The queue starts forming at nine in the morning, Pacific time.
A pull request lands. Then four more. Then forty. Each one needs a build, a test suite and a lint pass before any human reads the diff. At Anthropic, the volume of jobs feeding that machinery was 25 times higher in September than it had been six months earlier, as Claude took over most of the company's code output.
That is the moment continuous integration stopped being plumbing and became the constraint.
The bottleneck moved downstream
Aditya Jayaprakash watched the same pattern appear in customer conversations long before it showed up in a pitch deck. He had spent his engineering years at Faire, on systems where a bad capacity plan turns into a bad afternoon. Two friends from the University of Waterloo, Aayush Shah and Aditya Maru, had built distributed infrastructure at Cockroach Labs. In 2024 the three of them started Blacksmith on a bet that sounded eccentric at the time: that testing code, not writing it, would become the thing that limits how fast software ships.
Back then the argument was theoretical. Autocomplete was graduating into a coding assistant, and the volume of generated code was rising fast enough to be interesting. By 2026 it had become obvious. Agents now open pull requests in minutes, and every one of them still has to clear the same gates.
Writing code is no longer the slow part. Proving it is safe is.
The meter that ran out of road
Blacksmith sells one thing: a CI cloud. Customers change a line in a workflow file and their builds move onto machines Blacksmith owns, installed in data centres across the United States and Europe. The company claims jobs finish twice as fast at roughly 60% lower cost. It writes its own scheduler and its own caching layer, so the repetitive setup work that normally precedes every test run gets skipped rather than repeated.
The numbers behind that pitch moved quickly. In January, Blacksmith was running about 5 million CI jobs a week. It now runs more than 90 million. Annual recurring revenue is up 17 times over the same period, and revenue sits in the tens of millions of dollars. More than 10,000 companies are on the platform, roughly double the count in August.
Weekly validation load on Blacksmith
Up from 5 million jobs a week at the start of 2026, with ARR rising 17× over the same window. · Blacksmith, Sequoia, October 2026
CI is a strange workload to build a company around. Demand is spiky in a way that general-purpose clouds handle badly: a single customer can jump from zero to tens of thousands of CPUs and back within an hour, then repeat the trick the next morning. Renting that capacity from a hyperscaler means paying for elasticity you cannot control. Buying bare metal means eating the idle hours yourself. Blacksmith chose the second option and then wrote software to make the idle hours disappear.
A valuation ladder built in twelve months
The funding history reads less like a growth curve than a staircase. In May 2025 the company took a $3.5 million seed. Four months later GV led a $10 million Series A at a $60 million valuation. The Series B arrived in August 2026: $45 million led by Peak XV at $550 million. The Series C landed on 21 October, $66 million led by Sequoia at $1 billion, with Premji Invest joining and Peak XV, GV and Y Combinator returning. Total raised: $125 million.
| Round | Date | Amount | Valuation |
|---|---|---|---|
| Seed | May 2025 | $3.5M | — |
| Series A | Sep 2025 | $10M | $60M |
| Series B | Aug 2026 | $45M | $550M |
| Series C | Oct 2026 | $66M | $1B |
Blacksmith funding history, per company announcements and TechCrunch, October 2026
Validation, not generation, is where agentic coding now spends its wall-clock time — and that is a purchasable infrastructure problem.
Blacksmith's 17× ARR growth is a demand signal from engineering teams, not a marketing claim: 10,000 companies moved real CI workloads onto it.
The risk is structural, not competitive. If verification migrates into the agent's own loop, the dedicated CI cloud loses its reason to exist.
Sequoia's Isaiah Boone and Carl Eschenbach framed the bet in their announcement post as a shift in what software work even is: writing code is no longer the hard part, trusting it is. Premji Invest made a narrower argument, that the orchestration layer of CI was always thin and the infrastructure underneath was always the real business. Both readings point at the same thing. The layer that got repriced this year is the one that turns generated text into something a company is willing to merge.
Developers who used to push three or four pull requests a day are now pushing orders of magnitude more and their CI infrastructure can't keep up.— Aditya "JP" Jayaprakash, co-founder and CEO, Blacksmith
What ten megawatts actually buys
Blacksmith runs just under 10 megawatts of compute today. It plans to grow that footprint at least fivefold in 2027. That is a capital-intensive way to sell a developer tool, and it puts the company in the same queue for servers and memory as everyone else building AI infrastructure.
Jayaprakash has been candid about the operational side. Capacity planning is one of the hardest parts of the business, he told Fortune, because demand jumps every time a stronger model ships. When Anthropic released Claude Fable 5 in July, CI volume on Blacksmith rose 40%.
That detail matters more than the valuation. It means Blacksmith's revenue is levered to model releases it does not control, in both directions. A quiet quarter for frontier labs is a quiet quarter for validation load. A step-change in agent capability is an overnight capacity crisis.
The counter-case
The strongest argument against the CI cloud is that it solves today's version of the problem.
Some teams are already moving verification earlier, into the agent's own loop. Paul Stack, writing about his work at Swamp in August, described having the agent run the full verification cycle before it opens a pull request and produce an attestation of what ran; CI then validates the attestation instead of re-executing everything. Thirty-five pull requests in, he said, CI had not caught anything the earlier loop missed.
If that pattern generalises, the volume of work reaching a dedicated CI layer falls rather than rises. The counter-counter is cost: running builds on the same machine as the agent burns tokens and time, and most engineering organisations will not rebuild their pipelines around a practice that is still being invented. GitHub Actions remains bundled, familiar and free at the margin. Blacksmith's answer is speed and price, which is a real answer until the workload changes shape.
There is also the hardware bet. Owning servers gives Blacksmith control over performance and exposes it to the industry-wide shortage of compute and memory. The company is not renting flexibility; it is buying a fixed position and hoping demand stays on its side of the fence. As we wrote in October about Daytona's agent-native cloud, the winners in this layer are the ones who own the machine and can price the millisecond — until the workload moves and the machine does not.
Does the CI cloud survive the agent that verifies itself?
The constraint is organisational, not technical: most teams will not trust an attestation they cannot re-run.
🟢 Agent-native verification stays a niche practice (30%)
Consequences: CI becomes a standard infrastructure line item, priced per job like bandwidth.
🟡 Both layers grow, the boundary blurs (50%)
Consequences: Blacksmith's growth moderates but does not reverse; incumbents copy the pricing model.
🔴 Verification collapses into the model provider (20%)
Consequences: the standalone CI cloud gets squeezed between model vendors and hyperscalers, and the 10 MW becomes stranded capital.
The honest read on Blacksmith is that it found a real bottleneck and got to it before the incumbents noticed the bottleneck had moved. That is a good business. Whether it is a durable one depends on a question the round does not answer: how much verification stays outside the agent, and for how long.