AI's New Acceleration Phase

May 22, 2026 · Episode Links & Takeaways

Author

AI Daily Brief
May 22, 2026

AI TAKES OFF

A Week of Surprising AI Acceleration

Some weeks the big stories represent such an obvious change that it barely needs to be pointed out how much has shifted. This week was instead a week of surprising AI acceleration — where the individual stories add up to a whole that's much more than the sum of the parts, and where you can feel the acceleration almost more than intellectually recognize it. The frame: model development, policy, business model, consumer surfaces, and more — all accelerating at once.

Profitability Acceleration
The first profitable quarter for any foundation AI lab.
Anthropic is projecting its first-ever profitable quarter — and the first for any foundation model lab. There are caveats: the quarter isn't done, Anthropic counts gross revenue before partner distributions, and a compute discount from SpaceX helps the margin. But for most honest observers, those are quibbles. The old bubble narrative had shifted from "AI isn't useful" to "these labs can never serve tokens profitably." That narrative is now significantly harder to sustain. OpenAI also had a strong Q1 — generating about a billion dollars more in revenue than Anthropic, with token-hungry Codex a major driver, though Anthropic's revenue acceleration has since outpaced them.

Market Narrative Acceleration
NVIDIA beats again — and investors don't know how to deal with it.
NVIDIA delivered another record quarter, beating essentially every analyst estimate. As Patrick Moorhead put it, the challenge is that at a $5 trillion-plus market cap, investors don't know how to deal with it — if you take Jensen's forward demand pipeline at face value, this is an $8-9 trillion company, and people are afraid to move even on exceptional earnings. Even so, the broader market is increasingly gearing up for full-send mode on AI.

Token Shortage / End of the Subsidy Era
The flat-rate AI experiment is effectively over.
The shift away from one pricing paradigm to another — what many have called the end of the subsidy era — accelerated this week on multiple fronts. Google's "price cut" on its Ultra plan from $250 to $200 came bundled with a shift to usage-based billing for token-hungry use cases, not unlike the Anthropic change that caused so much discussion the week before. Microsoft canceling its Claude Code licenses added to the picture, with cost a real factor alongside the push to shore up GitHub Copilot. As Hedgie Markets put it, token-based pricing is forcing every enterprise customer to confront the actual cost of running these models at scale — and the number turns out to be far higher than flat-rate experiments suggested. On the other side of the equation, Cursor's Composer 2.5 emerged as one answer: according to Artificial Analysis, it performs comparably to Opus 4.7 and GPT-5.5 at 10 to 60x lower cost.

Compute Acceleration
Elon Musk fully settling into his role as AI compute czar.
Elon tweeted that SpaceX is now offering AI compute as a service at significant scale, is in discussions with other companies to do the same, and expects to eventually serve AI at "extremely high scale" via orbital data centers. Anthropic's Chief Compute Officer Tom Brown confirmed the company is expanding its SpaceX partnership to Colossus 2, not just Colossus 1. This changes the SpaceX IPO story considerably — rather than a referendum on Elon's various projects, SpaceX is positioning itself as an alternative NeoCloud with a unique orbital compute possibility that other providers simply can't match.

Tom Brown (X) Anthropic rents Colossus 2

Consumer Surface Acceleration
900 million monthly users — and AI is now baked into every Google surface.
The Gemini app has effectively closed the gap with ChatGPT, reaching 900 million monthly active users. Token consumption tells an even bigger story: overall tokens processed jumped 700% year-over-year, from 480 trillion last May to 3.2 quadrillion this year. Beyond the numbers, Google I/O introduced persistent AI agents directly into Search — users can set up agents to monitor ongoing queries (apartment listings, stock updates, sports scores) rather than searching one time. This turns Google Search from a one-time lookup into a persistent information-gathering layer, and may be one of the most significant things Google announced at I/O. Google is also taking the same principle into Docs, with Docs Live allowing voice-first document creation — an early signal of a broader voice-first, live interaction pattern that's likely coming to many different surfaces.

Andrew Curran (X) Gemini MAU topped 900 million
Google (X) Google Search goes agentic
Allie K Miller (X) Google Docs Live

Model Capability Acceleration
An 80-year-old math problem, solved with a general-purpose LLM and a simple prompt.
OpenAI announced that an internal model disproved a famous 1946 conjecture by Paul Erdős: how many pairs of points can be exactly one unit apart when placing N points on a plane? The prevailing answer had been the square grid arrangement. The model disproved this, using multidimensional mathematics flattened into a 2D plane to produce more pairs. Fields medalist Tim Gowers, who wrote the companion paper, called it "the first really clear example of AI solving not just an unsolved math problem, but a really well-known unsolved math problem." What makes it striking: nothing about this model was mathematically special — Noam Brown confirmed it was a general-purpose LLM with no specific training on this problem or mathematics — and the prompt was largely just a clear statement of the problem. OpenAI's Alexander Wei captured the feeling: "10 months ago, I was ecstatic that AI could win International Math Olympiad gold. Today, that excitement feels quaint." He added that math is a leading indicator — soon AI will begin autonomously producing landmark results in CS, physics, econ, and biology. One small fun footnote on the energy cost: Ethan Mollick calculated the Erdős solution took between 0.6 and 6.3 kilowatts of electricity and 3 to 31 liters of water — less than three almonds' worth of water.

Model Development Acceleration
"The next few years at the frontier of LLMs will be especially formative."
The biggest model development signal of the week was Andrej Karpathy — OpenAI co-founder, former Tesla AI lead, coiner of "vibe coding" — announcing he's joining Anthropic's pre-training team. What he's working on matters as much as the move itself: Anthropic's Nicholas Joseph confirmed Karpathy will be building a team focused on using Claude to accelerate pre-training research itself — what many call recursive self-improvement, or RSI. For anyone following Karpathy's auto research experiments, this is a natural continuation. The combination of who he is, where he went, and what he'll work on contributes to the overall sense of acceleration in a significant way.

Data Center Counter-Narrative Acceleration
The critics are being fact-checked — and the facts favor data centers.
Opposition to data centers is growing as a political force, but this week the counter-narrative picked up real momentum. A widely shared chart showed annual data center water use is less than a fifth of golf courses, about a tenth of almonds, and about a twentieth of lawns — which the Erdős energy discussion only amplified. Local media is also starting to tell a different story: a CBS affiliate in Richmond, Virginia profiled an electrician named Josh Price, who said AI infrastructure has been a huge win for tradespeople who previously struggled to find steady work. The conversation about data centers will be better when it's based on real facts and evidence, not fear and a generalized distrust of big tech.

Hedgeye (X) Almond water use vs Data centers
Nathan Learner (X) Data Centers are bringing solid blue collar jobs to rural areas

Policy Acceleration — California
A preview of what a 2028 Democratic presidential campaign sounds like on AI.
California Governor Gavin Newsom signed a first-of-its-kind executive order directing state agencies, academics, labor groups, and AI companies to develop policies around AI labor disruption — covering severance, employment insurance, retraining, and worker ownership models. The order is largely exploratory, prompting practical skepticism from Ramp's lead economist about whether unemployment insurance data can even detect AI-related layoffs. You can take it two ways: non-cynically, this is the governor of one of the most important states in the country tackling AI disruption head-on for the first time. Cynically, Newsom is definitely running for president in 2028, and this is a test run of the language and policy a major candidate will use on AI.

Policy Acceleration — Washington
David Sacks called the president, and an AI executive order died hours before signing.
A signing ceremony was prepared, invitations went to tech CEOs, a draft circulated to the press — and then, hours before it was set to happen, Trump pulled the order. He told reporters he didn't like certain aspects of it and thought it "gets in the way." Later reporting from Politico revealed that former AI czar David Sacks personally called the president that morning — unbeknownst even to White House staff — and argued that pre-release model review would slow innovation and hurt the US against China. Musk and Zuckerberg also reportedly called. Axios added that one source said Trump "just hates regulation" and that the whole thing was "just something doomers wanted." Trump's stated reasoning was consistently about China: "We're leading China, we're leading everybody, and I don't want to do anything that gets in the way of that lead." The fabled AI executive order will probably return in some form, but who knows what it'll say by the time it lands.

Closing: The Foothills of the Singularity
"When we look back at this time, I think we will realize we were standing in the foothills of the singularity."
One refreshing thing about Google I/O was that Demis Hassabis refused to participate in the AI doom cycle. In his closing keynote, he framed this moment not as an ending but as a beginning — arguing that this technology will be a force multiplier for human ingenuity and usher in a new golden age of scientific discovery. In many ways, that's the right note to end a week like this one on. It's a great message. It's our job to make sure it's true.