Summary

  • DigitalOcean reported $893.828 million of remaining performance obligations with a 3.7-year weighted-average life. Only $365.901 million, or about 41%, was expected to be recognised within twelve months.
  • At the same date, it disclosed $2.759287 billion of undiscounted fixed payments, primarily for co-location leases not yet commenced, plus $281.566 million for server finance leases that began in July. The $3.040853 billion total is a separate supplier-side clock, not the cost of the RPO.
  • Revenue grew 29% in the June quarter, but cost of revenue grew faster and GAAP gross margin fell from 60% to 55% as expansion costs arrived ahead of new data-centre revenue. Capacity conversion, not contract collection, is now the test.

A customer promise and a landlord promise are not the same asset

DigitalOcean ended June with two unusually large forward-looking numbers. One came from customers. The company assigned $893.828 million of transaction price to services it still had to perform. The other came from the infrastructure it expects to use: $2.759287 billion of estimated fixed payments, mainly for co-location leases that had not started, and $281.566 million of expected server and equipment finance-lease payments that were not yet on the June balance sheet.

It is tempting to put the two sides of that page into one ratio. The capacity payments total $3.040853 billion, about 3.4 times the RPO. That arithmetic is correct and the economic shortcut is not. RPO is future customer revenue measured under revenue-accounting rules. The facility and server figures are undiscounted supplier payments. They cover different assets, start at different dates and run for different periods. They are not matched contract by contract, and neither amount can be subtracted from the other to produce a margin.

The useful comparison is duration. DigitalOcean’s RPO had a weighted-average life of 3.7 years. The noncommenced co-location leases averaged 11.2 years. The server leases averaged 4.9 years. A customer commitment can expire before the building that supported it. A server can remain financed after the model, accelerator or workload mix has changed. The commercial task is to renew, refill or repurpose capacity after the first customer clock ends.

That is a more demanding reading than “bookings are strong”, but it is also more favourable than assuming a long lease is stranded. Infrastructure businesses create operating leverage by committing before all demand is recognised. The question is whether DigitalOcean has bought enough time and flexibility to make the gap productive.

RPO became larger because the commercial contract changed

The $893.828 million RPO was twelve times the $71 million reported for the comparable prior-year quarter. Management said DigitalOcean had signed its first nine-figure annual commitments with leading AI-native customers and extended weighted average contract life from 1.6 years to more than three years.

That is real evidence of customers willing to reserve a material amount of service. It is not automatic evidence of the same amount of new consumption.

DigitalOcean supplies the caveat itself. RPO does not include usage beyond contracted capacity. It can rise when a customer moves from usage-based purchasing to a commitment, even when the change does not create equal incremental revenue. Renewal timing, extra-capacity purchases and contract length all move the balance. The company expected to recognise $365.901 million within twelve months and the remaining $527.927 million later.

The split matters. About 41% of RPO sits inside the near-term recognition window; about 59% depends on later service. A three- or four-year contract can improve planning, support financing and justify an earlier capacity reservation. It also creates a performance obligation. DigitalOcean still has to make the infrastructure available, the customer still has to consume within the commercial structure, and revenue recognition still follows the service.

Individual contract quality is not public. The filing does not name the largest AI customers, allocate RPO among them or disclose every termination, service-credit or minimum-use clause. Nine figures establishes scale, not bargaining power. A large buyer can strengthen visibility and still negotiate price, capacity priority and remedies more forcefully than a fragmented developer base.

ARR is the current quarter repeated four times

Three metrics in the results release look similar and describe different states.

DigitalOcean reported $1.125 billion of ARR. Its definition is simple: multiply total revenue in the latest quarter by four. With June-quarter revenue of $281.184 million, the arithmetic produces $1.124736 billion before rounding. ARR is therefore a run-rate view of recognised activity, not a ledger of contracts that must be performed.

AI Customer ARR was $234 million. That is also an annualisation, and its subject is broader than AI-inference revenue. DigitalOcean includes all revenue from a customer that uses one or more AI or machine-learning offerings, including that customer’s infrastructure and platform services. Management said 85% of AI Customer ARR came from inference and core cloud rather than bare-metal GPU service. The statement supports a broader cloud attachment thesis; it does not reveal pure inference gross margin.

The company also reported $259 million of ARR from customers in its million-dollar category. The underlying customer groups are nested. The 45 million-dollar customers sit inside the 92 half-million-dollar customers and the 632 customers above $100,000. Their reported revenue shares—23%, 26% and 35%—must not be added.

Together the measures describe three stages. ARR shows the annualised pace of the latest quarter. AI Customer ARR shows that pace across accounts touched by the AI portfolio. RPO shows contracted service still to be recognised. None shows whether a particular megawatt is ready or profitable.

The cost arrived before the new rooms filled

The income statement already records the beginning of the capacity mismatch.

June-quarter revenue rose 28.6% to $281.184 million. Cost of revenue rose 44.2% to $126.522 million. Gross profit increased in dollars, but GAAP gross margin fell from 60% to 55%. Operating income declined to $29.371 million from $35.619 million.

DigitalOcean said the margin decline resulted from data-centre expansion costs incurred ahead of the revenue ramp. Compared with the prior-year quarter, depreciation and amortisation associated with infrastructure additions increased by $17.2 million, co-location cost by $12.4 million, third-party licences and partnership expense by $3.4 million, ancillary equipment cost by $2.9 million, and other net cost by $3.2 million.

This is not evidence that the expansion has failed. It is evidence that timing has become economically visible. A new facility starts generating rent, depreciation, power, licence and operating expense before every rack or accelerator carries a mature workload. If demand fills it, fixed cost is spread across more revenue and margin can recover. If commissioning or consumption slips, the same costs remain.

The 55% margin is therefore more useful than a promotional capacity number. It is where customer demand, supplier terms, fleet utilisation, product mix and price meet. DigitalOcean can report higher RPO while margin falls for sensible investment reasons. It cannot establish the return on that investment until utilisation and gross profit follow.

One megawatt does not specify an inference product

DigitalOcean said an incremental 20 MW of committed data-centre capacity was expected online in 2027 and 2028, taking total committed capacity to about 155 MW, with more capacity being pursued.

“Committed” is the important word. It does not say all 155 MW was energised, commissioned, occupied or available to a customer on 30 June. Nor does a megawatt specify GPU type, useful life, cooling efficiency, redundancy, network path, model throughput or the portion reserved for the company’s long-standing CPU, storage and database services.

DigitalOcean’s product design makes the allocation problem richer. Serverless inference pools requests and can share hardware across customers. Dedicated inference gives a buyer more explicit capacity and performance control. Batch inference moves work away from real-time demand. The Inference Router can choose among models and price or latency policies. Core cloud adds compute, storage, databases and network services around an AI account.

Those modes can improve utilisation. Batch jobs can fill quiet periods; routing can avoid an unnecessarily expensive model; core-cloud attachment can raise revenue per account. They also create different obligations. A dedicated endpoint cannot always be treated like a fully shared pool. A serverless service needs spare capacity for bursts. A third-party model route can alter cost without releasing a facility lease.

The commercial proof is not raw MW. It is delivered service at a margin: tokens, endpoints, storage, compute and support sold repeatedly without starving existing cloud customers or holding too much idle hardware.

The balance sheet has been strengthened with external capital

DigitalOcean held $767.026 million of cash and equivalents at June, up from $254.475 million at year-end. The increase did not come only from cloud operations. A March equity offering generated $887.888 million net, and $500 million of term-loan principal was repaid.

For the first half, operating cash was $156.889 million, almost unchanged from the prior-year period despite 26% revenue growth. Property and equipment purchases were $81.576 million, internal-use software spending was $11.785 million, and another $51.544 million of equipment was acquired through financing arrangements with an equal financing inflow.

At June, the balance sheet already carried $921.053 million of current and long-term debt, $577.720 million of finance-lease and equipment-financing liabilities, and $479.087 million of operating-lease liabilities. The $3.040853 billion highlighted here sits beyond those recorded amounts because the relevant facility commitments had not commenced and the specified server leases started in July.

In July, DigitalOcean issued more equity in a direct offering and used the proceeds to repurchase $471.828 million principal of its 2030 convertible notes. This can reduce one form of leverage while increasing the share count. It shows access to capital and active liability management; it does not make future capacity self-financing.

The company’s adjusted free cash flow was about $61 million in the quarter. DigitalOcean expressly warns that this measure is not residual cash available for discretionary use after debt and other obligations. That warning matters when long-lived payments are about to commence.

Four gates stand between a contract and a return

DigitalOcean’s growth case now has to pass four gates in order.

First, the customer contract must remain economically valuable. RPO establishes transaction price allocated to future service, but renewal, remedies, price and incremental consumption determine the quality of that commitment.

Second, capacity must commence and become usable. A signed facility lease is not power, cooling, network readiness or an installed server. A delivered server is not automatically the right accelerator for the workload that arrives.

Third, the customer must consume and DigitalOcean must perform. Only then does RPO become recognised revenue. The $365.901 million expected within twelve months is the first public conversion window.

Fourth, recognised revenue must produce enough gross profit and cash to justify facility, server, software, support and financing cost over their longer lives. The June margin decline shows why this gate cannot be assumed.

The favourable outcome is coherent: facilities arrive on schedule, the new customer commitments consume capacity, gross margin recovers as utilisation rises, cash generation grows, and renewals extend demand beyond the first 3.7-year clock. The adverse outcome has two forms. Capacity can arrive late and prevent DigitalOcean from serving contracted demand, or it can arrive on time and remain underused after customer architecture, model economics or demand changes.

RPO solves neither failure by itself. It gives DigitalOcean a stronger demand signal with which to make a physical decision. The return will be earned in the space between the 3.7-year customer promise and the 11.2-year building promise.

Sources