"Even Silicon Valley Says That AI Is a Bubble" is a March 12, 2026 Atlantic long-form feature by Lila Shroff, an Atlantic technology writer. It documents a shift in which a widening section of Silicon Valley has stopped denying that AI is a bubble and begun defending it as net-positive, then sets that defense against critics who argue the costs of a crash fall unevenly on those least able to choose them.
Argument
The essay describes a position that distinguishes "good bubbles" from "bad bubbles." On this view, good bubbles (railroads, dot-com) leave durable infrastructure behind even after they pop, accelerating socially-valuable build-out that disciplined capital allocation would never finance; bad bubbles (the 2008 housing bubble) destroy capital without producing offsetting assets. AI is presented by its defenders as a good bubble, on the expectation that a crash, if it comes, would leave the data centers, fiber, clean energy capacity, and chip-design talent in place.
Shroff treats this argument as having a named theoretical source rather than re-attributing it to its many promoters. The framework's formalization is Tobias Huber and Byrne Hobart's Boom: Bubbles and the End of Stagnation (2024), which argues that bubbles can be "good" (financing technologies society would otherwise underinvest in) or "bad" (destroying capital with no offsetting asset). Hobart told Shroff that in a bubble "a set of investments that you could never underwrite otherwise suddenly makes sense," citing railroads producing freight rail infrastructure and the dot-com bubble producing the fiber-optic backbone of the modern internet.
Defenders cited
Shroff assembles statements from figures who treat an AI bubble as acceptable or beneficial:
- Hemant Taneja, CEO of General Catalyst: "Bubbles are good. Some spectacular failures…a price worth paying for enduring companies that change the world forever."
- Jeff Bezos: AI is a "good kind of bubble."
- Sam Altman: AI will be "a huge net win for the economy" even if "a phenomenal amount of money" is lost.
- James Thomason: "Stop trying to make bubbles go away. The benefits of innovation outweigh the costs of volatility."
- Marc Andreessen and Peter Thiel: both praised Huber and Hobart's Boom on publication.
- Mary Daly, San Francisco Fed president: "even if the investors don't get all the returns…it doesn't leave us with nothing." The quote is from an October 2025 Axios interview.
- Ben Thompson (Stratechery): the AI-driven energy demand will spur clean-energy build-out.
Some of these statements were made at distinct points across 2025 and 2026 and the speakers' positions may have evolved since.
Counter-arguments
Shroff sets several critiques against the good-bubble defense, the central one being asymmetric incidence: the bubble is good only for those positioned to win from it. Tech billionaires can absorb crash losses voluntarily, while retirees holding 401(k)s tied to AI stocks cannot. She quotes Mark Zuckerberg saying that even if Meta "misspends a couple of hundred billion dollars…the risk is higher on the other side"; Shroff observes that Zuckerberg is structurally insulated whereas most Americans whose retirement is index-linked are not.
Howard Marks of Oaktree, who anticipated the dot-com crash, told Shroff that "if investors remained dispassionate, it would take a lot longer for a new unproven technology to be adopted," but added that "the investor doesn't say, 'Well, yes, I lost my money, but thank God it advantaged society.'" Bubble apologetics, on this reading, is easier for those with capital to lose voluntarily than for those whose retirement accounts crash involuntarily.
Carlota Perez, author of Technological Revolutions and Financial Capital, is cited describing the bubble as "the eye of a much larger hurricane that involves the whole financial world." Gita Gopinath of the IMF is cited estimating that an AI crash could wipe out roughly $35 trillion in global wealth.
A further counter concerns asset durability. Unlike railroad tracks and fiber-optic cables, which last decades, computer chips obsolesce in years; Shroff cites Epoch AI's "GPU frontier lifespan" data. On this argument the infrastructure-left-behind case is weaker for AI than for its historical analogs.
Scale anchors
Shroff anchors the scale of the build-out and the historical precedents with several figures:
- OpenAI valued higher than Toyota, Coca-Cola, and Disney combined.
- Big Tech 2026 capex on AI of roughly $650 billion, which she notes exceeds the GDP of most countries.
- The Panic of 1893, the fallout of the railroad bubble, when unemployment exceeded 10% for half a decade.
- The 2000 dot-com crash, which preceded a US recession.
Reception and caveats
The "good bubble" versus "bad bubble" distinction is itself a value-laden classification: what counts as a good bubble is determined after the fact by surviving infrastructure, which is the outcome the AI bubble's defenders are betting on. Shroff paraphrases the Huber and Hobart framework rather than contesting it on its own terms.
Relationships
- supports: AI Bubble Debate — major synthesis of the pro-bubble Silicon Valley position
- supports: Circular Financing in AI — connects bubble dynamics to the capex/revenue divergence
- supports: AI Data Centers — bubble would leave data-center build-out as durable asset (or might not, per the chip-obsolescence counter)
- contradicts: AI as Normal Technology — Shroff's piece is implicit evidence that the bubble framing is a sharper interpretive lens than the "normal technology" thesis, at least for capital-markets dynamics
- related: So, About That AI Bubble — Rogé Karma (The Atlantic, May 2026) — Karma's Atlantic piece argues burden of proof has shifted; Shroff's accepts the bubble framing but disputes whether it's "good"
- related: The AI Backlash Could Get Very Ugly — Lila Shroff (The Atlantic, May 13 2026) — same author's follow-on argument that bubble and crash dynamics will fuel the political backlash
Tracked claims
- Big Tech 2026 AI capex ≈ $650B — confidence high (cited across reporting).
- Huber & Hobart Boom (2024) is the canonical formalization of pro-bubble ideology in tech — confidence high.
- GPU frontier lifespan is shorter than rail-track / fiber-optic lifespan — confidence high (Epoch AI data on chip obsolescence).
- AI crash could wipe ~$35T in global wealth (Gopinath estimate) — confidence medium; single estimate, depends heavily on assumptions about contagion to broader markets.
- Most-optimistic-AI cohort is >$200K household income — see The AI Backlash Could Get Very Ugly — Lila Shroff (The Atlantic, May 13 2026) for the Quinnipiac source.