A $66 million Series C landed in my inbox this week, filed under "blockchain." I read it twice. There was no blockchain in it. No token, no chain, no wallet, no cryptographic primitive worth naming. Just a continuous integration platform called Blacksmith that makes GitHub Actions run faster, wrapped in a press release that some aggregator had tagged as Web3 news. Another rug pull? No. Something stranger โ a category error that tells you more about how capital narrates itself in 2025 than any token launch ever could.
This is not a story about a company. It is a story about a word being smuggled across a border because the border guard wasn't reading.
Blacksmith, with medium-high confidence, is blacksmith.sh โ a drop-in replacement for GitHub's hosted runners. If you have never sat inside a CI pipeline, here is the essential thing: every time a developer opens a pull request, a machine somewhere spins up, installs dependencies, compiles code, runs tests, and reports back. GitHub gives you these machines at $0.008 per minute for a two-core Linux box, up to $0.128 for sixty-four cores. Developers hate this. It is slow, it is metered, and the cache โ the mechanism that saves you from re-downloading the same dependencies every build โ is throttled to roughly ten gigabytes per repository.
Blacksmith's bet is almost embarrassingly simple: better hardware scheduling, better cache engineering, faster builds, cheaper bills. It is real engineering. It is also engineering that lives entirely inside someone else's house.
The round: $66 million, Series C, on top of roughly $14 million across seed and Series A, plus a $45 million Series B. The arithmetic is self-consistent โ about $125 million cumulative โ which is why I believe the press release is a genuine wire rather than an AI hallucination. But "17x ARR growth" and "90 million checks" arrive with no time window and no denominator. Numbers without denominators are not data. They are mood.
Let me start where the press release refuses to go: the economics.
Blacksmith bills per minute of compute. That means its cost of goods sold is hardware and cloud โ not the near-zero marginal cost that makes software beautiful. This is not a SaaS business wearing a SaaS costume; it is a compute reseller wearing a software layer. Gross margins in this configuration land somewhere around 50 to 65 percent, against the 75 to 85 percent benchmark investors reflexively apply to software. That single gap is the crack running through the entire valuation.
If you value this at 30 to 60 times ARR โ the SaaS multiple โ a $1 billion valuation implies ARR in the $17-33 million range. If you value it as infrastructure, at 15 to 25 times, the same $1 billion demands $40-66 million in ARR. The "17x growth" claim, with its implied small base, sits uncomfortably between those two pictures. The multiple and the growth rate are telling two different stories, and the press release never forces them to meet.
Then there is the technical architecture, which I find more interesting than the money. The real engineering sits in three places: the caching layer, micro-VM snapshotting for hot starts, and runner orchestration. Based on my own time reverse-engineering build systems and submitting patches to security libraries, I can tell you where the difficulty actually lives. It is not in clever algorithms. It is in the unglamorous, operational, hardware-economics tuning โ squeezing latency out of a VM boot, making a cache hit land in the right millisecond. This is operations-intensive engineering, and operations-intensive engineering is precisely the kind that a well-funded competitor can clone in eighteen months.
Here is the structural weakness nobody puts in the deck: Blacksmith has no API ecosystem of its own. It inherits GitHub's. The authentication flow, the log format, the cache API, the runner registration protocol โ all of it is defined unilaterally by GitHub. When GitHub changes an OIDC policy, Blacksmith adapts or dies. This is the same dynamic I have watched play out across the Layer 2 wars, where the winning rollup stack is never the technically superior one but the one that convinces the most projects to deploy on it first. Dependency on a platform's goodwill is not a moat. It is a leash.
And then the phrase that did all the work in this fundraise: "AI-generated code verification." It appears once. No product name, no accuracy figure, no customer. It arrives into a market already crowded โ CodeRabbit, Graphite, Qodo, Greptile, and GitHub's own Copilot code review. The AI layer is not a shipped asset. It is an option being priced as an asset, and its job is narrative, not technical: it exists to lift the addressable market above the ceiling that a CI-runner replacement can never break through. The CI/CD tooling market is perhaps $20-40 billion and permanently suppressed by GitHub's free tier. You cannot raise a billion-dollar round on "faster builds." You can on "AI that verifies code." Code speaks, but culture listens โ and right now the culture is listening for the word "AI."
One more economics note the wire skips: freemium. Free CI for open-source projects is a brilliant acquisition funnel and a cost black hole. GitHub already anchors the market at "free and unlimited" for public repositories, so Blacksmith's free tier competes against a price of zero while absorbing the build volume of projects with enormous CI appetite and zero willingness to pay. Open-source is also the highest-risk surface for abuse โ miners and volume farmers adore a free compute tier. Single-team CI spend has a natural ceiling: engineers times build frequency times unit price. That ceiling is exactly why the AI story had to be bolted on โ not for the technology, but to break an ARPU limit a per-minute reseller can never escape.
There is a quieter liability. A CI runner holds every customer's source code and build secrets. In March 2025, the tj-actions/changed-files compromise leaked secrets across a large number of repositories. That was a supply-chain wound. Blacksmith sits closer to the wound than almost anyone: a tenant-isolation failure or a cache-poisoning event would not damage one customer, it would damage all of them at once, irreversibly. This risk never appears on a balance sheet. It decides whether the company survives. Enterprise buyers will ask for SOC 2 Type II, penetration reports, EU data residency โ and every answer costs money the gross margin cannot easily spare.
Everyone reading this round is asking whether Blacksmith can beat GitHub. Wrong question. The real question is whether Blacksmith can stop being a line item in someone else's budget.
Notice what category the customer's money comes from. It is not "new tooling." It is "GitHub bill optimization." That framing lowers procurement friction to near zero โ and simultaneously nails the company permanently to the cost-cutting column instead of the value-creation column. It is a B2B2B existence: Blacksmith's revenue is parasitic on GitHub's enterprise customers, and its pricing power is sandwiched between cloud providers upstream (raise compute prices and the margin compresses) and GitHub downstream (cut official prices and customers evaporate).
The growth story has a second blind spot. That 17x is not decomposed, and it should be. How much came from the AI coding boom inflating the entire industry's build volume โ a tailwind every CI vendor rode โ versus genuine share capture? I would bet the tailwind is more than half. AI coding agents multiplied pull-request volume and CI minutes across the board. Blacksmith got bigger because the ocean rose, and the press release is happy to let you credit the boat.
And the customer base is fragile in a way that looks like strength. The core users are AI-native startups โ explosive CI usage, extreme cost sensitivity. They are also the cohort most likely to die or get acquired and consolidated away. Logo retention can look terrible while net revenue retention looks heroic, because the survivors spend more and the dead simply vanish. Funding rounds, I have come to believe, are not finance. They are anthropology โ studies of what a market is willing to believe about itself.
So here is the question I will be holding for the next eighteen months: when GitHub ships a genuinely fast hosted runner โ and it will, because the competitive pressure now demands it โ what is left of a company whose entire differentiation was "we are faster and cheaper than the default"?
The honest answer is: whatever the AI verification layer turns out to actually be. If it is real, Blacksmith becomes a software company and earns its multiple. If it is a slide, Blacksmith is a well-run arbitrage that the platform it depends on can close at will. The Cassandra complex is real, and I have been wrong before about timing โ but not, I think, about direction. Watch the next GitHub pricing page. It will tell you more than any term sheet.