This is definitely not a simple case of a top talent changing jobs; behind it is a life-and-death battle about how to play in the second half of AI.

Written by: Seed.eth

If you stop a random artificial intelligence engineer in Silicon Valley and ask him, "Who has the best technology in this field right now?" you will likely hear one name: Andrej Karpathy.

Many people might not be familiar with this name, but in the circles of programmers and AI, Karpathy is a legendary figure.

He is one of the founding members of OpenAI, who was later recruited by Musk to lead Tesla's Autopilot, and then returned to OpenAI, representing a super star with a natural following and top-tier technical prowess.

On May 19, Andrej Karpathy announced on his personal X that he officially joined Anthropic—OpenAI's most formidable and headache-inducing rival.

Interestingly, on the same day that Karpathy announced his job switch, OpenAI had just won a court case against Musk, lifting a legal burden that had lasted over a year.

Before they could pop the champagne to celebrate, their top brain turned around and walked into the competitor's door.

This is definitely not a simple case of a superstar changing jobs; behind it is a life-and-death battle about how to play in the second half of AI.

The "Secrets" Behind the Superstar's Job Change

Let’s turn the clock back to 2017. Musk was still sitting on the board of OpenAI while managing Tesla. He wanted to do self-driving but lacked an AI head who could command authority. So he set his sights on a young researcher at OpenAI—Karpathy.

Later, a leaked email revealed in court documents where Musk wrote: "The people at OpenAI will hate me for this, but it must be done." He also stated that Karpathy "is probably the second most skilled person in the world in the field of computer vision," second only to another giant (Ilya Sutskever, another legendary co-founder and former chief scientist of OpenAI).

Thus, Karpathy was "borrowed" by Musk. Note the term "borrowed"—because at that time, Musk was the boss on both sides, making the operation a bit gray.

Karpathy spent five years at Tesla, creating the core visual system for Autopilot. In 2022, he left Tesla, briefly returned to OpenAI, then left again to start his own AI education company, Eureka Labs.

Onlookers are entertained, but insiders see the nuance. Why would Karpathy leave a promising startup project to run to the competitor's house?

This time, Anthropic arranged for him to join the pre-training team. Simply put, pre-training is the phase where a large model "closes itself off" to eat box lunches and gain skills before its birth. This is also the most money-consuming, chip-consuming, and headache-inducing part of AI development.

In the past two to three years, the reason why AI large models have become increasingly intelligent is simple and brutal: desperately buy NVIDIA chips and burn money to stack computing power. Whoever has more money and chips has the stronger model.

But Karpathy is looking to do something different. Anthropic announced that Karpathy will lead a new team focused on "how to use Claude (Anthropic's own AI model) to help itself do research and train itself."

This is a new concept that Karpathy has been advocating recently; to explain it simply: stop blindly stacking chips and burning money, and let AI learn to be its own programmer and self-iterate.

For OpenAI, this move is a bit painful. Over the past year, there has been serious faction fighting within OpenAI, and a large number of top scientists and executives have left. Karpathy's foot vote indicates that in the eyes of these top scientists, the real hardcore innovation that can change the future of AI is happening at Anthropic, not at OpenAI, which is busy with commercialization and lawsuits all day.

Valuation Approaching 100 Billion, Competitors Strike Two Hard Hits

Karpathy's joining is just one of the "double strikes" that Anthropic launched against OpenAI on that day.

On the same day, Anthropic announced another big event: the global audit and consulting giant KPMG decided to adopt their Claude system for all employees. Anyone in finance and auditing knows that companies like this have extremely strict requirements for data security and accuracy. KPMG's adoption of Anthropic means that they have severely undermined OpenAI in the enterprise market.

And just a day earlier, Anthropic executed an even more stunning strategic acquisition. They spent about 300 million dollars to directly acquire a development tools company called Stainless and announced that they would gradually cease all services offered by that company.

Why is this move brilliant? Stainless produced critical "connector" tools for the AI era, allowing AI Agents to successfully communicate with various human software systems. Importantly, companies like OpenAI and Google previously heavily relied on the tools provided by this company.

By acquiring Stainless, Anthropic effectively monopolized a key link in the industry supply chain, while simultaneously cutting off supply to its competitors.

This combination of moves has sent the capital market into a frenzy. According to Bloomberg and other media outlets, Anthropic is currently undertaking a new financing round of 30 billion dollars, with a valuation skyrocketing to between 90 billion and 95 billion dollars.

Although there is still a distance from OpenAI's giant valuation of 850 billion dollars, Anthropic's growth rate has already made all investors sweat.

Why is this happening? The reasoning is not complicated.

OpenAI has been following a "first gather users, then monetize" approach. ChatGPT has 900 million active users per week, with many using it for free and only a few paying.

Anthropic has taken another route: from the start, it has focused on enterprise clients. They do not chat with you on your phone; they delve into your workflow—helping you write code, process documents, and conduct data analysis. Once your team gets used to it, switching becomes difficult.

Especially with the programming tool Claude Code they launched; it has only been live for a year but achieved annual revenues of 2.5 billion dollars, capturing over half of the programming assistant market. Currently, 4% of global GitHub code submissions are written using Claude Code. Just a month ago, this figure was 2%.

"Four Country Kill" in the AI Circle

By 2026, the AI sphere will no longer be simply dominated by OpenAI; the market will have turned into a "four country kill" scenario dominated by four major players:

• OpenAI: Valuation of 852 billion dollars, currently earning 2 billion dollars a month, with an annual revenue of up to 24 billion.
• Anthropic: Valuation racing towards 95 billion dollars, with its code tool Claude Code earning 2.5 billion dollars a year.
• Google: Also a considerable giant. Google is currently working on its own Gemini large model while acting as Anthropic’s second-largest computational supporter and investor. The latest financial report shows that just the shares they hold in Anthropic have given Google an unrealized gain of 36.8 billion dollars.
• Microsoft: A typical "bet on both sides" player. It holds about 8% of OpenAI, valued at over 220 billion dollars on paper; at the same time, it has invested 15 billion dollars into Anthropic to purchase its cloud computing power. Regardless of which side wins, Microsoft will collect computing taxes from behind.

Returning to today's explosive news. Karpathy's job switch reveals a major trend: the era of AI relying on who has more money and chips is likely coming to an end.

The upcoming competition will hinge on who can instruct AI to learn to evolve itself, and who can drive down the high costs of computing power.

Although OpenAI won the legal battle in court and cleared the obstacles to going public, it has already begun to lose absolute control over the core "talent balance." The second half of the AI industry, led by agents and automation, has just begun.