Bull vs. bear · is it a bubble? · 2026

Is AI a bubble?

The most-asked question about the whole build-out — and the one with the least honest answers, because it quietly mixes up two different questions. Separate them and the fog clears. Here's the bull case and the bear case, weighed with real numbers, and where the evidence actually points.

Question 1

Is the technology real & durable?

Almost certainly yes.

Usage and revenue are exploding off a real base — Anthropic alone went from ~$1B to a ~$47B run-rate in 18 months. People and companies use this every day. See sheet 05 →

Question 2

Is the spending a financial bubble?

Genuinely contested.

Capex is running years ahead of the revenue it needs to justify itself. That gap, not the technology, is where the real risk lives — and reasonable people disagree.

Keep those apart and most "AI bubble" arguments resolve into one or the other. The internet was unmistakably real in 1999 — and the dot-com bubble still wiped out trillions. Both can be true at once.
The tale of the tape · 2026
~$450B
AI-specific capex from the big-5 hyperscalers this year (~75% of ~$650B+ total)
$1.6T
projected annual AI capex by 2031 (Goldman Sachs baseline)
~$22B vs ~$13B
OpenAI's 2025 spend vs. revenue — and it has committed to ~$60B/yr of future compute (Oracle, 2027+)
~$400B
annual AI-asset depreciation — more than the hyperscalers' combined 2025 profit
The two cases, side by side
▼ The bear case

The revenue gap

Capex is racing ahead of income. By the Sequoia "$600B question" framing, the industry must eventually earn multiples of its chip spend to clear a return — and today's end-user AI revenue is far short.

GPUs rot fast

Unlike dot-com fiber (a passive 20-year asset), GPUs have a ~3–5 year useful life and depreciate ~20%/yr. Much of the build-out could be obsolete before it pays for itself.

Circular financing

Chipmakers investing in the customers who buy their chips, and debt secured against fast-depreciating GPUs, can flatter demand and concentrate risk if the music slows.

Physical limits already biting

~50% of planned US 2026 data centers are delayed or cancelled — a sign the build-out front-ran reality. See sheet 02 →

▲ The bull case

Paid from cash, not debt

The decisive difference from 1999/telecom: 4 of the 5 hyperscalers can fund this entirely from operating cash flow. No leverage cascade means no forced-seller spiral if growth wobbles.

The revenue is real — and vertical

This isn't eyeballs-and-no-sales. Anthropic grew ~80× year-over-year; token usage is up hundreds of fold. Demand is metered, recurring, and compounding. See sheet 01 →

Costs collapse, demand follows

Inference gets ~10× cheaper a year, which keeps unlocking new viable use cases faster than supply — the opposite of a product nobody wants. See sheet 05 →

Even a shakeout leaves the asset

If some players overbuild and fail, the compute, models, and know-how persist and get cheaper for everyone — exactly how fiber and railways played out after their busts.

History rhymes — but doesn't repeat
EpisodeWhat was realWhat still happened
Railway mania
1840s
Rail genuinely transformed the economy for a century. Investors were wiped out; the track stayed and was used for generations.
Dot-com / fiber
1999–2001
The internet was real and world-changing. ~$5T in equity value evaporated; over-built fiber was eventually lit and powered the 2010s.
Cisco
the picks-and-shovels play
Networking demand kept growing for decades. Cisco's stock took ~25 years to reclaim its 2000 peak (it finally did in Dec 2025) — a caution for assuming today's chip leader is a safe proxy for the trend.
The honest verdict

"The technology is real" and "the spending is a bubble" are not opposites.

The strongest bull point is structural: this build-out is paid for out of cash flow, so it lacks the debt-fueled fragility that turned past manias into crashes — and the underlying demand is real, recurring, and growing fast. The strongest bear point is timing: capex is running years ahead of revenue, on hardware that depreciates fast, with some circular financing masking the gap.

So the most defensible read isn't "no bubble" or "all bubble." It's that an infrastructure shakeout is plausible — even likely in pockets — where specific over-builders, speculative data centers, and circular deals get repriced, while the technology itself keeps compounding. A sturdier base than 1999, with a real gap that still has to close. Anyone selling you a confident one-word answer is selling something.

Note — This sheet weighs a contested, fast-moving question; it's analysis, not a prediction, and the figures move monthly. It deliberately avoids a yes/no verdict because the honest answer is conditional.

Sources — Goldman Sachs, Tracking Trillions (2026 AI capex ~$765B baseline → $1.6T by 2031). Hyperscaler 2026 capex guidance ~$635–690B, ~75% AI (company filings; IEEE ComSoc, Dec 2025). David Cahn / Sequoia, AI's $600B Question. OpenAI & Anthropic revenue run-rates (company disclosures; Epoch AI; VentureBeat; Fortune, 2026 — Anthropic ~$47B run-rate, ~80× YoY). GPU useful-life & ~20%/yr depreciation, ~$400B AI-asset depreciation vs hyperscaler profit (Guinness Global Investors; Allianz Research, Mar 2026). Circular-financing examples (CoreWeave SEC 8-K filings; NVIDIA–CoreWeave, Meta–CoreWeave, and Jane Street–CoreWeave agreements, 2026). Data-center delays — see sheet 02 sources.