I ran my citation scraper across the most-shared AI headline of the week. Thirty-one words of headline. Nine words of summary. One number โ $2.5 trillion โ and zero links to a primary source.
The script returned an empty array where the provenance should have been.
I have spent nine years pulling apart contracts that other people were paid to describe, starting with the 0x protocol library in 2017, when I spent eight weeks reverse-engineering their exchange contract while everyone else traded the token. The lesson repeats across every audit since: the most dangerous number in any system is the one nobody can point to. A price oracle that reports a value no participant can attest to is not a price. It is an opinion wearing a decimal point. The $2.5 trillion figure is currently functioning as an unattested oracle for one of the largest capital-allocation debates of the decade, and the headline it sits inside โ "Corporate America faces AI hangover as spending surges to $2.5T" โ is being routed through crypto feeds precisely because crypto readers are the audience most likely to act on it.
That routing is the story. Not the hangover. The pipe.
The Payload, Stated Honestly
The factual core of the piece is thin, and I will state it without decoration. US corporate AI spending is large, growing, and increasingly questioned. Returns are described as limited. Competition is described as intensifying. That is the entire payload. No methodology. No breakdown between capital expenditure and operating expenditure. No definition of whether $2.5 trillion is annual, cumulative, American, or global.
Everything credible in the claim comes from sources the article never cited. The MIT NANDA research on generative AI pilots โ roughly 95% produced no measurable P&L impact โ is the hardest public evidence for "limited returns," and it does not appear. Sequoia's framing of the infrastructure revenue gap, the "$600 billion question," does not appear. The documented collapse in inference pricing, with GPT-4-class capability falling more than 90% in roughly eighteen months, is the cleanest mechanical explanation for the word "competition," and it does not appear either.
So we have a conclusion with no derivation. In my line of work that is not a warning. That is a rumor with a press cycle.
What is verifiable is concentration. Global AI capital expenditure is not a broad corporate phenomenon. It is four balance sheets โ Microsoft, Alphabet, Amazon, Meta โ plus a short tail. Their combined capex moved from roughly $200 billion in 2024 to guidance above $300 billion for 2025 under public reporting conventions. When one buyer is 40% of your order book, sustainability is not a macro question. It is a counterparty question. The article never asks it.
And there is the source problem, which I flag without pretending it settles anything. Crypto Briefing publishes into a market that competes with AI for the same three scarce resources: venture capital, retail attention, and electricity. An "AI hangover" headline is structurally favorable to the asset class the outlet covers. That does not make the claim false. It makes it a claim with a holder. Discount the source, then test the content.
Decomposing the Container
Now let me do what the article did not.
$2.5 trillion is not a unit. It is a container. It holds at least four different things, and each one produces a different conclusion.
Capital expenditure buys assets โ accelerators, buildings, transformers, cooling loops. It hits the cash flow statement and the balance sheet, then drips into the income statement through depreciation across years. Operating expenditure buys capability โ API calls, seats, inference โ and hits the income statement immediately. Infrastructure build-out is largely the former. Enterprise software procurement is entirely the latter. Rolling both into one figure and calling it "AI spending" merges a long-duration asset bet with a quarterly subscription line. The risks are not comparable. A capex pause is a scheduling decision. An opex pause is a verdict.
If $2.5 trillion is annual and global, it describes the largest capital cycle in industrial history. If it is cumulative through the end of the decade, the annual figure is a few hundred billion and the urgency framing collapses. The article offers no way to tell which. A number without a denomination is not evidence. It is decoration.
The question that survives is narrower and harder: AI-attributable revenue across the entire economy has not converged on AI-attributable capex, and that gap is real regardless of what the $2.5T means. That is the audit finding. Not the total. The gap.
On the opex side, the funnel is measurable and ugly. Pilot participation rates among large enterprises sit high, often above 70%. Conversion to production deployments commonly lands under 30%. A pilot that never ships still consumes inference, seats, and integration labor. It simply never generates the P&L line that justifies the renewal. This is the shape of a spending category that can grow while value stays flat: high entry rate, low survival rate, and a renewal cliff that arrives twelve to eighteen months after signature. No headline number captures a renewal cliff.
The Depreciation Clock
Here is the part that actually decides the outcome, and the part no headline carries: the depreciation clock.
Take a representative ten-thousand-accelerator training cluster on H100-class silicon. Chips and networking land near $300 million. The shell โ power distribution, cooling, land, interconnection โ routinely adds $400 to $600 million before a single job is scheduled. Call it $800 million of assets placed in service.
Straight-line over four years, that is $200 million of depreciation per year before a dollar of electricity. Power at roughly 700 watts per accelerator with a realistic PUE around 1.3 puts the load near nine megawatts, which at industrial rates lands around $5 to $6 million annually, more in constrained grids. Add staff, spares, and the egress costs nobody budgets for.
Against that, rental economics. H100-class hourly pricing fell from roughly $8 to $2 across major marketplaces in about two years. At $2 per hour, 24/7, that cluster bills around $175 million annually at perfect utilization. Utilization is never perfect. Real-world model FLOP utilization on production workloads runs far below the theoretical ceiling, and idle accelerators depreciate exactly as fast as busy ones.
Run the arithmetic honestly and the cluster covers its own depreciation somewhere north of 85% sustained utilization, with margin thin enough that one soft year turns the asset into a liability. The asset's life is physics. The revenue's timing is a market. Mismatch them and depreciation does what a liquidator does โ only slower, and better disguised.
This is also where accounting becomes engineering. Several hyperscalers have extended assumed useful server life from three years toward five and six. That revision does not touch a single thermal cycle. It moves reported earnings. When a depreciation schedule is lengthened in the same period that lease terms soften and utilization compresses, the correct forensic response is not outrage. It is a question: what did the operating data show that the schedule does not? Code is law, but bugs are the human exception. Depreciation schedules are code written by humans with a P&L to manage.
Survivability is then a question of resource endowment, and the distribution is brutally uneven. Hyperscalers carry advertising or cloud cash cows that can absorb years of sub-cost returns without external financing. Pure model labs depend on the next round. Application-layer startups depend on the round after that. A sustained return shortfall does not distribute pain evenly โ it removes the financing-dependent tier first, then the mid-tier, then re-rates the survivors on free cash flow rather than on narrative. The hangover, if it arrives, is not a hangover for everyone. For part of the industry it is simply the end of the party.
Where This Touches On-Chain Capital
In 2026 I audited a protocol built for autonomous AI agents executing DeFi strategies. The risk surface was never the model. It was the input. Agent swarms consumed price feeds and acted inside the same block window, and I found a race condition in the validation path where a fast agent could nudge a feed during a high-frequency window and have the manipulation ratified by settlement logic. We built a formal verification model for temporal consistency, and the core team adopted it. The fix was not smarter agents. It was attested inputs.
The AI capital cycle has the same architecture and the same missing attestation. A $2.5 trillion headline with no source is a price feed with no oracle. Readers route capital through it. The ledger remembers what the wallet forgets โ every dollar of capex is a dated commitment that eventually surfaces as either revenue or a write-down, and no press cycle can reclassify it.
Which is why the on-chain market is worth watching here, for a reason most people miss. Tokenized compute markets and DePIN networks publish a live, timestamped, adversarial price for GPU-hours. That is an attestation the $2.5T narrative does not have. When on-chain compute spot prices compress while hyperscaler capex guidance expands, the spread measures where marginal capacity is actually clearing. It is a far better instrument than a headline, and I check it before I check the news.
Three Blind Spots
The metaphor is load-bearing, and it is wrong. "Hangover" is a consumption frame. It implies a binge, a morning after, a return to baseline. Infrastructure does not have hangovers. It has commissioning schedules. Electricity, rail, and the internet all overshot and undershot repeatedly across decades before their returns landed. Judging a ten-year general-purpose technology on a two-year ROI window is not skepticism. It is a units error.
The physical brake makes the metaphor worse. Interconnection queues for large loads in major US markets run three to seven years. You cannot binge on power you cannot connect. That constraint caps the downside in a way a demand-only model misses entirely, because the capex cycle is partially self-limiting while the grid refuses to accelerate.
There is a reflexive blind spot too. A narrative that capital is being wasted reduces the willingness of CFOs to sign pilot expansions. Fewer expansions produce lower measured returns. Lower returns confirm the narrative. The story becomes the mechanism. This is a bug that writes itself into the state, and it is the most plausible path by which an AI hangover becomes true rather than merely reported.
And the third blind spot belongs to my own readers. If your reason to buy tokens is that AI money will rotate into crypto, you are pricing one asset off a feed you have never audited. Capital does not rotate automatically. It rotates when a marginal buyer has a reason to move, and the observable is stablecoin supply and tokenized treasury growth โ not a headline about somebody else's capex.
Ignore the Total. Track the Attestations.
Hyperscaler quarterly capex guidance is the leading indicator, and the first downward revision will be the most informative print of this cycle. Order backlog and lead times at the accelerator suppliers are the second. On-chain compute pricing is the third, and the only one that updates continuously.
Treat the narrative itself as an unverified oracle. Demand provenance before you route capital through it.
The real question is not whether corporate America wakes up with a headache. It is whether you intend to keep pricing a trillion-dollar asset off a feed that has never once published its methodology.