The Week AI Grew Up

May 1, 2026 · Episode Links & Takeaways

Author

AI Daily Brief
May 01, 2026

Sign up for AgentOS, our latest free training course. It’s a platform and model agnostic program to help you build a personal agentic operating system that can evolve with you over time. The program is self-directed and project based, running over four weeks.

THEME OF THE WEEK

The Week AI Grew Up

This week marked a clear phase shift in the AI era. From business model overhauls to market reactions to new product paradigms, the through-line was unmistakable: AI is no longer a startup-era curiosity — it is critical global economic infrastructure, and everyone is being forced to grow up alongside it.

One favor to ask, this episode is a little more geared towards people who don’t listen daily but still want to catch up on the big stories of the week. If you could send a link to this episode or forward the newsletter to a friend, that would be a huge help.

GROWTH SPURT

The Demand Crunch and the End of the Subsidy Era

The first dimension of AI's growing up is a recognition of the demand crunch and a consequent shift in business models. GPU rental prices are up 40% over the last six months, the top two AI labs now generate close to $60 billion in aggregate annual revenue, and Dylan Patel of SemiAnalysis told Patrick O'Shaughnessy this week that even tier-two and tier-three labs are going to be sold out of tokens. Andy Jassy said Trainium is sold out. OpenAI CFO Sarah Friar called it a "vertical wall of demand," with compute as the bottleneck. In a world of agents and seemingly infinitely replicable intelligence, every token that can be produced will be sold.
Oguz Erkan (X) "What AI bubble?? GPU rental prices are up 40%"
Seeking Alpha Nvidia's H100 GPU rental prices surge nearly 40% in 6 months
Patrick O'Shaughnessy (X) Dylan Patel interview: "everyone is sold out"
Wall St Engine (X) Jassy: "Trainium is sold out"
Bloomberg OpenAI CFO Sees 'Vertical Wall of Demand' for Products

GitHub Copilot moves to usage-based billing
"The all-you-can-eat AI buffet is coming to an end."

Chief product officer Mario Rodriguez said a quick chat question and a multi-hour autonomous coding session can cost the user the same amount, and the current premium request model is no longer sustainable. On Microsoft's earnings call, Satya Nadella said any per-user business — productivity, coding, security — will become a per-user and usage business.

Anthropic resists the shift
Doing everything they can not to bite the usage-based bullet.

Over in Claude land, it seems like they're doing almost everything they can to avoid switching to a purely usage-based model, and that's clearly driving decisions around how third-party products like OpenClaw use their models.

AI Hits the Bottom Line

The growing up showed up in markets too. AWS was up 28% year over year — its best performance since climbing out of a trough in 2021. Azure was up 40%. Google Cloud absolutely spanked analyst estimates, growing 63% year over year, driving the second-biggest one-day jump in market cap in history. As the AI subsidy era ends, Google looks well-positioned: of the major labs, they have the best and most mature set of cheaper models that companies can turn to as they bring capital discipline to their token allocations.

Anthropic raising at $900B+
Stock already trading higher than OpenAI on secondary markets.

Bloomberg reported Wednesday that Anthropic has begun talks to raise at over $900 billion, putting them past OpenAI's last valuation of $825 billion. By Thursday, TechCrunch had the scoop: investors have just 48 hours to submit allocation requests, with Anthropic expecting to raise $50 billion. Some shares have reportedly traded at a trillion-dollar valuation on secondary markets. The logic isn't about the exact right multiple on revenue — it's a belief that there's about a half-dozen companies writing the story of the future, and basically no way they're not more valuable later than today.

Microsoft and OpenAI redraw the deal
OpenAI grew too big for any single cloud to fully serve.

Microsoft got royalty-free access to OpenAI's models for another half decade plus removal of the weird AGI clause that could have shut off their access on a whim. OpenAI is now free to sell through AWS and Google Cloud as well. As Razzo put it: this is simply a factor of OpenAI having grown too big for any single cloud to fully serve.

AI as Strategic Reality

Another dimension of the phase shift: AI is increasingly being treated as critical infrastructure, and the training wheels on AI policy have come off. At the start of the week, Axios reported the White House was working to unwind Anthropic's supply chain risk designation and redeploy Anthropic models in government — including Mythos in agencies, with executive order discussions around safe deployment. By the end of the week, the story had flipped: the White House moved to oppose Anthropic's plan to expand Mythos preview to 70 additional companies, citing national security concerns and apparent worries about whether Anthropic has the compute to serve that many entities without hampering government access. Anthropic says compute isn't a constraint. The White House isn't buying it. As Dean Ball puts it, this is effectively a licensing regime — informal and improvised, but a licensing regime nonetheless.

The Industrialization of Agent Harnesses

The last dimension of growing up is the product layer. As agents have become a dominant way people get value from AI, there's been a broader recognition that the harnesses in which models operate are a significant area for improvement. The shift from OpenClaw at the end of January — where you had to build everything by hand and wire it together — to today's baked-in, built-together products feels analogous to the move from the hobbyist PC era to the Apple II Plus era of personal computing.

Codex for (almost) everything
OpenAI bets on one interface for everyone.

Sam Altman called it a big upgrade and told people to try Codex for non-coding computer work. The new version asks what type of work you do — finance, product, marketing, operations, sales, data science, design, student — and tailors itself accordingly. Unlike Anthropic, which split technical and non-technical work between Claude Code and Claude Cowork, Codex is betting on one interface for everyone. I like the bet that knowledge workers will strive to be more technical to unlock their newly discovered wizard powers, versus the Cowork bet that they need special neutered tools.

Cursor SDK
Embedding Cursor agents directly into developer applications.

The Cursor SDK launched alongside a deep blog post on how Cursor continually improves its agent harness — testing, monitoring, repairing degradations, and customizing for different models.

Why I'm Not Covering the "Permanent Underclass" Piece

One part of AI growing up I'm not discussing is Jasmine Sun's NYT opinion essay "Silicon Valley Is Bracing for a Permanent Underclass." The reporting is thorough. But I think we're going to recognize that the skillset required to build great technology is very different from what's implicit in understanding how that technology will interface with the world. Many AI builders are first in line to see the power of these tools, and since so much of what they do has been transformed, they extrapolate to everyone — but they tend to miss how the real world outside Silicon Valley functions. I think you'll see a fork in the narrative, where we rely less on Silicon Valley priors and more on first principles thinking and the judgment of economists.

What You Should Try Building This Weekend

If you're not using Codex yet — or downloaded it months ago and haven't tried it lately — now is the right time. The absolute best way in is Riley Brown's pinned tweet: "Learn 95% of Codex in 28 minutes." Also worth checking out is Cursor. Six months ago the narrative was against them, but as people have appreciated the importance of harnesses, more and more are investing in their Cursor setup for the flexibility to move between models. Lenny Rachitsky tweeted this week that he's finding it more fun to work in Cursor than in the native Codex or Claude Code apps.

GPT Goes Goblin Mode

The quirky story of the week is, undeniably, goblins. On Thursday, OpenAI published "Where the Goblins Came From," prompted by a viral tweet from arb8020 noting that the GPT-5.5 Codex prompt had a duplicated line trying to stop it from talking about creatures. Starting with GPT-5.1, models began developing a strange habit of mentioning goblins, gremlins, and other creatures in their metaphors. OpenAI's conclusion: goblin references were an artifact of the "nerdy" personality trained to make cute creature references, which spilled over into other RL training. The implications are real — when models are built on top of other models, weird quirks from RL in one can have multiplying effects in others, with obvious consequences for how we think about alignment and safety training.