In May 2026, the tech world was so flooded with news that it was hard to keep up.
On May 7, Musk announced on X: xAI will no longer exist as an independent company; it is simply SpaceX AI, i.e., SpaceX's AI product. That day, SpaceX announced it would have a supercomputing cluster Colossus 1 with 220,000 Nvidia GPUs, all supplied to Anthropic—xAI's former competitor.
Two days later, Anthropic was reported to be preparing a funding round of up to $50 billion, with a pre-money valuation of about $900 billion; after completion, its valuation would approach $1 trillion—surpassing OpenAI to become the world's most valuable AI startup.
On the same day, ByteDance was reported to have AI infrastructure spending this year exceeding 200 billion yuan, at least 25% more than the plan at the end of last year, with a larger proportion allocated to domestic chips.
Wait, this is just what happened on May 9 alone.
If you feel a bit breathless, that's normal. Because this AI game is undergoing unprecedented changes.
---
Let's first talk about the world's most valuable AI company, which may soon no longer be called OpenAI.
On May 9, the Financial Times reported that Anthropic, the developer of the Claude series of large models, plans to raise up to $50 billion this summer, with a pre-money valuation of about $900 billion. If this funding round is completed as expected, Anthropic's valuation will reach nearly $1 trillion, overtaking OpenAI's current valuation of $852 billion.
How should we understand this number?
Let me give you a few reference points. Apple, the world's most valuable listed company, currently has a market value of about $2.3 trillion. An AI startup founded just over five years ago has a valuation reaching 40% of Apple's—this is an incredible speed in any industry cycle.
Even more astonishing is Anthropic's growth trajectory. In February this year, after completing funding, Anthropic's valuation was still $380 billion. In just three months, its valuation soared by more than 130%. Its annualized revenue is expected to soon exceed $45 billion, a fivefold increase from $9 billion at the end of 2024.
Five quarters, fivefold revenue growth.
This growth is not achieved by giving PowerPoint presentations. The Financial Times report noted that Anthropic's core growth comes from two products: the Claude Code tool for developers and the Cowork assistant for non-technical users. These two products are rapidly capturing the enterprise market, greatly narrowing the gap with OpenAI.
At the same time, Anthropic is also frantically locking down computing resources. In the past two months, it reached an agreement with SpaceX to obtain 220,000 Nvidia GPUs and 300 megawatts of computing support; it signed multi-billion-dollar computing agreements with Google, Broadcom, and AWS. Anthropic CEO Dario Amodei remarked at a developer conference: "ARR growth is exponential. We once thought it might gradually grow to 10x, but in the end we saw 80x growth."
80x.
So, in this AI arms race, no one thinks it's "enough."
---
Let's turn our eyes back to China.
On May 9, the South China Morning Post reported, citing people familiar with the matter, that ByteDance's AI infrastructure spending this year will exceed 200 billion yuan.
200 billion, not 20 billion.
This is at least a 25% increase compared with the plan at the end of last year (160 billion yuan). People familiar with the matter revealed that this increase is necessary—both because ByteDance is continuously increasing AI investment and because of the impact of rising memory chip costs.
There is another important development: ByteDance is reported to allocate a larger proportion of its budget to domestic AI chips.
Do you feel it? This is not just a budget issue for ByteDance. This is a signal—one of China's largest internet companies is voting with real money, betting on the domestic chip ecosystem.
Of course, if we broaden our view globally, ByteDance's 200 billion is not earth-shattering. In the same period, Google and Microsoft revealed that their full-year capital expenditures will each reach about $190 billion (about 1.3 trillion yuan), Meta raised its capital expenditure forecast for this year to $145 billion (about 1 trillion yuan), and Amazon maintained its $200 billion (about 1.36 trillion yuan) unchanged.
Converted, the AI capital expenditures of the four major U.S. tech giants add up to nearly 5 trillion yuan.
What kind of concept is this? In 2025, China's total social R&D expenditure was about 3.7 trillion yuan. The AI investment of four U.S. tech companies alone exceeds China's total annual R&D investment.
This is not a fair competition, but the starting gun has already been fired.
---
Facing such pressure, Chinese tech companies naturally will not sit idly by.
On May 9, Baidu released its new-generation foundation model—ERNIE 5.1. Two data points are worth noting.
First, cost control. According to Baidu's official introduction, the model uses "multi-dimensional elastic pre-training" technology and achieves leading foundational performance with only about 6% of the pre-training cost of industry models of the same scale. Total parameters are compressed to about 1/3, and activated parameters are compressed to about 1/2.
This is very interesting.
Over the past year or so, there has been a consensus in the industry: large model training is a "money-burning" endeavor—the larger the scale, the higher the cost. DeepSeek broke the myth that "compute equals everything" with innovation, and now Baidu has given its own answer on cost efficiency.
6% of the cost, achieving leading performance. If this data is true, it means China's AI has found its own methodology on the path of "doing big things while saving money."
The second data point: on the search leaderboard of the large model arena LMArena, ERNIE 5.1 ranked first in China and fourth globally with 1223 points, the only domestic model on the list.
Baidu stated that this search capability refers to a large model's ability to quickly retrieve, integrate, and generate from multi-source information. In other words, whether your AI assistant can quickly find the correct answer depends on this.
Of course, we must also clearly see the gap. On LMArena's text leaderboard, ERNIE 5.1 Preview once took first place in China, but still lags behind GPT-5.5 and Claude Opus 4.7. The gap is narrowing, but it has not disappeared.
On May 13, the Create 2026 Baidu AI Developer Conference will be held, and Robin Li will deliver a speech. At that time, more technical details of ERNIE 5.1 will be revealed.
Worth looking forward to.
---
If everything above is about "investment and catching up," then the story below is about "changing the rules of the game."
In May 2026, AI chipmaker Cerebras Systems disclosed IPO offering details, planning to raise up to $3.5 billion at a target valuation of $26.6 billion. This is already Cerebras's second attempt at an IPO—it filed for listing in October 2024 but withdrew due to a review by the Committee on Foreign Investment in the United States (CFIUS).
What should we look at in Cerebras's IPO?
It's not just about an AI chip company's IPO. It's about the restructuring of OpenAI's compute supply chain.
On May 6, OpenAI made a very interesting move—it brought together chip companies with clearly competitive relationships, including Nvidia, AMD, Intel, Broadcom, and Microsoft, to launch a network protocol for large-scale AI training clusters (MRC, Multi-Rail Compute).
On the surface, this is a supercomputing network collaboration. Deep down, it's OpenAI redistributing the pie.
In the past, the AI compute market was a classic case of "Nvidia takes all"—large model companies burned cash, cloud providers bought cards, startups queued for GPUs, and profits flowed to Nvidia, the shovel seller. But OpenAI doesn't want to forever be the miner bound by the "shovel."
What is its strategy? A compute combination punch.
Use Nvidia's high-end GPUs for training, bring in Cerebras's low-latency solution for inference, procure some GPUs from AMD, open up network protocols, and place bets on multiple cloud services among AWS, Azure, and Google Cloud. OpenAI is even pushing forward with self-developed chips.
What role does Cerebras play in this? Inference.
Cerebras's core competitiveness lies in its unique wafer-scale engine chip, the WSE-3. Traditional chips are cut from a whole wafer into many small pieces, each piece being a chip. Cerebras does the opposite, directly turning an entire 12-inch wafer into one giant chip—46225 square millimeters, equivalent to one-third the size of an A4 sheet of paper.
In inference scenarios—especially tasks requiring low latency such as long-text output, real-time interaction, code generation, and Agents—Cerebras's CS-3 system is 21 times faster than Nvidia's DGX B200, with both cost and energy consumption reduced to one-third.
21 times the speed, one-third the cost. The "solid monolith" of the compute market is being pried open.
This is a very important signal: AI infrastructure is moving from "one dominant player" to "specialized division of labor." Training, inference, networking, cloud distribution, application scenarios—each layer can have different players, and each layer can be repriced.
---
Having talked about "dividing the pie," let's talk about a story of "losing the pie."
Remember the bold words when Musk announced the founding of xAI in July 2023? "To understand the true nature of the universe."
At that time, xAI poached a group of top AI scientists from OpenAI, DeepMind, and Google. With 12 co-founders, the lineup could be called the "AI Avengers." In 2024, xAI deployed the Colossus supercomputer on a large scale, once one of the fastest AI training clusters in the world.
But by May 2026, what happened?
On May 7, xAI was officially merged into SpaceX. Musk said: "xAI will no longer exist as an independent company; it will just be SpaceX AI, i.e., SpaceX's AI product."
More ironically, on the same day SpaceX announced a partnership with Anthropic, supplying all 220,000 GPUs of Colossus 1's compute to Anthropic—xAI's competitor.
Then, the wave of departures began.
In the early hours of May 9, Zhuang Juntang, head of xAI pretraining, tweeted his resignation. He graduated with a bachelor's degree in engineering physics from Tsinghua University, a master's in statistics from Yale University, and a doctorate in biomedical engineering. He was a core contributor to GPT-4o, a major contributor to DALL-E 3, and Grok 2/3/4/4-fast/4.1/4.2/4.3 all came from his work.
On the same day, team member Xiuyu Li and front-end engineer Fraster Cook also announced their departures.
You see, one detail is quite interesting: below Zhuang Juntang's resignation tweet, the comments were all 🐐 (GOAT, greatest of all time). Colleagues were saying goodbye in this way.
In addition, according to The Information, xAI also conducted an internal layoff last week, affecting about 10 employees on the Grok model-related team.
Counting further back, among xAI's original 12 co-founders, only Musk remains today. Wu YuHuai left, Jimmy Ba left, Toby Pohlen left, Ross Nordeen left—Musk's AI startup partners all ran off.
Isn't this a lamentable story?
A once dazzling AI company, due to strategic wavering and resources ceded to a "competitor," went from star to abandoned child in less than a year.
Where did the core problem lie?
Let me try to summarize: when a company's founder simultaneously manages seven or eight companies in different fields (Tesla, SpaceX, X/Twitter, xAI, Neuralink, The Boring Company...), and AI happens to be a track that requires going all in—then your AI company is destined to be unable to compete with those companies that "only do AI."
Resources are limited, and attention even more so.
---
Having told the story of "losing the pie," let's tell one of "refusing to eat the pie."
Late at night on May 8, a piece of news exploded in venture capital circles: DeepSeek is seeking its first round of financing, raising up to 50 billion yuan, with a post-investment valuation of 350 billion yuan. This amount would set a new record for a single financing round by a Chinese AI company.
But what is more intriguing is an episode during the negotiations.
According to multiple sources, Alibaba reached a deadlock in negotiations for this financing round. Some market observers even told National Business Daily: "Alibaba probably did not negotiate."
Why?
An article from Phoenix Technology gives a pointed headline: "Big Tech Will Never Understand Liang Wenfeng."
There is a passage in the article that I really like:
Internet big tech will never understand Liang Wenfeng, just as old players struggle to keep up with a new game—they think this is a game where money alone gets you in, but in reality they may not even be at the same table.
Let me give you a few data points and you will understand. Among all the AI rising stars, DeepSeek is the one closest to original innovation. Liang Wenfeng advocates anti-KPI and anti-OKR, and this kind of "anti-big-tech organizational culture" is precisely the core methodology that big tech relied on for success over the past many years.
When big tech wants to use money and strategic resources to exchange for board seats, what is Liang Wenfeng doing?
On April 27, 2026, on the eve of intensive fermentation of financing news, DeepSeek quietly completed an equity structure adjustment. Through direct capital increase, Liang Wenfeng raised his shareholding ratio from 1% to 34%, controlling a total of about 84.29% of the company's equity. More directly, he planned to contribute up to 20 billion yuan in his own name, accounting for 40% of the total fundraising.
First use his own real money to tighten control, then open the door to welcome strategic investors—the signal conveyed by this sequence of operations could not be clearer: Liang Wenfeng welcomes capital, but will absolutely not tolerate capital taking over as host.
From big tech's perspective, "take money, give resources, lay out channels, and you focus on technology"—isn't this the best business in the world?
But in Liang Wenfeng's logic, the two words "being bound" may be an untouchable red line.
They do not dislike money; they are afraid of losing freedom.
---
At this point, let us string today's topic together and take a look.
In May 2026, the combined valuation/market value of China's top five AI companies (DeepSeek, Zhipu, MiniMax, Kimi, StepFun) had already reached 1.1 trillion yuan. Behind them stand multiple big tech and mid-sized tech companies such as Tencent, Alibaba, Meituan, Ant, Xiaohongshu, Xiaomi, and miHoYo.
Tencent has invested in almost all of the top five AI companies. Alibaba participated in investing in three: Kimi, Zhipu, and MiniMax. Meituan recently placed a heavy bet on Kimi.
But have you noticed? DeepSeek—the most scarce ticket—is where big tech hit the strongest wall.
It is not because the price could not be agreed upon. It is because the two sides' understanding of the "rules of the game" is simply not on the same channel.
---
Today, I want to share an observation with all of you:
AI is entering an era of "specialized division of labor," not an era of "winner takes all."
You see, OpenAI does not want to be bound by Nvidia and is splitting, layering, and repricing the computing power supply chain. Baidu produced a leading model at 6% of the cost, proving that "burning money" is not the only path. Anthropic proved with 80-fold revenue growth that AI commercialization is no longer a distant blueprint. ByteDance spending 200 billion to buy domestic chips proves that Chinese tech companies are joining the race in their own way. DeepSeek refusing to be bound by big tech proves that the logic of innovation can be harder than capital.
Every company is redefining the rules of the game in its own way.
2026 is not even half over, and the reshuffling of this card game has already begun.
Are you ready to place your bet?
---
*Sources cited in this article: Phoenix Technology, Zhidongxi, National Business Daily, IT Home, Financial Times, South China Morning Post, The Information, etc.*
Room 1409, Building C, Shenzhen Century Industrial Center, No. 743 Zhoushi Road, Hezhou Community, Hangcheng Sub-district, Bao'an District, Shenzhen
