OpenAI-style bubble talk is on a low ebb, so it's a good time to ask dispassionately: how big is the AI economy right now? Exponential View's new report says $175B annualized and growing 3x faster than any IT wave before it — while the headlines fill with identity checks, agent-neutrality bills, and a memory shortage nasty enough to earn the name Ramageddon.
The AI economy is real, big, and fast — more revenue-validated than any prior platform shift.
Exponential View audited over 1,000 AI companies and found demand backed by realized revenue, not just future promise: $110B banked over 12 months, a $175B run rate, and growth 3x faster than any IT wave before it. CapEx is enormous — $848B this year, $2T cumulative — but quarterly revenue now exceeds CapEx depreciation, older GPUs earn yields past their six-year life, and high-AI-spend companies have grown revenue 92% faster than non-adopters. The bubble question isn't settled, but the indicators are far more positive than the discourse suggests.
Code strings in the Claude app suggest Fable usage will be credit-based and billed separately from subscriptions — and that model access will require users to submit identity documents. One string reads: "Your credits will be added once your identity is verified."
— Max Weinberg, on the Fable identity-verification rumor. NLW notes the comparison is fair given how the government views these models — and that Dario Amodei himself said companies called Mythos a "super weapon" that should require a gun license. Despite the pushback, NLW is confident nearly everyone will hand over ID anyway.
Senator Mark Warner is preparing a 25-page discussion draft that protects third-party agents (send your own agent to shop on Amazon) and creates a "duty of loyalty" barring agents from favoring undisclosed partners — like a travel agent secretly preferring Hilton. It only covers consumer-facing agents, not internal enterprise workflows, and isn't expected to move this year.
— Senator Mark Warner. Warner frames the draft as a step toward a federal framework that promotes innovation while protecting consumers. NLW is of two minds: watch how much liability it imposes on agent providers, but the neutrality principles will likely be welcomed by many.
Governor Newsom announced the first statewide AI rollout: all state departments and local governments get Claude access at 50% off, plus free workforce training. Newsom cautioned that "AI should not replace the human work of government" — it should help workers move faster.
The Information reports Anthropic has renegotiated Amazon's wholesale, compute-hours pricing to token-based rates like every other large customer, starting next year. Amazon is now exploring cost savings by switching to OpenAI or its in-house Nova models — a sign the subsidy era is over.
Meta's Applied AI data-labeling team must now solve coding problems without Codex or Claude Code, for fear model outputs contaminate training data and trigger "serious escalations with partner companies." Distillation violates both labs' terms of service, so Meta is guarding against legal exposure as it builds its own MetaCode model.
— Chubby, on Meta's coding-agent restrictions. Chubby's summary of the bind every AI company will face: to build a better internal coding model, you must ensure you're not accidentally training or evaluating on rival model outputs.
Per the FT, Google imposed usage limits on Meta and other large customers to manage the compute crunch — restrictions that partly drove Meta to stop "token maxing" and push token efficiency. Meta had leaned on Gemini and Claude because they outperformed in-house Llama models.
AWS is hiking EC2 capacity block prices by 20% for Nvidia workloads, sparing only its own Trainium chips. It's another data point in a supply-constrained inference market where reservations, not on-demand rentals, are now the norm.
H100 spot prices are down 40% from their May peak, but SemiAnalysis says contract prices keep rising. Their read: serious buyers are locking in term capacity for production workloads, shifting opportunistic spot usage toward committed deployment — not a demand slump.
AI memory demand pushed Apple to raise prices up to 15% and Microsoft to hike Xbox prices; Lenovo says pricing will "never return to where it was last year." Micron raised prices 60% in three months, quadrupled them over a year, and is targeting 84% margins — putting it behind only Google and Nvidia among US firms.
Amid the memory crunch, Apple petitioned the Trump administration to buy memory from Chinese supplier CXMT — a firm on the Pentagon's blacklist over alleged military ties. Separately, a California class action alleges Samsung, SK Hynix, and Micron ran a memory cartel to inflate prices.
Exponential View's State of the AI Economy report, drawn from over 1,000 companies with confidence-weighted sources and deduplicated spend, finds AI companies banked $110B over 12 months at a $175B annualized run rate. Their top line: demand is real, big, and fast.
In 2023 the AI industry needed 180 days to add $1B in cumulative revenue. It's now gotten 90 times faster, adding each new billion in under two days.
Global semiconductor revenue is projected to hit $1.5T this year, roughly double last year's $792B. US electricity generation was flat from 2008–2024; it's now growing at 150% of the historical average, reaching nine terawatt-hours per month in annual growth.
CapEx will hit $848B this year and $2T cumulatively since 2020, still mostly balance-sheet cash but increasingly debt-funded. Since Q4 last year, quarterly revenues have exceeded CapEx depreciation, though not yet the cumulative bill.
A key bubble worry was that GPUs go obsolete before earning their return. The data suggests the opposite: rental yields show older GPUs delivering meaningful gains into years seven, eight, and even nine — a much healthier payback scenario.
The IT sector is ~9.4% of US GDP; AI revenue equals only 0.42% — but it's grown 3x versus Q1 2025 and 10x versus Q1 2024. Even heavy users barely dent their P&L: Uber spends about $1.5K per engineer, leaving enormous room to grow.
The chat-to-agents transition is multiplying token use — an agentic coding task can consume around 1,200x the tokens of a chat query. Global token volumes now exceed 30 quadrillion per month, growing 14x year over year.
From mid-2024 to mid-2026, the blended price per million tokens fell from $17 to $2 even as the capabilities index rose from 112 to 158, and intensity per request jumped from 12 to 36 tokens processed per output token. Price declines drive more use and make previously uneconomical apps viable.
Revenue is still concentrated in chips, but hosting is rising, model revenue is growing, and app revenue (companies like Cursor) is appearing for the first time. The app-and-model layer's share of AI revenue is up nearly 3x over the last year — even as labs push both down into infrastructure and up into apps.
Companies with no AI spend grew revenue roughly in line with nominal GDP (15–20%) over three years. Top-quartile AI spenders grew revenue more than 100% — a 92% growth differential. Meanwhile 33% of public companies now cite AI impact on earnings calls, up from ~10% in early 2023.
NLW allows the market can still get over-exuberant, but calls this the most insightful bubble tweet ever — from OpenAI's Rune. The indicators of return on CapEx, he argues, are far more positive than the average discourse suggests.