The AI Daily Brief: Artificial Intelligence News and Analysis - The AI Token Shortage Begins [AI Monthly Recap]

Episode Date: June 1, 2026

One of the most consequential AI months of 2026, May marked a major shift from the AI subsidy era into a new period defined by token scarcity, usage-based pricing, enterprise sticker shock, and a broader scramble for compute. NLW argues that the next phase of AI competition will be shaped by who can access, afford, optimize, and deploy AI tokens most effectively.Sign up for AI Executive Catchup: ⁠https://aiexecutivecatchup.com/⁠Brought to you by:KPMG – Research from KPMG and the University of Texas at Austin shows the highest-impact AI users treat AI like a reasoning partner — and those skills can be taught at scale. Learn more at ⁠⁠⁠⁠kpmg.com/us/Sophisticated⁠⁠⁠⁠Outsystems - Stop wondering how AI will change your business and start building the agents that will lead it - http://outsystems.com/Scrunch - The AI customer experience platform - ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠https://scrunch.com/⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠Zenflow Work - Agents for knowledge work - ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠https://zenflow.free/⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠Blitzy - Want to accelerate enterprise software development velocity by 5x? ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠https://blitzy.com/⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠AssemblyAI - The best way to build Voice AI apps - ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠https://www.assemblyai.com/brief⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠Robots & Pencils - Cloud-native AI solutions that power results ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠https://robotsandpencils.com/⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠The AI Daily Brief helps you understand the most important news and discussions in AI. Subscribe to the podcast version of The AI Daily Brief wherever you listen: ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠https://pod.link/1680633614⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠Our Newsletter is BACK: ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠https://aidailybrief.beehiiv.com/⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠Interested in sponsoring the show? [e-mail]

Today on the AI Daily Brief, we're recapping the month of May, one of the single most consequential AI months we've had in a very, very long time. The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI. All right, friends, quick announcements before we dive in. First of all, thank you to today's sponsors, KPMG, robots and pencils, Zen Coder, and Out Systems. To get an ad-free version of the show, go to patreon.com slash AI Daily Brief, or you can subscribe on Apple Podcasts. If you want to learn more about sponsoring the show, send us a note at sponsors at AIdailybrief.aI. Today is the first day of June. And while I don't always use the first of the month to look back and reflect on the month that was,

in this case, I think it's pretty important. We are now experiencing the second big AI transitional moment of 2026. Although you could argue that the first actually began in the end of 2025, in the November and December period where ClaudeCode and Codex were on the rise, and we got the series of models including Opus 4.5 and GPT52, which of course all came together to unleash the true agent era at the beginning of 2026. At this point, you've heard me talk ad nauseum about the fact that everyone went home for the holidays, started hacking around on ClaudeCode or with these new models,

and discovered that what you could do had changed fundamentally. That led into the open claw period, where all of a sudden people were getting their hands messy with harnesses in a new way, and just an absolute explosion of new behavior around AI. Not only were software engineers actually using agentic coding tools in a mainstream way, not just vibe coding prototypes and things like that, but actually pushing agent-created code into production. But the people who had previously just been knowledge work-style vibe coders

using tools like Lovable and Replit were moving to a way more advanced period, building much more extensive and complex applications in harnesses like Clod Code and Codex, or even spinning up and building entire agents and agentic systems thanks to tools like OpenClaw and later Hermes, really signaling that the times had changed. Now, one of the consequences of this shifting behavior is that the most relevant economic unit for AI companies ceased to be the seat and instead shifted to the token. And what I mean by that is that revenue for OpenAI and Anthropic was no longer constrained to what percentage of their users they could convert into paid seats either on the consumer or on the enterprise side,

but instead how much API revenue they were getting through people actually using and consuming tokens. API-based usage looks very, very different from an economic standpoint than seat-based usage. To put just a little personal example on it, When I dropped the personal context portfolio builder at contextportfolio.aI, turns out a lot of you wanted to use it, so much that it racked up about a $5,000 bill over the first six weeks or so of it existing. Compare that $5,000 to the $200 a month clawed seat that I had been paying for forever. You're talking about more than two years worth of Claude Max seats in spend with a single six-week project. Now, this is, of course, where the massive explosion

in revenue came from for the foundation model companies this year. OpenAI searched a $30 billion in ARR, and Anthropic went even farther, even faster, getting as we recently learned, all the way up to $47 billion in annualized run rate as of right now. To go from $3 billion in revenue, which was where they were at the beginning of 2025, to $47 billion in annualized revenue a year later, is just staggering. And the realization of what this meant, kind of where this month started. This was perhaps best captured in an article in the Atlantic called So about that AI bubble. And while the author themselves wasn't apologizing for getting it wrong or anything like that, the article itself served as a bit of a mea culpa for the Q4 period in which the media's

obsession was the idea of AI as a bubble. Now remember in Q4, the argument had never been that AI wasn't valuable. It's that the ability for the foundation model labs, i.e. the token sellers, to realize that value, felt to many as unlikely to be able to keep up with the cost of this incredibly extensive AI infrastructure buildout in the form of all these compute deals and all the things that we heard about throughout the back half of 2025. That starts to look very different when you see the type of growth numbers and frankly the type of pure raw revenue numbers that companies like OpenAI and Anthropic were putting up. And again, because it was not based on seats, which have a natural and imaginable,

cap, and was instead based on tokens, where we were seeing these numbers despite it being the very, very beginnings of us scratching the surface of how much AI we could use. A lot of people, to their credit, readjusted their priors about the possibility of an AI bubble, and really calibrated up their expectations of just how big this could all get. And this part of the story has never gone away, and one of the big themes throughout May was, as summed up in a single headline from the New York Times, how Anthropic got so big, so fast. They closed the month with a $65 million fundraising round, valuing them just under a trillion dollars.

And there was also a competitive dynamic to this, with the month seeing Anthropic racing out ahead of competitor OpenAI when it came to business adoption, according to statistics from Ramp. And we even got what Anthropic anticipates to be not only their first profitable quarter, but the first profitable quarter for any of the Big Foundation model labs. Once again, the psychological impact of achieving profitability, with this type of growth rate, with this type of expenditure, really reset people's expectations. And yet, as we're coming out of the month, we are once again in the midst of a massive shift in understanding. You could sum this up with a recent Axios article, AI sticker shock hits

corporate America. I believe that the period we are heading into now is one that is fundamentally defined by constraints and the second half of this month, and really the meta-trajectory of this month was all about starting to realize what that meant and what it was going to look like. Going back a couple months previously, Uber made headlines in April when its CTO shared that the company had burned through its entire 2026 AI budget in just four months. Now, on the one hand, when I saw this article, it didn't seem all that surprising to me in the sense that you have to think that those token budgets were being figured out, probably not even in the November or December period where the Opus 45 level models had started to really be

brought to bear, but before that, and so of course they weren't expecting the type of usage that we were going to see once Agentic AI really came online. And yet at the same time, this became the capstone story for a lot of different things going on. Every day it felt like we were getting some new token maxing story where some company or another was creating some sort of internal leaderboard, incentivizing people to consume as many tokens as they possibly could. I did a couple of episodes about this whole token maxing idea, actually providing more of a defense for it, even though I think that a lot of people's first instinct, which is completely reasonable, is to point out the truism of Goodheart's law that once you start to measure something,

it ceases to be a good measure because people just start to game the measure rather than whatever it was intended to measure. And what's more in the context of token maxing, you're measuring an input rather than an output, when inherently and in the long run, outputs are all that's going to matter. Now, my argument was, of course, about the value of experimentation and fact, the necessity of experimentation in a period where no one knows the best way to use these tools, but it would be very clear, very quickly, that there would be consequences of this idea of token maxing that would rear their ugly head soon. Once again, it was Uber that helped shift the conversation when, in an interview with Uber's C-O-O this time, the C-O shared a lot of skepticism

about how much value they had actually gotten from that AI budget that they had burned through in just a few months. This got reinterpreted and reduced to headlines like the information's, uncharacteristically oversimplified, Uber-C-O-O says AI lacks ROI, and really has brought up this whole conversation once again embodied in this AI sticker shock piece from Axios. Now, my contention is that we're in a secular shift from one business model paradigm of AI to another. In short, we're moving from an AI subsidy era to a token scarcity era. So what are those terms mean and what are the implications? Well, first of all, let's talk about the idea of an AI subsidy era.

The idea of the subsidy era is that for some time, especially the max level subscriptions from the labs, the $100, $200, $300 a month type of subscriptions, while perhaps being profitable for some portion of users were very, very frequently, very unprofitable. We don't know for sure, but estimates around the actual value of the tokens that you could theoretically consume on one of those $200 a month plans could sometimes be 10 or 20 times that $200 value. In other words, the most active power users of those max plans were sometimes getting $2,000, $4,000, $5,000, even $10,000 of value out of just $200 a month. And that was the AI subsidy. Now, there was a lot that was really great about that. I basically haven't for a second at any point over the last six months. Pause to consider the financial implications of any dumb idea I wanted to try on Godex or Claude

I'd just start building it. I'd just start releasing it. we have been in a letter-rip kind of place. Now, that may work for me, as an independent content creator whose job it is basically to do that and then share what I learn with you all, but for companies that starts to get a little bit tricky. On both sides of the equation, for the provider companies, there's only so long they can subsidize that type of usage

to the tune of 10 or 20x, and if they cease to subsidize it, there's only so long that the companies that are now paying for it on a per-usage basis can actually afford to do so. And that shift in business model is the first big implication of the AI subsidy era ending. Over the course of the month, we had a number of different companies announced that they were shifting from a flat seat sort of model to more usage-based billing. One of the first of those was GitHub co-pilot, who actually made the announcement at the very end of April. In their announcement post, they wrote,

Co-Pilot is not the same product as it was a year ago. It is evolved from an in-editor assistant into an agentic platform capable of running long, multi-step coding sessions, using the latest models and iterating across entire repositories. Agentic usage is becoming the default and it brings significantly higher compute and inference demands. Today, a quick chat question in a multi-hour autonomous coding session can cost the user the same amount. GitHub has absorbed much of the escalating inference cost behind that usage, but the current premium request model is no longer sustainable. A few weeks later at Google I.O., we got something similar,

where, yes, nominally the headline was that they had reduced the cost of their Premier plans, Gemini Ultra Plan dropped a $200, and they also introduced a new $100 plan, but they also introduced on top of that usage limits and usage-based billing on top of those limits, meaning that really, for a lot of power users, this was going to represent a big cost increase. Same with Anthropic, who specifically focused on billing around third-party tools, meaning that while the subsidy persists if you are using an Anthropic-specific harness like Claude Code, as soon as you move to a different type of harness or a different environment that's not owned by Anthropic,

you're shifting to per token billing with huge financial consequences that created frankly a bit of an uproar throughout the month. So, the shift in business model was one response to the AI subsidy era ending and the token shortage era beginning, but it's not the only one. One of the most important AI questions right now

isn't who's using AI, it's who's using it well. KPMG and the University of Tech at Austin just to analyze 1.4 million real workplace AI interactions and found something surprising. The highest impact users aren't better prompt engineers. They treat AI like a reasoning partner. They frame problems, guide thinking, iterate, and push for better answers. And the good news, these behaviors are teachable at scale. If you're trying to move from AI access to real capability, KPMG's research on sophisticated AI collaboration is worth your time. Learn more at KPMG.com

slash us slash sophisticated. That's KPMG.com slash us slash sophisticated. One thing I keep seeing in Enterprise AI, companies hedging across every cloud, every model, every framework, or paying a GSI for a pilot that never ends. The team's actually shipping, they've picked a lane, and they move fast. That's one of the reasons I like today's sponsor robots and pencils. They've gone all in on AWS. They're an advanced tier and AWS pattern partner and they ship production AI co-workers in 45 days. That's led to them doing some of the more interesting work I've seen on AI co-workers. And by that, I'm not talking about chatbots. I'm talking about actual agentic systems that sit inside a business architecture and do real work. That kind of

focus matters if you're an enterprise leader trying to get something real into production or an AWS rep trying to move a customer from interested to deployed. Request an AI briefing at robots and pencils.com. One conversation with robots and pencils and you'll know. So coding agents are basically solved at this point. They're incredible at writing code. But here's the thing nobody talks about. Coding is maybe a quarter of an engineer's actual day. The rest is stand-ups, stakeholder updates, meeting prep, chasing context across six different tools. And it's not just engineers. Sales spends more time assembling proposals than selling. Finance is manually chasing subscription requests. Marketing finds out what shipped two weeks after it merged. ZenCoder just

launched Zenflow work. It takes their orchestration engine, the same one already powering coding agents, and connects it to your daily tools. Jira, Gmail, Google Docs, linear, calendar, Notion. It runs goal-driven workflows that actually finish. Your stand-up brief is written before you sit down. Review cycle coming up? It pulls six months of tickets and writes the prep doc. Now, you might be thinking, didn't OpenClaught try to do this?

It did, but it has come with a whole host of security and functional issues, which can take a huge amount of time to resolve. Zen Coder took a different approach. SOC2, Type 2 Certified, curated, tighter security perimeter, enterprise grade from day one, model agnostic and works from Slack or Telegram. Try it at zenflow.3. This episode of the AI Daily Brief is brought to you by OutSystems, a leading agendic systems

platform built for the enterprise. Organizations all over the world are building, orchestrating, and governing agentic systems on the OutSystems platform and with good reason. OutSystems open and unified platform allows teams to architect, deliver, and scale governed agentic systems with agility. Teams of any size and technical depth can use OutSystems to build, deploy, and manage AI apps and agents quickly and cost-effectively without compromising reliability and security. Without Systems, you can rapidly launch ideas from concept to completion.

It's the leading Agendic Systems platform that is unified, agile, and enterprise proven, allowing you to accelerate growth, reduce operational friction, and deliver real enterprise impact with AI. OutSystems. Build your agentic future. In addition to that business model response, we've also seen a big uptick in the recognition that when it comes to really adopting the full capabilities of agentic AI well, companies are just going to need a lot of help. Already, even coming into this year, there was what we call a big capabilities overhang. In other words, a space between what the AI models could do and what most companies were actually getting out of them. But you sprinkle a little bit of agentic capability on top of that,

especially as it moves out of the realm of just coding and into the realm of every type of knowledge work. and that capability overhang has just completely exploded. So much so that this month, both OpenAI and Anthropic announced initiatives to more directly support enterprise-level transformation. The forms that they took are a little bit different. OpenAI announced their deployment company, which is a majority-owned but separate venture, to put forward-deployed engineers inside big clients,

while on the Anthropic front, they partnered with Blackstone, Helmand, and Freeman, and Goldman Sachs to also launch a separate as-yet unnamed Enterprise AI consulting or services firm, the guts of which are our friends at Fractional, so congrats to them. But in that one, Anthropic has a smaller stake than Open AI does in their deployment company. In either case, they both represent the same instinct, which is that in this new period of agentic change and token shortage, more support for enterprise deployment is going to be needed.

Yet at the same time, while these consulting and services lines may be a necessary budget line item, The core reality of the period that we're going into is one where companies have to be much more diligent about managing AI costs. And this is not going to be an easy thing to do. A lot of the companies that we had heard about doing some sort of token maxing experiment are now scrapping their AI leaderboards, with Amazon being the most recently announced example. And it's not just because of concerns around gaming those leaderboards or anything like that. It's also because of those shifts that we just saw in the business model where it's just too expensive to token max now. The fundamental and anchor characteristic of the world that we are moving into is one where there is a

structural shortage of AI tokens. There simply is not enough compute to produce all of the AI that people would want to consume, meaning that the cost of AI is going to be high with all sorts of different potentially problematic implications. A cool thing, however, is that we're already starting to see market-based responses to that. Cursor announced the next generation of their composer model Composer 2.5, and not only is it performing well, it's doing so at a much lower cost than Opus 47 and GPT-55. So part of response to the token shortage is market-based innovation to bring the cost of tokens down without sacrificing performance. Google seemed to recognize that this might be part of their play as well, giving lip service on the main stage at Google I.O.

this month to the idea of Gemini 3point Flash being a way for enterprises to cut costs. However, in practice, that's not really how it's playing out. As artificial analysis points out, Gemini 3.5 Flash costs about five times as much as Gemini 3 Flash, both based on higher token prices as well as higher token usage. Which isn't to say that Google might not have some better shots on goal when it comes to market-based solutions for token scarcity. Their Gemma series of models, which are their smallest and cheapest models, are actually seeing really fast adoption.

with the adoption of those models outpacing similar Chinese models and a sign that people and companies are adapting to this new economic reality. Now, one of the things that I would expect is some pretty serious price warring going on, with China running its tried and true playbook of artificially keeping prices low to create a competitive advantage, which it seems like might be happening around Deepseek, who has just made a recent temporary 75% price cut on their V4 model permanent. To be clear, that's not because Deepseek has figured out some way to serve those tokens at 75%

at the cost, it's because in a world of token shortage, a lot of companies around the world are going to be forced to look away from the state-of-the-art open AI and anthropic models to more affordable alternatives, and Deepseek wants to be right there scooping up that business. Another thing that's happening in the context of this token shortage is that everything regarding AI infrastructure is, as SWIX, Sean Wang put it, going vertical. Inference provider base 10 is raising a billion dollars at an $11 billion valuation, more than doubling its valuation from just one quarter ago, open router which can help developers automatically toggle between models

that have different tradeoffs in terms of cost, efficiency, performance, etc., raised $113 million Series B becoming an AI unicorn. And even more than that, we're seeing some big realignments in the broader infrastructure world as well. The most notable of these undoubtedly, and something that happened this may that I think will have fairly dramatic implications for the industry as a whole, is Elon shifting into a very different type of role vis-a-vis the AI industry.

Up till now, Elon's headliner role in the AI space has been as one, cheerleader of Grok, which candidly has never at any point really caught up to any of the leading models, and two, as main antagonist of Sam Altman and OpenAI. Now, Elon's lawsuit against OpenAI this month was thrown out based on Statue of Limitations but that wasn't the big thing that changed. The big thing that changed is that whether it was because of economic opportunity or an assessment of the reality of where Grok sat relative to the models, or simply wanting to stick it to Sam Altman, Elon decided to team up with Anthropic.

The first announcement was that SpaceX AI, which is the new XAI inside of SpaceX, basically SpaceX's AI division, would allow Anthropic to use Colossus 1 to provide additional capacity to Claude. Now, Anthropic has been severely compute-constrained throughout the year, causing major headaches for users, so this was very welcome news to Claude users and started to show this realignment happening. However, then just a couple weeks later, we found out that not only would Anthropic be using Colossus 1, which was SpaceX slash XAI's first big data center that they spun up at the end of 2024, but at least on a temporary basis, Anthropic would also be using

Colossus 2. In the span of just a couple of weeks, SpaceX became a neocloud, with absolutely massive implications for the upcoming IPO. I could talk about this basically endlessly, but I think Elon moving into a place where he is focused on a thing that he does better than just about anyone, which is building big, ungodly physical infrastructure, using that as his way to leverage and influence the AI race, by virtue of being a self-appointed czar of compute, and by providing a clear-line pathway between SpaceX as Neo-Cloud provider right now and future orbital data center provider just makes the space X IPO make so much more sense in context. First of all, it allows investors to get excited about a different part of the AI stack,

an increasingly important infrastructure part of the AI stack, as opposed to just investing in an also-ran in Grok. And by the way, for those of you who love Grok, I'm not trying to yuck your yum, there are lots of reasons that it still has value to many people, and I don't think Elon has given up entirely on it. I just think that the trend line is pretty clear at this point. And it's very clear that Wall Street is right now extremely excited about the infrastructure side and effectively the entire AI supply chain.

This month saw AI memory stocks absolutely surge with companies like SK. Hynix and Micron becoming trillion-dollar companies. And now even META is talking about the possibility that they could also become a cloud business as well. For the first time in a very long time, META's AI narrative wasn't completely freaking out investors, because if they can go sell back the $130 billion or whatever worth of compute that they're investing in at a premium, it significantly de-risks that big Kappex spend. I think the theme of the compute buildout as a response to the token shortage era is going to do nothing but grow in June,

especially surrounding the SpaceX IPO. Just to put one more fine point on this, when about a month and a half ago Elon started talking about orbital data centers, it was getting a lot of sci-fi blank stairs. Now, the narrative has shifted so much that Jeff Bezos is talking about orbital data centers not as a whether we can, or if we will, but instead saying that two to three years feels to him to be a little ambitious as a timeline for them. Anyway, watching what happens around the SpaceX IPO will be hugely instructive, I think, in how the market is digesting this token shortage. But before we wrap up, there are a couple other things that happen this month that I did

want to point out as well. It was relatively quiet in terms of new model releases. We did just at the end of the month get Claude Opus 4-8, but part of what was interesting about that to me was how much the emphasis is shifted from models alone to the harnesses they sit in. Creator and entrepreneur Riley Brown wrote, unless it's a major breakthrough in model capability,

I'm much more excited for super app updates like Codex and Claude Desktop. There's so much to be unlocked by making those surfaces better. That was his response to the release of Claude Opus 48, basically saying, I'm not interested if it doesn't come with a Claude Code update. Greg Eisenberg went farther saying, I didn't cover Claude Opus 4.8 on my pod because I don't think it's meaningfully better than GPT-55 as of May 29th. We're entering the era where model releases start to feel like iPhone releases. Remember when every new iPhone was a genuine leap?

Now it's a slightly better camera and you can't really tell the difference. That's where models are heading. 4-6 to 4-7 to 4-8. Each one is a little different. Nobody can agree if it's better or worse. The benchmarks say one thing, the vibes is another. The thing that actually matters right now is what's happening around the models. Claude code shipped dynamic workflows this same week,

and that genuinely changes what one person can build. And indeed, that was definitely the story of this month. We talked a little bit about this new dynamic workflows approach as part of the Opus 48 coverage. And of course, May was also the month that Slash Goal became a real primitive, jumping from Codex where it started to something that's in Claude Code as well.

If you want to learn more about Slash Goal, go check out the episode we did for this week's Long Read Sunday, which is a primer on how to use slash goal, especially for knowledge work. On the narrative side, May might go down as the month where Sam Altman and Dario Amade finally figured out that they probably shouldn't tell everyone that the thing that they were building, and which was making them unfathomably rich, was going to take

everyone's jobs and livelihoods and totally changed the world in ways that didn't all seem that particularly good. Now, I would say that the Dario reversal is much more nascent than the Sam reversal, with Sam actually putting in some time to articulate why he was changing his opinion, arguing that evidence had suggested that he had just overestimated how the transformation was going to happen, much to his delight. I think that this has opened up some new narrative space, which allows for a lot more nuanced to conversation around AI on a policy perspective for which I'm very thankful. Now, speaking of the policy side, you do see a lot of jockeying, particularly on the Democratic side of the aisle right now, for how they're going to interact with AI. And what's

interesting is that it's very clear that there's not one fully embraced approach yet. You've got the Bernie and AOC wing who are calling for data center moratoriums, but you now have Elizabeth Warren coming in, basically saying let's not stop AI with data center moratoriums, let's get our cut. She wrote an op-ed last week in time called Why We Need to Tax AI, and I think that the conversation about novel taxation structures like token taxes is going to become a more prominent theme in the months to come. Certainly when it comes to politics in a backwards-looking way, the story of May was all about the White House getting fully involved in model releases

surrounding the release or non-release of Anthropics mythos. And what's interesting about this, bringing it back to the token shortage meta-theme of moment is that not only is the White House thinking about cybersecurity issues, they are also positioning the U.S. government relative to the token shortage, with a story coming out at the very end of April and the very beginning of May, that part of the reason that they were opposing Anthropics plan to expand access to Mythos was that they knew that there was a token shortage and they didn't want other people using up the tokens that they might want to use. So what comes next? Certainly a huge theme this month is going to be the SpaceX IPO.

It's also likely that we'll get another open AI model pretty soon. Anthropic has explicitly said that some version of Mythos will be here in the coming week, so I would expect that in June as well. But overall, I think that a lot of the immediate term next period is going to be all about how to recalibrate for an era of token shortage. From business model changes to different approaches in the enterprise to new policies, we are in for a big shift, and one where I think that there are going to be serious advantage opportunities that can accrue especially to enterprises who figure out how to the

how to manage this more quickly and more efficiently than others. We will discuss exactly how I think companies can come at this token shortage in episodes to come. But for now, that's going to do it for today's AI Daily Brief. I'm back from traveling tomorrow and we will be back with our normal formats. For now, however, appreciate you listening or watching. And until next time, peace.