A revenue chart that goes vertical does not tell you who pays. It tells you who gets charged.

A friend who runs infrastructure at a 200 person SaaS company called me on Monday. He had just gotten the quarterly bill from his AI vendor through an enterprise contract. He told me the number. I asked him to repeat it.

The number ran high. Not "we are building a foundation model" high. High enough that he had a week of his life mapped out to write a proposal to bring some of the workloads on prem so the company could keep the AI features without the bill getting written into the next earnings call.

That conversation kept me thinking. Anthropic's annualized revenue went from roughly $1 billion in late 2024 to $30 billion as of April 2026. A 30x in 16 months. That is not a SaaS curve. It is not a marketplace curve. It is a curve that comes from a small number of contracts each writing a number with a comma in it.

So I went looking for the line items. Not the press releases. The places the money actually flows from. Here is what I found.

What just happened, in numbers

The revenue trajectory:

• Late 2024: ~$1 billion ARR
• Mid 2025: ~$5 billion ARR
• Late 2025: ~$15 billion ARR
• April 2026: ~$30 billion ARR

In April, Anthropic passed OpenAI on revenue for the first time. The two companies shifted from "duopoly with one obvious leader" to "duopoly where the lead changes hands by quarter."

Behind the headline number sits a much smaller story than most people want to make it. The breakdown that leaked out across coverage points to roughly three quarters of revenue coming from API calls, the rest from Pro and Max subscriptions plus smaller enterprise integrations. Within the API revenue, a handful of customers account for an outsized share. AWS, through Bedrock. Microsoft, through the new Copilot integrations. Three hyperscalers and four large coding tool vendors who resell Claude wrapped in their own products.

When Anthropic says they hit $30 billion ARR, the honest reading is: a handful of large enterprise contracts, a handful of large coding tool vendors paying through the nose for Opus 4.7 access on behalf of their users, and a tail of API and subscription revenue from everyone else.

The thing nobody is saying

Individual developers do not write the $30 billion check. Their employers do, twice removed.

Trace the path. A senior engineer at a mid sized company opens GitHub Copilot, picks Claude Sonnet 4.6 from the model dropdown that GitHub added last month, and asks it to refactor a function. That call goes to GitHub. GitHub forwards it to Anthropic. Anthropic charges GitHub the per token API rate. GitHub charges the company a Copilot Enterprise seat plus, starting June 1, premium request budgets. The company charges, ultimately, the customers of its product through whatever line item is closest to "engineering labor cost".

Every leg of that chain takes margin. The model call costs $0.04 in raw inference. By the time three companies have wrapped, billed, and amortized that call, the end customer pays a multiple of the original cost.

Anthropic gets the smallest cut by percentage. They get the largest cut by absolute dollars, because the volume runs enormous and the gross margin on inference at scale runs high. The wrappers earn the rest. The end users notice nothing because the cost lives buried in seat fees and monthly minimums.

That is the mechanism. The $30 billion is not magic. The number reflects the visible part of an iceberg of indirect billing that runs through every developer tool that touched a Claude API key in the last 12 months.

Numbers that matter

A few benchmarks to anchor the scale:

Anthropic now ranks larger by revenue than any independent data company in history. They got there in two years from product launch. The closest comparable: the early growth of AWS, which took eight years to reach the same scale and had Amazon's retail business funding it.

The other number that matters: gross margin. Public analyst estimates put inference gross margin at large hyperscale operators in the 50 to 70 percent range. If Anthropic sits at the lower end, $30 billion ARR generates roughly $15 billion in gross profit. That covers a lot of training compute. Not yet a lot of net profit, because training the next model class still costs enough to consume most of the gross.

Which is where the pressure on the consumer subscriptions comes from. Pro plan subscribers cost more to serve than they pay, on average, when they use Claude Code heavily. Enterprise customers pay per token and are profitable per call. The math points in one direction. Optimize for the segment that pays per use. Defend the consumer segment as a brand and onboarding play, but do not let it grow faster than the gross can carry.

The honest take

The thing that makes this market different from past compute waves: cost scales hyper in both directions. A single team using Claude Code aggressively can spend more in a month than they did on cloud the entire previous year. A single agent loop run wrong can spend a thousand dollars overnight. The unit economics of an inference call now drive the unit economics of software.

Exhilarating if you build the wrappers. Uncomfortable if you write the checks. The companies my friend works with are realizing that the line item labeled "AI tooling" will keep growing every quarter for the foreseeable future, because the tools are getting better at burning tokens, not just at writing code.

Two things follow.

First, the wrapper layer holds the next big margin compression. Every coding tool that resells Claude or GPT sells on convenience, not on inference cost. The convenience is real. The convenience is also not defensible. Anyone with two days and an API key can build a thinner wrapper for their own team that captures 80% of the value. The wrappers know this. That explains why GitHub, Cursor, and Windsurf have all announced model dropdowns and pricing tiers in the last 90 days. They are racing to build platform features around the model so that switching costs grow before the customer notices the bill.

Second, the agentic workflow becomes the new unit. Inference cost per chat message converges across vendors and approaches a floor. Inference cost per agent run varies by ten times depending on how you write the agent. The next wave of optimization work happens here, and individual engineering teams can actually move the needle on their own bills without waiting for the vendor to drop prices.

What I am doing about it

Three concrete things, in case they help.

I cancelled the AI seat subscription that I was using for chat and switched to direct API access via a small script. The gross saving is small. The visibility is large. I now know exactly what each conversation costs.

I built a multi model router that picks the cheapest model that can do the job. About 60 percent of my queries route to Haiku. The bill dropped 41 percent in the first month with no perceptible quality loss on the routed traffic.

I started logging the cost of every AI call in our internal projects, broken out by feature. The first week of data was an unflattering surprise. One feature accounted for half the bill. We cut its budget by limiting context length and the entire feature still works. Nobody noticed.

None of this is novel. Old infrastructure habits applied to a new compute primitive. The only reason these moves feel novel: the AI vendor pitch has read "do not worry about the cost, the wrapper handles it" for so long that thinking about the bill counts as a contrarian position.

The closing

The headline number is real. The story behind it runs smaller and stranger than it sounds. A handful of contracts with hyperscalers. A handful of coding tool vendors paying through the nose to put Opus 4.7 in front of their users. The rest: a fast growing tail of API and subscription revenue from individual developers and small teams.

Anthropic earned the number. They built the better model, they shipped the better agent, and the market rewarded them with a revenue chart that reads more like an oil discovery than a software company. None of that comes off as wrong.

Worth keeping in mind: a feedback loop funds the chart. Better model. More usage. Higher bill. Higher bill funds more training. More training produces a better model. The loop runs as long as the customers keep paying without flinching. The day they flinch will be the day the wrapper market starts compressing and the routing layer becomes the most valuable real estate in the AI stack.

I think that day comes sooner than the press releases suggest. The pricing tremors of April 2026 are the early reading. Watch the bill.

---

Written by **GDS K S* (thegdsks.com), building Glincker.*
If this was useful, follow me on X / @thegdsks. I write about the parts of the AI stack vendors keep off the pricing page.