Sovereign AI Compute: How to Finance a GPU Fleet Without Betting the Company

For most of the last decade, cloud spend was an operating expense — a line that scaled with usage, could be optimised with reserved instances or savings plans, and rarely appeared in a board conversation as a strategic risk. That framing broke in 2024 and shattered in 2026. For AI-native companies, compute is now the largest, most capital-intensive, and most volatile input on the income statement, and the finance leaders who continue to treat it as opex are the ones being surprised by their own margins.

The scale of the shift is often obscured by cloud-provider billing structures that spread costs across long-term commitments, private pricing, and burst usage. Netted out, a Series B AI infrastructure company today typically spends 45 to 65 percent of gross revenue on compute — a ratio that would have been considered structurally unfundable in the SaaS era. The businesses that survive the next twenty-four months are the ones that have moved compute out of the P&L abstraction layer and onto the balance sheet, where it belongs, and are financing it with the instruments used to finance infrastructure: project debt, offtake agreements, and residual-value insurance.

The founders who resist this framing usually do so because they associate infrastructure finance with slowness — with lawyers, covenants, and quarterly reporting packages that feel incompatible with an eighteen-month product horizon. The counter-argument is that the dilution being suffered to fund compute out of equity capital is far worse than any covenant, and that the covenants themselves impose exactly the discipline that stops a founder from over-committing to reserved capacity on a demand curve they cannot yet see.

The first is a compute-collateralised term loan issued into a bankruptcy-remote SPV. The parent company contributes equity to the SPV, the SPV takes on senior debt against the GPU hardware, and the SPV licenses compute back to the operating company at a rate that services the debt and generates a modest residual. This is the same structure used to finance aircraft, container ships, and data centres for four decades. The innovation is not the structure; it is the willingness of specialty credit funds to underwrite depreciating semiconductor assets, which arrived in mid-2025 and is now a genuine market with real price discovery.

The second is offtake-anchored financing. If your compute is being consumed by a named enterprise customer under a multi-year contract, that contract can be pledged as collateral in a structure that looks almost identical to a receivables facility. Rates are dramatically lower — we are seeing 8 to 11 percent on senior tranches versus 15 to 22 percent on pure hardware-backed paper — because the credit is on the offtaker rather than on the residual value of the chips. The friction is that the offtake contracts have to be genuinely committed, with take-or-pay language, and most enterprise buyers resist this framing until the fourth or fifth conversation.

The third is sovereign co-investment. The UAE, Saudi Arabia, France, and increasingly India are running programmes that will match GPU capex up to a ceiling — typically 40 to 60 percent — in exchange for local hosting, local hiring commitments, and a right of first refusal on domestic government workloads. The paperwork is genuine and the timeline runs 90 to 120 days from first conversation to first tranche, but the equity dilution avoided by taking sovereign capital instead of a growth round can be the difference between a founder-owned outcome and a founder-diluted one at exit.

Every generation of accelerator introduces the same finance problem: the prior generation impairs faster than the depreciation schedule predicts. The H100s bought at $32,000 per unit in early 2024 are trading in the secondary market at $18,000 to $22,000 in mid-2026, and the utilisation rate on H100 clusters is falling as workloads migrate to H200 and B200 hardware. If your books still carry these units at straight-line depreciation over four years, you are overstating asset value by 25 to 35 percent, and any covenant tied to a leverage ratio is quietly closer to breach than the reported numbers suggest.

The correction is not exotic. Refresh the mark on your GPU fleet monthly using secondary-market comparables, adjust depreciation to a utilisation-weighted curve rather than a time-weighted one, and pre-agree with your lenders a mechanism to re-covenant when the mark-to-book delta crosses a threshold. Lenders who have done this before will accept the mechanism because the alternative — a covenant breach negotiation under duress — is worse for them too. Lenders who resist the mechanism are lenders who do not understand the asset class, and taking capital from them is a strategic error regardless of the headline rate.

The founders who are getting this right are also modelling the option value of a mid-life refinancing. When the next generation of hardware lands and the market discovers that your existing fleet is still 80 percent as productive at 60 percent of the amortised cost, there is a window to refinance the underlying debt at materially better terms. Missing that window because your finance function is still on quarterly cadence is a common and expensive mistake.

The single most expensive commercial mistake we have seen AI-native companies make in 2025 and 2026 is signing multi-year reserved-capacity contracts with hyperscalers on the strength of a demand curve extrapolated from a six-month window. The contracts look attractive on paper — 30 to 45 percent discounts against on-demand pricing — and they solve an immediate supply anxiety. The problem is that they convert a variable cost into a fixed cost at exactly the moment when the trajectory of your revenue is least legible.

We have now seen four separate cases in the last twelve months where a company signed a three-year reserved-capacity deal representing $40 to $120 million of committed spend, watched their end-market flatten six months in, and spent the next eighteen months either paying for compute they did not use or attempting to resell the capacity into a secondary market that did not exist at the price they needed. The equity value destroyed in these situations dwarfs any conceivable discount on the headline rate.

The discipline is straightforward and unpopular: never commit reserved capacity beyond the horizon over which you can defend your demand forecast to a hostile board member. For most Series B and C AI companies, that horizon is six to nine months, not thirty-six. Yes, you will pay more per GPU-hour on the marginal tranches. Yes, this feels like leaving money on the table. The founders who leave that money on the table are the founders who still have a company in 2028.

The pre-work that determines whether a compute-financing conversation goes anywhere in three meetings or stalls for six months is straightforward and finite. First, a fleet inventory: every unit, its acquisition cost, its current book value, its utilisation over the last twelve months, its physical location, and its expected residual under two scenarios. Second, a revenue attribution: what percentage of the fleet is generating revenue under committed contracts, and what percentage is speculative capacity. Third, a scenario model: how the leverage ratios evolve under three demand cases, with sensitivity to the two variables that matter most, which are utilisation rate and end-market price per token.

The tell of a company that is not ready is when any of those three artifacts is more than a week old, or when the numbers in them do not tie to the audited financials without a reconciliation. Lenders are not looking for perfection; they are looking for a finance function that can produce these views in a day when asked, because that is the same finance function that will produce a covenant compliance certificate on time every quarter for the life of the facility.

If you are within six months of a compute-financing decision, the highest-leverage move is to build the fleet-level view now, when the pressure is low, rather than in week three of a term-sheet negotiation. The CapMaven Capital Structure Diagnostic maps the exact structure — SPV, offtake, sovereign, or hybrid — that fits your specific fleet composition and demand profile, and produces the artefact stack that a specialty credit fund will accept without a second data request.