Don't focus on what AI can do. Focus on what AI companies are becoming.
May 4, 2026, federal court in San Jose, California.
On the fourth day of the trial, the gallery was no longer as crowded as it had been on the first day. But today's witness was heavyweight enough—Greg Brockman, president of OpenAI, Sam Altman's "golden partner," was brought to the witness stand.
It should have been a minor case: Musk sued OpenAI, accusing it of betraying its original nonprofit mission and asking the court to block its "illegal transformation." But this trial turned into the most dramatic public interrogation in the AI industry.
Brockman immediately displayed a typical "top-student-style" defense. Musk's lawyer Steven Molo asked him: Does OpenAI often use purple boxes to mark key information? Brockman said: No, it's not that special.
Then Molo, in front of the jury, projected an internal document onto the screen—in black and white it read: "Purple boxes are used in OpenAI internal documents to highlight important information."
This was the famous "Battle of the Purple Box" (Why are we fighting about the fucking purple box?—a court reporter later wrote).
Why is a purple box so important?
Because it illustrates a core problem: OpenAI habitually is dishonest about key information.
Let's rewind to 2015.
That year, Greg Brockman resigned as CTO of Stripe and sent Sam Altman a message: "I want to do an AI project." Altman replied: "Funny, I've been thinking about that too."
The two had a meal. Elon Musk arrived an hour late, but they talked enthusiastically. The three people's consensus was simple: Google DeepMind was dominant in AI, and that was too dangerous. They wanted to build an open-source, nonprofit AI lab for the "benefit of all humanity."
Thus OpenAI was born. Mission statement: "Build safe artificial general intelligence (AGI) and ensure its benefits benefit all humanity."
How noble. How moving.
But by today in 2026, what did Brockman admit in court? His OpenAI stake is worth about$30 billion. Molo asked him: Since you say $1 billion is enough, why not donate the remaining $29 billion to OpenAI's nonprofit arm?
Brockman fell silent. Then began to circle around.
This is the biggest paradox in the AI industry:A company that claims to "serve all humanity" has made its co-founder worth $30 billion.
Musk's lawyers dug deeper. They pulled out Brockman's 2017 diary:
"btw, another realization: stealing the nonprofit from him (Musk) is wrong. Converting to a public benefit corporation without him. That would be pretty morally bankrupt."
"Maybe we should just convert to a for-profit company. Making money for us sounds great."
"Can't say we promised nonprofit. Also can't say we promised. If three months later we become a public benefit corporation, that's lying."
Diaries don't lie. The words Brockman himself wrote are 100 times more honest than any "I don't think so" he said in court.
But this trial is not just about Brockman's diary. It also exposed a deeper industry change.
At the same time as the trial, a blockbuster news story broke:OpenAI shelved plans to spin off its hardware and robotics divisions.
The Wall Street Journal reported that Sam Altman had considered a structure similar to Google Alphabet—separating the core AI business from "side businesses" such as robotics and hardware, in preparation for an IPO. Robotics and hardware, in Altman's view, were "side quests."
Wait a minute. In 2015 they said nonprofit, open, for humanity. In 2020 they said they had to be for-profit because they needed compute. In 2024 they said they would do hardware and robotics "on the path to AGI." In 2026 they say those are just "side quests"?
Strategy is not what you want to do, but what you choose not to do.
OpenAI went from nonprofit → capped profit → fully for-profit → spin-off IPO → cutting non-core businesses. This path clearly shows: when a company faces IPO pressure, the first step is always to "focus on what can generate the most revenue."
This is not a criticism. This is a business rule.
OpenAI's transformation is not an isolated case. In fact, in May 2026 the AI industry is undergoing three deep splits:
On April 30, Anthropic published an important paper—"Introspective Adapters." Simply put, it lets AI models "confess themselves": the model can not only give an answer, but also say why it gave that answer. If the model is "lying," it will tell you it is lying.
This is a major breakthrough in the field of AI safety. But at the same time, Anthropic is also facing increasing commercial pressure. It is rumored that its valuation has exceeded 60 billion US dollars.
The deeper the safety research goes, the more computing power is needed; the more expensive the computing power, the greater the commercialization pressure. This is a death spiral.
DeepSeek continued to push forward in 2026, and its R1 model has already caught up with GPT-4 on multiple benchmarks. But what is more noteworthy is that the DeepSeek team recently proposed and open-sourced "Thinking With Visual Primitives"—a brand-new multimodal reasoning paradigm.
In the United States, OpenAI, Google, and Anthropic are increasingly leaning toward closed source. The reason? Safety.
On one hand they say "open source threatens safety," while on the other they hide their most core models behind a paywall. This tension is tearing apart the entire AI community.
The paper trends at CVPR 2026 show that visual AI is shifting from "looking real" to "being physically correct." The work of Liang Xiaodan's team at Sun Yat-sen University shows that merely generating good-looking videos is no longer enough; AI must understand the laws of physics.
AI has not replaced you, but it makes you work overtime every day—because you not only have to complete your own job, but also learn to use AI tools.
General AI shines in academia, while specialized AI quietly changes everything in actual work. But this change is slow, painful, and uneven.
On May 4, there was also a piece of news that many people overlooked: The New York Times reported that the White House is drafting an executive order on AI regulation.
Strange, right? The Trump administration abolished Biden-era AI safety regulations as soon as it took office. So why start regulating again?
The reason is simple: after Anthropic's "Mythos" model was released, CISA (the Cybersecurity and Infrastructure Security Agency) was excluded, triggering national security concerns. The intelligence community began to worry: "If a devastating AI cyberattack really happens, how severe would the political consequences be?"
So now the White House plan is: the government can get priority access to new AI models, but it will not block public release. Implementation will be carried out by a working group composed of industry and government officials.
This is a typical "want this and also want that": they want the economic growth brought by AI capabilities, but they do not want to bear the political consequences of AI getting out of control.
Regulation is always lagging. But when regulation begins to catch up belatedly, it shows that this technology has truly reached a tipping point.
Coase said: when transaction costs are low enough, resources will flow to the most efficient users. AI has reduced the transaction cost of knowledge to almost zero.
But at the same time, AI has also greatly increased the transaction cost of computing power—Nvidia GPU prices have soared, and the cost of building data centers often runs into billions of dollars. This in turn forces AI companies toward centralization and for-profit operation.
The marginal cost of knowledge is approaching zero, while the marginal cost of computing power is soaring. This is the fundamental reason why the "open-source ideal" of the AI industry is being crushed by "commercial reality."
In the 19th century, British economist Jevons discovered that after the efficiency of steam engines improved, coal consumption not only did not decrease, but surged. Because improved efficiency made steam engines cheaper, more people used them, and the total amount went up.
The same is true for AI. Every breakthrough that makes AI more efficient will bring more AI usage; more usage means more demand for computing power; more demand for computing power means more centralized and more expensive AI infrastructure.
Do not expect AI to become cheaper. It will become better and more widespread, but it will not become cheaper. Because it will consume more resources.
There is a detail in this trial that many people overlooked: after Musk left OpenAI, he told Brockman that he wanted to build an AGI competitor inside Tesla. Brockman relayed Musk's exact words in court: "The probability of OpenAI succeeding—zero. But we must have a counterbalancing force against Google/DeepMind."
Today, at least three "trillion-level" players are racing to pursue AGI: OpenAI (backed by Microsoft), Google DeepMind, and xAI (Musk).
When competition for AGI turns from a technological race into a geopolitical arms race, no one can stay out of it.
If this article only lets you remember three things, I hope they are these three questions:
First, is your AI strategy "chasing trends" or "building moats"?
OpenAI cut its hardware and robotics businesses not because they were bad, but because they did not fit the IPO path. Your company has limited resources. What do you choose to do and not do in AI?
Second, do you understand the risks to you of the "AI arms race"?
If the cost of AI infrastructure continues to soar, what will the landscape look like five years from now? Is your current AI investment an "investment" or a "ticket"—merely obtaining the qualification to enter this track?
Third, are you prepared for AI's "trust crisis"?
Brockman's diary exposed a fact: when the interests are large enough, even the noblest mission statement will be diluted. Do you trust your AI vendor? Are you prepared to be deceived by AI?
Back to that courtroom.
When Brockman was asked "What do you do at OpenAI," he gave an answer that made the whole room collapse:
「I do all the things.」
I do everything. What a typical millennial CEO line.
But behind this sentence lies a brutal truth:When AI is so important that it requires "doing everything," no one really knows what they are doing.
OpenAI doesn't know. Musk doesn't know. The White House doesn't know.
And it is precisely this "not knowing" that makes this industry both incredibly fascinating and extremely dangerous.
This is not an article about what AI can do. This is a story about how AI will change the nature of business.
Room 1409, Building C, Shenzhen Century Industrial Center, No. 743 Zhoushi Road, Hezhou Community, Hangcheng Sub-district, Bao'an District, Shenzhen
