디지털오션(DOCN) 2026년 3분기 실적 발표회: AI 성장 및 전망치 상향

핵심 요약

• 디지털오션의 2026년 3분기 매출은 3억 2,100만 달러로, 전년 동기 대비 40% 증가하며 5분기 연속 성장이 가속화되었습니다.
• 회사는 사상 최대인 1억 6,000만 달러의 순증 ARR을 추가했습니다. 총 ARR은 약 12억 8,000만 달러에 달해 전년 동기 대비 40% 증가했으며, AI 고객 ARR은 288% 늘어난 3억 7,600만 달러를 기록했습니다.
• 수익성은 견조하게 유지되었습니다. 조정 EBITDA는 1억 4,400만 달러로 45%의 마진을 기록했으며, Non-GAAP 희석 EPS는 0.45달러를 나타냈습니다.
• AI 추론 토큰 처리량은 전분기 대비 약 10배 증가했습니다. 경영진은 AI 고객 ARR의 90% 이상이 추론 서비스와 코어 클라우드에서 창출되었으며, 추론 서비스는 전년 동기 대비 600% 가까이 성장했다고 밝혔습니다.
• 디지털오션은 실적 전망을 상향하여 2026년 4분기 매출 성장률 44%~45% 및 2026년 전체 성장률 약 34%를 제시했습니다. 경영진은 이제 2027년 매출 성장률이 최소 60%에 이를 것으로 예상합니다.
• 회사는 분기 말 기준 75메가와트(MW)를 약간 상회하는 가동 용량을 보유했으며, 2027년 예정된 약 75메가와트 및 2028년 30메가와트를 포함해 약 180메가와트의 총 약정 용량을 확보하고 있습니다.

주요 재무 데이터

GAAP 순손실에는 5,700만 달러의 전환 유도 비용(주당 0.46달러)이 포함되었으며, 이는 디지털오션의 2030년 만기 전환사채 일부 매입과 관련이 있습니다. 경영진은 이 일회성 비용을 제외했다면 GAAP 순이익이 흑자였을 것이라고 말했습니다.

고객당 지출 규모가 큰 고객들이 성장을 지속적으로 견인했습니다. 연간 10만 달러 이상을 지출하는 고객의 ARR은 133% 증가해 전체 ARR의 41%를 차지했습니다. 연간 100만 달러 이상을 지출하는 고객의 ARR은 273% 증가해 전체 ARR의 30%를 차지했습니다.

사업 및 영업 실적

디지털오션은 이번 분기의 매출 호실적이 빠르게 확장하는 추론 서비스, AI 워크로드와 연계된 코어 클라우드 사용량 증가, 견조한 일라스틱 컴퓨트(elastic-compute) 가동률, 단기 용량 가격 상승, 에이전트 제품의 초기 도입 덕분이라고 설명했습니다. 경영진은 멤피스 데이터 센터가 예정보다 몇 주 일찍 개설된 데 따른 영향은 일부분에 불과하다고 말했습니다.

토큰 처리량은 전분기 대비 약 10배 증가했습니다. 디지털오션은 지난 4월 이후 37개 모델을 출시했으며, 4,000개 이상의 고객사가 추론 라우터를 사용했습니다. 회사는 OpenRouter 상의 주요 오픈 웨이트(open-weight) 모델 분야에서 상위 추론 제공업체로 꾸준히 이름을 올리고 있으며, 하루 평균 5,000억 토큰을 훌쩍 넘는 처리량을 기록하고 있다고 밝혔습니다.

회사는 또한 매니지드 에이전트, 에이전트 학습 서비스, 에이전트 데이터 서비스를 출시했습니다. 이들 제품은 상용 에이전트 워크로드를 위해 설계된 보안 런타임 환경, 도구, 메모리, 스토리지 및 데이터베이스를 결합한 것입니다. 경영진은 65일 동안 70개 이상의 플랫폼 리리스를 진행했다고 보고했습니다.

이번 분기 동안 플랫폼 채택이 더욱 광범위해졌습니다. 4개 이상의 제품을 사용하는 고객이 전체 매출의 약 60%를 창출했으며, 여기에는 연간 10만 달러 이상을 지출하는 대부분의 AI 고객 계정이 포함되었습니다. 경영진은 AI 고객 매출의 90%가 코어 클라우드 제품도 함께 이용하는 고객에게서 나왔다고 밝혔습니다.

디지털오션은 전체 인프라에서 평균 가동 메가와트당 약 1,800만 달러의 ARR을 창출했습니다. 경영진은 최신의 토큰 효율적인 하드웨어 도입, 가동률 향상, 가격 정책, 고마진 컴퓨트·스토리지·데이터 서비스의 채택 확대에 따라 이 지표가 상승할 것으로 기대하고 있습니다.

리치먼드, 캔자스시티, 멤피스 시설은 모두 2026년 중에 예정대로 또는 예정보다 일찍 가동을 시작했습니다. 회사는 3분기 중 2028년용으로 20메가와트를 추가로 확보하여, 총 약정 용량을 약 180메가와트로 늘렸습니다.

경영진 가이던스

경영진은 높아진 전망치에 대한 자신감이 메가와트당 ARR 증가, 제품 결합 판매(attachment) 확대, 그리고 추가적인 데이터 센터 용량에만 의존하지 않고도 컴퓨팅 및 데이터 소비를 촉진할 수 있는 토큰 및 에이전트 워크로드의 확장을 반영한다고 설명했습니다.

리스크 및 주시 영역

• 용량 확충 실행: 2027년 중 약 75메가와트가 가동에 들어갈 예정입니다. 구체적인 시기는 데이터 센터 제공업체, 장비 공급업체, 하드웨어 벤더와의 협의에 따라 달라집니다.
• 매출총이익률 압박: 시설이 본격적으로 가동되기 전에 신규 데이터 센터 임대 및 감가상각이 시작되어 단기적으로 매출총이익률에 부담을 줍니다.
• 가격 정상화: 애널리스트들은 업계의 추가 공급과 스팟 가격 하락에 따른 영향 가능성을 제기했습니다. 경영진은 광범위한 플랫폼 채택과 에이전트 활용 사례 확대로 단순 컴퓨팅 가격에 대한 의존도가 낮아질 것이라고 설명했습니다.
• 자금 조달 요구사항: 디지털오션은 7억 8,500만 달러 규모의 장비 리스 한도를 확보하고 10억 달러 규모의 장내 주식 발행(ATM) 프로그램을 개설했습니다. 회사는 해당 주식 발행 프로그램을 이용할 수도, 하지 않을 수도 있으며 자금 조달 비용, 주식 희석, 레버리지 간의 균형을 맞출 것이라고 밝혔습니다.
• 하드웨어 투자: 경영진은 CPU 및 GPU 조합에 따라 최신 GPU 인프라에 대한 2027년 설비투자(CapEx)가 메가와트당 3,000만 달러대 중반이 될 수 있다고 시사했습니다. 이는 공식 가이던스가 아닌 참고용 벤치마크로 설명되었습니다.

애널리스트 Q&A 하이라이트

플랫폼 차별화: 경영진은 디지털오션이 단일 추론이나 GPU 제품이 아닌 통합 AI 네이티브 클라우드를 통해 경쟁하고 있음을 강조했습니다. 고객은 GPU, 토큰, 코어 클라우드, 에이전트 서비스를 통해 진입한 후 런타임, 스토리지, 데이터베이스 등 다른 제품으로 확장할 수 있습니다.

코어 클라우드와 AI의 융합: 경영진은 일반적인 디지털 워크로드에 AI 기능이 점점 더 통합됨에 따라 AI 고객과 기존 코어 클라우드 고객을 구분하기가 어려워지고 있다고 밝혔습니다. 코어 클라우드 성장 역시 AI 관련 수요 및 제품 결합 효과를 보고 있습니다.

용량 및 하드웨어 로드맵: 디지털오션은 2027년 내내 약 75메가와트의 추가 용량이 순차적으로 가동될 것으로 예상하고 있으며, 용량 증설에 따라 하반기에 매출 성장이 가속화될 것으로 전망합니다. 최신 세대 GPU는 메가와트당 토큰 산출량과 ARR 밀도를 개선할 것으로 기대됩니다.

자금 조달 전략: 회사는 보유 유동성, 장비 금융, 회전 한도 대출, 선택적 ATM 프로그램을 통해 2027년 계획된 성장을 지원할 유연성을 충분히 확보하고 있다고 설명했습니다. 장비 및 자산담보부 금융이 기본 전략으로 유지됩니다.

장기 마진: 경영진은 신규 시설 오픈 시 단기적인 매출총이익률 압박을 예상하고 있으나, 데이터 센터가 안정화됨에 따라 영업 레버리지와 가동률 상승이 보다 견조한 정상 상태(steady-state) 수익성을 뒷받침할 것이라고 말했습니다.

실적 발표 컨퍼런스 콜 전문

전체 실적 발표 컨퍼런스 콜 녹취록

경영진 발표

Operator

Hello, everyone. Thank you for joining us, and welcome to the DigitalOcean's Third Quarter 2026 Earnings Conference Call. After today's prepared remarks, we will host a question-and-answer session. [Operator Instructions] I will now hand over the conference over to Josh Beer, Senior Director of Investor Relations. Josh, please go ahead.

Joshua Baer

Thank you, and good morning. Thank you all for joining us today to review DigitalOcean's Third Quarter 2026 Results. Joining me on the call today are Paddy Srinivasan, our Chief Executive Officer; and Matt Steinfort. Our Chief Financial Officer. For those of you following along and accompanying slide presentation is available on the webcast. Before we begin, -- the SEC's EDGAR filing system is experiencing technical difficulties that have prevented us and other companies from filing our earnings release this morning, but it is available on our Investor Relations website.

Let me remind you that certain statements made on today's call may be considered forward-looking, which reflect management's best judgment based on currently available information. Our actual results may differ materially from those projected in these forward-looking statements, including our financial outlook. I direct your attention to the risk factors contained in our SEC filings as well as those referenced in today's press release that is posted on our website.

DigitalOcean expressly disclaims any obligation or undertaking to release publicly any updates or revisions to any forward-looking statements made today. Additionally, non-GAAP financial measures will be discussed on this conference call. Reconciliations to the most directly comparable GAAP financial measures can be found in today's earnings press release. As well as in our investor presentation that outlines the discussion on today's call. A webcast of today's call is also available in the IR section of our website.

And with that, I'll turn the call over to Paddy.

Padmanabhan Srinivasan

Thank you, Doris. Good morning, everyone, and thank you for joining us today. Q3 was a fantastic quarter at DigitalOcean. We launched critical capabilities that help customers build Agentic systems that recursively self-improve grew token volume by an order of magnitude and saw customers adopt our AI native cloud as a unified platform. This led to tremendous results, and we are raising our revenue outlook for both full year 2026 and 2027.

Let me start with 4 key takeaways. First, growth continues to accelerate. Revenue grew 40% year-over-year, our highest in a decade and our fifth consecutive quarter of acceleration. We added a record $160 million in ARR and we beat our guidance on every key metric, continuing to deliver durable and profitable growth.

Second, customers are adopting our AI native cloud as a complete platform. AI native applications require both tokens and agents, and we are a leader in both. Token volume grew roughly 10x quarter-over-quarter. We shipped managed agents agent learning services and agent data services, all critical capabilities that let agents run, learn and recursively self-improve. Combined with our core cloud primitives, they make our AI native cloud complete and unique.

In Q3, AI customer revenue came from customers that also consume our core cloud perimeters. They're not looking for GPU capacity. They're building and deploying production agent applications on the whole platform. Third, our software continues to make our megawatts more valuable. Our growth is driven more by yield per megawatt than just by the quantity of megawatts we activate.

We already deliver industry-leading ARR for active megawatt. This premium is evident in our current results and our differentiated software platform will continue to widen our returns as we scale. And finally, in this noisy environment, we are different by design. We offer a complete platform rather than a point feature.

We deliver durable and profitable growth. Our customer base is large and diversified rather than concentrated in a handful of buyers. Our profile is unlike the growing list of bare metal focused infrastructure providers. As concerned grows about large-scale gigawatt data center and power deployments, we're turning up megawatts of Tier 1 data center capacity on time or early.

Demand continues to far outstrip our supply, and we continue to work on securing high-quality capacity for future expansion. As concerned grows about customer concentration risk or is minimal as concerned growth about funding AI growth, we sit with minimal leverage and a flexible balance sheet. We are different by design with a more complete platform offering, a diversified customer base and a differentiated financial profile.

All of this together is why we are raising our outlook. We expect Q4 revenue growth of 44% to 45% and expect to deliver 60% or more revenue growth for the full year 2027. I will now drill into these key takeaways. First, we continue to accelerate growth, and the growth is durable and profitable. Q3 revenue was $321 million, up 40% year-over-year, more than $14 million above the high end of our guidance and up from 16% just a year ago. We added a record $160 million ARR in the quarter, up 266% year-over-year.

Total ARR ended the quarter at approximately $1.28 billion, up 40% year-over-year. AI customer ARR reached $376 million, growing 288%. The guidance beat came from rapidly growing infant services, more core cloud attached to AI workloads, strong utilization of our elastic compute fleet, higher prices on short-term capacity and early agent adoption. The beat came from software and platform with only a small part of the beat coming from our Memphis data center, which we brought online a few weeks early.

We continue to deliver durable and profitable growth. Adjusted EBITDA margin was 45% and non-GAAP earnings per share was $0.45, both above guidance. Adjusted operating income margin was 23%. The Second, customers are increasingly adopting our whole AI native cloud. AI center of gravity has moved from people chatting with models to agents doing real work and delivering outcomes. This shift has created 2 new slip streams of value.

First is token flow. Agents think in tokens, lots of them. Third-party estimates expect agent AI to lift global token consumption 24x by 2030. The second slip stream is AgenticFlow. Agents bring together -- it's run time, tools, memory and learning loop that turn thinking and intelligence into finished work. They require new perimeters and a cloud that is built for agents, not humans, we believe there is a massive opportunity in servicing this rapidly scaling flow.

These 2 slip streams are inseparable. Every agent session fans out dozens of model calls each consuming and generating significant tokens. Tokens only become valuable when an agent acts on them to deliver tangible business outcomes. Most providers serve only 1 of these flows. We are a leader in both. Our inference engine is a leader in token flow, and our managed agents platform leads an agentilow.

That combined strength delivers platform sales and accelerates the AI native flywheel. Let's take a closer look at token flow. Our token volume grew 10x quarter-over-quarter. We operate at the pace of the market. Our model catalog has grown rapidly with 37 model launches since April -- more than 4,000 customers have used our inference router to optimize dollar per unit of intelligence. And on open router, the largest AI gateway and our spillover for idle cycles -- we are routinely a top inference provider for many of the leading open weight models, averaging well over 0.5 trillion tokens per day.

We are also pushing the innovation frontier in AgenticClow. We recently shipped what we believe is the most advanced agency platform in the market built on 3 pillars: managed agents, which includes CPU-based secured sandboxes and harness on time, agent learning services for context and memory and age and data services with high-performance storage and advanced databases.

The platform is evolving rapidly with over 70 releases in 65 days and it is the first platform built for agents that recursively self improves with every run. These flows drive customers to adopt multiple layers of our AI native cloud, regardless of which of the 4 front doors they come through. Customers that come for GPUs, stay for tokens, age and sandboxes and run time Customers that come for agents expand into storage, tokens and more customers that start with tokens quickly expand to agents and cloud perimeters.

And finally, core cloud customers are starting to consume inferencing and our building agents. Let me bring this whole platform flywheel to life with a customer example. Amplitude, which powers product analytics for some of the world's largest brands is building its new products, amplitude Wave for self-improving products on our AI native cloud. We power it with agent on time, dynamic tool access and inferencing. We value our full platform and its total cost of ownership and for 1 of its core workloads the open model on -- do runs about 67% cheaper than a Frontier model.

This is a public software company choosing our full stack for 1 of their new products from day 1. And Amplitude is just 1 example of customers adopting multiple layers of our stack in Q3, customers on 4 or more products drove roughly 60% of our total revenue and included the majority of our $100,000-plus AI accounts. Agents are also now direct customers of dissolution.

They do not just run on our cloud. They increasingly select it, sign into it and deploy on it. Already Roughly half of our unique visitors to digital otion.com are agents and more than 37,000 unique customers have deployed on DIO through an agent this year. When an agent autonomously chooses where to build, it chooses the cloud, it can read, reason and deploy to without a human in the loop, and agents are increasingly choosing distillation. Serving agents as builders and as buyers are 2 sides of the same AI native cloud, and we are increasingly winning both.

Third, our software continues to make megawatts more valuable. Our growth is driven more by yield per megawatt than just by the quantity of megawatts we activate. We delivered $18 million in ARR per average active megawatt. This is not a vanity outlier metric. This is the return we are generating today on our entire fleet on currently deployed technology generation at still ramping utilizations in newly deployed data centers.

We already earn industry-leading returns and have multiple levers that we believe can drive our returns higher. First, we can rapidly capture increasing prices with our consumption and short-term contracted revenue model. We've already closed multiple individual deals well north of $20 million in ARR per megawatt.

Next, we expect to generate more ARR per megawatt as we deploy newer, more token efficient technology generations. We can also drive higher utilization, leveraging our elastic compute capacity. Finally, as we just saw, AI customers are pulling through higher-margin compute, data and other platform perimeters. Maximizing utilization with an elastic compute fleet and driving platform adoption at scale takes very different muscles than servicing bare metal contracts for a handful of large customers.

It is 1 thing to sell spot capacity at higher prices for a week or a month. It is a very different thing to maximize utilization and yield at scale. Doing so, requires fleet management across a diverse set of technologies, serving a vast and ever-changing catalog of models, operating across global time zones and workload patterns and delivering multiple products all with hyperscaler level security and SLAs.

This requires significant cloud operations experience, and we've been operating a cloud at scale for almost 15 years. All of this contributes to our ability to drive our yields up over time. We continue to innovate and build a cloud that agents and AI natives will run on. Said plainly, our software makes our megawatts more valuable. And we continue to steadily add incremental data center capacity to take advantage of our strong returns.

While data center and power risk is always present, our execution on new data centers has been terrific in 2026. Richmond launched in Q1, Kansas City launched in Q2, both ahead of plan. MEMS followed in Q3, also ahead of schedule. All 3 planned 2026 data centers are now live and ramping. Our strategy of bringing on tens of megawatts facilities with Tier 1 data center operators is paying off.

And as demand continues to outstrip supply, our track record of turning up capacity on time gives us confidence to keep investing to meet the demand. In Q3, we secured an incremental 20 megawatts of committed capacity from our Tier 1 data center partners for 2028. This brings total committed capacity to approximately 180 megawatts as we continue to actively seek high-quality incremental capacity.

Finally, we are different by design, a different platform, a different target customer a different financial profile and capacity strategy. We are a complete cloud different from many AI infrastructure providers. Customers come to us for our whole platform not just for bare metal GPUs. We have minimal customer concentration with our top 25 customers representing only 27% of our revenue and no single customer in high single-digit percentage of revenue.

Self-built is not a desirable or realistic future path for our customers, where it may be for hyperscalers or frontier labs. We are not locked into long-term bare metal contracts in our consumption model and elastic compute fleet enables us to capture rising prices as the market moves. We have operated a full stack cloud delivering hyperscale SLAs and strong security while servicing hundreds of thousands of customers for more than 15 years.

We benefit from but are not entirely reliant on the massive AI tailwinds with a stable, highly profitable and growing $900 million ARR core cloud business. Finally, we are profitable. We're generating cash, and we carry minimal leverage. Very few companies combine this financial profile with our growth outlook. With momentum continuing to grow, we are again raising our outlook for both 2026 and 2027. We expect revenue growth of approximately 34% for the full year 2026 and to exit this year at 44% to 45% growth in Q4.

We -- for 2027, we are increasing our outlook from 50% plus to 60% or more of a materially higher 2026 base. I want to be clear about what is driving our confidence. It is the software levers I just described, higher yield per megawatt, higher attach and new token and agentic workloads that drive compute and data consumption without requiring new data center capacity. Our growth and upside is not tied solely to how many megawatts we activate -- it is also driven by our software.

With that, I will turn it over to Matt.

Matt Steinfort

Thanks, Paddy. Good morning, everyone, and thanks for joining. I'll walk through our terrific Q3 results our balance sheet and funding and our updated outlook. Starting with revenue. Q3 revenue was $321 million, up 40% year-over-year, more than $14 million above the high end of our guidance range. Our record $160 million ARR was up more than 3.5x year-over-year, bringing total ARR to approximately $1.28 billion.

Growth continues to come from our higher-spending customers. ARR from our 100,000-plus customers grew 133% year-over-year to 41% of total ARR. And ARR from our 1 million-plus customers grew 273% and now represents 30% of ARR AI customer ARR reached $376 million, up 288%. Importantly, 90% plus of our AI customer ARRs from inference services and core cloud with inference services up nearly 600% year-over-year.

Turning to margins and profitability. We continue to deliver durable and profitable growth. GAAP gross margin in the quarter was approximately 50% and -- as expected, gross margin has declined as we brought on new data centers and GPU capacity online, but more importantly, our operating leverage remains strong as we efficiently scale revenue. Adjusted EBITDA was $144 million, up 45% year-over-year at a 45% margin, well above guidance of 38% to 39%.

Adjusted operating income was $74 million, a 23% margin. Non-GAAP diluted net income per share was $0.45 compared to our guidance of $0.28 to $0.30. On a GAAP basis, we reported a net loss of $43 million or $0.37 per share. Net loss is entirely attributed to a $57 million or $0.46 per share induced conversion expense that was related to the partial repurchase of our 2030 convertible notes.

Excluding that onetime expense, we would have reported positive GAAP net income. Trailing 12-month adjusted free cash flow was $220 million, a 20% margin. compared to $178 million and 21% in the third quarter of last year. Moving on to the balance sheet and funding. The way we have carefully managed our balance sheet has turned out to be a competitive advantage in an environment where financial discipline is increasingly valued. We have approximately $1.2 billion in total liquidity, and we ended the quarter with $845 million in cash, cash equivalents and marketable securities and roughly $400 million available under our revolving facility.

We also ended Q3 with net leverage of only 1.2x. And we continue to strengthen the balance sheet and create incremental capacity to fund our growth. In September, we secured a $785 million equipment lease facility at attractive rates. in a debt market that increasingly favors strong financial discipline. Today, we also proactively put in place a $1 billion at-the-market equity program. While we may or may not use this program, we believe it is a proactive and prudent step that positions us to efficiently access equity funding should we decide to do so.

We also plan to retire the remaining $312 million of our 2026 convertible notes on or before maturity in December. We will continue to appropriately balance the cost of capital, dilution and leverage as we grow. We have proactively delevered to create borrowing capacity for incremental data center GPU and CPU investments to fuel our growth. With our profitable financials and flexible balance sheet, we believe we have access to a wide range of funding solutions. The primary funding strategy for our build-out remains to match the timing of our cash flows with revenue. With that said, we have the ability to augment this equipment and asset-backed financing strategy with other sources of debt and equity as needed as we continue to optimize our cost of capital and operate within our target leverage trend.

Our product platform is clearly different. Our target customer base is clearly different. And our financial profile and balance sheet strategy are now increasingly proving to be different as well. Turning to guidance. we are raising both our 2026 and 2027 revenue growth outlook. For the fourth quarter of 2026, we expect revenue of $348 million to $351 million, representing 44% to 45% year-over-year growth.

We expect adjusted EBITDA margin of 42% to 43% and non-GAAP diluted net income per share of $0.30 to $0.32 on approximately 128 million to 129 million weighted average fully diluted shares. For the full year 2026, we now expect revenue of $1.208 billion to $1.211 billion, representing approximately 34% year-over-year growth. We expect adjusted EBITDA margin of 41% to 42%, adjusted free cash flow margin of 20% to 22%; and non-GAAP diluted net income per share of $1.60 to $1.62 on approximately 122 million to 123 million weighted average fully diluted shares.

We now expect revenue growth of 60% or more for the full year 2027. This is a meaningful increase to the projected dollar revenue in 2027 as this higher projected growth rate is also off a higher projected 2026 base. Since our initial guidance in February, we have raised the midpoint of our 2026 guide by $120 million. On top of this, we've increased our 2027 growth rate outlook from 30% to 60% plus, an increase of over $0.5 billion at the midpoint.

We'll provide more detailed '27 guidance, including projected margins and cash flow on our fourth quarter call in February. We like how well positioned we are heading into 2027, and that strong outlook extends into 2028. With that, I'll hand it back to Paddy.

Padmanabhan Srinivasan

Thank you, Matt. Before we move to Q&A, let me recap what we shared today. First, our growth is durable, profitable and accelerating. 40% revenue growth up from just 16% a year ago, a record $160 million of incremental ARR. 45% adjusted EBITDA margin, 23% adjusted operating income margin and strong cash generation. .

Second, customers are adopting our AI native cloud as a whole platform. inference engine token volume grew roughly 10x quarter-over-quarter. We shipped managed agents agent learning services and agent data services. In Q3, 90% of AI customer revenue came from customers that also consume our core cloud primitives. They're not looking for just GPU capacity. They're building and deploying production agency applications on the whole platform.

Third, our software continues to make our megawatts more valuable. Our growth is driven more by yield per megawatt than just by the quantity of megawatts we activate. [ 18 million ] ARR per megawatt across our entire fleet today with differentiated software that we believe will enable us to increase our leading returns as we scale.

And finally, we are different by design in an increasingly noisy landscape, durable and profitable growth. a differentiated platform with increasing customer adoption, minimal customer concentration risk, lower leverage, a strong balance sheet and additional funding flexibility. With this growing momentum, we are again raising our outlook. We expect to grow revenue 44% to 45% in Q4 2026 and expect full year revenue growth of 60% or more in 2027.

I have never been more confident in where DigitalOcean is headed. The Agent Kira is here. It rewards exactly the platform we have built. We are accelerating and excited for what's ahead. With that, let's open it up for questions.

Operator

[Operator Instructions]

Your first question comes from the line of Raimo Lenschow with Barclays.

질의응답

Raimo Lenschow

Congrats for me. That was a great quarter and great guide. A question I had for you, Paddy, was you increasingly talk about how it's not just about offering capacity. It's about kind of giving more and you talked about the inference engine, et cetera. How do you see that when you talk with customers as a differentiation like a lot of people talk about inferencing engines or like trying to optimize things.

But like -- how does it come across in your customer conversation? What are you seeing there? And then what's the customer demand? Or like what's the pipeline for you guys in terms of signing more customers there?

Padmanabhan Srinivasan

What we are hearing from customers is, as I was describing in my prepared remarks, there are 2 main entry points for AI native workloads -- the first is, obviously, they're very token hungry. But increasingly, over the last 90, 100 days, what we're seeing is customers are starting to deploy production agentic workload. And agents actually know are full-fledged production workloads that need a lot of tokens, but they need a lot more than just token. .

They need a secure run time, which some people refer to as sandboxes or some kind of a secure execution environment within which agents run. -- agents are performing actual work. So they need a lot of tools to execute and access the Internet and other data sources. So we have this feature called agent gateways, action gateway, which offer more than 20,000 tools at agent's disposal.

We also are making available a lot of different context and memory primitives for agents to recursively self-improve. We have also rebuilt a data substrate, specifically for agents, which include high-performance storage and very advanced databases and so forth. So agents are AI native cloud. And that's what is very exciting for us because we are just starting to see agents being pressed into production bow and beyond just coding as a core workload, we are seeing a proliferation of personal agents.

We are now starting to see agents take over, media production and generative media. So we are starting to see agentic workloads take over a lot of different core business workflows. The example that I was giving on amplitude is a great example of a company using Agentic workflows to improve digital products. So the use cases are just proliferating and we are very proud to be the most advanced agent-type platform in the market. So we think we are really, really well positioned.

Operator

Your next question comes from the line of Kingsley Crane with Canaccord Genuity.

William Kingsley Crane

Really encouraging to see the momentum in the business. So my understanding is in the past that you have a large amount of core cloud customers that weren't really quite ready to adopt AI. And it seems like that may be changed. I mean you also noted that customer revenue comes from customers consuming Core Cloud, core called AR within that base grew 17%.

And -- but if you look at just the rest of the business, I think AR, excluding AI is growing around 10%, 11%. So can you just help us break down that disparity there? And just thinking about if there's been an acceleration in the core cloud customers now adopting AI.

Matt Steinfort

Ken, that's a great question. It's actually getting harder to tell who an AI customer is versus a non-AI customer, to your point, we've seen a tremendous amount of growth in our -- what we traditionally called the core cloud funnel with new customers coming in to our M1 to M12. And a lot of that is being driven by customers that are deploying Open Color Hermes or some of the identic frameworks. And those customers are coming in and they're not declaring their an AI customer.

They're just using AI to help drive their business. Also, some of our largest, what I would say, traditional core cloud customers are actually very active in the AI ecosystem, providing other capabilities and services to AI customers. So we're having a harder time distinguishing who's an AI customer and who isn't -- but what we are seeing is the core cloud is accelerating. Just like you said, if you just look at it from a product standpoint, traditional core products, our basic droplets are expanding. We're seeing a lot of attached to our AI customers from the core side of the house.

So it's -- I think the tailwinds that we're seeing are across both sides of the business. And again, the nice thing about our business and 1 of the things Paddy highlighted from a different standpoint is we have this $900 million business of core cloud that's growing. It's wildly profitable and it generates a lot of cash. And that's accelerating at the same time we've got this tremendous opportunity in front of us. with newer AI-oriented products. So we're very encouraged by the trends that we're seeing.

Padmanabhan Srinivasan

Yes, I'll just add 1 thing. One cool thing we did this quarter, Kingsley, is we bundled a lot of the things that I rattled off that agents need into a new package called agent Rolet, which makes it super easy for a non-AI customer to just start consuming different perimeters that makes them agent. So -- so we're doing a lot of things.

But to echo what Matt just said, it's becoming increasingly hard to distinguish between who is AI, who is not because every digital workload is going to require intelligence and they're all becoming AI workloads.

Operator

Your next question comes from the line of Gabriela Borges with Goldman Sachs.

Gabriela Borges

Pat, I appreciate all of the color on how strong demand is. I want to ask you about the other side of the equation. So there's a tremendous amount of supply that's coming online for the industry over the next 12 months. And we also know that the efficiency of models, the efficiency of compute continues to go up.

So my question for you is that at some point, we're going to see supply demand normalization such that spot prices come down. Maybe just give us your historical perspective. How do you think about what the implications are for the industry and for DigitalOceans model, when spot prices peak and start to normalize?

Padmanabhan Srinivasan

So for us, as I was describing, it is not just about the megawatts we energize. It is about the yield we are getting from the fleet that we have, both on GPUs, but also CPUs. And as I was describing, identic workloads require more than just tokens or more than just the cycles, right? Agents require the full breadth of the platform because -- the expectation from agents are that they actually deliver positive business outcomes, not just spin up tokens.

So from that perspective, I feel we are very, very favorably position as the supply demand normalizes at the price of tokens fall, price of compute, fines and equilibrium. I think the amount of use cases we're already seeing personal agents exploding onto the scene, I spent most of last week talking to physical intelligence companies that are already leveraging our platform and none of the mainstream reporting actually even covers most of those companies, and that's another massive workload that is coming on or will come on over the next few years.

So the amount of use cases in the size of the pie is accelerating far more than some of these supply-demand equilibrium that you just referenced. And we are very favorably positioned because we are not a single feature platform. As I explained in my prepared remarks, is there are just so many different base by which customers enter our platform. And once they come into the platform, like 90% of AI customer revenue comes from companies that are using other cloud perimeters and we have very little customer concentration also, right?

We have very few TP-only customers. Most of our customers are leveraging multiple products. In fact, 60% of our revenue comes from companies that are using 4 or more products. So that the power of a full platform adoption, a full platform sales.

Operator

Your next question comes from the line of Param Singh with Oppenheimer.

Paramveer Singh

Paddy, maybe I wanted to kind of peel back the cover of that $18 million ARR you talked about Typically, of course, this is on the cloud side, got a better run rate. So I want to understand what's coming from the CPU versus GPU architecture. And what's the mix of CPMG your current install as of megawatts? .

Matt Steinfort

Thanks, Pam. It's Matt. The $18 million in ARR per megawatt is a fleet-wide number. So it's literally total ARR across all products across all of our kind of average active megawatts. It's depressed even a little bit in Q3 because we turned Memphis up and only got a little bit of revenue there, but we counted the portion of the quarter where it was active. And the important thing is, if you think about it from a going-forward basis, that number is only going to improve.

And it's on existing generation of technology that have already been deployed that have lower cost of CapEx per megawatt than what you're hearing when you hear some of the quotes out there about what people say they're getting in incremental ARR per megawatt. So for us, the important number is the total. Can you drive an effective yield across your entire portfolio and maximize the value of that investment.

And we believe the best metric for that is the total yield you're getting on your megawatts. So it's a combination, as we've talked about, it's a combination of incremental growth on our AI products, our core cloud products, but it's across the entire fleet, across our entire customer base. And again, we expect that number to increase, and we have a number of levers that we're pulling to be able to drive that number up.

Padmanabhan Srinivasan

And as Matt said, as customers come for tokens and GPUs and start expanding into agentic run times and data services, high-performance storage, these are much higher margin products on our -- in our fleet. So we expect that to go up. And we are already reporting evidence of this platform attached. .

Operator

Your next question comes from the line of Wamsi Mohan with Bank of America.

Wamsi Mohan

Nice results. You guys had very solid execution in your so far data center starts that came in earlier just given a lot of noise around potential delayed starts at data centers, how are you thinking about maybe the cadence of incremental megawatts, especially in 2027 and what's baked into guidance there? .

And can you maybe share some sense of where your GPU fleet stands between older and newer generation GPUs as you're pointing to sort of token efficiency is another area that can accelerate revenue growth?

Matt Steinfort

Thanks, Wamsi. We've been very, I'd say, diligent and disciplined in terms of adding incremental data center capacity. We tend to work with Tier 1 data center providers that have been in this business for decades that have very, very good track records of execution and delivering on time and very good supply chain kind of management. So that's been kind of the part of the reason for our success in '26, turning up capacity we'd expect '27 to be very similar.

The incremental, call it, 75 or so megawatts of capacity that will turn up next year will come on over the course of the year, just like we brought multiple data centers online this year, and we're very confident in our ability to hit the time line that underpins the guidance that we gave -- so it will be very similar. Growth will accelerate during the year. So you'll see capacity come online, which will drive acceleration. The second half will grow faster than the first half.

From a GPU standpoint, the other nice thing about bringing on data centers in, call it, 10, 20-megawatt chunks is we can make discrete decisions about which technology generations we put in each data center. We're not -- we took down a gigawatt and we have to order all the gear for that, and we have to take today's generation of GPUs -- so as you look at 2027, we're likely to take where we were installing, say, B-300 generations this year.

We'll be installing GB 300s or and even evaluating maybe varirubin or looking at other vendor similar technologies that we could deploy -- so that's part of the driver for an expectation, to Param's question, of increased ARR per megawatt is right now, we're getting that industry-leading ARR per megawatt on kind of the older generations of GPUs. And as we install the newer generation to 2027, that's only going to drive the ARR per megawatt up because they're more token efficient.

There's more CapEx, clearly, that's involved in that. But the returns that we generate on that CapEx are very similar. So you would expect the ARR per megawatt to be rising with those newer generations of technologies.

Operator

Your next question comes from the line of James Fish with Piper Sandler.

James Fish

I think the big concern in this space is around financing. And I know you guys talked about like available capacity and all of that. But -- as we start thinking about '27 here, how much financing do you guys think you need beyond the $1 billion ATM that you're announcing here and why the decision to use equity over further finance leases. .

Matt Steinfort

Is, great question. So our balance sheet is a competitive advantage for us right now. We have an incredibly flexible balance sheet. We've been proactively delevering the very proactive and opportunistic putting in place an ATM today just gives us option value. We believe that with 850-something million in cash and $75 million equipment finance facility that we've drawn.

We've got $400 million sorry, we haven't drawn that we've raised. We've got $400 million of revolver -- and now we have an ATM as option value. We believe that we have the ability to fund all of our growth in 2027 -- but that's -- to me, it's the option value that we've created. Our primary funding strategy remains to pay for equipment over time to match our revenue for our costs with our revenue.

And we believe that we can continue to tap that market and get financing at attractive rates. But the nice thing about the flexibility we created is we're not time box done anything. We have the ability to do that as we need it over the course of 2027 and 2028, we have a lot of flexibility. And so we feel very, very good about where we sit, particularly in this noisy market and this funding environment.

We believe that we've got 2027 covered but our strategy will be to continue to use equipment financing to be able to grow the business. And so that -- the structure we have gives us the flexibility to pursue that.

Operator

Your next question comes from the line of Sanjit Singh with Morgan Stanley.

Unknown Analyst

Great. This is Ryan Lance on for Sanjit. I guess just any additional color you can provide around the incremental capacity set to come on. And in 2027 and 2020? I'm just trying to understand like how much of the capacity currently contemplated in 2027 is already tied to contracted demand?

Matt Steinfort

So from a -- 2 questions in there on the customer and the data center side. So on the data center side, we've got just over 75 megawatts of capacity that's active today. That was with our Memphis facility. -- we've contracted for, signed for another 75 or so megawatts that will come on in 2027 and another 30 that will come on in 2028, bringing our total committed signed capacity from a data center standpoint to about 180 megawatts.

So -- and we're actively out there continuing to look for incremental capacity, again, with Tier 1 data center providers with a similar strategy as we've executed. As we've said and as Patti highlighted, our model is different in the market. We don't go and sign 5-year contracts that are bare metal with hyperscalers or 1 of the frontier models still on the top 2 frontier models. -- at bare metal pricing.

We that come to us for capacity as they're building their own applications. The demand for that is well in excess of our supply. And we continue to be able to be super picky about who we bring on its customers and make sure that we were very comfortable with them from a funding standpoint, from a an economic model standpoint, and viability and how effective are they at reaching end customers, we can be super picky.

And as you've seen with our data centers that we brought online in 2026 very soon after we bring them online within a couple of months, we're ramping at fairly high utilizations, and we expect to do that with the pipeline that we have for the '27 capacity and the '28 capacity as well.

Operator

Your next question comes from the line of Mark Zhang with Citi.

Mark Zhang

So great to see the demand strength, but maybe profitability certainly stood out this quarter as well. I understand you're not giving margin guidance on '27 yet, but can you just maybe give us a sense of the levers in the model sort of the main buckets of puts and takes on the OpEx and the cost structures going into next year.

How much operating leverage in the model is there to continue to capture going forward?

Matt Steinfort

That's a great question, Mark. We feel really good about the profitability, really strong EBITDA margins, really strong operating income margins. And as you highlighted, there's a lot of operating leverage as we scale. Gross margin, clearly, those will be pressured as we bring on new data centers and the depreciation hits and the data center lease hits before we start ramping revenue. but that has been mitigated by the efficiencies that we've been able to drive in the business on the operating leverage side.

So we feel very good about the margins, as you said, we'll provide more guidance in an outlook when we get to February and we provide the full year. But we continue to believe that we can drive strong profitability from an operating income standpoint that we can generate positive adjusted free cash flow and that we can grow the business responsibly with the right amount of leverage within our target range and -- and as we grow, those margins only get better, right, as you mature the data centers and you're not taking that upfront hit from a ramp standpoint.

So from a steady state standpoint, the margins are very, very strong.

Operator

Your next question comes from the line of Jason Ader with William Blair.

Jackson Ader

Matt, can you comment on the CapEx per megawatt on the incremental capacity that you're adding? And then just kind of framing the return on invested capital -- and then just, Paddy, I wanted to give you an opportunity to just comment on the current environment and just the fear and kind of overall nervousness out there in the market for your space? .

Padmanabhan Srinivasan

Yes. Maybe I can start. We are not nervous. We are very confident because we are -- as I explained, we are different by design. We have very little customer concentration. We have a whole platform that we have proof points now that customers are not just coming to us for a GPU or CPU or for a token. They're coming and expanding and using our whole platform, which is very, very unique and differentiated in the market.

And the financial profile we have is also very differentiated. As Matt mentioned, our balance sheet is a competitive advantage for us. right? We have a core cloud business approaching $1 billion, growing really well, and all those customers are digital native companies. And their existence depends on making themselves AI-centric -- so they have a lot of appetite in consuming our AI platform services.

So we are well positioned, and we feel this is a great opportunity for us over the next couple of years where things find equilibrium, we feel we are very, very well positioned to execute both from a strategic and platform perspective but also from a financial point of view.

Matt Steinfort

And Jason, to answer your question on CapEx per regular one, we don't guide to that, and it depends on the mix of CPU and GPU that you put in place. But -- but if you're looking for kind of benchmarks, we talked about what generations are coming from a GPU standpoint. And if you look at like a GB 300 equivalent, you're probably in the mid-30s from a CapEx per megawatt -- and again, it will depend on whether you're -- how much CPU you're putting in alongside that, et cetera. And then if you fast forward to the next generation and you start to look at varirubin, which is still early, -- now you're probably starting with a 4. It's something probably closer to 40 a megawatt. So if you look at our project for 2027, I would say it's likely that it's something in the mid-30s -- but again, that's not a guide that we could -- it could be slightly different than that, depending on the mix of infrastructure that we put in. that's probably...

Padmanabhan Srinivasan

And also to state the obvious, with that new generation or next-generation comes a much higher yield as well. .

Matt Steinfort

Yes, we're underwriting at the same return on invested capital. So we are in the same -- we have the same thresholds. We are in the same return or similar returns because we can generate way more tokens with those new generations. So the density, the ARR per megawatt density is just going up. and the return that you're getting on that invested capital is the same or better in the cases of some of the advancements that we're talking about that are software-driven.

Operator

Your next question comes from the line of Jackson Ader with KeyBanc Capital Markets.

Jackson Ader

I was curious on what you feel like is within your control and what you feel like is left to the colocation providers in terms of when capacity actually come online in Memphis came on a few weeks early, and that's great. But like, literally, what can you do in the interim? What are you in control of to maybe bring some of the capacity into the first half of '27 versus second half or maybe even into '26?

Matt Steinfort

It's a partnership with the data center provider. So clearly, we're jointly engineering and designing the build-outs and then sourcing the components that are required, both on their side to build out the powered shell and our side to get the GPUs and the CPUs and all the associated equipment. We work -- we have an incredibly strong team. We've continued to add really strong talent in that area.

And we're working every day to try to bring those on at what we call speed of light, which is as fast as we possibly can. What I'll tell you though is we're very confident in our guidance and the outlook for 2027 -- as you saw with 2026, and you've seen from us as a management team will be prudent and appropriate when we give you guidance these are numbers and dates that we believe that we can meet and beat.

And -- but in the background, clearly, we're working very aggressively to try to create upside on that and bring on the facilities earlier. And it's -- again, it's a partnership with the data center providers. We have very strong relationships with multiple global OEMs on the GPU side. We've got fantastic relationships with the with NVIDIA and AMD, and we feel very good about our ability to kind of meet and beat our -- the expectations that we set.

Operator

Your next question comes from the line of Patrick Walravens with Citizens JMP.

Unknown Analyst

Congratulations on the quarter. This is the com for Pat. I remember being at Deploy earlier this year when inference router was announced and the room broke into applause. Since then, it feels like more companies have been launching model routers. So my question is, has more model routers come to the market how does digital ocean differentiate itself and what decisions you guys to compete effectively in the space? .

Padmanabhan Srinivasan

Yes. Great. Good question, Nick. We are in a great position if we are 3 to 4 months ahead of the rest of the players in the market. That's a great position to be in. to months in AI is like 3 to 4 years and the rest of software. So we'll take that. But the story here is not just on 1 feature, right?

Yes, infant routing is important. And I think I mentioned 4,000-plus customers are already using it. And every traction on our platform is measured in thousands, if not tens of thousands of customers, right? That's the power of the distribution we have. And for us, it's a whole platform story. So someone coming for tokens quickly expands to agents and sandboxes and storage and databases and so forth.

So for us, for a router to be efficient, right? There's a lot of confusion between routers and just simple load balancers or traffic multiplexers. -- or is a true router in the sense that it finds the right model for the task an agent is trying to accomplish, which is very different from AI gateway. So that's number one.

Number two, for a router to perform its actions you need a catalog of not just the best AI model, but the diversity of AI models. You don't need everything to be routed to a frontier model. There are decision models, which are called System 1 models. -- there could be near frontier open weight models. There could even be cases where you don't even need a model, what you need is to execute a tool, and we have 20,000 of them available for agents.

So those are all very distinguishing attributes of our routers. And we are also so advanced in our capabilities that we are actually customizing it and adapting this for specific customer workloads because the way you do routing for coating is very different from say, doing routing for a company that is doing media generation on the fly. So we are very advanced in our capabilities even though a lot of companies are announcing it, that's all great.

And I want to keep bringing this back to our platform is the story, not a single feature in the platform.

Operator

Your next question comes from the line of Rob Oliver with Baird.

Robert Oliver

Paddy, on the disclosure around 75% of 100,000-plus AI accounts using 4 or more products, appreciated that. just wanted to get a sense from you for, I guess, what kind of uplift you're seeing from as they adopt those new products? And I assume that number predates the Open Intelligence Summit where you guys showed some pretty tremendous innovation around managed agents, action gateway droplets.

So just wanted to get a sense from you when you talk to customers, where that number -- and maybe help us frame and perhaps that's Matt with some of how that uplift, how we should think about that?

Padmanabhan Srinivasan

Yes, I appreciate the question, Rob. Yes, you're absolutely right that these numbers are over a larger or a longer window of time. And obviously, -- we are not slowing down in terms of the velocity of innovation on our platform. So the number -- I mean, we care about the number, but we care about the density of the workload and the depth of their adoption of our platform.

So yes, the breadth of products and modules that they're adopting is important, but I think the depth of the workload on our platform is more important for us because that just makes the customers stickier and we look at the expansion, the workload is driving on us. And that's why I spent a lot of time explaining how customers are entering through 4 distinct front doors on our platform. They come for GPUs.

They may come for tokens. They come for core cloud or they are increasingly coming for agent platform. And in all of these cases, they come for 1 thing and they stay on for many other things. That is the power of our platform, and that's the flywheel in action. So we really are looking forward to adding more stuff and some of the new stuff that we announced in Open Intelligence Summit, we are still the only player in the market that has recursive self-improvement for agents.

We are still the only company that has 0 ATL, near real-time data substrate for agents. So there are a lot of things that we are still the only provider that is providing this -- and that comes from the quality of customers that we have on our platform that are pushing the envelope. So we feel like -- with every quarter, this software lead is only going to expand, and we are building lapse of leads, if you think about it from a racing perspective, we have a tremendous lead and we feel really good that we are adding value to our customers, and they are repaying us by adopting the whole platform and not just a single feature.

Operator

We have reached the end of our Q&A session. This concludes today's call. Thank you so much for attending. You may now disconnect.