Summary

  • Palantir committed to spend at least US$5.6 billion under an amended third-party cloud agreement across ten contract years through 29 February 2036. Annual minimums range from US$268 million to US$979 million, but the exact yearly sequence is not public.
  • The amendment terminated the agreement’s previous payment obligations. It therefore replaces the predecessor minimum of at least US$1.95 billion through September 2033; the two headlines must not be added or treated as a recognised debt refinancing.
  • Q2 cost of revenue rose US$104 million year over year, primarily including US$89 million more third-party cloud-hosting cost. Revenue grew 93% to US$1.935 billion. The durable test is whether supplier consumption keeps producing customer value without eroding margin or architecture choice.

The agreement replaced a clock rather than adding a second one

Palantir’s quarterly filing contains one sentence that changes the duration of its infrastructure economics. In March 2026, the company amended an unnamed third-party cloud-services agreement and committed to spend at least US$5.6 billion over ten contract years through 29 February 2036. The annual minimums range from US$268 million to US$979 million.

Those endpoints describe a corridor, not a schedule. Dividing the total by ten produces a simple average of US$560 million a year, but Palantir did not disclose an even payment path. Nor did it publish which year carries US$268 million, which approaches US$979 million, what usage qualifies, how prices change with volume, whether a shortfall creates a credit, or what termination would cost.

The contract also has a reset clause with analytical consequences. Palantir says all previous payment obligations related to the agreement ended when the amendment was signed. Its 2025 annual report had described the predecessor as at least US$1.95 billion over ten contract years through September 2033. At year-end, US$79.2 million of a US$170.2 million minimum for the contract year ending September 2026 had been satisfied.

The new US$5.6 billion headline is about 2.87 times the old US$1.95 billion headline. That arithmetic does not establish 2.87 times more capacity, usage or price. The terms and horizon changed together. The old and new promises should not be added, because the filing expressly says prior payment obligations were terminated. “Terminated” also does not prove that Palantir received forgiveness of a recognised liability or booked an accounting gain. The disclosure describes an executory purchase contract, not a debt exchange.

The first observable bill is already in cost of revenue

The supplier’s identity is not the most important missing name. The more useful question is where consumption appears in Palantir’s operating accounts. In the second quarter, cost of revenue increased by US$104 million from a year earlier. Palantir attributed most of that movement to an US$89 million increase in third-party cloud-hosting services, alongside higher stock compensation and lower subcontractor expense.

That figure does not belong entirely to the amended agreement. It covers third-party hosting services in aggregate and may include providers, products, workloads and usage periods outside the March contract. It nevertheless establishes the mechanism: cloud is not an abstract off-balance-sheet promise. As workloads run, hosting enters the cost of delivering Palantir’s software and services.

Revenue grew faster. Q2 revenue rose 93% to US$1.935 billion. U.S. commercial revenue reached US$764 million and U.S. government revenue US$809 million. GAAP operating income was US$912 million, a 47% margin. On the current evidence, higher cloud cost has not prevented exceptional operating leverage.

That is a starting observation, not a ten-year verdict. The amendment reaches 2036, while customers, product configurations and model workloads can change much faster. A contract year can be economically successful even when its minimum is large if qualifying use supports products that customers adopt and renew. It can be wasteful even when the minimum is met if low-value workloads are run merely to clear the floor.

The correct numerator is not only dollars consumed. It is productive, eligible consumption. The denominator is not only supplier price. It includes customer revenue, gross-profit dollars, deployment quality and the cost of preserving alternative architectures.

Customer obligations cannot be poured into the supplier account

Palantir reported US$4.9 billion of remaining performance obligations at 30 June. It expects approximately 43% to become revenue within 12 months, 36% in the following 13 to 36 months and the remainder later. Contract liabilities were US$1.1 billion, and US$604 million of first-half revenue had been in the opening contract-liability balance.

These are customer-side measures. RPO is revenue expected from contracted performance obligations that remain unsatisfied. Contract liabilities generally reflect consideration received or due before the related performance is complete. Neither is a bank account dedicated to cloud purchases.

The US$5.6 billion supplier minimum is US$700 million larger than quarter-end RPO, but that subtraction has no coverage meaning. One measure spans ten contract years of purchases; the other is a point-in-time customer-revenue ledger shaped by contract duration, cancellation provisions and revenue-recognition rules. Future customer contracts can enter RPO after the cloud amendment is signed. Existing RPO can convert long before the final supplier year.

Keeping the ledgers separate does not make them unrelated. It makes the bridge testable. Palantir needs customer revenue and cash generation to outrun the cost of useful infrastructure over time. Investors need to see hosting-cost growth beside revenue, gross and operating margin, not a manufactured ratio between incompatible headline balances.

A long runway can become a long architecture boundary

A ten-year minimum can improve supplier economics. It may secure discounts, predictable availability, engineering coordination or a foundation for rapidly expanding workloads. Palantir’s growth and profitability give it more room than a weak buyer to absorb uneven annual floors. At quarter-end it had no debt outstanding under its revolving facility and US$500 million of undrawn commitments.

Duration also transfers leverage. The provider controls availability, technical interfaces and parts of the price and service boundary. Palantir controls how workloads are designed, which tasks run where, how efficiently software uses infrastructure, and how cloud cost is reflected in customer pricing. Customers decide whether the resulting applications remain worth buying.

The filing does not say that the agreement reserves physical GPU capacity. It does not identify a cloud, disclose workload allocation or state that every Palantir deployment runs under one provider. Those absences matter because supplier concentration cannot be measured from the minimum alone.

The practical exit question is architectural. If products and operational processes become built around provider-specific services, migrating a workload may require more than moving data. It can require rebuilding identity controls, networking, observability, deployment tooling and performance assumptions. The commitment can therefore be both a purchase obligation and an incentive to make the chosen path work.

Sources