Anthropic Just Reset AI Expectations

May 21, 2026 · Episode Links & Takeaways

Author

AI Daily Brief
May 21, 2026

HEADLINES

OpenAI Races to File for IPO

The Wall Street Journal reports that OpenAI has engaged investment bankers and expects to file IPO paperwork as soon as this Friday — confidentially, meaning full financial disclosure won't come until much later. Sources say OpenAI has set a goal of being ready to IPO by September. The resolution of Elon Musk's lawsuit cleared the way, removing the unlikely but real risk that it could have unwound OpenAI's for-profit conversion. Up until now, Anthropic had been expected to go public first, but with Anthropic said to be pulling together one last private round, OpenAI may beat them to the punch — and there's a real question about whether the sequence matters if you think three $1–2 trillion IPOs could stretch public market liquidity.

Trump's AI Executive Order Takes Shape

Major labs have been briefed on a new AI executive order that could arrive by end of week. The White House's Office of the National Cyber Director held a session Tuesday with CEOs from OpenAI, Anthropic, and Google in attendance. The order is aimed at establishing a voluntary framework for pre-release disclosure and testing of frontier models — with the White House floating a 90-day advance notice period and the labs pushing for just 14 days. A 90-day window would be a meaningful constraint; it's been less than two months since the first Mythos rumors surfaced, and that already feels like an eternity. The executive order is also expected to direct the Pentagon to harden critical systems within 30 days and instruct Treasury to establish an AI clearinghouse pairing frontier labs with critical industries. The whole framework feels a lot like Anthropic's Project Glasswing rollout of Mythos becoming formal government policy.

OpenAI Guaranteed Capacity: AI Billing Goes Cloud-Style

OpenAI launched a new program allowing enterprise customers to make one-to-three-year capacity commitments in exchange for discounts and certainty — making AI billing look a lot more like cloud infrastructure than SaaS. This is a direct response to the compute crunch: Uber's CTO made headlines in April when they burned their entire annual token budget in four months, and Box's Aaron Levie says token strategies are now a hot topic for CIOs and CFOs. As a side benefit, locked-in multi-year contracts give OpenAI rock-solid ARR numbers heading into an IPO.

Cursor's Composer 2.5 Makes a Dent on Coding Benchmarks

Cursor's new Composer 2.5 model has jumped to third place on Artificial Analysis's coding agent index — behind Opus 4.7 on max and GPT-5.5 on extra-high, but ahead of both on medium settings, and at 10–60x lower cost. As token efficiency becomes one of the biggest constraints for enterprise deployments, this positions Composer as a serious contender.

OpenAI Offers $2M in Tokens to Every YC Startup

OpenAI offered $2 million in tokens to every startup in the current Y Combinator batch in exchange for equity — a move that's less like free AWS credits and more like headcount cash at this point. Tyler Bosmeny called it a "mic drop moment." Altman framed it as a bet on token-maxing as a new mode of building.

MAIN STORY

Anthropic Just Reset AI Expectations

A single week produced a hire that rattled the entire AI industry and financials that made AI skeptics look for cover. Taken together, Karpathy joining Anthropic's pre-training team and the company's first-ever profitable quarter aren't just milestones — for a growing number of observers, they're signals that we may be entering a new phase of the race entirely, one where recursive self-improvement isn't a distant theoretical and Anthropic is pulling ahead of the field.

ANTHROPIC ARRIVES

Andrej Karpathy Joins Anthropic
"The next few years at the frontier of LLMs will be especially formative."
Karpathy — OpenAI co-founder, former Tesla Vision director, coiner of "vibe coding," and one of the most respected AI researchers alive — announced Tuesday he's joining Anthropic's pre-training team. The move caught people off guard not just because of who he is, but because of where he went. Anthropic's Nicholas Joseph confirmed he'll be building a team focused on using Claude to accelerate pre-training research itself — putting him squarely at the center of the recursive self-improvement question. The tea leaves had been there: on Sarah Guo's podcast in March, Karpathy had said that anyone working outside a frontier lab would inevitably start to drift from the frontier, and Noam Brown had publicly asked why he wasn't at one. Now he is — just not at OpenAI.

Business Insider Andrej Karpathy's move to Anthropic is a major win for the AI giant in the talent wars
CNBC Anthropic hires OpenAI co-founder Andrej Karpathy, former Tesla AI leader
Axios OpenAI co-founder Andrej Karpathy joins Anthropic
Andrej Karpathy (X) Personal announcement post
Noam Brown (X) March question: why is he not at a frontier lab?
Milk Road AI (X) Karpathy on Sarah Guo's show: judgment drifts outside the labs
Noam Brown (X) Happy he's at any frontier lab pushing the field forward

The RSI Reading
For many, this hire signals recursive self-improvement is close.
The more sophisticated take on the Karpathy hire isn't about talent wars — it's about what it signals for where the frontier is heading. Karpathy will be using Claude to help accelerate pre-training research itself, which is essentially a bet that the recursive self-improvement loop is real and close. Freda Duan at Altimeter laid out the argument clearly: we're used to SOTA rotating between labs every few months, but RSI — where each better model helps build the next one — could break that pattern entirely. TMT Long Short and Shanu Matthew both read the hire the same way: this is a move you make when you think you're on the final sprint to the endgame. Swyx summarized the feeling more tersely.

Anthropic's First Profitable Quarter
The first profitable quarter for any foundation AI lab — years ahead of schedule.
Financials shared with investors in Anthropic's current fundraising round show the company is forecasting $10.9 billion in revenue for Q2, up from $4.8 billion last quarter, with a small operating profit of $559 million. That's an annualized rate of $44 billion and — critically — it would be the first profitable quarter for any foundation model lab. Coming into this year, Anthropic had projected it wouldn't be profitable until 2029; OpenAI wasn't expecting it until 2030. There are caveats: Anthropic counts gross revenue before partner revenue shares, which potentially inflates the headline number, and part of why they're profitable is that compute is sold out and they can't spend more even if they wanted to. But as Derek Thompson put it, a profitable quarter at a $44B annual run rate while rationing compute suggests revenue could hit $100B or more with sufficient inference capacity.

WSJ Mind-Blowing Growth Is About to Propel Anthropic Into Its First Profitable Quarter
The Information Anthropic Projects Turning an Operating Profit in Second Quarter
CNBC Anthropic set to hit $10.9 billion in revenue during second quarter, source says
Derek Thompson (X) On what the numbers imply with unconstrained compute
Mr. Ratable (X) On the absurdity of Anthropic's revenue relative to legacy SaaS
Gavin Baker (X) On why today's infrastructure build differs from the 2000 telco overbuild

NVIDIA Earnings: Another Record Quarter
"The largest infrastructure expansion in human history is accelerating at extraordinary speed."
NVIDIA's Wednesday earnings reinforced the demand signal, with revenue hitting $81.6 billion against estimates of $78.9 billion and data center revenue growing 92% year-over-year. For the first time, NVIDIA broke out data center revenue between hyperscalers (46% of the total, growing 12% quarter-over-quarter) and other customers — and said it's gaining share among the hyperscalers, putting to bed the idea that Google's TPUs are making a dent. Huang also acknowledged NVIDIA has effectively exited the Chinese market, ceding it to Huawei. The stock fell 3% after hours anyway, which analyst Patrick Moorhead attributed to the difficulty of pricing a $5 trillion company even on exceptional results. And Huang couldn't help but call out Anthropic specifically as a driver of growth.

Anthropic and SpaceX: The Colossus 2 Deal
A $45 billion compute commitment that makes Colossus SpaceX's biggest revenue source.
Anthropic Chief Compute Officer Tom Brown announced Wednesday that Anthropic is deepening its SpaceX partnership, scaling up GB200 capacity in Colossus 2 throughout June. Colossus 2 is substantially larger than the original Colossus — which is now fully committed to running Claude. The SpaceX IPO prospectus then revealed the price tag: $45 billion over three years, or roughly $1.25 billion a month. That instantly makes the Anthropic deal SpaceX's largest revenue source, surpassing Starlink's $11 billion in 2025 and adding 80% to SpaceX's revenue based on last year's figures. Musk's post made clear he's settled into his new role as compute czar — noting SpaceX is in discussions with other companies to offer the same service at scale. The strategic picture this creates for OpenAI was summarized neatly by Signull: Anthropic is now a profitable competitor with potentially infinite compute, and Elon can't walk away from the deal without a fiduciary crisis.