DigitalOcean Just Proved AI Infrastructure Can Be Profitable — While the Giants Drown in Debt

DigitalOcean Just Proved AI Infrastructure Can Be Profitable — While the Giants Drown in Debt

I watched a CNBC interview this week that made me want to stand up and applaud. Paddy Srinivasan, the CEO of DigitalOcean, sat down on "The Exchange" and talked about how his company is pursuing the AI infrastructure trade — and doing it profitably. Not "profitably" the way hyperscalers use the word, where you ignore free cash flow and pretend a stock price is a business model. Actual profitability. Real margins. Positive cash flow.

And then he dropped numbers that should embarrass half this industry.

The Numbers — This Is What Profitable AI Infrastructure Looks Like

Let me tell you what DigitalOcean just reported for the second quarter. Revenue of $281 million, up 29% year over year, above the high end of its own guidance. Adjusted EBITDA of $114 million — that's a 40% margin. Adjusted operating income at a 24% margin. And $61 million of adjusted free cash flow in the quarter, with trailing twelve-month free cash flow of $175 million, or 17% of revenue.

Stop and read those numbers again. A cloud company, growing 29%, converting 17% of revenue into free cash flow. That isn't fantasy accounting. That's a real business. They also retired $472 million in convertible notes and cut net leverage to 0.7 times — while still funding expansion. In an industry where everyone's balance sheet looks like a hostage note, this one reads like a spreadsheet a banker would actually sign off on.

The AI part is where it gets interesting. AI customer ARR hit $234 million, up more than 200% year over year. Inference services grew almost 800% and now represent over 70% of total AI customer ARR. The inference engine added more than 6,000 customers, and token volume grew 30x in 60 days. Remaining performance obligations — the contracted revenue sitting in the pipeline — hit $894 million, up more than 12 times year over year, with a 3.7-year average life. Incremental ARR was a record $93 million in the quarter, nearly triple what they added in the same quarter last year. They raised full-year guidance to about 30% growth and are talking openly about 50%+ growth in 2027.

What the CEO Actually Told CNBC

Here's the part that matters for anyone running infrastructure for a living. Srinivasan wasn't on TV bragging about GPU counts or gigawatts. He was talking about the length of the AI cycle — and about where the demand is actually coming from.

His projection: by 2030, the world chews through four to five quintillion AI inference tokens a year. That's five followed by eighteen zeros. Today we process about 50 trillion tokens a day; he's projecting that to pass 500 trillion daily by the end of the decade — a tenfold increase. And here's the kicker: roughly 70% of those tokens won't come from humans at all. They'll come from AI agents talking to other AI agents. Agentic workloads use about four times more CPU and fifteen times more tokens than traditional AI interactions. Goldman Sachs is singing from the same hymn sheet — its May report projects monthly token consumption hitting 120 quadrillion by 2030, a 24-fold increase.

Now, I've been around long enough to treat five-year projections with suspicion. The exact number doesn't matter. What matters is the direction, and the shape of the demand. This agent-to-agent inference traffic is exactly the kind of workload that does not need a hyperscaler. It needs distributed, accessible, cost-efficient compute — which is what DigitalOcean built with its AI-Native Cloud platform launched in April. That's the whole game: don't chase the same whale customers as the giants. Build for the millions of builders the giants can't be bothered with.

Now Look at the Other Side of the Street

I want to put this in context, because the contrast is the story. Same week, remember, Oracle is sitting on a BBB- credit rating after burning through negative $23.7 billion in free cash flow on $55.7 billion of capex. The market has already priced $334.5 billion in data center debt this year, with maybe $1.65 trillion in off-balance-sheet obligations nobody wants to count. Hyperscalers are pushing capex toward $200 billion a year apiece and watching their stocks get punished for it — Alphabet dropped 4% after raising its plan to $195-205 billion, even while beating estimates.

DigitalOcean went the other direction. It guided to positive free cash flow on every metric for 2026. It's not selling its future to buy gigawatts today. And the market is rewarding the difference: the stock is up more than 365% in a year, and analysts recently raised price targets by nearly 90% to $172.

Same AI gold rush. Two completely different balance-sheet philosophies. One of them is a business. The other one is a bet.

The Counter-Argument — It's Not All Roses

Let me be fair, because this isn't a flawless story. DigitalOcean's net dollar retention is 102% — a three-year high, but still low for a growth cloud, and they've stopped leading with it as a headline metric. That tells you expansion within the existing base is still the hard part. Their top 25 customers now represent 20% of ARR. That's concentration risk, and in this market one big customer walking away moves the needle.

There's also the supply chain question. They're at 155 megawatts of committed capacity, and management is honest that getting megawatts online on time is the operational fight of this decade. And the roughly 30% list price increase on GPU fleets that helped Q2 — that's a pricing lever that works while demand is desperate, but it's not a law of nature. If the overbuild correction cools GPU demand, that lever gets harder to pull.

Still, when your worst problems are "growing so fast we can't build capacity fast enough" and "a few of our customers are really big," you're in a much better place than "our free cash flow is negative by twenty-three billion dollars."

What This Means for Independent Hosting Providers

If you run an independent hosting or colocation business, this quarter is validation of the model you've been defending for years. Let me give you the takeaway in four pieces.

First, the AI demand curve is real — but it's shifting from training to inference. Training needs billion-dollar GPU clusters. Inference runs everywhere, at every scale. That's your lane. Position for inference workloads, agent traffic, and the long tail of builders who can't get a meeting with Azure.

Second, profitable growth is a competitive weapon. DigitalOcean is proving a mid-size cloud can grow 29% with a 40% EBITDA margin while hyperscalers drown in negative free cash flow. When the credit cycle tightens — and it will — the operators with positive cash flow and low leverage get the pick of the market. That's you, if you run your business that way.

Third, watch token economics, not just hardware. The unit of demand is changing. Agent workloads burn 15 times the tokens and four times the CPU of a human chat session. Capacity planning based on users is obsolete. Plan for machine-generated traffic, because that's where the exponential lives.

Fourth, reprice with confidence. If a 30% GPU price increase sticks at DigitalOcean, your colo rates have room too. The market is telling you demand is outstripping supply. Don't be the last operator on your street to raise rates.

The Bottom Line

Here's what I keep coming back to. Everyone in this industry got hypnotized by the hyperscaler capex circus — $200 billion here, $700 billion there, debt to the moon. And buried under all that noise, a cloud company most people still think of as "the cheap VPS guy" just posted a 40% EBITDA margin on 29% growth with positive free cash flow. That's not a fluke. That's a blueprint.

The AI infrastructure trade doesn't have to be a loss-leader. It can be a business. The question was never whether AI compute demand is real — it's whether you can make money serving it without betting the company. DigitalOcean just showed you how. Buh, you can keep waiting for the hyperscalers to figure it out. I'm not holding my breath.

— Allan Ali, Founder

This article was produced with AI-assisted research and editorial support. Reporting is based on sources cited in the article.