On a Bloomberg Television segment this month, Citi's Tyler Radke did something sell-side analysts rarely do: he described a mechanism rather than a mood. Microsoft, he observed, folds security software and Copilot into an Office seat priced below $100 a month, while competitors charge several hundred. That is not a product comparison. It is a claim about who controls the invoice.
The headline that followed — AI is splitting software into winners and losers — is accurate and nearly useless. What held my attention was a quieter line near the end, explaining why neoclouds sold off: not weak demand, but bond yields. Follow the money, not the noise. The demand story is loud, quotable, and already priced. The financing story is silent, and it decides what survives.
Radke's core claim is that AI is not creating new software revenue so much as reallocating existing IT budgets. Microsoft wins by bundling — the marginal inference cost absorbed into a subscription the customer already pays, so the perceived price of Copilot is zero. Adobe and smaller application vendors absorb the other side of that ledger. Palantir and Snowflake join Microsoft on the winner list; semiconductors, he notes, have seen their worst performers outrun software's best. Salesforce, after roughly six large acquisitions, is now judged on whether it can accelerate organic growth rather than on what it builds.
Two caveats belong up front, because I spent years auditing documents that looked exactly like this one. The item circulated through a crypto-native outlet republishing a television interview, with the transaction details — a reported $47.50-per-share cash acquisition of Braze — unverifiable against any filing I can reach. A claim that Salesforce rose roughly 20% after its last earnings also does not match the price action I remember. Neither breaks the argument. Both should be checked against an 8-K before anyone sizes a position.
But the argument does not need those numbers. It rests on something more durable: the divergence between where AI demand is contracted and where AI capacity is financed.
Three business models are being flattened into one word. Microsoft sells a platform tax — distribution across hundreds of millions of Office seats, with AI as a mechanism to raise average revenue per user. Palantir sells forward-deployed engineers and deep institutional integration, which means revenue recognized project by project, lumpy and renegotiated. Snowflake sells metered data gravity. Their moats have almost nothing in common, and treating them as a single "AI winner" basket is a category error that will be paid for in 2027.
The genuinely instructive case is the neocloud. Radke reports 2026 capacity sold out. That single fact carries three meanings, and the coverage carried one.
Sold out means demand has contractual backing, which is a real rebuttal to the "AI is a hallucination" camp. Sold out also means growth is capped by capital: if you cannot finance new capacity at an acceptable rate, the 2027 revenue curve flattens by arithmetic, not by competition. And sold out means the price on those contracts was locked at a cyclical high, before anyone knew what a GPU's economic life actually is.
Put those together and you get the structural flaw almost no one is pricing: these firms are issuing high-coupon debt to buy rapidly depreciating assets in order to service long-dated fixed-price contracts. That is duration mismatch wearing a growth multiple. When I reverse-engineered payment protocols in 2017, the failures I documented were rarely bad code. They were governance structures that assumed a refinancing would always be available. The mechanism has not changed; only the collateral has.
The depreciation assumption is the part I would put in bold on any risk committee memo. If a GPU's true economic life is shorter than the three-to-six-year accounting schedule — and the pace of architectural turnover suggests it may be — then reported profit across the entire AI infrastructure layer is systematically overstated. That is not a bear thesis. It is an accounting question with a date attached, and it will surface in an earnings season, not in a think piece.
Then there is the loop. Chips fund model labs, model labs buy cloud, cloud anchors neoclouds, neoclouds order more chips. A portion of "demand" is manufactured by the investment that services it. I watched the same circularity in 2020, when yield-farming incentives were counted as organic TVL, and I watched it unwind in eleven days. My stablecoin research that year tracked remittance corridors in Latin America where a two-cent peg deviation meant a family's grocery budget. Abstract liquidity is never abstract at the endpoint.
For crypto readers, the translation is direct. The AI-token complex — GPU DePIN, inference marketplaces, agent-payment layers — is financed by the same credit conditions and priced against the same depreciation assumptions. When I designed an on-chain verification framework for AI-generated content this year, the hard problem was never cryptographic. It was economic: who pays for continuous attestation, and what happens to that budget when the cost of capital doubles?
Beneath all of it sits a pricing shift the software commentary keeps missing. AI agents are priced per task or per outcome; software has been priced per seat. If a single assistant replaces four subscriptions, the vendor's revenue model does not compress — it inverts. The same inversion is coming for token "utility" designs built on the assumption that more users means more fees.
The consensus narrative is that AI is eating software. I think that misassigns the cause.
Radke himself supplies the correction without noticing: when capital is expensive, large incumbents outlast small ones. That is a statement about interest rates, not about model capability. AI is the accelerant; the high-rate regime is the fire. If credit spreads compress by a hundred basis points, a meaningful share of today's "structural divergence" converges within two quarters — and the analysts who called it technological destiny will call it a rotation.
The second blind spot is the bundler's self-harm. Embedding Copilot into an enterprise tier raises revenue per user this year and teaches the entire market that AI should arrive free with something else. Microsoft is simultaneously the largest beneficiary and the largest destroyer of the industry's AI pricing ceiling. The European scrutiny over Teams bundling was a preview; the Copilot version is larger and, so far, unexamined.
Volatility is the tax on impatience. The impatient are currently paying it in the wrong asset — buying demand visibility while ignoring solvency plumbing.
The signal to watch this cycle is not a benchmark score or a model release. It is the high-yield spread, the take-or-pay ratio inside neocloud contracts, and the depreciation footnote in the next 10-Q. Every cycle has a lender, and this one has not yet been asked to testify.
The question worth sitting with: when the financing leg tightens, does the market discover that AI demand was real all along — or that it was simply well-funded?
