Blacksmith Aims to Fix AI Coding's New Bottleneck: Slow Continuous Integration

One of the company's co-founders explains that every time a developer writes code, they have to run a set of checks to make sure nothing is broken. That process is called CI, and Blacksmith writes infrastructure to run those checks very fast.

The founder's framing is that the speed of these checks determines how quickly code can move forward, which is the context for the rest of the discussion.

The co-founder says the company has grown 20 times in the last year. He says Blacksmith recently raised its Series C, led by Sequoia, at a billion-dollar valuation, and that the round totaled about $66 million.

The founder also describes the founding team as outsiders: none of them were platform engineers or worked in DevOps, and none had written a CI pipeline before. He attributes their approach to being naive enough in the right way not to kill a good idea before it meets the market.

The founder says developers now write more code than ever with agentic coding tools. He argues that writing 100 times as much code is not very useful if it cannot be shipped at that speed. As a result, the bottleneck is no longer writing software but validating it, which usually means waiting on CI.

In his description, shipping depends on making sure the tests pass and that the team has confidence that what it ships actually works.

The co-founder says he always wanted to start a company in Silicon Valley. In high school, he spent time reading Wikipedia pages about technology companies, how they were founded, and what they did, and he wanted to build something of his own.

He chose the University of Waterloo because of its startup program and its many technical founders. Blacksmith is his second application to Y Combinator, and the three founders were friends from undergraduate study who were all in Toronto at the time and wanted to start a company.

The co-founder says that at essentially every company they worked at, CI was poor. It was super slow, and managing that infrastructure was a pain across the board. The founders felt they had a better way.

He says they were initially skeptical that the idea could work, but their conviction grew as they did more research. Two co-founders were database engineers, and he was working on search infrastructure.

The co-founder says that from the outside, they saw the reason CI infrastructure was so hard to manage was that the shape of the workload differs from what hyperscalers such as AWS, GCP, and Azure are good at. He describes the mismatch as the cause of the pain.

From this, the founders concluded that the real way out was to build a cloud that is extremely good at running CI workloads, which he frames as using the best tool for the job.

The co-founder says the team realized CI workloads resemble desktop applications, particularly video games run on a typical gaming PC. These workloads care less about multiple threads of execution and more about one single thread running very fast.

This similarity is why the company uses gaming-style CPUs for its infrastructure.

The co-founder says that when the company started, it rented servers from a German data center services company known for hosting Minecraft and Call of Duty game servers. As its scale grew, Blacksmith began to own more of the stack.

He says the company now operates under much larger agreements with data center providers, effectively buying megawatts of power rather than individual machines.

An investor, who identifies as a Y Combinator figure and a Blacksmith board member, says he invested twice. He led the initial round out of Y Combinator on a SAFE, and a few months later led the Series A and joined the board.

He says he found it very interesting that the company used gaming CPUs, because he had never seen anyone else try that solution for CI. He said it surprised him that someone would put many of these machines in one place, since such hardware is usually kept at home for a single gaming setup, while Blacksmith needed to buy around a hundred of them for compiling.

The co-founder says a large part of making CI fast is skipping redundant work that has already been done before. At a high level, developers refer to this as caching.

He says a big part of Blacksmith's special approach is moving cache artifacts close to the compute nodes that run the jobs.

The co-founder says the first 200 customers were extremely hard to win. Initially, the team asked friends working at companies to give the product a try, and asked for logo rights to put on its site to build credibility. The founders then personally contacted CTOs, VPs of engineering, and platform engineers.

The earliest customers were small companies like Blacksmith itself. They cared about moving fast, wanted fast CI, and had little red tape. He says the only real way to win over engineers is to perform and solve a problem they probably could not solve themselves, and he notes that engineers are discerning and like to try many new tools, so they are not always the stickiest customers.

The co-founder says that without any sales team, the company was able to win large contracts in the five- and six-figure range, because Blacksmith made it so easy to optimize and run CI as fast as possible. He says customers can switch all of their CI to Blacksmith in one click.

The co-founder says LangChain recently onboarded to Blacksmith after previously running its CI on GCP. On GCP, developers routinely waited 15 to 20 minutes just to acquire the compute needed to run a CI job.

On Blacksmith, he says, the company effectively exposes an infinite pool of compute, so the wait dropped to anywhere from 10 to 15 seconds. He says this significantly improved the speed at which LangChain's developers could push code.

The co-founder says a big push for Blacksmith now is to make sure that when any engineering team thinks about CI speed or gaining trust in the code it generates, they immediately think of Blacksmith.

To that end, the company is rolling out a large marketing campaign across San Francisco and New York, taking over 15 subway stations in New York and wrapping a subway train in its bright yellow color.

The co-founder says that when he joined, the team was about five people running a small cluster of a couple hundred servers. Today the scale is many orders of magnitude larger, and the team has 30 people, including 12 engineers.

He attributes this to a few key decisions. The company relies on a lot of bespoke software, including software to manage fleet growth and bespoke observability at every layer of the stack. It also keeps the product's surface area small and manageable, scaling it only with the size of the team rather than giving in to bespoke customer requirements. He says this focus lets a very small team scale to a very large number of machines and customers.

The co-founder says one of the largest challenges is keeping up with demand. Every time a new LLM is released, he says, CI usage spikes massively as people rush to test it and ship more code. Because it takes time to order and install machines, the team has to get technically creative, using better algorithms for placing jobs, packing jobs more efficiently, and adopting more efficient network topologies.

He also describes a power surge in Phoenix that took out both the primary and backup cooling systems at one of the company's biggest data centers, right after the Series B announcement early in the summer. Customers saw longer startup times for their jobs during the outage, which he says underscored the importance of geographic distribution.

The co-founder says the company wants to help developers have confidence in their code changes. One product it is working on is autonomous QA, which gathers artifacts such as screenshots or video snippets that build confidence in a change.

He argues that with the amount of code being produced, developers will not have time to check whether things work, so they will look for evidence instead, and Blacksmith aims to give them confidence to merge code faster. He says the goal is to remove the bottleneck and let developers ship software as close as possible to their speed of thought.