Two pieces of news spread through Silicon Valley almost simultaneously.
One: Anthropic is expected to achieve $10.9 billion in sales this quarter, reaching quarterly profitability for the first time.
Another: OpenAI is accelerating its IPO process, planning to confidentially file its prospectus in the coming weeks at the earliest, with a potential listing in the fall, and a valuation that could reach a trillion dollars.
Upon the news, SoftBank Group's stock price soared nearly 20% intraday, with its market value rising approximately 240 billion RMB in a single day.
One has just touched the profitability line, while the other, still in the red, is rushing to go public. Looking back at the personnel change two days ago, the logic becomes clear—
On May 19th, former OpenAI co-founder Andrej Karpathy announced on X: he is joining Anthropic full-time.
This is no ordinary job change.
Today's OpenAI is already the largest AI company by volume in the capital markets.
It just completed a $122 billion financing round at an $852 billion valuation.
Japan's SoftBank's Masayoshi Son, ignoring internal executive opposition, concentrated over $60 billion to bet on OpenAI.
But inside the company, something else is happening:
Of the 11 co-founders who signed the startup agreement in that humble office back in the day, only two remain—CEO Sam Altman and President Greg Brockman.
Capital is piling up, but core founders are dwindling.
The reasons behind this go beyond a simple explanation of "philosophical differences"; it's more like the result of a clash over strategy, competition for computing power, and a game of giants.
Who is Karpathy? Why Did He Choose Anthropic?
To understand this, one must first grasp Karpathy's position in the AI industry.
In the eyes of top investors, he is not just a technical manager but more like a key figure who can directly influence R&D pace—whichever company he joins, that company's model iteration speed changes.
The 39-year-old Karpathy does have a standout resume.
While pursuing his PhD at Stanford under Fei-Fei Li, he helped create Stanford's first deep learning course.
But what truly made him famous was his five years at Tesla.
He left OpenAI to join Tesla in 2017, and briefly returned to OpenAI in 2023.
In 2017, Musk, then an OpenAI board member, bypassed OpenAI management and directly recruited Karpathy to Tesla, responsible for AI and autonomous driving vision. Court evidence shows this move displeased OpenAI at the time.
At Tesla, Karpathy did far more than write papers.
He built the autonomous driving engineering system from scratch, including assembling a data labeling team and deploying neural networks onto Tesla's self-designed chips.
The tech circle's trendy concept of "Vibe Coding" in recent years was also popularized by him.
So, what will he do at Anthropic now?
The answer: Join the pre-training team to use Claude to accelerate the pre-training of the next-generation model.
Simply put, OpenAI currently trains large models mainly by brute-forcing computing power—massive amounts of NVIDIA GPUs running simultaneously, competing on who can afford more electricity and hardware costs.
What Karpathy aims to do at Anthropic is to have Claude help accelerate the training process itself.
If this path succeeds, the training cost of large models will drop significantly.
Karpathy's choice actually signals something: from the perspective of those actually doing engineering work, the path of simply burning money on computing power is nearing its end; using models to assist training is a more realistic direction.
Compute Consumption and the "WeWork Shadow"
The successive departure of core talent is often related to the company's operational direction.
Today's OpenAI has transformed from an early non-profit research institution into a company bearing revenue pressure.
As of February 2026, OpenAI's annualized revenue exceeded $25 billion.
But compute costs are growing even faster.
According to a 2024 Reuters report citing insider predictions, OpenAI might face up to $14 billion in losses in 2026, with positive cash flow not expected until 2029. This prediction has not been updated or confirmed.
Compute power is a heavy asset with rapid depreciation. To control losses, OpenAI began cutting unprofitable projects.
The Sora video project was shut down in March this year because it reportedly burned about $1 million per day in server costs, with user growth falling short of expectations.
The OpenAI for Science division, established in 2025, also saw its team split and merged into other product lines.
These adjustments, on one hand, are to comply with the requirements for transitioning to a "Public Benefit Corporation (PBC)" in 2025, and on the other hand, are preparations for the IPO. But for the scientists who joined driven by technological ideals, the company's priorities have changed.
And it is at this moment that Son chose to double down.
Over the past year, SoftBank has channeled over $60 billion into OpenAI through various means.
There is significant internal controversy at SoftBank about this.
Several executives privately believe concentrating this much capital on a single private company is excessively risky.
To raise funds, SoftBank sold off some assets, including NVIDIA shares. Simultaneously, the Vision Fund cut about 20% of its staff, tilting resources towards the AI track.
What SoftBank executives fear is a repeat of the WeWork debacle.
Back then, Son was enamored with WeWork's business story, ultimately losing tens of billions. According to Bloomberg, some insiders used the term "starstruck" to describe Son's attitude towards Altman this time—eerily similar to his attitude towards WeWork's founder back in the day.
After investing $60 billion, SoftBank did not secure a substantive board seat at OpenAI. But Son had already missed the last internet wave; he is unwilling to miss AI again. In his view, these losses are the price to pay for a ticket to "base intelligence."
And when the news of OpenAI's IPO came out, SoftBank's market value rose by 240 billion RMB in a single day—at least for now, this bet hasn't lost.
Musk's Compute Play
The one best at causing trouble in this game is still Musk.
He is one of OpenAI's earliest co-founders and now its most direct competitor.
In May this year, Musk lost his lawsuit against OpenAI for deviating from its original purpose, on grounds of the statute of limitations.
But the trial disclosed much information: the one who originally wanted to turn OpenAI into a for-profit company was none other than Musk himself.
He had calculated the math—Mars colonization needs about $80 billion, and controlling an AGI company was his way to raise funds.
Failing to gain control, he chose to exit, stop funding, and simultaneously poached Karpathy.
Although he lost the lawsuit, Musk soon took action on the compute front.
In early May, Musk announced the merger of xAI into SpaceX. Subsequently, SpaceX leased its Colossus 1 computing center in Memphis, Tennessee—equipped with over 220,000 NVIDIA GPUs—to Anthropic as a whole. SpaceX's IPO prospectus shows the total value of this lease is between $40 and $45 billion.
Just months ago, Musk publicly called Anthropic "misanthropic and evil" on X.
But before business interests, positions can be adjusted at any time.
Musk pinpointed OpenAI's weak spot—compute power.
Leasing the computing center to Anthropic, on one hand, generates hefty rent, and on the other hand, indirectly strengthens the power of OpenAI's competitor, putting pressure on OpenAI.
Anthropic's Fearsome Comeback
With ample compute power, Anthropic's performance is indeed accelerating.
In April 2026, Anthropic announced its annualized revenue exceeded $30 billion, surpassing OpenAI (approximately $25 billion) in scale for the first time.
By May 21st, the Wall Street Journal further disclosed: Anthropic is expected to achieve $10.9 billion in sales in the second quarter, reaching quarterly profitability for the first time.
For reference, it took Salesforce over twenty years to reach a comparable revenue scale. Anthropic, from its founding in 2021 to now, took less than five years.
More crucial is cost control.
Anthropic's product line is relatively focused, mainly on enterprise-level code generation and AI agents, without venturing into C-end video generation and other fields. Its model training costs are estimated to be only about one-fourth of OpenAI's.
Higher revenue, lower expenditure—that's Anthropic's current advantage.
For someone like Karpathy, who has long focused on engineering implementation, this difference is persuasive.
From Karpathy's choice to the compute-power game among giants, this round of competition sends a signal: the threshold for large-scale model foundational training is already very high, making it difficult for ordinary entrepreneurs to find opportunities in the general model domain. More pragmatic paths are either to focus on specific B-end scenarios like Anthropic does—such as using AI to solve workflow problems with clear willingness to pay, like code generation; or to find niche opportunities in directions like AI-assisted training, synthetic data, etc. Compute costs determine who can survive this round; that's the most fundamental calculation.
(This article was first published on TMTPost APP, author | Silicon Valley Tech_news, editor | Linshen)
