AI Coding Tools Flood CI Pipelines, Fueling Demand for Blacksmith’s Bare-Metal Alternative

The validation bottleneck in software development has moved from annoyance to emergency. AI coding assistants like Cursor, OpenAI’s Codex, and Anthropic’s Claude Code have collapsed the time developers spend writing code. But every line they generate still needs to be built, tested, and merged — a process that still runs on the same cloud infrastructure teams have used for years.

Blacksmith, a two-year-old startup that owns and operates its own fleet of servers rather than renting from Amazon, Google, or Microsoft, just closed a $66 million Series C at a $1 billion valuation led by Sequoia Capital. The round comes barely two months after a $45 million Series B led by Peak XV Partners at a $550 million valuation, bringing total funding to roughly $124.5 million.

The numbers illustrate the pressure. Blacksmith ran 5 million validation checks per week at the start of 2025. Today it runs 90 million. Annual recurring revenue has grown 17-fold over the same period, now sitting in the tens of millions, according to cofounder and CEO Aditya Jayaprakash. More than 10,000 companies use the platform, up from 5,000 in August. Customers include Granola, Handshake, Supabase, and LangChain. The largest accounts spend more than $1 million annually.

Jayaprakash told Fortune most of the new capital will purchase additional compute capacity. The company employs about 40 people across San Francisco and New York and plans to reach 75 by year-end.

The core differentiator is hardware ownership. While GitHub Actions, CircleCI, and others run workloads on rented cloud instances, Blacksmith buys its own machines — tens of thousands, Jayaprakash says — and installs them in data centers worldwide. It writes its own orchestration software to eliminate the repetitive setup that slows each test run on shared infrastructure.

That approach matters because AI has inverted the economics of continuous integration. Developers once split their time roughly evenly between writing code and waiting for builds. AI tools have made the first half “extremely easy,” Jayaprakash said. “But now the bottleneck has shifted to the second piece: how do I actually validate all this code?”

Even model providers feel the strain. Anthropic disclosed last month that its internal validation workloads grew 25-fold in six months as Claude took over most code generation. When Anthropic released Claude 3.5 Sonnet in July, Blacksmith saw a 40% spike in CI volume almost immediately.

Owning hardware gives Blacksmith control over performance but exposes it to the same compute and memory shortages affecting the entire industry. Capacity planning is “one of the hardest parts of the business,” Jayaprakash said, because demand jumps unpredictably with each major model release.

A longer-term question looms: whether AI model vendors will absorb validation into their own platforms. GitHub, owned by Microsoft, already sells hosted runners for CI. Cursor launched an agentic code-review product in March. Codex and Claude Code can execute a project’s tests as they work. Jayaprakash argues the relationship is complementary — better models produce more code, which requires more validation. “Usage on us actually goes up,” he said.