May 2026 newsletter
The AI industry’s quiet admission that its core technology is getting more expensive, not less, is this month’s most revealing headline, buried under the usual buzz. A newsletter mentions “AI got expensive” in passing, but this isn’t a minor trend; it’s the fundamental economic pivot the entire ecosystem has been dreading. We’ve spent years celebrating scaling laws, assuming that throwing more compute at ever-larger models was the only path to better performance. Now, the bill is arriving. The i
Analysis
The AI industry’s quiet admission that its core technology is getting more expensive, not less, is this month’s most revealing headline, buried under the usual buzz. A newsletter mentions “AI got expensive” in passing, but this isn’t a minor trend; it’s the fundamental economic pivot the entire ecosystem has been dreading. We’ve spent years celebrating scaling laws, assuming that throwing more compute at ever-larger models was the only path to better performance. Now, the bill is arriving. The implication is brutal: the era of endlessly subsidized, loss-leading AI APIs is ending. Companies that built their entire value proposition on cheap, magical intelligence are facing a reckoning. Are we building indispensable tools, or just costly conveniences?
Into this sobering landscape steps Anthropic, reportedly having a “really good month.” But what does “good” even mean in this context? If it means securing another round of funding at a preposterous valuation, that’s not a measure of product success but of investor faith in a long-term bet. If it means technical breakthroughs, where are the earth-shattering demos that change everything? The newsletter hints the recent model releases were “disappointing,” which is the real story. We’ve hit a wall. The incremental improvements in the latest GPT-4-class models don’t justify the astronomical costs. Anthropic’s “good month” might simply be better marketing of a marginally better product in a market suddenly terrified of its own pricing. Being the “less irresponsible” option in a field of profligate spenders isn’t the same as winning.
This stagnation is where the real tension lies. The parade of conferences and podcast appearances feels increasingly like a distraction. It’s the industry talking to itself, reinforcing its own importance while the core promise—that the next model will be qualitatively different—rings hollow. Meanwhile, on the ground, the interesting work is happening away from the billion-parameter arms race. The launch of Datasette Agent and progress on Datasette itself point toward a different future: one of curated data, precise tools, and practical utility. This is where intelligence becomes actionable, not just impressive in a sandbox demo. It’s the contrast between a general-purpose oracle that costs a fortune to query and a specialized system that can actually perform a task reliably. One is a science experiment; the other is a tool.
The split is becoming clearer. On one side, you have the cloud AI oligopoly, where costs are spiraling and model improvements are feeling diminishing returns. Their business model depends on you being perpetually awed and dependent. On the other, you have the emerging ecosystem of open-source tooling and agent frameworks. This side isn’t trying to sell you “AGI in a box.” It’s trying to help you connect to your own data, automate your own workflows, and—crucially—control your own costs. The newsletter’s focus on its author’s own practical projects is a quiet testament to this shift. The most engaged people in this space are moving from prompting a monolith to building with components.
So, where does this leave Anthropic and its peers? Their “good month” is a stay of execution, not a victory. They are caught between the demands of their massive investors for a path to profitability and the reality that their core technology is becoming a commodity whose price is going up, not down. Their moat is not in being the smartest—OpenAI will always have a press release for that—but in being the most trusted or the most integrated. That’s a defensive play, not an offensive one. The real innovation is migrating from the model labs to the application layer, where people are grappling with the expensive reality of these tools and building something useful anyway.
The future isn’t about who has the biggest model. It’s about who can make intelligence cheap, reliable, and specific enough to matter. Right now, the big labs are losing that race by their own design. They built a product that’s too expensive to scale and too unreliable to replace human judgment entirely. The next chapter will be written by the toolmakers, the data engineers, and the pragmatists who treat AI not as a magic answer, but as a powerful, costly, and sometimes frustrating component in a larger machine. The hype cycle is over. The utility cycle has just begun.
Disclaimer: The above content is generated by AI and is for reference only.