Today on the AI Daily Brief, we are talking about why only AI training can save the economy. The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI.
Now, today we're talking about a theme which has been pretty much ever-present in my entire journey with AI, which I think is now more existentially important not just for the AI industry, but for the economy as a whole than it has ever been. I'm talking about AI training, AI education, upskilling, whatever you want to call it. The process by which we help people close the capability gap between what AI could be doing for them and the value that they are actually getting out of it.
Now, this is about as bombastic a title as you're ever going to get on the AI Daily Brief, but I'm going to try to stand on business for this one. The short of the argument is that we are in a world where the relationship between AI lab revenue growth and AI infrastructure buildout is the defining relationship of the American economy, and where in that context we will increasingly find ourselves caught between, on the one hand, the AI labs' need for ever-increasing growth in token usage, and on the other hand, increasing scrutiny and limitations from enterprises. My belief is that the only way to solve for both — to provide both the labs what they need and the enterprises what they need to keep the whole party going — is training. So let me lay out the argument for you guys.
Part one of the argument is that the American economy just is the AI trade. AI investment is not a sector story. It is the growth story. In Q1 of this year, GDP grew at 2% annualized, with AI-driven investment contributing about 75% of the increase. AI data centers, hardware, and networking hit 1.4% of US GDP in Q1 2026, doubling from 0.7%, and making AI infrastructure the leading driver of US private investment growth. Data from the St. Louis Fed suggests that AI investment accounted for 39% of marginal GDP growth over the trailing four quarters, which is bigger than the tech sector's 28% contribution at the peak of the dot-com boom. What's more, excluding these investments, growth in the first half of 2025 would have been 0.1% annualized — a near standstill. In 2026 alone, big tech's AI capex spend will pass $800 billion, which some, like AI czar David Sacks, have argued could represent a 2.5% GDP tailwind this year and a 3% GDP tailwind next.
Now, this infrastructure spend isn't coming from nowhere. It was justified initially by the belief in the importance of AI in the future, and as time goes on, it is increasingly justified by specifically revenue growth from the labs. That's the contract. As long as token consumption keeps rising and rising fast enough, the capital keeps flowing. In fact, if you think about the difference between Q4 of last year and the first half of this year in terms of the popular market narrative — last year, from about mid-August all the way through December, the biggest discussion on Wall Street was about an AI bubble. And there were all sorts of different approximate reasons for that: comments from Sam Altman, the MIT air-quotes report that said that 95% of pilots were failing. But there was something much bigger underlying it that wasn't about narratives, but was instead about math — specifically seat math.
In short, $20 to $200 a month times the number of addressable seats among knowledge workers was not enough revenue to justify trillions of dollars of infrastructure spending. The TAM of AI when AI is sold as seats just wasn't going to cut it. The shift, of course, is that seats have, in the era of viable agents, ceased to be the core unit that matters when it comes to AI economics. The shift that we have all lived through is from an assisted, seat-based paradigm to an agentic, usage-based consumption paradigm. Per-person economics move from $20 to $200 a month to potentially thousands of dollars. And the revenue evidence is clear. Anthropic had this insane API surge that catapulted it to a $30 billion annual revenue run rate, which then jumped all the way up to $47 billion by late May. This was driven not by all sorts of new people paying for $20 or $200 a month Claude accounts, but an insane amount of usage of Claude Code. And Anthropic wasn't the only AI company experiencing this shift. OpenAI's revenue also jumped significantly in the first quarter, aided and embedded by their Claude Code competitor, Codex, which is their vehicle for delivering agentic tokens. In the beginning of the year for Anthropic, the number of enterprises spending a million dollars a year jumped from 500 to more than a thousand in under two months.
But of course, this came with consequences. The big theme for the last month or so on this show has been the shift from the token subsidy era to the token scarcity era. Now, we don't know how much exactly the labs have actually been subsidizing their accounts, but recent estimates from SemiAnalysis estimate that on the Claude $200 a month plan, the max possible spend was approximately $8,000 a month worth of tokens. And on the max ChatGPT plan, the max possible spend was up at $14,000 a month. Now, even if these numbers aren't correct, even if they're significantly off, you're still talking about just absolutely huge subsidy. As the amount of AI being consumed increases, while those infrastructure projects lag in their capacity to increase the availability of AI — both because of delays that they're facing, but also because it just takes a long time to bring that capacity online — the tried-and-true rules of economics apply, and market pricing and incentives start to shift.
The end of April, beginning of May is really when we started to see this happen. GitHub Copilot was one of the first to announce a move to usage-based billing, specifically calling out the fact that agentic sessions were just fundamentally different than the way that they had built the previous pricing model around. At Google I/O, they announced a whole bunch of new pricing, actually bringing down the price of some of the premium tiers, but also adding usage limits for the first time, after which you get shifted over to the API — in effect, hiding a usage-based shift behind a decrease in the base plan. One of the biggest dust-ups between developers and Anthropic happened when they started to shift all usage that happened on third-party harnesses — so basically anything outside of Claude Code or Cowork — to usage-based billing as well. And very quickly, the consequences started to hit home on the enterprise side of the equation as well. We've been living through 2025 assisted AI budgets meeting 2026 agentic AI reality.
Obviously, Uber has been the big held-up example — first making news for blowing through their entire AI budget in the first four months, and eventually moving to a $1,500 a month cap per employee. Other advanced AI-using companies like Walmart did something similar. This is what led me to argue, as I did on Twitter at the beginning of the month, that every AI business is now and for the foreseeable future, in some way, shape, or form, a token efficiency business. And you see so many examples of this. One is you're seeing a lot of the harness companies start to really emphasize model routing — in other words, a more sophisticated approach to routing certain tasks to cheaper, lower-cost models, leaving the most state-of-the-art models and highest cost for only the most important tasks. After Factory announced a new model routing feature at the beginning of June, CEO Matan Grinberg said that $13 million had been saved so far in the first 30 days of private preview. Now, other companies are just shifting models entirely. DeepSeek came in as Ramp's top-trending SaaS vendor, as companies like AI startup Lindy have shifted off expensive American models and towards those cheaper Chinese alternatives. In addition to those cheaper Chinese alternatives, you're also seeing a lot of companies experiment with tactics like post-training to roll their own alternatives with a specific industry or functional focus. Cursor's Composer 2.5 is performing at Opus 4.7 and GPT-5.5 type of levels at a tenth of the cost. And even vertical companies like Harvey are experimenting with more complex structures that can use post-trained versions of open models like Kimi K2.6 in concert with more advanced models like Opus to perform both at a higher level and at lower cost.
Now, all of this got us to the freakout that I focused on for Friday's episode — the Citadel Securities note, which showed the Silicon Data LLM token expenditure index starting to roll over and point downward. Now, as I explained on Friday, this chart actually isn't showing what people thought it was. Specifically, it has nothing to do with overall demand or overall token volume or overall token expenditure. Instead, the index tracks the average price that customers are paying for a million tokens. But because their data comes entirely from third-party router companies — i.e., not the labs themselves — it's biased towards companies that are actively seeking out lower-cost alternatives. Still, it does tell part of the same story where companies, especially leading companies, are looking for cost advantages for the first time.
And so we get to this equation: what the labs need versus what enterprises will pay. Now, already, even in their private market context, the relationship between the labs' continued growth in token consumption, expressed as revenue, and the amount of capital that is available for the infrastructure buildout is incredibly important. However, when Anthropic and OpenAI IPO, the intensity of the public market pressure to show every single quarter massive — and I mean massive — growth in token consumption is hard to overstate. Already, we live in a world where it doesn't matter that Nvidia grows its revenue by a significant amount; it has to dramatically outperform market estimates, or else its stock price tends to go down after quarterly reporting. That sort of phenomenon is going to be even more aggressive around the leading labs, meaning that they will have to do whatever it takes to get people consuming more tokens.
Now, on the other side of the equation, it's not that enterprises don't want to spend money on tokens. It's not that they don't want to get value out of AI. But increasingly, if the CFO set becomes the most important force in the decisions being made around AI, there are some pretty serious implications for the AI companies' ability to achieve that never-ending growth that is so important not just to their bottom line, but to our overall economy. Now, already you've seen the labs acknowledge that their initial ideas that agents were just going to show up and all of a sudden take over a huge portion of knowledge work have kind of slammed up against the reality of human institutional inertia. Perhaps the best expression of this is that over the last six weeks, both OpenAI and Anthropic have launched major consulting efforts focused around forward-deployed engineering. Now, I applaud those efforts. I think they're a great step, but I think that they're only part of the transformation we're going to see in how the labs think.
In short, there are two big realizations that are going to be hurtling towards OpenAI and Anthropic specifically. The first is that as they dig in with all of their FDEs, they will discover that a huge portion of the value of AI is not going to come from a set of centrally planned agents built by the FDEs in concert with the software engineers inside enterprises. Instead, the value will come from many diverse knowledge workers of all different stripes building and using agents. There is a bottoms-up agent experimentation that is going to absolutely be required for companies to get the most value out of AI, and that's not going to come from FDE efforts alone.
Now, the second realization, if we're allowing ourselves to be a bit more cynical for a moment, is that even if some folks inside labs don't believe realization one and don't think that the right paradigm is every knowledge worker off experimenting with their own agents, there's a fairly decent chance that the token growth pressure will force them to act as if that's true anyway. In other words, they'll figure out that they won't be able to hit quarterly token growth with a strategy that only gives leverage to a select few, and that demand expansion will require everyone building and using agents. My prediction then is that over the next six to twelve months, we will see dramatic increases in lab investment in enablement, training, and expanding the user base and depth of usage. Put differently, if everyone copies Uber and sets spending limits at $1,500 a month per employee, then the labs have to do whatever it takes to get every employee spending $1,500 a month, and then having the impact of that $1,500 be so high that it makes sense for the enterprises to increase those limits.
Now, this is not going to happen on its own. One of the things that I am most worried about when it comes to enterprises moving into this token efficiency period is that that type of thinking and those types of caps come with a hidden cost. Caps don't just limit spend. They shape what gets attempted. Budget scrutiny, even if completely understandable, will push enterprises and individuals within enterprises towards basic productivity type of use cases and away from the big, unseemly experiments that are required for the next generation of economic value to be created. I call this the known ROI bias. If people aren't given permission and structure and sandboxes and encouragement to go out and see what new things they could create with a fleet of agents, they're just going to try to do today's work a little bit faster or a little bit cheaper. This will not unlock the full value of AI. In fact, it will limit severely the value that the world gets out of AI. And perhaps more pertinently for my predictions for the next six to twelve months, it will limit the amount of tokens that the AI labs can sell.
And so we come back to this equation, and my argument that the single and only thing that can solve for the needs of both of the categories of parties on either side of this equation is AI training resources at mass scale and high quality to move people from assisted to agentic AI and to help them learn to use agentic AI to uncover the next generation of use cases that unlock value that make the input costs seem negligible.
Unfortunately, the state of AI education is abysmal. It is actually just an insane market failure how little high-quality AI training and education has come about in the last few years. An EY AI survey from this year found that only 28% of organizations have managed to empower AI employees to utilize AI to actually change any sort of business processes. DataCamp did a survey of more than 500 enterprise leaders and found that while video courses are the most common AI training format, they produced, as DataCamp puts it, awareness without confidence and adoption without judgment. The World Economic Forum notes that the half-life of skills is at a critical point, meaning that when it comes to AI education, content decays before a course catalog can even ship. And guess what? Things were already tough when we were just talking about prompt engineering. Now that we're in the world of agents and agent management and agent building, it is massively more complicated. It's also incredibly more important.
Even if you don't agree with me about the intrinsic economic importance of AI training, I think at this point it is very hard to deny that we aren't in the midst of a secular shift in what knowledge work consists of. Simply put, we are moving from a paradigm in which we do things to one in which increasingly we oversee synthetic intelligences that do those things for us. Prompting assisted AI was a new skill, but it was not a new knowledge work primitive. Managing agents, on the other hand, is a new knowledge work primitive that every single knowledge worker in the future will need to be skilled in. This is a lot closer to management training than it is to software training.
Now, obviously, if you've been watching my moves this year, it's been pretty clear that I'm thinking about AI education and training a lot. I've now released three different free self-directed programs: the AIDB New Year's program, Claude Camp in the midst of the open Claude craze, and Agent OS, which is sort of an updated agentic operating system program that takes a lot of the pieces of Claude Camp and moves them into the Claude Code/Codex type of paradigm. Part of the reason why I've wanted to experiment with these sort of free self-directed programs is that I think that the scale of the need for this training is mass scale. There need to be more free programs. There need to be more paid programs. There need to be more programs of every stripe in between. And to be clear, there are some good educational resources out there. AI entrepreneur and educator Riley Brown is constantly pumping out really, really great how-to videos. I often recommend his video about learning 95% of Codex in just half an hour as a great place for people to start with that tool. And you do have some companies like AIDB sponsor Section who are doing their damnedest to try to close the capability gap, but it's just not enough, and we're going to need more.
So, what to do with this? Well, for my part, you'll be hearing a lot more about some new initiatives coming soon, particularly with Superintelligent, where we're going to be in some ways returning to our roots quite soon. And more broadly, this is just a drum that I'm going to be beating a lot more. I think everyone has a role to play in this. But I think the most critical role, as you can probably tell, is going to have to come from the labs themselves. So, Anthropic, OpenAI leaders, if you are listening to this, whether it's now or three months from now or six months from now, I can almost guarantee these are conversations you're going to be having. If you want some ideas for what you can do with this and how you can, in the process, not only help your companies, but yes, save the entire American economy, well, you know how to reach out.
For everyone else, keep being the shining examples that you guys are for the communities that you're a part of. People's sense of the possible is shaped by what's around them. And AIDB listeners, I have found over and over again, are the portion of the population who are helping everyone else see how powerful and exciting and dynamic and good AI can be. So, keep it up, and I'm here to help. For now, that's going to do it for today's AI Daily Brief. Appreciate you listening or watching as always, and until next time, peace.
