The Annual AI Slowdown Panic is Here

May 27, 2026 · Episode Links & Takeaways

Author

AI Daily Brief
May 27, 2026

Take our AIDB for Teams survey and let us know how we can help you share the podcast within your organizaiton.

HEADLINES

DeepSWE: The Coding Benchmark That Finally Matches Reality

A new coding benchmark called DeepSWE is getting rave reviews from developers who say it's the first one that actually lines up with how these models feel to use in the real world. Built by DataCurve, it addresses the core problems with SWE-Bench — memorization and trivial tasks — by constructing problems from scratch that require real-world workflows like parsing repos, multi-file edits, and long-context reasoning. On the initial run, GPT-5.5 was head and shoulders above the field at 70%, with GPT-5.4 and Opus 4.7 in a close race for second at 56% and 54% respectively. Chinese models look significantly further behind than other benchmarks suggest — Kimi K2.6 tops that group at just 24%, with DeepSeek V4 at 8%.

AI Leaders Are Walking Back the Jobs Apocalypse

A portion of AI leaders are finally changing their tune on the jobs apocalypse — whether because of a genuine update on what's likely to happen, or a belated recognition of how terrible the messaging has been. This week Sam Altman reinforced his new talking points, telling an interviewer he no longer believes AI will produce the kind of jobs apocalypse "some of the companies in our space advocate or talk about," and admitting his intuitions about entry-level white-collar job displacement were simply off. Last week, Goldman Sachs CEO David Solomon published a New York Times op-ed making a similar case — noting that AI has displaced 16% of entry-level tasks at his own firm, but arguing that, like past technological revolutions, AI will ultimately create more jobs than it destroys by enabling better products rather than just cheaper ones.

The Inference Layer's Funding Surge

The AI industry's shift from training to inference is now showing up clearly in startup valuations. Baseten is closing in on a $1B fundraising round at an $11B valuation — more than double its value from just three months ago — after seeing annualized revenue triple from $200M to $600M in Q1 alone. OpenRouter also became a unicorn this week, closing a $113M Series B at a $1.3B valuation led by Google's CapitalG, with Menlo Ventures reporting the company now processes 100 trillion tokens per month — a 5x jump in six months.

MAIN STORY

The Annual AI Slowdown Panic Is Here (A Little Early)

Every summer like clockwork, a slowdown panic takes hold — the shape changes each year, but the pattern is consistent. In 2023 it was ChatGPT's first down month. In 2024, the pre-training wall. In 2025, that MIT study claiming 95% of AI projects fail, arriving just as GPT-5 underwhelmed. Each time, the panic fades. This year's version is arriving early, and its shape is becoming clear: token budgets are running out, the AI subsidy era is ending, and the professional critics are weaving that into a full-blown bubble narrative. It's worth understanding both what's real and what's being overblown.

AI’S OUT FOR SUMMER

The Token Crunch Is Real — And Actually Healthy
The golden age of free experimentation is over; here's why that's okay.
The first half of this year was extraordinary — agents became real, agentic revenue exploded, and OpenAI and Anthropic hit $30B and $45B run rates respectively. But it was built on a subsidy model that was never going to last. Companies like Uber burned through their full-year token budgets in four months. The US government is now rationing access to the most powerful models. What's happening isn't the popping of a bubble — it's the market finding sustainable pricing. A world where companies continue subsidizing usage is actually more bubble-prone than one where people pay real prices.

The Uber Story and the VS Code Chart
Two data points being wildly overhyped by AI skeptics.
Uber's COO said this week that heavy token spending didn't correlate with more useful consumer features shipped — and the professional AI critics had a field day drawing a straight line to catastrophic bubble failure. Separately, a viral chart of VS Code AI coding assistant installs showed a plateau over recent months. But developer Simon Willison pushed back: the plateau likely reflects a platform shift away from VS Code entirely, toward CLI tools like Codex, which have surged from 100,000 NPM installs per day in January to over a million per day now.

The Macro Picture Doesn't Support a Bubble
Supply is tripling; demand is growing 10x. That's not a bubble popping.
Research firm Epoch AI put numbers around the token supply/demand gap: global inference capacity is more than tripling each year, but demand for tokens is growing roughly 10x per year. GPU rental prices are still up 2x from four months ago. CNBC's Deirdre Bosa raised the concern that companies switching to cheaper models could threaten OpenAI and Anthropic's pricing power — a real question worth watching — but the underlying demand signal doesn't look like a market that's losing faith in AI's value.

What's Actually Worth Worrying About
The end of free experimentation has real costs.
The loss of cheap, open-ended agent experimentation is a genuine problem — not because it signals a bubble, but because experimentation is how we figure out what agents are actually good for. The implication of agents isn't just doing existing work faster; it's doing entirely new types of things. That requires non-technical people being able to play freely. There's also a real AI inequality risk: as the gap widens between what well-resourced organizations can access versus everyone else, the compounding advantages could be significant.

Cursor's Composer 2.5 and Gemma 4
Market innovation in response to the token crunch.
One response to token scarcity is model efficiency innovation. Cursor's new Composer 2.5 has jumped to third place on Artificial Analysis' Coding Agent Index — behind only Opus 4.7 Max and GPT-5.5 — while costing 10 to 60 times less than those models. Meanwhile, Google's small cheap model Gemma 4 is seeing adoption that outpaces Chinese models like Qwen 3.5 and 3.6, a dynamic Latent Space's Swyx flagged as underreported.

New Dynamics, New Challenges, No Slowdown in Sight
For those paying close attention, these slowdown panic periods are actually an opportunity. While everyone else opts out for a couple of months hoping this whole thing finally goes away, the window is wide open to get ahead. The token crunch is real, the subsidy era is over, and there are genuine conversations worth having — about agent debt, about thoughtful adoption, about what agentic work actually looks like at sustainable prices. That's not a bubble popping. It's a market maturing. And as always, we'll be tracking it here.