Sequoia Capital led Blacksmith’s $66 million Series C on October 21, 2026, valuing the San Francisco-based continuous integration startup at $1 billion — a unicorn crossing that happened in under two years from founding, and one that Sequoia is betting on because AI coding agents have turned a decades-old infrastructure problem into an acute crisis. The round draws in Premji Invest as a new investor alongside returning backers Peak XV Partners, GV, and Y Combinator, and brings Blacksmith’s total funding to $125 million.

The valuation expansion is steep by any measure. Just two months ago, Blacksmith’s Series B at $550 million put it at that figure. The new round doubles that figure in a single news cycle. What drove it: CI jobs grew from 5 million per week in January 2026 to more than 90 million per week by October — an 18-fold surge in under ten months — and annual recurring revenue grew 17 times in the same period.

What AI Coding Agents Did to CI Infrastructure

Continuous integration has existed as a software practice for more than 35 years. Grady Booch coined the term in 1991; Kent Beck formalized it inside Extreme Programming in 1997; the first open-source CI tool, CruiseControl, arrived in 2001. All of those CI tools, from Booch through CruiseControl, rest on a simple principle: every time a developer merges code into a shared branch, an automated system builds and tests it. The value is detecting errors within hours instead of weeks.

That value scales directly with build speed. CI infrastructure that takes 45 minutes to validate a change leaves developers waiting, and waiting developers stop submitting frequently — defeating the point. But for decades, the rate at which humans produced pull requests kept that problem manageable on general-purpose cloud machines.

AI coding agents ended that equilibrium.

Tools like Anthropic’s Claude Code, OpenAI’s Codex, and Cursor have not merely accelerated code production — they have restructured the demand curve for CI in a way that human commit patterns never did. A human developer submitting three or four pull requests per day creates a predictable trickle of CI work. An AI agent running an iterative debugging loop — submit code, watch CI, diagnose failure, amend code, resubmit — can trigger 10 to 20 CI runs per task. The cost multiplier is not commit volume; it is plan steps per agent session.

GitHub’s published data illustrates the scale: AI-agent-opened pull requests on the platform grew from approximately 4 million in September 2025 to more than 17 million in March 2026 — a more than fourfold increase in six months. Weekly compute minutes consumed by GitHub Actions grew from 500 million in 2023 to 2.1 billion in a single week in 2026. GitHub’s infrastructure showed strain publicly, with multiple platform incidents in 2026 as demand patterns shifted.

Anthropic’s own internal CI infrastructure grew 25-fold in the six months after Claude Code launched in March 2026, according to a disclosure Blacksmith cited in its Series C announcement. CEO Aditya Jayaprakash described watching the same dynamic arrive across his customer base.

“Since Claude Code took off in March, almost every conversation we’ve had with a large engineering team has started the same way,” Jayaprakash said. “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. We’re grateful to our customers who are trusting us to solve this bottleneck, and excited to partner with Sequoia to solve it at scale.”

The deepest implication — not captured in the raw job-volume numbers — is that as AI agents become more autonomous, the cost structure shifts further. An agent that can run longer planning loops before escalating to a human creates more CI work, not less. CI infrastructure spending, not LLM API cost, may emerge as the larger line item in AI-era software development budgets.

The Engineering Stack Behind the Speed Claim

Blacksmith’s claim — twice as fast as GitHub-hosted runners, at up to 60 percent lower cost — is not a marketing abstraction. It rests on four specific architectural decisions that address the root cause of why general-purpose cloud runners fail at CI’s spike-and-drain demand shape.

The CPU choice: desktop silicon over server silicon. Blacksmith runs CI jobs on AMD Ryzen 7950X processors with boost clocks reaching 5.7 GHz. GitHub’s hosted runners use AMD EPYC server-grade chips — designed for parallel multi-threaded workloads, not for maximizing single-thread throughput. Most CI work — compilation, test execution, linting — is single-thread bound or poorly parallelized. Higher single-core clock speed translates directly into shorter wall-clock run time for CI, and Blacksmith deliberately chose desktop-optimized silicon over server-optimized silicon for this reason.

The storage choice: local NVMe over network-attached disks. GitHub-hosted runners store caches on network-attached storage — AWS EBS — meaning every cache read and write crosses a network hop. Blacksmith collocates cache storage on PCIe NVMe SSDs physically attached to the same bare-metal host as the job. Local NVMe retrieval reaches approximately 400 megabytes per second; network-attached cache retrieval runs closer to 100 megabytes per second.

The difference matters because CI workflows spend 30 to 50 percent of their total run time pulling and pushing caches — npm package trees, Docker image layers, Gradle dependency graphs. A 4× storage speed improvement on 30 to 50 percent of the workload compresses run times materially. CI jobs are also ephemeral by nature — if a host fails, the job reruns from scratch — which means the redundancy that justifies network-attached storage in stateful applications provides no benefit in CI, and local NVMe’s non-redundant design is architecturally safe for this use case.

The isolation choice: Firecracker microVMs per CI job. Each CI job on Blacksmith runs inside an ephemeral Firecracker microVM — a lightweight virtual machine built on KVM (Kernel-based Virtual Machine) that AWS open-sourced and uses internally for AWS Lambda. Firecracker boots in approximately 5 milliseconds, fast enough to provision a fresh VM per job without meaningfully increasing job startup latency. Every job gets hardware-level isolation from every other job on the same physical host, including pull requests from untrusted contributors running potentially adversarial code.

This isolation matters because CI platforms are high-value security targets: they routinely hold access tokens, environment variables, cryptographic keypairs, and cloud credentials. CircleCI’s January 2023 security incident exposed exactly these categories of secrets to attackers during a roughly two-week breach window, urging all customers to rotate any CI-stored credentials. Firecracker’s KVM-based isolation ensures a compromised job cannot reach another job’s memory space or credentials.

The orchestration: Redis queues and Go agent daemons. When GitHub sends a webhook signaling a new CI job, Blacksmith’s backend validates the webhook signature, parses job metadata and matrix variables, and enqueues the job in a Redis queue. Go agent daemons running on bare-metal hosts poll that queue, spin up the appropriate Firecracker VM, execute the job steps, stream logs back to GitHub, and report completion status — all within the standard Blacksmith’s architecture for GitHub Actions interface, meaning customers see no difference in how they monitor jobs.

The cost model is counterintuitive until you see the arithmetic. A bare-metal host with a high-frequency desktop CPU costs more per second than GitHub’s shared cloud instances. But if that host finishes a CI job in half the time, the net cost per job is identical or lower, depending on the workload — and Blacksmith claims typical jobs run not just twice as fast but at up to 60 percent lower total cost. LangChain’s infrastructure engineer Joaquin Borggio put the real-world figure at around 40 percent lower cost in production. “Borggio reported roughly 40 percent savings — and that doesn’t even account for the engineering time we’ve saved,” Borggio said.

The gap between Blacksmith’s headline claim (60 percent) and LangChain’s production figure (40 percent) is worth noting: it likely reflects workload variability. Cache-heavy workflows, where the 4× storage-speed advantage applies most directly, will see larger savings than test-light workflows, where the CPU advantage dominates. Organizations considering a migration should benchmark their specific workflow mix rather than applying either number uniformly.

Sequoia’s Thesis: Infrastructure Is Where AI Coding Margin Concentrates

For Sequoia, the investment signals where durable value accrues in the AI-era software stack. Isaiah Boone, the Sequoia partner who led the deal, framed the thesis explicitly.

“JP and team understood early that as AI agents wrote more code, the infrastructure to build and test it would need to change,” Boone said. “What stood out to us was their deep systems expertise and speed of execution. That has translated into a product that helps their customers test and ship their software faster. We believe that foundation gives Blacksmith room to build much more as the demands of software development change.”

The investment logic fits a pattern Sequoia and other infrastructure-focused VCs have articulated since late 2025: AI coding tools are picks-and-shovels demand generators for the infrastructure layer that validates their output. Writing code is comparatively cheap; ensuring that code is correct before deployment is computationally intensive, latency-sensitive, and directly in the path of the AI-agent commit loop that runs tens of times per task.

The macro numbers support the framing. The AI-powered software testing and quality assurance market was estimated at nearly $12 billion in 2026 and is projected to grow at a compound annual rate approaching 27 percent through 2031. If even a fraction of the AI-generated code volume increase demands dedicated CI validation infrastructure — rather than general-purpose shared cloud capacity — the addressable market for purpose-built CI clouds expands substantially beyond what legacy estimates would suggest.

From University of Waterloo to Unicorn in Two Years

Blacksmith was founded in 2024 by three University of Waterloo alumni: Aditya “JP” Jayaprakash, Aayush Shah, and Aditya Maru. Jayaprakash had previously worked as an engineer at Faire; Shah and Maru were early engineers at Cockroach Labs.

The company completed Y Combinator’s Winter 2024 batch and raised a $3.5 million seed round in May 2025. Blacksmith’s seed and Series A history shows a $10 million Series A, also led by GV, in September 2025 that closed in 14 days — an unusually fast close that signaled early investor demand. The Series B, led by Peak XV Partners at a $550 million valuation, was raised in March 2026 but not publicly announced until August — meaning the two rounds bookending the unicorn threshold were separated by roughly seven months in practice.

More than 10,000 companies now use the platform, including Mercury, Granola, Handshake, LangChain, Ashby, Clerk, and Supabase — up from roughly 800 at the Series A announcement in September 2025 and more than 6,000 at the Series B. The company currently manages just under 10 megawatts of compute across data centers in the United States and Europe and plans to grow that footprint fivefold by 2027.

Alongside its core CI infrastructure, Blacksmith has developed Codesmith, an AI coding agent on Blacksmith that automatically identifies and fixes failed code checks. The extension positions the company to sell not only to human developers submitting code for validation, but potentially to the AI agents doing the submitting — a closed loop in which Blacksmith’s infrastructure processes the output of AI coding tools while Blacksmith’s own AI tool diagnoses and repairs failures in that output.

Migration to Blacksmith requires a one-line change to a GitHub Actions workflow file, which reduces switching friction to near-zero for teams already on GitHub’s CI platform.

Does Blacksmith Have Anything Rivals Cannot Copy?

Blacksmith competes directly with GitHub’s hosted CI runners, AWS CodeBuild, Google Cloud Build, and Azure Pipelines on the hyperscaler side, and with developer-tool incumbents like CircleCI and a growing cohort of AI-native startups targeting the same bottleneck.

The company’s competitive position rests less on any single patent or proprietary technology — Firecracker is open-source, NVMe SSDs are commodity hardware, AMD Ryzen is available to any buyer — and more on operational execution: the ability to run CI at scale reliably across tens of thousands of customers, manage the spike-and-drain demand shape that causes general-purpose clouds to fail, and do so with infrastructure the company has been building for two years.

The 18-fold growth in CI job volume over ten months is the clearest signal of whether that execution is actually working. Ten thousand customers and 17× ARR growth in under a year suggest the product is doing something real in production. The harder question — whether the moat deepens as competitors build similar infrastructure — will likely be answered by whether Blacksmith’s data-center density, cache topology, and job orchestration continue to compound into operational advantages that newly-entering players cannot replicate quickly.

The $66 million raised in this round will primarily fund scaling compute capacity to 50 megawatts by 2027 — a fivefold expansion from its current just-under-10 MW footprint — and growing engineering and go-to-market teams in New York and San Francisco.

Frequently Asked Questions

Why do AI coding agents increase CI costs so much more than human developers?

A human developer typically submits a pull request and waits for CI to pass or fail before taking action. An AI coding agent runs an iterative loop: submit code, read CI output, diagnose the failure, modify the code, and resubmit —often 10 to 20 times per task. Each loop iteration triggers a full CI run. The cost multiplier isn’t the volume of code an agent produces, but the number of plan steps in its debugging cycle. As AI agents become more capable and autonomous — running longer unattended loops — this multiplier grows, making CI infrastructure spending a larger share of AI-era software development costs. The underlying reason is structural to CI’s design: every submitted change must be built and tested from scratch in an isolated environment, with no shortcut for “I already tested this exact state.”

What is a Firecracker microVM and why does it matter for CI security?

Firecracker is an open-source virtualization technology built on KVM (Kernel-based Virtual Machine) and originally developed by AWS for running Lambda functions. It creates a lightweight virtual machine that boots in approximately 5 milliseconds and provides hardware-level memory isolation between jobs — meaning code running in one CI job cannot access the memory, credentials, or environment variables of any other job on the same physical host. This matters for CI specifically because CI platforms routinely store access tokens, cryptographic keypairs, and cloud credentials in the build environment. The Firecracker microVM open-source project is available for review and public use. CircleCI’s 2023 breach exposed CI secrets during a two-week window, demonstrating why hardware-level job isolation is a meaningful security property rather than an abstract one.

How do I know if my team’s CI costs are being driven by AI coding agents?

Check whether your GitHub Actions compute-minute usage has grown faster than your team headcount or total PR volume over the past six to twelve months. If costs are growing while team size is flat, the most likely cause is AI coding agents running multiple CI loops per developer task. Within your CI logs, look for large numbers of jobs triggered by bot accounts or agent-tagged commits — GitHub’s API exposes the PR author’s account type. If agent-opened PRs represent more than 20 percent of your CI volume but you only see a small headcount increase, purpose-built CI infrastructure is worth a cost-per-minute comparison against your current spend. Migration to Blacksmith’s architecture for GitHub Actions requires only a one-line change to a GitHub Actions workflow file for existing GitHub Actions users.

What happens to Blacksmith’s cost advantage if GitHub upgrades its own runner hardware?

GitHub’s architecture ties its hosted runner infrastructure to AWS EBS for storage — a constraint rooted in the multi-tenant shared-infrastructure model that gives GitHub portability and redundancy guarantees it needs across millions of repositories. Switching to local NVMe would require GitHub to restructure the hosted-runner storage architecture in ways that would affect its entire runner product, not just CI performance. The AMD Ryzen 7950X CPU gap is easier for GitHub to close by updating its hardware generation, and it likely will over time. Blacksmith’s response to that scenario is data-center density and job-orchestration software that continues to improve with scale—and Codesmith, the AI agent layer, as a product expansion that makes Blacksmith a participant in the agent loop, not just its infrastructure.