The Contracted Backlog & Switching CostsNarrow moat

If CoreWeave has anything resembling a durable moat, it is here: the enormous contracted backlog and the switching costs that come with running customers' AI workloads. CoreWeave has signed long-term, take-or-pay commitments that had reached roughly $99 billion by early 2026 — up from around $67 billion just a few months earlier¹ — from customers including OpenAI, Microsoft, and Meta. This backlog is the strongest single argument that CoreWeave is more than a momentary beneficiary of a compute shortage: it represents years of contracted future revenue, giving the business a visibility and a foundation that a pure spot-market reseller would lack. Combined with the real friction customers face in moving established AI workloads, the backlog is the sturdiest element of CoreWeave's thin moat — and, in the same breath, its most concentrated and most scrutinized.

The backlog's value is genuine. Take-or-pay contracts commit customers to pay for capacity whether or not they use it, converting CoreWeave's business from uncertain spot demand into contracted, multi-year revenue and underpinning the debt financing that funds the GPU fleet. A backlog approaching $100 billion, against annual revenue still in the single-digit billions, implies years of growth already committed, and it is the basis for the bull case that CoreWeave will grow rapidly and predictably as it converts those commitments into revenue. Switching costs reinforce the hold: once a customer has its models, data, and workflows running on CoreWeave's infrastructure, moving them elsewhere involves real cost, disruption, and risk, so contracted customers tend to stay through the contract term.

But the reservations are substantial, and they are why even this, the best of CoreWeave's advantages, supports a thin moat. First, the backlog is extraordinarily concentrated: it leans on a small number of enormous customers — Microsoft, which was roughly two-thirds of 2025 revenue, plus OpenAI and Meta — whose commitments dominate the total, so the visibility it provides is only as reliable as those few counterparties. Second, backlog is a promise, not cash: it must be converted into revenue over years, and it carries execution risk (CoreWeave must build the capacity to fulfill it), credit risk (the counterparties must remain willing and able to pay), and the ever-present possibility of renegotiation as the market shifts. Third, several of those counterparties are entangled with CoreWeave through the circular-financing web, and one — OpenAI — is itself a heavily loss-making company whose ability to honor tens of billions in commitments depends on its own uncertain funding.

The switching costs, too, are real but weaker than they appear. AI workloads are increasingly built on portable software layers, and the largest customers — the hyperscalers and best-resourced labs — are precisely the ones most able to move workloads or to build their own capacity when contracts expire. The lock-in holds within a contract term but reopens at renewal, when a customer that has built its own infrastructure or found a cheaper alternative is free to leave.

So the contracted backlog is CoreWeave's most credible claim to durability — a genuine, enormous, multi-year book of committed revenue that distinguishes it from a pure spot reseller and underpins its whole financing model. But it is concentrated in a few powerful, entangled, and in one case financially fragile counterparties; it is a promise that must be executed and could be renegotiated; and the switching costs behind it are real but cyclical. It is the sturdiest part of a thin moat, the place an investor should look hardest for evidence of durability — and the place where concentration and counterparty risk are most acute.