AI Token Spending Just Fell 20% From Its Peak — Here's Why That Matters for the Entire Infrastructure Boom

After surging through early 2026, the index turned down sharply in late May and early June ^BT. Macro strategist Andreas Steno Larsen flagged the downturn as a serious warning, urging investors to watch the index closely ^BT.

The timing is important: the index peaked just as the hyperscalers were accelerating their 2026 capital expenditure plans to unprecedented levels ^LFF.

Token Prices Have Collapsed 90%+ Since 2023

The decline in the spending index is happening against a backdrop of extraordinary price compression. Apollo Global Management's Torsten Slok reported in a June 2026 presentation that the price per token has fallen about 90% since 2023 ^ALF. Other sources put the decline even steeper: 98% from GPT-4's March 2023 launch to mid-2026 ^STCT.

The paradox, well-documented by economists, is that cheaper tokens led to much higher total spending — a classic Jevons paradox. As tokens got cheaper, companies used far more of them, driving aggregate expenditure higher. The Silicon Data index roughly doubled from late 2025 to its May peak precisely because token consumption surged ^FAL.

But the May peak appears to be a turning point. Bain & Company analysts noted that while token prices fell by half from December 2024 to December 2025, token consumption surged by 450% in the same period — suggesting the volume growth that offset price declines may now be decelerating ^FB.

The $600 Billion Capex-to-Revenue Gap

The token spending index rollover comes at a precarious moment for AI infrastructure investment. The five largest US cloud and AI infrastructure providers — Microsoft, Alphabet, Amazon, Meta, and Oracle — collectively committed to spending between $660 billion and $690 billion on capital expenditure in 2026, nearly doubling 2025 levels ^FAMC.

Sequoia Capital partner David Cahn has calculated that the AI ecosystem needs approximately $600 billion in annual revenue to justify current infrastructure spending levels ^FFA. As of mid-2026, actual AI-attributable revenue is estimated at $50–150 billion annually — implying a gap of 4–13x ^U. Allianz Research found that US Big Tech capital expenditure intensity has risen to roughly 23% of revenue, more than double pre-ChatGPT levels ^A.

Stanford's 2026 AI Index confirms the broader tension: AI company revenue is rising at historically fast rates, but compute costs and infrastructure spending are also reaching record levels ^H. The gap, as multiple analysts have noted, is widening rather than narrowing ^FFM.

Meta's Cloud Business: A Sign of Overcapacity?

The most concrete signal that excess capacity may be building came on July 1, 2026, when Reuters and Bloomberg reported that Meta is developing a cloud infrastructure business to sell access to AI computing power and models to external customers ^RBSC.

Analysts predict Meta's total expenditure could reach $145 billion by 2026, a figure that surpasses the GDP of several smaller nations ^F. Meta CEO Mark Zuckerberg had previously told shareholders that entering the cloud computing market was "definitely on the table" if the company ended up with excess capacity from its data center expansion, and confirmed that companies approached Meta "almost every week" to buy spare compute ^GTIS.

Reuters framed the move as part of a broader push by technology companies to seek returns on costly AI investments amid worries about overspending ^A. Meta's stock jumped as much as 9% on the news, as investors welcomed a plan to monetize what might otherwise be idle infrastructure ^CS.

But the move also signals a potential problem: Meta may have more AI computing capacity than its internal products require. And if Meta enters the cloud market, it would add more supply to a sector where prices are already under severe pressure — token prices have fallen 90% since 2023 without any sign of stabilization ^ALF.

Pricing Power Under Pressure

Taken together, these data points paint a concerning picture for AI companies' pricing power. Token prices have collapsed while the spending index has rolled over. Infrastructure spending continues to accelerate. And one of the largest AI infrastructure investors is exploring ways to sell excess compute externally — adding supply to a market already facing pricing headwinds.

Bain & Company provided a useful framing: token costs for AI-using companies are currently only about 1–2% of headcount costs, with some scenarios imagining that rising to 20–30% ^B. That implies AI providers need either much higher usage, stronger pricing power, or both, to convert massive infrastructure investment into durable revenue growth ^ALB.

How the Pieces Fit Together

The core mechanism is this: AI companies spent enormous sums on infrastructure assuming usage and revenue would scale fast enough to justify the buildout. Instead, competition and efficiency gains have pushed token prices down about 90%, the token spending index has rolled over from its May peak, and Meta is exploring ways to sell excess AI compute externally ^RBLFFF.

The market is therefore reassessing whether AI infrastructure spending can be converted into durable, high-margin revenue ^HL.

The Verdict

The Silicon Data LLM Token Expenditure Index rollover is not a definitive signal that the AI boom is over. A softer index could simply reflect a shift toward cheaper models, or a temporary pause as enterprises digest their AI investments ^LF. But when combined with a 90% decline in token prices, a $600 billion revenue gap, and Meta's need to monetize excess compute, the data raises serious questions about whether the current level of AI infrastructure spending is sustainable.

The single most important leading indicator, as one analyst noted, is that quarterly AI revenues only first exceeded quarterly depreciation in Q4 2025 — and that was before the 2026 CapEx wave (which roughly doubles the asset base) hits the depreciation schedule ^U. The industry just crossed the most basic break-even line on its accounting depreciation. The market is now watching to see whether demand can catch up to supply before the next wave of depreciation arrives.