I originally wanted to write an article about the financing boom of domestic large models.
But after carefully reviewing what happened in the past week, I discovered a more important signal—
The underlying logic of the AI industry is undergoing a structural shift.
This is not just as simple as a company raising 50 billion, nor as straightforward as a company stopping the use of Nvidia chips.
Behind this is a battle over"computing sovereignty".
And China is completing a silent transfer.
Let me start with my deepest impression.
Do you know how much DeepSeek founder Liang Wenfeng hates VCs?
According to Leiphone reports, this programmer-entrepreneur from High-Flyer Quant has said the same thing to every investor who came knocking over the past two years—"VC money is a burden". Tencent came to talk, Alibaba came to talk, and almost every top venture capital firm in the market is on his list of closed doors.
He is not pretending. Liang Wenfeng himself has a fund of 2 to 3 billion, which is his own money. He once stated publicly that he has 40,000 to 50,000 cards in hand, and 4 to 5 billion is something he can come up with, without needing external institutions to pay.
But it is exactly such a person who "doesn't lack money" who, one month after releasing DeepSeek V4 on April 24, 2026, suddenly launched DeepSeek's first-ever external financing round—50 billion RMB。
I thought carefully about what this reversal means.
It means Liang Wenfeng saw a goal more important than "spending sparingly"—scale。
The release of DeepSeek V4 is a watershed. Trillion-parameter MoE architecture, full open source under the MIT license, and most critically—the entire process from training to inference runs on Huawei Ascend, no longer relying on Nvidia chips。
Nvidia CEO Jensen Huang had a very direct comment:"Once DeepSeek is the first to launch on the Huawei platform, it will be a catastrophic result for the United States."
Before V4, the market saw DeepSeek as an extremely efficient model company—low training cost, strong model capabilities, and an influential open-source ecosystem.
After V4, the market sees DeepSeek asthe most critical model node in domestic AI infrastructure—it was the first to fully open up the complete path of "domestic computing power + large models."
So that 50 billion financing is not life-saving money.
It is DeepSeek's pricing anchor.
Liang Wenfeng himself put in 20 billion, with the National Big Fund leading the investment. If you don't invest, others will. The valuation was directly pulled from 10 billion USD a month ago to 45 billion USD. This is not financing; this is redefining this company's position in the national AI infrastructure landscape.
At the same time as DeepSeek's financing, a major event also happened on the other side of the ocean.
On May 5, 2026, Anthropic committed to paying Google Cloud approximately200 billion USDover the next five years, to obtain 5 gigawatts of TPU computing power and cloud services.
What does 200 billion USD mean?
If you convert it into cash, it could buy two entire Intels.
Why would Anthropic do this?
The answer is just two words:Cost。
Nvidia sells a card, and the gross margin can reach an astonishing 70% or more. After cloud vendors buy the cards, they have to mark them up and rent them to AI companies. Google's approach is "self-developed chips + Broadcom customization → self-built data centers → directly providing computing power services." This "farm-to-table" supply chain allows Google to turn the excess profits that Nvidia earns directly into huge discounts for Anthropic.
According to insiders, the unit price of computing power provided by Google isa full 40%-50% lower than Nvidia's solution。
In other words, Anthropic is not "taking sides," but doing a simple elementary school math problem—the same effect, half the price, would you use it?
But the significance of this matter goes far beyond saving money.
This isthe first time a top AI company has abandoned Nvidia's general-purpose computing power on a large scale and over a long period, turning instead to self-developed chips.
Chip investor Wang Yue commented that this is a textbook-level milestone in the process of "de-Nvidia-ization."
On one side, China has made it through the Huawei Ascend route; on the other side, top American AI companies are defecting from Nvidia. Two seemingly opposite paths point to the same endpoint—computing power is changing from a "general-purpose commodity" into "customized infrastructure"。
On May 7, Infinigence AI announced that it had secured another over 700 million yuan in financing.
The company's CEO Xia Lixue said something that I think gets to the core change of this era:
"In the Token economy era, AGI infrastructure plays a role like the refining plant in the petrochemical industry chain. It converts energy into digital oil (Token), outputting basic resources for various AI terminal applications and services."
Token economy. This is the most underestimated new consensus in today's AI industry.
The last time I heard a similar analogy was when cloud computing was just emerging, when someone compared data centers to power plants. Now Infinigence AI compares AGI infrastructure to a refining plant and Tokens to digital oil—this analogy is more precise because it maps out the upstream and downstream:
Energy → Computing Power → Token → Productivity
Infinigence AI released an AI productivity formula:
AI Productivity = Intelligence Scale × Token Production Efficiency × Token Value Conversion
The power of this formula is that it makes "AI capability" quantifiable, comparable, and optimizable for the first time—just as Moore's Law defined the chip industry back then.
Following this logic further, you will find: DeepSeek's 50 billion financing is grabbing land for a refining plant, and Anthropic spending 200 billion to lock in Google TPU is signing a long-term crude oil supply contract. They are doing the same thing—seizing the core infrastructure of the Token economy era。
After all this macro narrative, let's talk about a scenario that ordinary people can also empathize with.
On April 22, 2026, during Tesla's first-quarter earnings call, Musk admitted a fact that broke the hearts of 4 million Tesla HW3 owners:
"HW3 really does not have the capability to achieve unsupervised full self-driving. Its memory bandwidth is only one-eighth of HW4's."
Wait, I remember that in 2019 he said, "No new hardware is needed; existing vehicles can achieve L5 through OTA upgrades."
From 2015's "fully autonomous driving within two years," to 2019's "by the end of 2020 you'll be able to sleep in your car," to 2021's "unsupervised version by the end of the year," to repeated delays in 2022-2024—11 years, more than 10 public promises。
Then, in 2026, all HW3 owners were told: the FSD software package you bought for $8,000 to $15,000 will never be usable in your car for its entire life.
This is not because the software was not written well. This is a physical defect at the hardware level.
HW3 uses a 14nm process with 144 TOPS of computing power; HW4 uses a 7nm process with 720 TOPS. This gap cannot be filled by writing code.
The reason this case is important is that it reveals a cruel truth in the AI industry—
At the foundation of the Token economy, hardware is your ceiling. No matter how excellent the software is, it cannot run at 7nm speed on a 14nm track.
Back to the original question: what kind of migration is taking place in the underlying logic of the AI industry?
Putting together these events from the past month, what I see is actually a clear triangular structure:
First path: Self-developed chips + binding with major clients
Representatives: Google + Anthropic. Use your own chips, sign your own clients, bypass Nvidia. This path suits giants with ecosystem capabilities.
Second path: Domestic computing power + open-source ecosystem
Representatives: DeepSeek + Huawei Ascend. Use domestic chips to run the full chain, open source in exchange for ecosystem. This path suits excellent teams with strategic patience.
Third path: Multi-heterogeneous + neutral platform
Representative: Infinigence. Do not take sides with any chip manufacturer, focus on being a "computing power refining factory," managing and outputting the computing power of different chips in a unified way. This path suits middle-layer players building infrastructure.
Three paths, three players at different levels, all doing the same thing—reducing dependence on a single chip。
10 years ago, everyone talked about "de-IOE" (removing IBM/Oracle/EMC). Today, the AI industry is talking about "de-Nvidia-ization." The underlying business logic is the same—no company is willing to hand its lifeblood to a single supplier.
Back to Liang Wenfeng. If I had to describe DeepSeek's 50 billion with one word, I would say: transfer.
It's like a high-speed train switching from one track to another. At the moment of switching tracks, the speed must drop, the carriage will shake, and passengers will feel uneasy. But once the track switch succeeds, you will find a brand-new, longer, and faster runway ahead.
What China's AI is doing is exactly such a silent transfer.
From "using others' chips to make the fastest models" to "using our own chips to build our own ecosystem."
From "a small and beautiful tech dark horse" to "national-level AI infrastructure."
From "a capital game judged by funding amounts" to "a productivity race judged by Token output capacity."
Transfer comes at a cost. 50 billion is a cost. Liang Wenfeng himself putting in 20 billion is him using his own money to vote for this transfer.
And the endpoint of the transfer is not whose model has the largest parameters, nor whose funding amount is the highest. It iswho can, at the lowest cost, with the highest efficiency, and at the largest scale, convert computing power into Tokens, and then convert Tokens into real productivity。
This is the ultimate question of AI business.
And the answer to this question is, at a speed visible to our naked eyes, moving from negotiation tables in Silicon Valley to offices in Zhangjiang Shanghai, Zhongguancun Beijing, and Nanshan Shenzhen.
Room 1409, Building C, Shenzhen Century Industrial Center, No. 743 Zhoushi Road, Hezhou Community, Hangcheng Sub-district, Bao'an District, Shenzhen
