Neoclouds: The Debt-Fueled Business of Renting GPUs

Strip the branding and a neocloud is a leveraged asset business, closer to aircraft leasing than to software:

• Buy accelerators with borrowed money — often loans collateralized by the GPUs themselves plus assignment of customer contracts. CoreWeave pioneered multi-billion-dollar GPU-backed facilities; the model spread across Nebius, Lambda, Crusoe, and dozens of smaller players.
• Pre-sell capacity on multi-year take-or-pay contracts to a concentrated customer set — AI labs, hyperscalers overflow-buying, and enterprises locked out of allocations. CoreWeave's reported backlog ran $55–66 billion in 2026, anchored by OpenAI commitments and a $14B Meta deal.
• Race depreciation. The core wager: rent out each GPU generation profitably before the next one (Blackwell → Rubin, shipping H2 2026) crushes its market rate. Accounting lives depreciate over ~5–6 years; market rental rates for a given generation historically decay much faster once successors land. Every quarter of deployment delay is pure margin erosion.

The numbers that define the category leader, from Q1 2026 filings: revenue $2.08B (+112% YoY); ~$24.9B total debt; quarterly interest expense of $536M — roughly 46% of adjusted EBITDA; 2026 capex guidance of $31–35B against $12–13B of expected revenue. That last ratio is the whole business model in one line: spending nearly 3x revenue on new capacity, funded by debt and pre-sold contracts, in a bet that demand outruns depreciation. It is not a scam; it is infrastructure finance at maximum aggression — the railroad model with silicon rails and a two-year technology clock.

The financial architecture of the GPU boom has a loop in it, and naming it plainly matters because loops amplify both directions:

• Nvidia invests in neoclouds (and AI labs) that buy Nvidia chips. Its equity stakes, capacity guarantees, and supply allocations to CoreWeave and peers mean the chip vendor participates in financing its own demand — the pattern rhymes with Lucent and Nortel's vendor financing circa 1999, the canonical cautionary tale, though today's version is equity-heavy rather than receivables-heavy, an important softening.
• Customer concentration is extreme. A neocloud's revenue typically hangs on a handful of counterparties — for CoreWeave, the top two have exceeded 70% of revenue in some periods, and much of the industry backlog chains back to a single company's (OpenAI's) own famously pre-profit spending commitments. Backlog is only as good as the counterparty's funding.
• Hyperscalers are customers and competitors simultaneously — renting neocloud capacity today while building their own (and custom silicon) for tomorrow. The neocloud thesis requires the capacity shortage to persist; the customers are spending $725B this year to end it.

Mid-2026 gave the thesis its first real stress test: CoreWeave's stock fell ~48% from its highs as investors repriced interest burden and Rubin-transition risk, even as SoftBank launched a competing neocloud venture — capital still pouring into the category while the market marks down its economics. Both things being true at once is exactly what a maturing infrastructure cycle looks like.

The neocloud model becomes intuitive when you run it like a loan-financed vehicle purchase, because structurally it is one:

A GPU server as a financed asset (illustrative):
Purchase price (8-GPU node, installed):     $350,000
Financing: 5 years @ ~9% (GPU-backed debt)
Monthly payment (EMI):                      ~$7,270
Rental income @ $2.20/GPU-hr, 85% util:     ~$10,900/mo
Power, space, ops (~25%):                   -$2,700/mo
Net margin per node:                        ~$900/mo  (thin!)
→ At 65% utilization: net margin ≈ -$1,600/mo (underwater)
→ If next-gen chips cut market rates 30%:   deeply underwater

Three lessons fall out of the arithmetic, and they generalize to any leveraged-asset business:

• Utilization is everything — the gap between 85% and 65% booked hours is the gap between a business and a bankruptcy. This is why take-or-pay contracts (customer pays whether they use it or not) are the industry's load-bearing legal instrument.
• The interest rate is a first-order input. At 46% of EBITDA, CoreWeave's interest line means its true competitor isn't just other clouds — it's its own cost of capital. Every 100bps of rate movement swings the model more than most operational decisions.
• Technology transitions are refinancing events. When Rubin ships, every Hopper- and Blackwell-backed loan in the industry gets an implicit mark-to-market. Lenders know this, which is why GPU debt carries premium rates despite investment-grade counterparties — the collateral has a half-life.

For companies buying compute, the neocloud era created a genuine three-way choice:

• Hyperscaler (AWS/Azure/GCP): highest unit price, deepest integration, easiest procurement, elastic. Right for spiky workloads and anyone whose data already lives there.
• Neocloud: typically 30–60% cheaper per GPU-hour for committed volume, bare-metal performance, fastest access to new silicon. Right for sustained training/inference at scale — with counterparty diligence now a real line item (who financed the cluster you're renting, and what happens to your reservation in their restructuring?).
• Own: lowest theoretical cost at high, steady utilization — but you inherit the depreciation race, the power procurement problem, and the ops discipline. The break-even against renting typically requires 60%+ sustained utilization for multiple years, which most companies overestimate having.

The uncomfortable meta-lesson of the neocloud phenomenon: in an infrastructure boom, the scarcest skill is not deploying capital but pricing risk. The same GPU generates a fortune at 85% utilization on pre-sold contracts and a write-down at 60% on spot — and the difference was decided in the financing structure before the first rack powered on. That's as true for a company weighing a $350k node as for a $25B-debt neocloud; the EMI math is identical, only the zeros change.