Insights · Market
The Neocloud Reckoning: The AI Boom Has Found Its Cost of Capital
The neocloud selloff has stopped being an equity story and become a credit story. The question the market is now asking is simple: can renting GPUs generate enough cash, soon enough, to service the debt that built the data centers?
From a CoreWeave problem to a sector problem
Innes opens with the tape. CoreWeave fell more than 15% in a single session, and the pain spread across the Goldman Sachs neocloud basket: TeraWulf, Cipher, IREN, Hut 8, Nebius, and Applied Digital all sold off hard. Nvidia, the supplier at the top of the chain, declined far less. His read is that investors have begun treating the stress as structural to the neocloud model rather than as an accident specific to one company.
The revenue gap, in numbers
The essay leans on a set of sell-side and analyst estimates that frame the scale of the problem.
| Metric | Figure | As cited from |
|---|---|---|
| Cumulative revenue needed through 2030 to cover capex, financing, and a 12% ROIC | ~$8.5T | Gartner |
| Cumulative revenue projected through 2030 | ~$3.6T | Gartner |
| Implied gap | ~$4.9T | Gartner |
| AI infrastructure spend potentially exposed to impairment by 2030 | >$3.5T | Gartner |
| Neocloud providers that could stop operating independently by 2030 | >80% | Gartner |
| Capex per gigawatt of AI compute | ~$42B | Goldman Sachs |
| 2028–2030 revenue needed to earn 15% on compute installed by the six largest US spenders in 2026–27 | ~$1.4T | Goldman Sachs |
Innes is careful on the 80% figure: most of those operators, he notes, are more likely to be acquired or merged than to go bankrupt. The point is not collapse but a change of ownership.
Leverage meets a fixed calendar
The sharpest section is on balance sheets. According to the essay, CoreWeave's total debt rose from roughly $15 billion to roughly $52 billion in a year, and quarterly interest expense climbed from about $267 million to about $640 million. D.A. Davidson's Gil Luria is quoted estimating that the company borrows at roughly 9% while earning about 1% on its investment.
Morgan Stanley, meanwhile, counts 20 project-finance-style data center bonds in the high-yield market, totaling about $45.6 billion, backing more than 4 GW of capacity and resting on just eight tenants. That is a lot of fixed obligations concentrated on a small number of counterparties.
The core thesis
Innes's central argument is about timing, not demand. Scarcity pricing for short-term compute will not last. Hyperscalers can afford to wait for AI monetization to mature because they fund buildouts from operating cash flow. Leveraged neoclouds cannot wait, because their debt payments arrive on a schedule that does not care when customers start paying.
“The neocloud has a coupon date.”
Stephen Innes, The Dark Side of the Boom
The likely outcome, in his view, is that the AI boom continues while the financing and ownership of its infrastructure change hands, with the strongest balance sheets absorbing the assets of the weakest.
Our view
Why this matters at the intersection of compute and energy
We read this essay as confirmation of the thesis AICE Capital Partners is built on. Three things stand out.
- The binding constraint is power, and power sets the cost of capital. At roughly $42 billion per gigawatt, a site without firm, affordable energy is a stranded asset with a coupon attached. Projects that solve interconnection and generation up front are the ones that can borrow on sane terms.
- Sequential development is what creates the repricing risk. Operators that bought GPUs first and chased megawatts second are now carrying interest on idle capacity. Developing energy, infrastructure, and compute as one project removes the gap where debt accrues and revenue does not.
- Consolidation is an opportunity for integrated capital. If most neoclouds change hands by 2030, the buyers will be those who can underwrite both the power and the compute. That is the entire stack, and it is where we operate.
The boom is not ending. It is being refinanced. Talk to us if you are building, powering, or funding AI capacity and want a partner that works across all of it.