DigitalOcean's AI Growth Is Real. The $234M Number Isn't.

On September 8, DigitalOcean stock jumped 13% — closing near $127 — after its CEO and CFO walked an investor audience through the company's AI-inference pitch at a Goldman Sachs conference. It was the third straight week of the move, leaving the shares near $131, up roughly 173% this year, on a market cap of about $15.4 billion.

The story being bought is clean. DigitalOcean — long a cheap, friendly cloud for developers and small businesses, the kind of company you remember from the days when a $5-a-month "droplet" ran your side project — has pivoted into an "AI-native" cloud, and the market has decided it is winning. The single number everyone reaches for to prove it is AI Customer ARR of $234 million, up 212% year over year.

That number is doing a lot of work in the bull case. It is also not the number most people think it is.

The number in the headline

"ARR" — annual recurring revenue — is an annualized snapshot of how much a company is on track to bill, usually one quarter's revenue multiplied by four. DigitalOcean discloses "AI Customer ARR" by taking total revenue from its "AI customers" in the most recent quarter and quadrupling it. Where it stops being a normal number is in the definition of an "AI customer."

DigitalOcean defines AI Customer Revenue as total revenue from any customer using one or more of its AI/ML offerings — "inclusive of their revenue from IaaS and PaaS/SaaS offerings during the period." Read that again: if a customer runs one large language model on DigitalOcean but also hosts their whole website, databases, and app there, 100% of that customer's bill counts as "AI Customer Revenue."

So the $234 million is the aggregate bill of DigitalOcean's AI-adjacent customers, not the amount those customers actually spend on inference. The pure inference revenue is a slice inside it. And inside that slice is the genuinely remarkable number — inference services revenue grew 762% year over year — but a 762% rate off a tiny base is still a small dollar amount, and it is not the figure in the headline. The fastest-growing, most impressive number (inference) is buried; the headline number is the customer-level total, and it flatters the size of the real AI business.

This is not an accusation of fakery. The 212% is real within its own definition. It is a metric-identity problem: the stock is being traded on a number that is meaningfully larger than the inference workload revenue it is standing in for.

The execution is not the problem

Here is where the skeptic owes a concession, and it is a big one. Strip away the marketing language and the operating evidence is unusually concrete, which is why this is not a keynote-to-be-shrugged-off situation.

Remaining performance obligation — the backlog of contracted revenue not yet recognized — reached $894 million, up from $71 million a year ago, roughly 12x. The company landed its first nine-figure annual customer commitments and stretched the weighted average contract life from 1.6 years to over three. It commissioned three new data centers ahead of schedule and has locked in about 155 megawatts of committed capacity, with another 20 megawatts lined up for 2027 and 2028.

The tell that this is demand, not inventory, is pricing: DigitalOcean has raised list prices on some GPU offerings by about 30%, on the back of supply tightening. A cloud does not raise GPU prices because it wants to; it does it because the racks are full.

And the "is this engineering-achievable" test has at least one named, measurable pass. Character.ai, a production AI customer, reports 2x inference throughput and a 50% improvement in cost efficiency running on DigitalOcean's AMD-powered inference platform. That is a per-unit economics result, not a press-release adjective. For that customer, the platform works.

The point is to be honest about what the evidence is: the pivot is not a PR stunt, and the execution so far is the kind that usually survives contact with a real earnings call. The skepticism belongs elsewhere.

The cost side the ARR number hides

ARR is a revenue concept. It does not tell you what it costs to deliver the compute, and in this business the two move apart.

To fund roughly 80 megawatts of new capacity coming online through 2027 and early 2028, capital spending is stepping up sharply, and the consequence is already visible in the guidance. DigitalOcean guided full-year 2026 adjusted free-cash-flow margin to 11%–13%, down from a 22% adjusted FCF margin in the second quarter itself. The same quarter posted a 40% adjusted EBITDA margin and 13% GAAP net income — so this is not a company losing money, it is a company deliberately spending cash to buy the capacity it needs to be the company it claims to be. Total debt rose to about $2.19 billion even after it retired $472 million of 2030 convertible notes, leaving pro-forma net leverage around 0.7x.

The margin that has to hold is the one management keeps pointing at: core cloud runs at roughly 70% gross margin, and 85% of AI Customer ARR now sits in higher-margin inference and core-cloud services rather than bare-metal GPU leasing (up from 43% a year ago). That is the whole bet in one line — use megawatts of hardware, but capture software-grade margin by pricing on tokens and agents instead of by the hour. If that 70% survives the GPU capex, the conversion works. If the hardware bill outruns the token bill, the 22% FCF margin was the last good look at the old, capital-light DigitalOcean.

What 15x revenue is actually pricing in

At roughly 15x trailing revenue, 65x trailing earnings, and about 48x trailing EBITDA, the multiple has already done the work the ARR number is asked to do. The market is not paying for the small-business cloud; it is paying for the inference business to become the company. Seventeen analysts cover the stock with zero sell ratings and a mean price target of $175; the bull model that circulates in the trade values it well above $500 by 2030 on exactly the software-margin conversion described above. The upside is real, and it is not free — it is priced.

Which brings the attribution question the headline skips. Is DigitalOcean's gain challenger strength, or incumbent mis-positioning, or both? The honest read is a mix that matters. Its genuine edge is that the hyperscalers built their AI stack for the training era and for enormous enterprise deals, while inference for mid-market and AI-native startups is a different workload — and "simpler, predictable, cheaper per token" is a real total-cost win, the 50% Character.ai cut being the unit proof. But "cheaper and simpler" is the kind of moat a hyperscaler erases with a price cut or a good-enough product. Part of what DigitalOcean has captured may be hyperscalers' failure to compete for this niche, which is share a bigger player can reclaim, not a structural advantage that compounds.

So can the stock keep up? The answer does not ride on the next ARR headline. It rides on whether inference-specific revenue — the small number, not the $234 million customer total — compounds out of a small base into a large, software-margin one while carrying a much heavier capex load, and whether that 70% core-cloud margin survives the GPU spend. The multiple already assumes yes. The investor's job now is to track the slice the market is confusing with the whole.