Does the demand prove the value?
Anthropic runs at $30–40B in annualised revenue. Is that proof that AI creates value people will pay for — or proof that capital is chasing itself?
Revenue at this scale doesn't happen by accident
Anthropic grew from near-zero to a $30–40B annualised run rate in roughly three years — faster than any enterprise software company in history. OpenAI's consumer subscription business is the fastest-growing consumer product ever. Enterprises don't renew nine-figure API contracts out of hype; they renew because the output is worth more than the invoice.
Google, Microsoft, Amazon, and Meta are each committing $100B+ in 2026 capex. These are the most capital-disciplined infrastructure buyers on Earth, and they are all making the same bet with their own money.
The users aren't using it the way the revenue implies
Anthropic's own research (March 2026) documents a massive gap between what models can do and what people actually deploy them for. OpenAI's January 2026 data shows "power users" use reasoning capabilities 7× more than the median paying user. The typical user understands the technology — and still doesn't trust it enough to hand it real work.
If the median paying customer uses a fraction of the capability they're paying for, current revenue measures willingness to try, not durable value. The dot-com era also had real revenue — Cisco was profitable the whole time. Revenue proves demand existed; it doesn't prove it sustains.
Is the money real?
Half a trillion dollars of cloud backlog traces back to two startups funded by the sellers. Is that how infrastructure always gets built — or a $1.65T house of cards?
Railways and fibre were "overbuilt" too — and then the world ran on them
The 1840s railway mania bankrupted most of its financiers — and built the network that industrialised Britain. The 1990s telecom overbuild destroyed shareholder value — and laid the fibre that made the internet economy possible. In both cases, the capital was misallocated and the infrastructure was indispensable. Both things were true at once.
Unlike dot-com, the assets here are physical: chips, data centres, power plants. Even in the bear scenario, the compute doesn't vanish — it reprices and gets absorbed, the way dark fibre did. Dario Amodei's defence: one player has capital, the other has confident future revenue, and neither can fund the buildout alone. That's not fraud; that's project finance.
The loop pays for itself, and the debt is off the obvious balance sheets
NVIDIA invests $100B in OpenAI; OpenAI commits $300B to Oracle; Oracle buys $40B of NVIDIA chips; CoreWeave sells OpenAI $22.4B of capacity backed by a $6.3B NVIDIA guarantee. Dollars leave one door and return as revenue through another. Half of more than $2T in cloud backlog at four hyperscalers comes from two startups those same hyperscalers fund.
Meanwhile Prof G Media / Bloomberg's July 2026 analysis puts $1.65T in hidden AI debt — leases, special-purpose vehicles, and commitments that don't appear as conventional debt. Oracle's backlog jumped 438% in a year. When the loop's revenue depends on the loop's investment, "growth" is a financing decision, not a market verdict.
Does it actually work?
Strip away the finance. In the coding editor, the hospital, the research lab — is AI producing real, measurable output gains?
Coding, medicine, and science show gains no previous tool matched
Randomised trials show AI pair-programming tools completing defined tasks substantially faster (the GitHub Copilot RCT: ~55% faster on a specific task). Frontier labs report internal engineering throughput gains of 30–50% on some workstreams. In medicine, AI assistants are clearing radiology and dermatology backlogs; in science, AlphaFold-class tools compressed decades of protein-structure work into months, and model-driven materials discovery has produced experimentally validated candidates at scale.
This is not demo-ware. It is peer-reviewed, replicated, and — crucially — visible in the usage data of the companies paying for it. The gains are uneven, but they are real.
The gains are real, and they are arriving faster than the labour market can absorb them
Since 2023, US employers have announced 175,796 job cuts citing AI (Challenger, Gray & Christmas). Entry-level hiring in exposed occupations has collapsed even as senior hiring holds — the classic signature of automation eating the bottom rung of the career ladder. Previous technology waves displaced tasks over decades; this one is compressing the same transition into years, possibly quarters.
The bull case says "new jobs appear." Historically true — but the historical transitions involved technologies that complemented average cognitive labour. If the machine substitutes the median skill itself, the historical analogy may not transfer. The burden of proof is on the optimists to show the new jobs, not to assume them.
Can the trajectory hold?
The entire valuation stack assumes improvement continues and scale keeps compounding. What are the physical and economic limits?
AI is now doing the science that makes AI better
The strongest long-run argument isn't chatbots — it's recursion. Models are already contributing to their own improvement pipeline: automating evaluation, generating training data, assisting chip design, and running agentic research loops. DeepMind's AlphaEvolve (2025) used models to discover faster matrix-multiplication algorithms — a genuine, peer-verified research contribution made by AI.
Combine that with the energy curve: inference cost per token has fallen orders of magnitude in three years, and efficiency gains (distillation, sparsity, better hardware) are compounding faster than demand. If intelligence becomes a reproducible input, every sector's R&D gets a multiplier. That is not a bubble — that is an industrial revolution priced early.
The energy bill is real, and the product may become a commodity
Data-centre electricity demand grew 17% in 2025 (IEA), with AI as the driver; grid interconnection queues run years long, and labs are resorting to on-site gas turbines and nuclear PPAs because the grid cannot keep up. A technology whose marginal unit requires a power plant is not scaling like software — it is scaling like heavy industry, with heavy-industry economics.
Meanwhile the moat is leaking: open-weight models close the gap with each release cycle, inference is becoming a commodity market, and the January 2025 DeepSeek shock showed a frontier-adjacent model built for a fraction of the assumed cost. If intelligence commoditises, the trillion-dollar capex stack rests on pricing power that doesn't survive contact with an open alternative.
Where the evidence actually sits
Same claims, plotted by evidentiary strength. Green bars: core claims of the FOR case. Red bars: core claims of the AGAINST case. Length = our confidence rating (see methodology).
Reading: the bear case's structural claims (circularity, debt, energy) are better documented than the bull case's forward-looking claims (recursion, absorption) — but the bull case's micro claims (task productivity, discovery) are as solid as anything on either side. The disagreement is mostly about extrapolation, not observation.
Eight arguments. No verdict.
Deliberately. Every number above is an input, not an answer. The question — durable value or historic bubble — depends on assumptions about growth, improvement, energy, and time. Set those assumptions yourself and watch the probabilities move.