DigitalOcean’s $894 Million Backlog Rewrites the Old Cloud

DigitalOcean’s remaining performance obligation jumped to $894 million in the second quarter, more than 12 times the year-earlier book, as AI customers signed multi-year deals. Revenue rose 29% to $281 million, and the company raised its 2026 sales outlook after the August 4 report. Those contracts stretch average life to 3.7 years and pull the old droplet vendor into the same capacity race as larger clouds.

An $894 Million Backlog With a 3.7-Year Life

The figure that changed the shape of the quarter is not the $281 million of sales. It is the remaining performance obligation of $894 million, up from $71 million in the second quarter of 2025. Of that book, $366 million is due to be recognized over the next 12 months, which leaves $528 million sitting further out.

A year ago the typical contract still ran 1.6 years. Chief Financial Officer Matt Steinfort told analysts the new average life is 3.7 years, and he warned the RPO line will be lumpy as more of those deals land. Annual run-rate revenue closed at $1,125 million, up 29%, so the unearned book is now in the same neighborhood as a full year of sales.

This traction is evident in the material commitments we secured during the quarter, which collectively increased remaining performance obligations to $894 million, up more than 12x year-over-year with a 3.7-year average life. While changes to RPO will be lumpy, these commitments add visibility, and we expect to secure more of them in the future.

Matt Steinfort, Chief Financial Officer, Q2 2026 earnings call

On July 7, weeks before the print, DigitalOcean had already told investors it expected RPO above $800 million after signing multiple nine-figure annual commitments for inference and cloud products. The August 4 number cleared that preview. Remaining performance obligation is not cash in the door. It is contracted work the company has not yet recognized, and it does not capture usage beyond committed capacity.

THE CONTRACT BOOK VERSUS A YEAR AGO

The backlog is the tell. A self-serve cloud that billed by the month does not print a 12-times jump in unearned contracts unless the customer has stopped acting like a hobbyist spinning up a droplet.

The Inference Engine Turned Tokens Into a Sales Channel

DigitalOcean launched its Inference Engine in late April as part of what it now calls an AI-native cloud. Chief Executive Padmanabhan Srinivasan, who goes by Paddy, said more than 6,000 customers had used it by quarter-end, customer count grew close to 60% month over month, and token volume rose 30 times over 60 days. Open-weight models went from about 15% of token volume at launch to close to 75%.

INFERENCE ENGINE THROUGH QUARTER-END

• Customers: More than 6,000 have used the engine since the late April launch.
• Token volume: Early customers lifted consumption about 30 times in 60 days.
• Open weights: Share of tokens climbed from about 15% at launch to close to 75%.
• OpenRouter traffic: The company said it serves more than 20 billion tokens a day on that gateway, up more than 330% over 60 days.

Inference services, the non-bare-metal slice of the AI stack, grew close to 800% year over year and now make up over 70% of AI customer ARR. Srinivasan said 85% of the $234 million in AI customer ARR, itself up 212%, comes from inference and core cloud rather than rented metal. About 1,400 inference customers already send traffic through the Inference Router, which picks models by cost, latency, and task instead of a hardcoded endpoint.

The product page sells serverless, batch, and dedicated inference behind one OpenAI-compatible API, plus a catalog of 70-plus models. Serverless is prepaid; if the balance hits zero, requests stop. Batch jobs promise results inside 24 hours at up to 50% less than real-time. Dedicated endpoints bill by GPU-hour, with an NVIDIA B300 eight-pack listed at $83.10 an hour on the public price sheet.

THREE WAYS THE ENGINE RUNS A MODEL

• Serverless: Pay-as-you-go tokens for live APIs and agents, billed against a prepaid balance.
• Batch: Asynchronous jobs with a 24-hour window, isolated from production rate limits.
• Dedicated: Reserved GPUs, bring-your-own models, and scaling without a Kubernetes project.

Srinivasan framed the pitch as a shift he called valuemaxxing: the right model at the right cost for each task, measured in business outcomes per dollar. Tokenmaxxing, throwing the largest frontier model at every request, inflates the bill until tokens become cost of goods. On the call he put it in one line: the model may be free, but making it useful and serving it well is the product. That is the software layer he is trying to sit on top of the megawatts.

The Kimi K3 launch on July 27 showed how the catalog is being used as a funnel. DigitalOcean said it was the only full-stack cloud to ship as a day-zero partner and picked up over 400 net new customers in the first week. Fourteen day-zero launches have landed since April. The thing builders keep asking for is not another GPU SKU. It is the ability to run an agent for hours without standing up servers, which is exactly what the engine is priced and packaged to do.

Who Is Writing the Nine-Figure Checks?

The growth is concentrated. ARR from customers spending $1 million or more reached $259 million, up 214%, and now accounts for 23% of total ARR, versus 9% a year earlier. The count of those accounts rose 73%. Revenue from the $500,000-plus tier grew 160% and is 26% of sales, while the count of those accounts rose 35%. The $100,000-plus group is 35% of revenue after 98% growth, even though the number of those accounts rose only 9%.

WHERE THE SPEND IS CONCENTRATING

Srinivasan said the higher the spend, the faster the cohort has grown, for eight quarters in a row. Record incremental ARR of $93 million, up 191%, came with that mix. About 70% of AI customers that already spend $100,000 or more have attached a core cloud product to the workload, which is the flywheel the company wants: inference first, then storage, databases, and agents on the same bill.

The book is still spread. The top 25 customers were only 20% of ARR, and Steinfort said that share will rise only modestly as the nine-figure deals ramp. These are not brick-and-mortar enterprises buying a seat license. Steinfort described founders who are deeply technical and who jointly develop with DigitalOcean’s engineers. Net dollar retention hit 102%, a three-year high, and the company is dropping it as a headline metric because new AI logos, not expansion inside a stable SaaS base, now set the growth rate.

Named design partners on the call included OpenCode, which Srinivasan said has more than 7.5 million monthly active developers and now uses the Inference Engine as well as Droplets; Daytona, running sandboxes for AI-generated code; Vercel, wiring the engine into its AI gateway; and OpenRouter, used as both a traffic dial and a customer-acquisition channel. More than 680,000 customers sit on the platform. The dollars, though, are migrating to a few thousand AI-native accounts that sign for years.

155 Megawatts, Most of It Already Spoken For

Long contracts only work if the power shows up. DigitalOcean added incremental 20 megawatts of capacity due in late 2027 and early 2028, taking committed capacity to about 155 megawatts. Srinivasan said the majority of that total should be online by the end of 2027, and the company is still hunting for more.

THE MEGAWATT BUILD

1. First quarter of 2026: Richmond data center launches ahead of plan, and the company lifts committed capacity to about 135 megawatts after securing roughly 60 megawatts across four sites for 2027.
2. Second quarter of 2026: Kansas City, about 10 megawatts, launches ahead of target. Most new 2026 capacity is already allocated to named customers or the token fleet before the halls go live.
3. July 7, 2026: Management announces another 20 megawatts for late 2027 and early 2028, taking the committed total to about 155 megawatts.
4. Second half of 2026: Memphis, the remaining 15 megawatts of the 2026 program, stays on the second-half schedule.

Srinivasan’s line on the call was that software makes megawatts more valuable. Steinfort went further on the math. He said incremental ARR per megawatt should rise as inference and core-cloud attach replace bare metal, even though the old general-purpose cloud once cleared more than $22 million of ARR per megawatt. Newer NVIDIA and AMD kits cost more per megawatt and push more tokens through the same hall, so both the bill and the revenue potential go up.

List prices on several GPU generations rose about 30%. Steinfort said a lot of that had already been worked into renewals or by pulling capacity from customers who would not pay up and sending it to the token fleet. Srinivasan said the $93 million of new ARR in the quarter had very little help from that price move. The 2026 guide now bakes the higher list in, which is one reason the company thinks it can exit the year at 35% growth or more.

The constraint is time, not demand. Steinfort told Bank of America’s Wamsi Mohan that 2027 revenue is still tied to the exact dates halls turn on, which is why formal 2027 guidance is waiting. The directional message was upside to the prior “50% or more” 2027 growth comment, sitting on a 155-megawatt map that does not fully light up until next year.

Operating Income Fell While Revenue Accelerated

GAAP did not celebrate with the backlog. Net income attributable to common stockholders was $35 million, down 4%, for a 13% net margin. Operating income was $29 million, down 18%, for a 10% operating margin. Adjusted operating income rose 9% to $67 million, but the adjusted operating margin slipped to 24% from 28% a year earlier. Adjusted EBITDA was $114 million, up 27%, and the 40% margin is the number management keeps pointing at.

Cash followed the same split. Operating cash rose to $110 million, yet the margin on that cash narrowed to 39% from 42%. Adjusted free cash flow rose only to $61 million, a 22% margin versus 26% a year earlier. Trailing-twelve-month adjusted free cash flow margin is 17%. For the full year the company now guides adjusted free cash flow margin of 11% to 13% of revenue, well below the second-quarter rate, as the 2026 halls and the 2027 kit show up in the cash flow statement before they show up in sales.

In July the company retired about $472 million of its 0.00% convertible senior notes due 2030, funded by a concurrent registered direct offering. Management said the share issuance is offset by the retired notes and by an intended repurchase of about 500,000 shares. Steinfort put pro forma net leverage at 0.7 times after that equitization. Full-year diluted weighted average shares are guided at about 122 million to 123 million; the third-quarter range is 126 million to 127 million. The convert clean-up cuts future debt risk. It does not reverse the fact that a droplet business is now financing GPUs like a utility.

Full-year revenue is now $1.170 billion to $1.180 billion, up 30% to 31%, with adjusted EBITDA margin of 38.5% to 39.5% and non-GAAP diluted earnings of $1.35 to $1.40 a share. Srinivasan said he wants 35% or more growth by the fourth quarter and still sees a path to more than 50% in 2027. Adjusted EBITDA is holding near 40% in the quarter. GAAP operating income is not, and the free-cash-flow guide for the year is the bill for the megawatts.

After the Print, Cheaper Cache Reads and a Sold-Off Stock

The August 4 beat did not hold as a victory lap in the tape. Shares sold off in the weeks after the report even as the company kept shipping models on day zero, which is the tell that investors are underwriting the mix shift and the capex, not the 29% growth rate. DigitalOcean was added to the Russell 1000 in the same quarter. Index inclusion does not fund a data hall.

On September 2, Srinivasan posted that Anthropic’s Claude Fable 5.1 was live on the Inference Engine the day it shipped. Cache reads now cost $0.25 per million tokens, 75% less, while input and output stay at $10 and $50. He argued that a long-running agent spends most of its tokens re-reading its own context, so that cut lowers total cost about 25% on typical workloads and up to 45% on context-heavy agent jobs. The Inference Router is cache-aware: moving a session with 90,000 of 100,000 tokens already cached onto a “cheaper” model can cost 1.6 times more because the new model re-processes the pile.

https://x.com/paddix/status/2094967671661985799

That post is the second-order story in miniature. The company is no longer trying to win on the sticker price of an H100. It is trying to win on routing, caching, and synthesis so that a nine-figure inference customer can stay for 3.7 years without lighting its own cluster. GLM-5.3 and GLM-5.3-Flash followed in late August. Managed Agents Runtime Services, which Srinivasan described as Firecracker microVMs that resume in as little as 200 milliseconds, is in private preview. An AI Builder Summit is set for October 13 in San Francisco.

Third-quarter revenue is guided at $304 million to $307 million, up 32% to 34%, with adjusted EBITDA margin of 38% to 39% and non-GAAP earnings of $0.28 to $0.30 a share. The contracted book is already larger than a year of the old cloud. The next test is whether Memphis, the 2027 megawatts, and those nine-figure customers convert the $528 million that sits past the next 12 months without taking another bite out of the operating line.

Disclaimer: This article is news reporting and analysis of DigitalOcean’s second-quarter 2026 results and related product updates, and it is for information only. It is not investment advice, a recommendation to buy or sell DOCN or any other security, or a forecast of future returns. Readers should consult a licensed financial adviser or other qualified professional who can review their own holdings, tax position, and risk tolerance before acting on any figure in this piece. Revenue, margins, remaining performance obligation, share count, and guidance reflect company statements as of the August 4, 2026 earnings release and may change with later filings and results.