The whole internet is going wild over Noam joining, but OpenAI’s loss report has added another page.

| AI Going Against the Grain

Ultraman posted on X: “Noam is one of the people I most wanted to work with when I founded OpenAI. In just 10 years. Worth it.”

OpenAI’s Chief Research Officer Mark Chen immediately announced: Noam Shazeer will serve as the Head of Architecture Research.

The whole internet lined up to cheer: “With the father of the Transformer in charge, the next superintelligence is locked in.”

In the same week, another document also circulated in the circle: audited financial data was fully disclosed for the first time. OpenAI’s 2025 revenue was $13.07 billion, with an operating loss of $20.92 billion. After accounting for a one-time non-cash provision related to the architecture restructuring, the net loss was nearly $39 billion. Even if you exclude that non-recurring accounting phantom loss, the real cash operating bleeding is still a bottomless pit. In the first quarter of 2026, cash burn was $3.7 billion—more than half of the revenue in the same period.

So don’t rush to celebrate “OpenAI is stable.” This isn’t a technical hype story—it’s just a check written on the back of a $20.9 billion operating loss. What OpenAI bought has nothing to do with the future; it’s just a page of a story meant for the next round of investors.

Talent is leaving, stars are filling the gaps

Noam’s resume really is dazzling. In 2017, he was the core author of “Attention Is All You Need,” and a key contributor to Transformer, MoE, and T5. In 2021, he left Google to found Character.AI. In 2024, Google brought him back with a $2.7 billion technology licensing agreement, appointing him co-head of Gemini. Less than two years later, he left again.

With $2.7 billion, Google proved one thing: money can buy a person’s time, but it can’t buy the soil that makes them stay. Now OpenAI plans to try again—this time with equity.

But the soil at OpenAI isn’t any better for pure research than Google’s. Over the past three years, the company has been running a talent swap: co-founder Karpathy left, Ilya Sutskever left, John Schulman left, and Jan Leike, head of the super-alignment team, left as well. The core founding team has largely departed, leaving very little behind in the core decision-making layer.

According to publicly available industry data, in 2021 research positions accounted for 23% of OpenAI’s total hiring; by 2024, that figure had dropped to 4.4%. Internal researchers’ assessments were blunt: the team’s focus has fully shifted from “exploratory research” to “product iteration.” The switching of a few people is only the surface. Research soil is being squeezed inch by inch by product KPIs—and the truth can no longer be hidden.

What he has to face isn’t a lab starting from scratch. The kind of system that even can’t accommodate Karpathy—who could only do “personal projects”—is exactly the mess Noam is inheriting. He’s just filling the hole left after the Karpathys left.

Buy people at a sky-high price—can’t solve the accounting predicament

Everyone is discussing what new architecture Noam can bring to OpenAI. But OpenAI’s current predicament has nothing to do with “missing someone who can write Transformers.”

The financials are there to see. In 2025 alone, R&D expenses paid Microsoft $10.59 billion for compute leasing. Full-year R&D costs were $19.18 billion, inference compute costs were $7.5 billion, and sales and marketing spend was $5.73 billion. On the other side are 900 million weekly active users, and only 50 million paying users. Massive free traffic is nothing but a cost bottomless pit—the bigger the user base, the heavier the compute bills.

Even OpenAI itself is trying to save money: leaked documents show that to control costs, it has reduced the Sora video model and cut non-core businesses. One side is shutting down businesses to save money; the other side is spending sky-high amounts to buy talent. That’s an anxious procurement.

The anxiety isn’t only OpenAI’s; the industry’s direction has already changed.

Microsoft’s own Copilot Cowork has already abandoned its unlimited pricing model due to high costs and switched to pay-as-you-go. Reports also say it is considering integrating DeepSeek V4 as a lower-cost option. Even Microsoft’s GitHub is seeking support from AWS because of AI compute shortages. When the main player’s own compute pool isn’t enough, who will foot the bill for OpenAI’s next sky-high compute cost?

Cook has publicly issued a warning: AI hype has driven storage chip prices up by four times since 2024, and the upward trend is expected to continue through 2027. The next-generation iPhone price may therefore rise by $270. These numbers are more like industry barometers; the real hard accounting is inside OpenAI. Compute hardware costs are still climbing, and efficiency improvements from model architecture optimization can’t keep up with the speed of hardware price increases. Noam can design a more efficient model structure, but he can’t fix the CFO’s nightmare. For every additional free user, there’s another rigid compute bill.

If the technical side can’t provide the answer, only the capital side can.

OpenAI is in a critical window preparing for its IPO, with valuation talks reaching as high as $1 trillion. Underwriters need a story with enough impact to support the valuation, and “the father of the Transformer leading R&D” is exactly the kind of narrative the capital market loves. The accounting for this investment—how much model performance improves—is not the focus; what matters is whether the roadshow PPT can add another page of highlights.

Retail investors see “the father of the Transformer joins” and get fired up, thinking the technical moat has just thickened. But institutional investors, looking at a $20.92 billion operating loss in 2025, will only ask three most practical questions:

Can this person make free users suddenly willing to pay?

Can he get Microsoft to discount the $10.59 billion compute bill?

Can he bring the departed research backbone back to turn around?

If none of that works, then his value is only a footnote in the valuation story. The biggest winners are always early shareholders and underwriters. Add one more story about “a genius at the helm,” and the risk of a $20.9 billion operating loss can be transferred—courteously—to the secondary market buyers.

The outcome of the two routes is already starting to emerge

What is Anthropic doing at the same time?

In enterprise adoption statistics reported by multiple organizations, Anthropic’s share has climbed to the 35% to 40% range, significantly narrowing the lead versus OpenAI, and even surpassing it in some samples. More importantly, its customer mix is such that about 80% of revenue comes from enterprise customers. Among Fortune 100 companies, many have already put Claude on their procurement list. The company reportedly secured its first profitable quarter in its history, and has reportedly secretly submitted an IPO application. Anthropic opened an office in Seoul, handling NAVER and Nexon orders—earning cash flow.

Anthropic didn’t dig up the “father of the Transformer,” and it didn’t rely on a single genius to prop up its valuation. It relies on an enterprise-appropriate, neutral positioning; controllable Token costs; and a deep binding of Claude Code to development scenarios. What enterprise customers want has little to do with whether the parameters are high or low—they want a “policy that won’t cause trouble” and a bill that can be calculated.

Google DeepMind’s Hassabis hasn’t moved in ten years, and Google hasn’t spent $2.7 billion to “redeem” him back from the outside. The real soil for innovation can’t be bought by paying billions to purchase stars every year. Only by growing talent internally and retaining research capability can the soil remain alive.

Anthropic proves one thing: profitability has little relationship with individual genius; the foundation lies in business soil.

OpenAI also proves one thing: individual genius alone can’t cover up business-soil stagnation—especially when you’re buying people at sky-high prices while watching the people you trained leave.

Conclusion

The next-generation architecture outcomes led by Noam will take at least 1 to 2 years to deliver. Based on the burn pace of $3.7 billion in a single quarter in 2026 Q1, annual cash consumption won’t be less than $14.8 billion. This doesn’t include the inevitable increase in marketing and compliance expenses before the IPO.

Google spent $2.7 billion to rent one person for two years. Now OpenAI wants to rent again with equity. The difference isn’t the payment method. Back then, Google at least had a profit-and-loss statement to back it up; for OpenAI, this money is written on the back of a $20.9 billion operating loss, discounted by IPO valuation bubbles.

The ledger doesn’t care about geniuses. The ledger only cares: how many more quarters do you still have to wait?

LAYOUT REFERENCE (source): total_lines=75, non_empty_lines=38, blank_lines=37