Written by Ralph Sun
The useful lesson from 2022 is not that a fixed 35% decline must happen. It is that a growth market can reset sharply when investors stop paying for distant revenue and start demanding proof of profit, cash generation and capital discipline.
That is the question now facing artificial intelligence. The technology can remain important while a large part of its market value is repriced. Those two outcomes are not contradictory.
The 2022 comparison is more demanding than the headline suggests
The 2022 growth-stock selloff was not a straight line down, and it was not finished in a few weeks. QQQ, a liquid Nasdaq-100 proxy, closed at $403.99 on November 19, 2021. It later closed at $260.10 on December 28, 2022: a 35.6% decline measured from closing price to closing price over 404 calendar days. QQQ did not regain that November 2021 closing level until December 15, 2023.
That history matters because it separates a useful analogy from an easy slogan. A similar AI reset could include short-covering rallies, earnings-driven rebounds and sharp selloffs that trap both bears and late buyers. It would not need to be a single crash. Nor would a later recovery erase the damage done to companies whose balance sheets or business models fail the capital test.
The relevant analogy is therefore a multiple compression cycle. In 2022, rates rose, the discount rate applied to distant cash flows increased, and the market became less patient with companies valued primarily on revenue growth. The next AI reset, if it happens, may be driven by a comparable shift: investors deciding that model capability, user growth and fundraising are no longer enough without credible returns on compute, data-center and distribution investment.
The AI business model needs to earn its capital
The AI buildout has real demand. It also has unusually demanding economics.
Reuters reported that Anthropic told prospective IPO investors that third-quarter revenue reached $16.8 billion, up 45% from the prior quarter, and that the company was adjusted-profitable in the quarter. The same report said its prospectus described a $42 billion loss in 2025 and at least $518 billion of cloud, computing and infrastructure commitments over a decade.
The point is not that Anthropic or any other named lab must fail. It is that rapid top-line growth does not settle the question of who earns an adequate return after the cost of training, inference, cloud capacity and financing is included.
Gartner’s scenario analysis makes the mismatch explicit. It estimates that AI providers will spend nearly $1 trillion on infrastructure in 2026 and another $6 trillion between 2027 and 2030. In Gartner’s forecast, more than $3.5 trillion of AI infrastructure spending could be exposed to impairment by 2030 if the present capital-spending trajectory continues and revenue does not support acceptable returns.
That does not mean every AI company relies only on fundraising. Diversified hyperscalers can fund AI investment from cloud, advertising, search, software and other cash flows. But smaller frontier labs, neoclouds and single-product providers face a much harder version of the same question: can the next dollar of investment earn more than its cost before the underlying hardware is superseded?
A market can believe in AI while refusing to finance every version of the AI business model. That is the distinction investors should keep in view.
Inflation is a warning light, not a replay button
The comparison with late 2021 also needs discipline. September 2026 CPI rose 3.7% year over year, with core CPI up 2.6%. In November 2021, headline CPI was up 6.8% year over year. September 2026 final-demand PPI rose 6.3% year over year, below the 9.6% pace recorded in November 2021 but still elevated, with a 1.3% monthly increase.
Those figures support a partial analogy, not an identical one. Producer-price pressure is meaningful today, while headline consumer inflation is materially lower than it was near the late-2021 market top.
More importantly, a fresh high in the level of CPI or PPI is not a standalone top signal. Price indexes tend to make new nominal highs over time when inflation is positive. The Bureau of Labor Statistics explicitly notes that index-point changes are affected by the level of the index and that percent changes are usually the more useful way to analyze movements.
For markets, the relevant questions are whether inflation is accelerating or decelerating, how broadly it is distributed, whether policy rates must stay restrictive, and whether earnings expectations can absorb higher financing costs. A record index level by itself does not answer any of them.
A bear market would likely proceed through several market regimes
The most credible AI bear case is not “AI turns out to be useless.” It is that the market prices too much future profit too soon, while the sector’s capital needs remain high.
The first stage would be a valuation reset. The most expensive model, infrastructure and application names could fall as investors move from revenue multiples toward measures of gross margin, free cash flow, debt capacity and the cost to serve each additional user.
The second stage would be differentiation. Companies with real distribution, proprietary data, durable software integration or existing cash flows could hold up better than firms dependent on external capital. The market would stop treating every AI exposure as interchangeable.
The third stage would be consolidation. Some capital-constrained providers could need to recapitalize, merge, sell assets, shrink their compute commitments or exit independent operations. Gartner specifically forecasts that cost pressure and weak differentiation could force more than 80% of neocloud providers to cease independent operations by 2030 in its scenario. That forecast concerns neoclouds, not a prediction that named mega AI companies will collapse.
This distinction matters. A few failures or restructurings would reduce excess capacity and subsidy-driven competition, but they would not automatically make all survivors attractive investments. Less competition can improve pricing discipline for token inference and cloud capacity. It can also raise costs for customers and slow adoption if the remaining suppliers use their pricing power aggressively.
Why a recovery could come before the narrative feels comfortable
A bear market can create the conditions for a better next cycle. Lower valuations can make future returns less dependent on perfection. Consolidation can reduce overcapacity. Lower hardware and inference costs can widen the set of applications that make economic sense.
Physical AI—robots, industrial automation, autonomous systems and other embodied applications—could become a more important source of demand if deployment starts producing measurable productivity gains. But it should not be treated as an automatic sequel to a model-company shakeout. Physical AI has its own bottlenecks: hardware reliability, safety, integration, regulation and customer willingness to redesign workflows.
The recovery case is therefore not “everything falls, then AI immediately resumes its old leadership.” It is more demanding. Investors would need evidence that application revenue is catching up with infrastructure spending, that token and compute economics are becoming durable, and that the surviving companies have defensible distribution or product advantages.
What would make the 2022 analogy wrong
The bear-market scenario weakens if AI providers show that revenues are scaling faster than their capital burden, and if gross margins, utilization and free-cash-flow trajectories improve at the same time. It also weakens if inflation pressure eases without requiring a restrictive policy response, allowing growth-stock valuations to remain supported.
Conversely, the risk grows if capital expenditure keeps rising while application revenue and compute utilization disappoint; if financing terms tighten for AI infrastructure; or if high-priced companies miss the margin milestones embedded in their valuations.
Investors should not confuse volatility with a broken long-term trend. The 2022 example shows that a 35%-plus growth-stock drawdown can be followed by new highs. It also shows that the path is rarely clean, that the timing cannot be known in advance, and that a recovery in an index does not rescue every company inside the theme.
