DigitalOcean (DOCN)
Cloud infrastructure, managed services, and AI tooling for developers and growing technology businesses.
Something off on this page? Send us feedback.
DigitalOcean rents computers by the hour to the people who build software — small teams who want a server running in a minute, without a salesperson or a contract. That cheap, self-serve business still pays most of the bills. Now the company is pouring money into data centers and AI chips so those same builders can run artificial intelligence on it, and it is betting the next chapter on them.
Item facts: FY2025 10-K + H1 FY2026 filings · year ends Dec 31, from filings, earnings calls and company pages.
Judgment weights, not filed revenue
The band summarizes business focus and direction. ~ marks estimates.
9 in detail · 14 more below
Droplets
Plain rented servers, from four dollars a month, sold to more than 680,000 paying customers. Much the biggest line, and the thing AI customers end up attaching to — but it is also the slow-growing part.
Competes with EC2 and Lightsail (Amazon) · Linode Compute Instances (Akamai) · Flat-rate virtual servers (Vultr)
In plain English
A Droplet is a computer you rent instead of buy. You choose how much memory and disk you want, type in a card number, and a minute later there is a machine on the internet that is yours until you switch it off.
Billing runs by the hour, starting around four dollars a month, so a weekend project costs pocket change while a busy business quietly spends thousands. More than 680,000 paying customers keep the lights on this way, and the roughly 21,000 of them spending over five hundred dollars a month brought in about sixty percent of last year's revenue.
Managed Kubernetes and App Platform
Two ways to stop babysitting servers: a manager that spreads an app across many machines, and a service that publishes straight from a customer's code. Management lists it among the pieces AI work now demands.
Competes with Elastic Kubernetes Service (Amazon) · Google Kubernetes Engine (Google) · Linode Kubernetes Engine (Akamai)
In plain English
Picture a restaurant kitchen on a busy night. Kubernetes is the head chef: it decides which station cooks what, restarts anything that burns, and calls in extra hands when orders pile up — except the stations are rented machines and the dishes are pieces of one app.
DigitalOcean runs that head chef for customers and charges for the machines underneath. App Platform is the simpler door: point it at your code and the site goes live. Both earn less as a price tag of their own than as a reason to keep more rented capacity busy.
Inference Engine
Customers send text to a ready-made AI model and pay only for what comes back — from five cents a million pieces of text. Up roughly 800% in a year off a small base, and now most of what AI customers spend.
Competes with Token Factory inference service (Nebius) · Bedrock (Amazon) · Baseten inference platform (Baseten)
In plain English
Running a modern AI model yourself means buying scarce chips, keeping them fed, and paying for them at three in the morning whether anyone is using them or not. The alternative is to send your question to somebody else's machine and pay only for the answer. That is what this sells.
DigitalOcean keeps a catalogue of more than seventy-five models ready to answer, plus a traffic controller that picks a cheaper or a faster one depending on what was asked. Charges start at five cents for a million pieces of text — the chunks models are billed by — so a small app pays cents and a busy one pays like a utility customer.
Multi-year AI commitments
Signed promises to spend rather than money earned: $893.8M at mid-2026, more than twelve times a year earlier, with about $366M of it due within twelve months. Conversion waits on data centers opening.
Competes with Multi-year capacity contracts (CoreWeave) · Enterprise committed-spend deals (Amazon Web Services)
In plain English
Most of this company's revenue is a card swipe that can stop tomorrow. This is the opposite — AI companies signing up years ahead for capacity, with the average promise running close to four years. One of them committed to over a hundred million dollars a year; DigitalOcean has not said who.
The money only becomes revenue as the machines actually run, so the pile is a to-do list, not a bank balance. Management says demand sits well ahead of the capacity it can build, which makes the opening dates of new sites the thing that decides the pace. The twenty-five largest customers are still only about a fifth of the run rate.
Managed Agents
Tools for software that acts on its own — doing errands, calling other programs, remembering what it tried. Almost no revenue yet; how it gets charged for is due at the company's October event.
Competes with Bedrock AgentCore (Amazon) · AI SDK and AI Gateway (Vercel)
In plain English
An agent is a program you hand a goal instead of a command: find the cheapest supplier, then place the order. To do that it needs a safe place to work, a memory of what it has already tried, and permission to use tools like search and other people's software.
DigitalOcean sells that workshop — sealed workspaces, starter kits, and a way to see what the agent actually did. Barely anyone is paying much for it so far. The point today is that agents chew through the AI models and rented machines on the floors below, and those do get billed.
Storage and Networking
Where customers' files, backups and private links live. It rarely wins a customer alone, but it lifts the bill on every rented server and makes leaving a chore.
Competes with S3 and EBS (Amazon) · R2 object storage (Cloudflare) · Linode Object Storage (Akamai)
In plain English
Renting a server is like renting an empty room — you still need somewhere to put your things. This is the cupboard, the loading door and the corridors: file and disk storage, backups and snapshots, private networks, and the plumbing that moves data in and out.
It is sold by the gigabyte next to the servers, so nearly every bill grows a little. The quiet value is stickiness: once a company's files and backups sit here, moving elsewhere means moving all of it. Management says the newer AI work needs more of this too, because programs that act on their own have to keep notes somewhere.
Managed Databases and Knowledge Bases
DigitalOcean runs the customer's database for them — backups, patches, midnight failures. Around six percent of revenue, and management counts it as one of three things that make customers hard to prise away.
Competes with RDS and Aurora (Amazon) · Pinecone vector database (Pinecone) · Linode Managed Databases (Akamai)
In plain English
Every order, message and password an app handles lives in its filing room, and that filing room has to answer every single time. Keeping one healthy means backups, security patches and a plan for the night a disk dies — chores DigitalOcean takes over for a monthly fee.
Since April 2026 it also holds the reference material AI models look things up in, along with a search store for that material priced from twenty dollars a month while it is still being tested in public. The pitch is plain enough: software can be moved, but nobody wants to move their data.
Cloudways
Managed hosting for web agencies and small firms, bought in 2022. The one line with no AI wind behind it — it pays its way and has stayed out of the story for over a year.
Competes with WP Engine managed hosting (WP Engine) · Kinsta managed hosting (Kinsta)
In plain English
Somebody has to keep the small-business website up. Cloudways does that for web agencies and small firms — largely WordPress sites — running them on DigitalOcean's own machines for a monthly fee, which is why every extra customer here uses hardware the company already owns.
The market is crowded and price-conscious and grows slowly. Management once called this business a drag on how much existing customers grow their spending, and it has not come up on an earnings call since mid-2025. The newer wrinkle: an assistant shipped in 2025 that diagnoses server and site problems on its own.
GPU Droplets and Bare Metal GPUs
Raw AI chips rented by the hour, from about 76 cents. Demand outruns supply and list prices went up roughly 30% — yet the company is steering customers up to the model service instead.
Competes with CoreWeave Cloud (CoreWeave) · GPU cloud instances (Lambda) · EC2 GPU instances (Amazon)
In plain English
These are the expensive chips that do the heavy arithmetic behind AI, rented by the hour much as you would rent a cement mixer — about 76 cents an hour to start. Some customers want the machine bare, with nothing of DigitalOcean's on top of it.
It will sell that, but would rather not. By management's own yardstick, bare machines earn far less from the same electricity than the services stacked above them, and the share of AI customers' spending that arrives this way fell to fifteen percent from thirty a year earlier. The chips come from both NVIDIA and AMD, and management says it is close to not minding which.
DropletsPlain rented servers, from four dollars a month, sold to more than 680,000 paying customers. Much the biggest line, and the thing AI customers end up attaching to — but it is also the slow-growing part.Plain rented servers, from four dollars a month, sold to more than 680,000 paying customers. Much the biggest line, and the thing AI customers end up attaching to — but it is also the slow-growing part.
In plain English
A Droplet is a computer you rent instead of buy. You choose how much memory and disk you want, type in a card number, and a minute later there is a machine on the internet that is yours until you switch it off.
Billing runs by the hour, starting around four dollars a month, so a weekend project costs pocket change while a busy business quietly spends thousands. More than 680,000 paying customers keep the lights on this way, and the roughly 21,000 of them spending over five hundred dollars a month brought in about sixty percent of last year's revenue.
Competes with EC2 and Lightsail (Amazon) · Linode Compute Instances (Akamai) · Flat-rate virtual servers (Vultr)
Managed Kubernetes and App PlatformTwo ways to stop babysitting servers: a manager that spreads an app across many machines, and a service that publishes straight from a customer's code. Management lists it among the pieces AI work now demands.Two ways to stop babysitting servers: a manager that spreads an app across many machines, and a service that publishes straight from a customer's code. Management lists it among the pieces AI work now demands.
In plain English
Picture a restaurant kitchen on a busy night. Kubernetes is the head chef: it decides which station cooks what, restarts anything that burns, and calls in extra hands when orders pile up — except the stations are rented machines and the dishes are pieces of one app.
DigitalOcean runs that head chef for customers and charges for the machines underneath. App Platform is the simpler door: point it at your code and the site goes live. Both earn less as a price tag of their own than as a reason to keep more rented capacity busy.
Competes with Elastic Kubernetes Service (Amazon) · Google Kubernetes Engine (Google) · Linode Kubernetes Engine (Akamai)
Inference EngineCustomers send text to a ready-made AI model and pay only for what comes back — from five cents a million pieces of text. Up roughly 800% in a year off a small base, and now most of what AI customers spend.Customers send text to a ready-made AI model and pay only for what comes back — from five cents a million pieces of text. Up roughly 800% in a year off a small base, and now most of what AI customers spend.
In plain English
Running a modern AI model yourself means buying scarce chips, keeping them fed, and paying for them at three in the morning whether anyone is using them or not. The alternative is to send your question to somebody else's machine and pay only for the answer. That is what this sells.
DigitalOcean keeps a catalogue of more than seventy-five models ready to answer, plus a traffic controller that picks a cheaper or a faster one depending on what was asked. Charges start at five cents for a million pieces of text — the chunks models are billed by — so a small app pays cents and a busy one pays like a utility customer.
Competes with Token Factory inference service (Nebius) · Bedrock (Amazon) · Baseten inference platform (Baseten)
Multi-year AI commitmentsSigned promises to spend rather than money earned: $893.8M at mid-2026, more than twelve times a year earlier, with about $366M of it due within twelve months. Conversion waits on data centers opening.Signed promises to spend rather than money earned: $893.8M at mid-2026, more than twelve times a year earlier, with about $366M of it due within twelve months. Conversion waits on data centers opening.
In plain English
Most of this company's revenue is a card swipe that can stop tomorrow. This is the opposite — AI companies signing up years ahead for capacity, with the average promise running close to four years. One of them committed to over a hundred million dollars a year; DigitalOcean has not said who.
The money only becomes revenue as the machines actually run, so the pile is a to-do list, not a bank balance. Management says demand sits well ahead of the capacity it can build, which makes the opening dates of new sites the thing that decides the pace. The twenty-five largest customers are still only about a fifth of the run rate.
Competes with Multi-year capacity contracts (CoreWeave) · Enterprise committed-spend deals (Amazon Web Services)
Managed AgentsTools for software that acts on its own — doing errands, calling other programs, remembering what it tried. Almost no revenue yet; how it gets charged for is due at the company's October event.Tools for software that acts on its own — doing errands, calling other programs, remembering what it tried. Almost no revenue yet; how it gets charged for is due at the company's October event.
In plain English
An agent is a program you hand a goal instead of a command: find the cheapest supplier, then place the order. To do that it needs a safe place to work, a memory of what it has already tried, and permission to use tools like search and other people's software.
DigitalOcean sells that workshop — sealed workspaces, starter kits, and a way to see what the agent actually did. Barely anyone is paying much for it so far. The point today is that agents chew through the AI models and rented machines on the floors below, and those do get billed.
Competes with Bedrock AgentCore (Amazon) · AI SDK and AI Gateway (Vercel)
Storage and NetworkingWhere customers' files, backups and private links live. It rarely wins a customer alone, but it lifts the bill on every rented server and makes leaving a chore.Where customers' files, backups and private links live. It rarely wins a customer alone, but it lifts the bill on every rented server and makes leaving a chore.
In plain English
Renting a server is like renting an empty room — you still need somewhere to put your things. This is the cupboard, the loading door and the corridors: file and disk storage, backups and snapshots, private networks, and the plumbing that moves data in and out.
It is sold by the gigabyte next to the servers, so nearly every bill grows a little. The quiet value is stickiness: once a company's files and backups sit here, moving elsewhere means moving all of it. Management says the newer AI work needs more of this too, because programs that act on their own have to keep notes somewhere.
Competes with S3 and EBS (Amazon) · R2 object storage (Cloudflare) · Linode Object Storage (Akamai)
Managed Databases and Knowledge BasesDigitalOcean runs the customer's database for them — backups, patches, midnight failures. Around six percent of revenue, and management counts it as one of three things that make customers hard to prise away.DigitalOcean runs the customer's database for them — backups, patches, midnight failures. Around six percent of revenue, and management counts it as one of three things that make customers hard to prise away.
In plain English
Every order, message and password an app handles lives in its filing room, and that filing room has to answer every single time. Keeping one healthy means backups, security patches and a plan for the night a disk dies — chores DigitalOcean takes over for a monthly fee.
Since April 2026 it also holds the reference material AI models look things up in, along with a search store for that material priced from twenty dollars a month while it is still being tested in public. The pitch is plain enough: software can be moved, but nobody wants to move their data.
Competes with RDS and Aurora (Amazon) · Pinecone vector database (Pinecone) · Linode Managed Databases (Akamai)
CloudwaysManaged hosting for web agencies and small firms, bought in 2022. The one line with no AI wind behind it — it pays its way and has stayed out of the story for over a year.Managed hosting for web agencies and small firms, bought in 2022. The one line with no AI wind behind it — it pays its way and has stayed out of the story for over a year.
In plain English
Somebody has to keep the small-business website up. Cloudways does that for web agencies and small firms — largely WordPress sites — running them on DigitalOcean's own machines for a monthly fee, which is why every extra customer here uses hardware the company already owns.
The market is crowded and price-conscious and grows slowly. Management once called this business a drag on how much existing customers grow their spending, and it has not come up on an earnings call since mid-2025. The newer wrinkle: an assistant shipped in 2025 that diagnoses server and site problems on its own.
Competes with WP Engine managed hosting (WP Engine) · Kinsta managed hosting (Kinsta)
GPU Droplets and Bare Metal GPUsRaw AI chips rented by the hour, from about 76 cents. Demand outruns supply and list prices went up roughly 30% — yet the company is steering customers up to the model service instead.Raw AI chips rented by the hour, from about 76 cents. Demand outruns supply and list prices went up roughly 30% — yet the company is steering customers up to the model service instead.
In plain English
These are the expensive chips that do the heavy arithmetic behind AI, rented by the hour much as you would rent a cement mixer — about 76 cents an hour to start. Some customers want the machine bare, with nothing of DigitalOcean's on top of it.
It will sell that, but would rather not. By management's own yardstick, bare machines earn far less from the same electricity than the services stacked above them, and the share of AI customers' spending that arrives this way fell to fifteen percent from thirty a year earlier. The chips come from both NVIDIA and AMD, and management says it is close to not minding which.
Competes with CoreWeave Cloud (CoreWeave) · GPU cloud instances (Lambda) · EC2 GPU instances (Amazon)
Named in filings, launches and programs
- FunctionsProductSmall pieces of code that run only when something calls them, billed only while they run — the quick way to glue an app together.
- MarketplaceEcosystemOne-click installs of other companies' software onto a rented server, saving customers the setup.
- Container RegistryProductPrivate storage for packaged app images, sold alongside the Kubernetes service that deploys them.
- UptimeProductWatches a customer's site from the outside and raises an alert when it stops answering.
- InsightsProductDashboards showing what the machines and apps are doing; one of the 80-plus releases shipped since April 2026.
- 1-Click ModelsProductPuts an open model from Hugging Face or DeepSeek onto a customer's machine in a single step.
- Jupyter NotebooksProductBrowser notebooks for experimenting with code and models — the surviving piece of the Paperspace purchase, still listed in the annual report.
- HatchCustomer programFree credits and perks for startups, with the benefits widened in January 2025.
- DigitalOcean Partner ProgramEcosystemA partner program launched in October 2025.
- Partner Network ConnectServicePrivate links between DigitalOcean and other clouds, added April 2025, for customers keeping work in more than one place.
- Laravel VPSEcosystemLaravel's own server-provisioning product, which sets customers up on DigitalOcean; launched October 2025.
- Persistent Systems partnershipCustomer programA strategic AI partnership announced in December 2025.
- Cloudways CopilotProductAn assistant inside the hosting business that diagnoses server and site problems; available to everyone since August 2025.
- Katanemo LabsBrandBought in April 2026; its technology underpins the traffic controller that decides which AI model answers a request.
FunctionsProduct
Small pieces of code that run only when something calls them, billed only while they run — the quick way to glue an app together.
MarketplaceEcosystem
One-click installs of other companies' software onto a rented server, saving customers the setup.
Container RegistryProduct
Private storage for packaged app images, sold alongside the Kubernetes service that deploys them.
UptimeProduct
Watches a customer's site from the outside and raises an alert when it stops answering.
InsightsProduct
Dashboards showing what the machines and apps are doing; one of the 80-plus releases shipped since April 2026.
1-Click ModelsProduct
Puts an open model from Hugging Face or DeepSeek onto a customer's machine in a single step.
Jupyter NotebooksProduct
Browser notebooks for experimenting with code and models — the surviving piece of the Paperspace purchase, still listed in the annual report.
HatchCustomer program
Free credits and perks for startups, with the benefits widened in January 2025.
DigitalOcean Partner ProgramEcosystem
A partner program launched in October 2025.
Partner Network ConnectService
Private links between DigitalOcean and other clouds, added April 2025, for customers keeping work in more than one place.
Laravel VPSEcosystem
Laravel's own server-provisioning product, which sets customers up on DigitalOcean; launched October 2025.
Persistent Systems partnershipCustomer program
A strategic AI partnership announced in December 2025.
Cloudways CopilotProduct
An assistant inside the hosting business that diagnoses server and site problems; available to everyone since August 2025.
Katanemo LabsBrand
Bought in April 2026; its technology underpins the traffic controller that decides which AI model answers a request.