How Big Is the AI Economy?

June 30, 2026 · Episode Links & Takeaways

Author

AI Daily Brief
June 30, 2026

HEADLINES

Fable's KYC Return

More signs of a Fable relaunch have emerged — but they come with strict new controls. AI leaker M1 Astra posted code strings from the Claude app suggesting usage will be credit-based rather than subscription-included, and that model access will require users to submit identity documents to Anthropic. Haider's skepticism is understandable — "no sensible person" giving over their ID to use a heavily guardrailed model — but as Max Weinbach noted, this may be the only path forward given how the government views these models. Dario himself compared Mythos to a super weapon that should require a gun license, so perhaps this is exactly where things were always headed.

Senator Warner's AI Agent Bill

Senator Mark Warner is preparing to unveil a sweeping new regulatory framework for AI agents — and while it's just a discussion draft that isn't likely to move this year, it signals where Washington is headed on the question of agent behavior. The bill would protect third-party agent access on major platforms and enshrine a "duty of loyalty" preventing agents from secretly prioritizing their corporate creators over the users they serve — think a travel booking agent steering toward Hilton due to an undisclosed partnership. There's a real risk that well-meaning liability provisions end up functioning as a backdoor ban on agent providers. But the agent neutrality principles at the core of this bill are probably going to be welcome to a lot of people, and at minimum it's worth not dismissing out of hand.

California Cuts a Deal for Half-Price Claude

Governor Gavin Newsom has announced a first-of-its-kind agreement with Anthropic, giving all California state departments and local governments access to Claude at 50% off, along with free workforce training and technical support. The deal explicitly isn't framed as a response to Washington's ongoing tensions with Anthropic — though the timing is hard not to notice. Newsom's framing was careful: AI should help government workers move faster and solve problems more effectively, not replace them.

Amazon Paying More for Anthropic

Anthropic has renegotiated its sweetheart deal with Amazon, which had been structured around raw computing hours at effectively a wholesale rate. Starting next year, Amazon will pay token-based pricing like every other large Anthropic customer — a change that applies to internal use and products like Alexa for Shopping. The reporting also surfaced real acrimony between the two companies: Anthropic was frustrated late last year by slow Bedrock feature velocity, and on Amazon's side, some engineers have reportedly been distilling Anthropic models proactively over fears prices would eventually rise. Both sides denied any problems — but the end of the AI subsidy era is clearly reshaping the economics of these foundational partnerships.

Meta's Distillation Problem

Meta has placed strict limits on how engineers in its Applied AI division — its recently established data labeling initiative — use Claude Code and Codex. The concern isn't cost: it's distillation. Internal memos warned that using these tools on certain tasks could contaminate training data and trigger "serious escalations with partner companies." As Chubby put it, this is the distillation trap — the more companies rely on frontier models to build their internal AI infrastructure, the harder it becomes to prove where the intelligence actually came from.

Google Caps Meta's Gemini Use

According to the Financial Times, Google imposed usage limits on Meta and other large customers in March, citing a compute crunch that even the largest tech companies are struggling to stay ahead of. Meta was hit hardest given its exceptionally high token demand — the restrictions were reportedly part of why it stopped tokenmaxing and pushed staff toward greater token efficiency. Meta has since shifted focus toward Muse Spark, its own in-house model released in April.

AWS Raises GPU Prices 20%

AWS has raised prices on EC2 Capacity Blocks — its reserved GPU compute product — by 20%, adding a useful data point to an increasingly complicated picture of where AI inference pricing is actually headed. H100 spot prices are down 40% from their May peak, but Semianalysis argues that's not the demand signal it might appear to be: falling spot prices alongside rising contract prices reflect serious buyers locking in term capacity, not a broad demand slowdown. The spot and contract markets are telling two very different stories right now.

RAMageddon and the Chipmaker Backlash

Spiking memory prices are driving a search for a scapegoat — and the AI industry is the obvious first target. Apple raised prices by up to 15% last week, Microsoft followed with Xbox price hikes, and Lenovo declared at a recent conference that memory pricing will never return to last year's levels. The Wall Street Journal describes it as a massive transfer of cash to memory chip makers: Micron is running at 56% gross margins and targeting 84% by year end, which would put it third among US companies by profit margin behind only Google and Nvidia. Apple has now reportedly petitioned the Trump administration for clearance to buy memory from Chinese supplier CXMT — currently on the Pentagon's blacklist — a clear signal of how acute the crunch has become.

MAIN STORY

How Big Is the AI Economy?

The team at Exponential View has just released one of the most rigorous attempts yet to measure the actual size of the AI economy — going through reports on over 1,000 companies, weighting sources by confidence, and carefully deduplicating to avoid counting the same dollar twice. The headline finding: demand is real, big, and fast, with AI companies having banked $110 billion over the past 12 months and running at a $175 billion annualized rate. This is more than a vanity metric — it's an empirical case that the sector is growing three times faster than any previous IT wave, and that the bubble debate is increasingly out of step with the underlying numbers.

The Big Numbers
$110B banked; $175B annualized rate; 90x faster revenue velocity.
In 2023, it took the AI industry 180 days to add a billion dollars in cumulative revenue. It now takes less than two days — a 90x improvement. The sector is growing three times faster than any previous IT wave.

The Compute Super Cycle
Semiconductors doubling; the US power grid finally waking up.
Global semiconductor revenue is projected to reach $1.5 trillion this year, up from $792 billion last year. After nearly 16 years of flat US electricity generation post-financial crisis, AI is now driving annual growth at 150% the historical average — reaching nine terawatt-hours per month, compared to six terawatt-hours per month between 1950 and 2008.

CapEx vs. Revenue
The build-out is finally paying for itself on a flow basis.
Starting in Q4 last year, quarterly AI revenues began exceeding CapEx depreciation. Total CapEx has reached $848 billion this year and $2 trillion cumulatively since 2020 — a sum quarterly revenues haven't retired yet, but are now running ahead of on a depreciation basis. Critically, older GPUs are generating returns well into years seven, eight, and nine — past the standard six-year depreciation horizon — which significantly improves the economics of the infrastructure bet.

Room to Run
AI revenue is still less than half a percent of US GDP.
The IT sector represents 9.4% of US GDP; AI revenue is equivalent to 0.42%. Even with AI revenue tripling relative to GDP since Q1 2025 and growing 10x since Q1 2024, the runway ahead is enormous — and current enterprise AI spend, like Uber's roughly $1,500 per engineer, barely registers on a P&L.

Token Economics
Tokens are getting cheaper while consumption explodes.
Global token volumes have crossed 30 quadrillion per month and are growing 14x year-over-year. The blended price per million tokens has dropped from $17 in mid-2024 to $2 today, even as average request intensity has tripled. The shift to agents is accelerating consumption sharply — an agentic coding task consumes around 1,200 times the tokens of a basic chat task. And energy monetization per gigawatt has roughly doubled since mid-2024, even as per-token revenue falls.

Where Value Is Accruing
The app and model layer is the fastest-growing slice of the stack.
Revenue remains concentrated in chips, but the mix is shifting — the app and model layer is up nearly 3x over the past year. The report draws a comparison to digital advertising's shift from untracked banner ads to pay-per-click: token-based pricing is the mechanism that enables the next wave of economic accountability and growth up the stack, even as labs simultaneously push down into infrastructure.

The Business Case
High AI spenders are outgrowing everyone else by 92 percentage points.
Companies in the top 25% of AI spenders by share of revenue have grown revenue more than 100% over the past three years. Companies with no AI spend have grown roughly in line with US nominal GDP — between 15 and 20%. A third of public companies are now citing AI impact on earnings calls, with 20% offering quantified claims. The gap in outcomes is the most concrete evidence yet that the investment is paying off.