method.

A node can contribute GPUs, a company can contribute data centers, a team can provide model services, and the network is responsible for organizing these resources and settling through verifiable rules.

The biggest difference from traditional cloud computing is not whether there are Tokens but rather the difference in resource ownership and organization methods.

In the past, computing power belonged to a large cloud vendor.

In the future, a new model may emerge: computing power belonging to thousands of nodes, but coordinated by a single network.

This is precisely what DePIN is most期待的地方.

4. The real collision point between AI and Web3 may be the "computing power market" rather than "AI tokens"

Of course, things are not that simple.

If we really want to organize the globally dispersed GPUs into an AI computing network, the first thing to address is the performance disparity. The performance of RTX 5090, 4090, A100, H100, and various domestic acceleration cards are completely different, and the video memory, bandwidth, network capability, and stability also vary. You cannot simply assume that "one GPU is a standard computing unit."

Next is the network issue.

Training a large model requires a large amount of high-speed data exchange between GPUs. Connecting thousands of GPUs dispersed around the globe does not equal having a super data center. Network latency, bandwidth, task scheduling, and failure rates can all directly affect efficiency.

On top of that, there is the data security issue that companies care about the most.

Is a financial company willing to entrust its sensitive data to an unfamiliar node for processing? Is a large AI company willing to place its core model on a completely unfamiliar machine? If these questions cannot be solved, then DePIN can at most take on some low-sensitivity and divisible computing tasks.

Finally, there is one very practical problem: the economic model.

If a node contributes a GPU but the monthly earnings are not enough to cover the electricity, depreciation, and maintenance costs, then even the most attractive Token economics are meaningless.

Thus, truly valuable DePIN projects will not just tell you, "plug in GPUs and earn Tokens daily."

What they really need to establish is a computing market that can operate long-term.

In this market, some provide computing power, some purchase computing power, some are responsible for scheduling, some validate tasks, and others are responsible for solving security and settlement issues.

Once all these aspects are established, Tokens may transition from "marketing tools" to genuine network economic tools.

In other words, the truly worthwhile aspect of participating in Web3 may never have been about "AI tokens," but about the entire infrastructure supporting AI.

5. When AI starts to compete for ownership of computing power, Web3 may welcome a real opportunity

Now, looking at DeepSeek, Xiaomi, and DePIN together, the story becomes much clearer.

DeepSeek begins to develop AI chips to better control its computing costs; Xiaomi is laying out AI acceleration chips to enable more endpoints to acquire computing能力；而 DePIN 的思路则是把原本分散在世界各地的 GPU 和计算设备组织起来。

These three routes may seem different, but they ultimately point to the same question: who owns the computing power in the AI era, and who manages it?

The last twenty years of the internet represent a process of increasingly centralized infrastructure evolution. Data concentrated in the cloud, services concentrated on platforms, and computing power centralized in large data centers.

AI’s development may be pushing this concentration to new heights.

Because large models are more reliant on computing resources than traditional internet applications.

If this trend continues, then the most crucial companies in the future may not only be model companies or chip companies but those that can truly master, organize, and allocate computing power.

This provides Web3 with a very unique path.

Web3 does not necessarily need to recreate ChatGPT, nor does it need to produce another "AI concept coin." It is more likely to participate at another layer: networking the computational resources from the real world.

From this perspective, the future "miners" may also change.

In the past, miners contributed the computing power needed for blockchain consensus; in the future, nodes in certain computing networks might contribute AI inference, model execution, data processing, and other real tasks.

In the past, people bought mining machines to extract a Token.

In the future, people may buy GPUs to become part of AI infrastructure.

The greatest difference between the two is that the former creates a digital asset, while the latter serves an already existing and continuously expanding industry.

So, DeepSeek making chips and Xiaomi entering the AI accelerators market is not just about adding two more players in the domestic chip race.

They are more like reminders to the entire market: the competition in AI is shifting from the model layer to the infrastructure layer.

And when computing power becomes the most essential production material in the AI era, the new business models generated around computing power may have just begun.

Perhaps in a few years, looking back at today's changes, what truly alters the AI industry is not a model with bigger parameters or a chip with higher scoring performance.

What matters is that humanity has begun to seriously consider a question for the first time: can computing power be shared, jointly contributed, and settled by people and machines around the world like the internet?

If the answer to this question is ultimately positive, then the real opportunity for Web3 may finally not be to recreate a financial casino but to participate in building the next generation of AI infrastructure.

After all, models can become cheaper, and applications can become more widespread, but one thing will not disappear out of thin air.

For AI to think, it must compute.

And for computing, power must be expended.