Blacksmith Deploys Gaming Chips to Speed AI Code at $1B Valuation

Y Combinator

As AI tools multiply software output, continuous integration testing has replaced code authoring as the primary development bottleneck, exposing traditional enterprise cloud servers as architecturally ill-suited compared to high-clock-speed consumer gaming hardware.

AI-assisted summary of The $1B Company Fixing AI’s Software Bottleneck — original source. Check the original for full context.

Source published: 2026-11-02

Key ideas

Blacksmith grew revenue 20x over the past year and raised a $66 million Series C at a $1 billion valuation led by Sequoia Capital, targeting the developer testing bottleneck created by AI-assisted coding. (0:00)
The company's core architectural insight is running continuous integration on high-clock-speed consumer gaming CPUs rather than traditional enterprise multithreaded server chips. (2:00)
With only 12 engineers out of a 30-person team, Blacksmith maintains massive distributed server fleets by refusing bespoke feature requests and applying algorithmic bin-packing. (6:46)
Infrastructure demand directly mirrors AI releases, surging immediately whenever major labs drop new foundational models that generate higher volumes of pull requests. (7:15)

Detailed notes

## Background Published by Y Combinator on YouTube on November 2, 2026, this inaugural episode of 'Built to a Billion' highlights Blacksmith (YC W24), a startup building ultra-fast cloud infrastructure for continuous integration (CI). The video description outlines how the founders leveraged gaming CPUs to eliminate testing bottlenecks for AI-generated code, growing 20x in a year and raising a $66 million Series C at a $1 billion valuation.

## TL;DR As agentic AI tools cause code output to skyrocket, continuous integration testing has emerged as the central engineering bottleneck holding back software deployment. Blacksmith addresses this by bypassing traditional cloud multithreading in favor of high-performance consumer gaming processors, which maximize single-threaded clock speed alongside aggressive local caching. Scaled by a lean team of 30, the platform cuts test queue wait times from 20 minutes down to seconds for customers like LangChain. Managing high-stakes physical data center hurdles, the founders are expanding into visual autonomous QA to let developers ship software as fast as they can think.

## Atmosphere & style The video combines pragmatic, engineering-grounded candor with the electric momentum of a hypergrowth startup retrospective. The founders speak with casual, self-effacing humor about their outsider naivety—such as leasing servers originally intended for Minecraft and Call of Duty—contrasted against serious technical grit. Watching it feels like getting an inside peek into a fast-moving war room where high-level architectural contrarianism meets the messy physical realities of racking hardware, power surges, and subway ad takeovers.

## Bulletin - [00:00](https://www.youtube.com/watch?v=NVzgt2AEvDg&t=0s) Blacksmith grew revenue 20x over the past year and raised a $66 million Series C at a $1 billion valuation led by Sequoia Capital, targeting the developer testing bottleneck created by AI-assisted coding. - [02:00](https://www.youtube.com/watch?v=NVzgt2AEvDg&t=120s) The company's core architectural insight is running continuous integration on high-clock-speed consumer gaming CPUs rather than traditional enterprise multithreaded server chips. - [06:46](https://www.youtube.com/watch?v=NVzgt2AEvDg&t=406s) With only 12 engineers out of a 30-person team, Blacksmith maintains massive distributed server fleets by refusing bespoke feature requests and applying algorithmic bin-packing. - [07:15](https://www.youtube.com/watch?v=NVzgt2AEvDg&t=435s) Infrastructure demand directly mirrors AI releases, surging immediately whenever major labs drop new foundational models that generate higher volumes of pull requests.

## Key takeaways - [00:00](https://www.youtube.com/watch?v=NVzgt2AEvDg&t=0s) AI code generation shifts the engineering bottleneck from writing software to running continuous integration tests and validating changes. - [01:15](https://www.youtube.com/watch?v=NVzgt2AEvDg&t=75s) Traditional hyperscalers like AWS, GCP, and Azure are architecturally misaligned with CI workloads, which require raw single-threaded performance rather than wide multi-threading. - [02:00](https://www.youtube.com/watch?v=NVzgt2AEvDg&t=120s) Blacksmith initially built its cloud fleet by leasing consumer gaming CPUs from German data centers originally dedicated to hosting Minecraft and Call of Duty servers. - [03:46](https://www.youtube.com/watch?v=NVzgt2AEvDg&t=226s) Co-locating cached build artifacts directly adjacent to active compute instances eliminates redundant test execution and recompilation cycles. - [04:10](https://www.youtube.com/watch?v=NVzgt2AEvDg&t=250s) Early customer acquisition depended on manual cold outreach to engineering leaders after exchanging pilot access for initial logo rights with peer startups. - [05:10](https://www.youtube.com/watch?v=NVzgt2AEvDg&t=310s) Frictionless one-click migrations enabled rapid enterprise adoption, slashing compute acquisition wait times for teams like LangChain from 15–20 minutes on GCP down to 10–15 seconds. - [06:15](https://www.youtube.com/watch?v=NVzgt2AEvDg&t=375s) Blacksmith pairs high-performance cloud hardware with aggressive transit advertising in San Francisco and New York City to establish category dominance. - [06:46](https://www.youtube.com/watch?v=NVzgt2AEvDg&t=406s) A lean team of 12 engineers supports thousands of customers across megawatts of infrastructure by strictly resisting bespoke customer feature requests. - [07:15](https://www.youtube.com/watch?v=NVzgt2AEvDg&t=435s) Spikes in global CI workload volume occur in direct tandem with new foundational LLM model releases from major AI labs. - [07:40](https://www.youtube.com/watch?v=NVzgt2AEvDg&t=460s) Physical data center disruptions, such as a major power surge in Phoenix knocking out dual cooling systems, forced rapid engineering of geographic redundancy. - [08:10](https://www.youtube.com/watch?v=NVzgt2AEvDg&t=490s) Blacksmith is expanding beyond raw speed into autonomous QA by capturing visual and behavioral proofs like automated screenshots and video recordings. - [08:50](https://www.youtube.com/watch?v=NVzgt2AEvDg&t=530s) The long-term objective of CI optimization in the AI coding era is enabling developers to ship software as close as possible to the speed of thought.

## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=NVzgt2AEvDg&t=0s) The CI Bottleneck and Gaming CPUs

Every time software engineers push code, continuous integration (CI) systems run extensive suites of tests and checks to ensure nothing breaks. With the emergence of agentic coding tools, developers generate software at unprecedented volumes, shifting the core engineering bottleneck from authoring code to validating it. While writing a hundred times more code is possible, shipping it remains stalled if teams are forced to wait on painfully slow CI pipelines to verify test suites. Blacksmith was founded to address this bottleneck by building dedicated cloud infrastructure optimized specifically for running CI workloads at maximum speed, enabling the company to grow 20x within a single year and secure a $66 million Series C funding round at a $1 billion valuation led by Sequoia Capital.

```mermaid flowchart TD accTitle: Blacksmith CI Architecture vs Traditional Cloud accDescr: Flowchart contrasting traditional hyperscaler multithreaded compute with Blacksmith's single-thread optimized gaming CPUs and local caching.

subgraph Hyperscalers["Traditional Cloud (AWS / GCP / Azure)"] A["Standard Server Fleet"] -->|"Optimized for"| B["Multi-threaded Workloads"] B -->|"Result"| C["Slow CI Execution & Queuing"] end

subgraph BlacksmithCloud["Blacksmith Infrastructure"] D["Gaming PC Processors"] -->|"Optimized for"| E["Maximum Single-Thread Speed"] E --> F["Localized Cache Artifacts"] F -->|"Result"| G["Instant Compute & Fast CI Runs"] end ```

The founders arrived at the problem as outsiders rather than traditional platform or DevOps engineers; two worked as database engineers while the third worked on search infrastructure after graduating from the University of Waterloo and working at Faire. None of them had ever authored an internal CI pipeline, but having experienced the universal frustration of slow CI across every prior workplace, they suspected that existing hyperscalers like AWS, Google Cloud Platform (GCP), and Microsoft Azure were architecturally ill-suited for CI workloads. One of the founders reflected on this outsider advantage: “When I think about why we were one of the only ones to to approach the problem this way and to think of it this way, I think the part of the answer here is just being naive enough in the right way to not kill a good idea before it meets the market.”

Analyzing the workload characteristics revealed that CI jobs behave much like desktop software and video games: they prioritize raw, blazing-fast single-threaded execution rather than spreading tasks across dozens of weaker threads. Acting on this realization, Blacksmith bypassed standard server hardware and began renting machines equipped with consumer gaming CPUs from a German data center provider traditionally known for hosting *Minecraft* and *Call of Duty* servers. As demand expanded, the company moved from leasing individual gaming boxes to signing large-scale data center contracts where they procure megawatts of power directly.

- **Founders' Background:** Classmates from the University of Waterloo based in Toronto; applied to Y Combinator a second time with Blacksmith after working in database and search infrastructure. - **Valuation and Growth:** Scaled revenue 20x over twelve months, culminating in a $66 million Series C round at a $1 billion valuation led by Sequoia Capital. - **Hardware Thesis:** CI testing cycles rely heavily on raw single-core clock speeds rather than wide multi-threading, matching the performance profiles of consumer gaming processors over enterprise server chips.

---

[03:46](https://www.youtube.com/watch?v=NVzgt2AEvDg&t=226s) Winning Over Developers and Instant Scaling

The unconventional choice of hardware initially surprised tech investors. An early investor who backed Blacksmith twice—first on a SAFE right out of Y Combinator, and later leading their Series A and taking a board seat—recalled the sheer shock of seeing gaming hardware racked in a cluster: “It was even kind of a shock to me that somebody would even put a bunch of these things in one place. That was sort of the weird thing. You usually would have it at home for your like single, you know, your big box you have under your desk and you're like, you know, you're playing whatever game. But yeah, being like, 'No, we need to have buy like a hundred of these things and use them to compile.' Like that that's where I was like, 'Oh, that's that's kind of different.'” Beyond the CPU clock advantage, Blacksmith tackled software efficiency through intelligent caching, aggressively moving cached build artifacts right next to active compute nodes to eliminate repetitive compilation and redundant test execution.

Acquiring the first 200 customers required manual grind and personal outreach. The founders started by asking friends at peer companies to pilot the platform in exchange for logo rights, which established early social proof on their website. Armed with that credibility, the founders initiated cold outbound campaigns targeting CTOs, VPs of Engineering, and platform engineers. Early adopters were primarily agile startups with lean teams and zero bureaucratic red tape, though engineers proved to be a highly discerning customer base prone to experimenting with new tooling while remaining difficult to retain unless the platform delivered clear, unmatched performance.

Despite having no formal sales team, Blacksmith began closing sizable five- and six-figure enterprise contracts thanks to a friction-free onboarding experience that allowed engineering teams to migrate all CI workflows in a single click. A prominent example is LangChain, which previously executed its CI pipelines on GCP:

- **LangChain Wait Time Reduction:** Under GCP, LangChain developers routinely waited 15 to 20 minutes merely to acquire the underlying compute instances before tests could even start; switching to Blacksmith's virtually infinite compute pool slashed that wait down to 10 to 15 seconds. - **Marketing Blitz:** To establish default name recognition for continuous integration speed, Blacksmith launched an aggressive visual campaign across San Francisco and New York City, taking over 15 NYC subway stations and wrapping an entire transit train in the company’s signature bright yellow branding.

---

[06:46](https://www.youtube.com/watch?v=NVzgt2AEvDg&t=406s) Scaling the Fleet and Autonomous QA in the AI Coding Era

Blacksmith scaled its operations from an initial team of five people overseeing a cluster of a couple hundred servers to a 30-person company with only 12 engineers managing orders of magnitude more infrastructure. This leverage relies on bespoke automation software that monitors every layer of the fleet, paired with strict product discipline that avoids bespoke customer feature requests: “And secondly, we are ruthless about keeping the surface area of the product small and manageable and only scaling it with the size of the team rather than giving into one of bespoke customer requirements.” This discipline allows a compact engineering organization to support thousands of customers across massive distributed server fleets.

Keeping pace with runaway hardware demand has emerged as one of the platform's stiffest technical hurdles. Every time an AI lab releases a new foundational LLM, Blacksmith observes an immediate, massive spike in CI job volume as developers globally rush to generate and test code changes. Because ordering, delivering, and racking physical server hardware entails lead times, engineers rely on sophisticated bin-packing algorithms, efficient job placement heuristics, and bespoke network topologies to maximize utilization across existing hardware.

The physical realities of data center operations were laid bare early in the summer immediately following the company's Series B announcement:

- **Phoenix Power Surge:** A major electrical surge in Phoenix knocked out both the primary and backup cooling infrastructures in one of Blacksmith’s largest regional data centers, degrading startup latencies for customer CI jobs and prompting rapid investment in broad geographic redundancy. - **Autonomous QA Vision:** Blacksmith is building automated QA features that collect visual and behavioral verification artifacts, such as automated screenshots and video recordings, validating that code changes render properly. - **The Speed of Thought:** As AI models multiply the volume of pull requests, developers will lack the bandwidth to manually inspect every line of generated code, making automated trust mechanisms essential: “We want to remove that bottleneck and allow developers to basically bring shipping software as close as they can to their speed of thought.”

## Notable quotes > [06:46](https://www.youtube.com/watch?v=NVzgt2AEvDg&t=406s) We are ruthless about keeping the surface area of the product small and manageable and only scaling it with the size of the team rather than giving into one of bespoke customer requirements.

## Mentioned AWS, Azure, Blacksmith, Call of Duty, Faire, GCP, LangChain, Minecraft, Sequoia Capital, University of Waterloo, Y Combinator

## Video description Blacksmith (YC W24) is building a faster cloud for continuous integration (CI), helping developers test and ship code as AI coding tools generate more of it.

As software production accelerates, testing and validation are becoming the new bottleneck. That shift has helped Blacksmith grow 20x over the last year, and the company recently raised a $66 million Series C at a $1 billion valuation.

The founders started as outsiders to DevOps and built their first infrastructure using the same high-performance CPUs found in gaming PCs, renting servers originally designed for Minecraft and Call of Duty. Today, they run a much larger fleet with a team of just 30 people, helping companies cut CI waits from minutes to seconds.

In this first episode of Built to a Billion, they share how they found the idea, won their first 200 customers, and why faster testing will become even more important as agents write more of the world’s code.

https://blacksmith.sh

Chapters: 00:00 — Making CI Faster With Gaming CPUs 03:46 — Winning Over Investors and Developers 06:46 — Scaling CI for the AI Coding Era

Apply to Y Combinator: https://www.ycombinator.com/apply Work at a startup: https://www.ycombinator.com/jobs

[Source](https://www.youtube.com/watch?v=NVzgt2AEvDg)

中文简报

Blacksmith以10亿美元估值部署游戏芯片加速AI代码测试随着AI工具成倍提升软件产出，持续集成测试已取代代码编写成为首要开发瓶颈，暴露出传统企业云服务器在架构上不如高主频消费级游戏硬件适配的问题。随着智能体AI工具推动代码产出激增，持续集成测试已成为阻碍软件部署的核心工程瓶颈。Blacksmith通过绕过传统的云多线程方案，转而采用高性能消费级游戏处理器来解决这一问题，该方案将单线程主频与激进的本地缓存发挥到极致。在仅30人的精简团队运作下，该平台为LangChain等客户将测试排队等待时间从20分钟缩减至数秒。在应对高难度的实体数据中心挑战的同时，创始人正向视觉自主QA领域拓展，旨在让开发人员能够以思维般的速度发布软件。 Blacksmith在过去一年收入增长了20倍，并以10亿美元估值完成了由红杉资本领投的6600万美元C轮融资，瞄准AI辅助编程带来的开发者测试瓶颈。 (0:00) 该公司的核心架构洞见在于，利用高主频消费级游戏CPU而非传统的企业级多线程服务器芯片来运行持续集成。 (2:00) 在30人的团队中仅有12名工程师的情况下，Blacksmith通过拒绝定制化功能需求并应用算法装箱技术，维持着庞大的分布式服务器集群。 (6:46) 基础设施需求直接映射AI模型的发布情况，每当主流实验室推出生成更多拉取请求（PR）的新基础模型时，其需求便会立即激增。 (7:15)

Reading summary

As AI code generation accelerates software production, continuous integration testing has emerged as the central engineering bottleneck. Blacksmith addressed this constraint by running CI pipelines on consumer gaming processors optimized for single-threaded speed rather than standard enterprise cloud hardware, reaching a $1 billion valuation. Gaming Hardware Architecture for Single-Threaded Workloads (0:00) Traditional hyperscalers like AWS, Google Cloud Platform, and Microsoft Azure optimize their fleets for multi-threaded enterprise workloads. However, continuous integration (CI) test suites depend primarily on raw, single-threaded execution speed, closely resembling the workload profile of desktop applications and video games. Founded by University of Waterloo graduates with backgrounds in database and search infrastructure rather than platform engineering, Blacksmith leveraged an outsider perspective to rethink CI infrastructure from first principles. The company initially bypassed traditional enterprise servers to lease machines equipped with consumer gaming CPUs from a German data center host known for running Minecraft and Call of Duty servers. To maximize execution speed, Blacksmith paired high-clock-rate gaming processors with intelligent caching that places build artifacts directly adjacent to active compute instances. As demand scaled, the company transitioned from renting individual host machines to executing direct megawatt-scale power contracts, driving a 20-fold revenue increase over twelve months and securing a $66 million Series C round led by Sequoia Capital at a $1 billion valuation. Self-Serve Migration and Enterprise Adoption (3:46) Deploying gaming CPUs in data center racks initially surprised early investors, who were accustomed to seeing consumer hardware strictly beneath personal desks rather than in clustered compilation fleets. To gain market traction, the founders initiated cold outreach to engineering executives after securing initial pilot logos from peer companies. Despite maintaining no formal enterprise sales team, Blacksmith began closing five- and six-figure contracts by eliminating onboarding friction through a single-click migration process. The platform's performance differentiation attracted early-stage startups and high-growth AI organizations sensitive to pipeline delays. For example, framework provider LangChain previously endured 15 to 20 minutes of compute queue times on GCP before test runs could commence; migrating to Blacksmith's compute pool reduced that waiting interval to between 10 and 15 seconds. To reinforce brand recognition across tech hubs, the company supplemented word-of-mouth adoption with high-visibility campaigns, including wrapping a New York City subway train and advertising across 15 transit stations. Fleet Automation and Physical Infrastructure Strain (6:46) Blacksmith maintains a compact team of 30 employees, relying on just 12 engineers to operate its large distributed server footprint. The company achieves this operational leverage by building proprietary automation software across the fleet and strictly rejecting custom feature requests that expand the product's maintenance surface. Every major foundational LLM release triggers an immediate, sharp surge in customer CI runs as developers produce more automated code, forcing the company to balance hardware order lead times using custom bin-packing algorithms, placement heuristics, and proprietary network topologies. Scaling physical infrastructure also exposed real-world operational vulnerabilities. Early in the summer following its Series B round, an electrical surge in Phoenix disabled both the primary and backup cooling systems in one of Blacksmith’s core regional data centers. The failure degraded startup latencies for customer CI pipelines, prompting the company to accelerate investments in broader geographical redundancy to withstand local facility failures. Automating Verification for High-Volume AI Coding (6:46) The broader software delivery bottleneck is shifting from manual coding toward code review and validation. Because generative AI tools can produce hundreds of times more code than human engineers can manually inspect, validation infrastructure must evolve beyond traditional unit test execution. To prevent pull request queues from stalling software delivery, Blacksmith is expanding into autonomous quality assurance. This vision involves generating automated behavioral and visual verification artifacts, such as recorded videos and screenshots, to confirm that applications render and function as intended. The company aims to provide trust mechanisms that allow engineering teams to validate and deploy software as fast as AI systems generate it. Blacksmith has proven the commercial viability of tailoring physical hardware to specific developer workloads, yet the company remains exposed to the physical constraints of data center operations and power delivery. As agentic code creation multiplies testing volume, maintaining continuous hardware availability while expanding into autonomous QA will test whether specialized infrastructure can keep pace with exponential software generation.

精炼阅读摘要

根据 Y Combinator 的专访内容，初创公司 Blacksmith 针对 AI 辅助编程带来的海量代码验证瓶颈，采用消费级游戏 CPU 与就近缓存重塑持续集成（CI）基础设施；该公司在一年内实现营收 20 倍增长，并完成由红杉资本领投、估值达 10 亿美元的 6600 万美元 C 轮融资。单线程芯片破解 CI 验证瓶颈 (0:00) 随着智能体编码工具的兴起，软件工程师生成代码的速度大幅提升，开发的主要瓶颈已从编写代码转向运行 CI 系统的测试与验证环节。传统超大规模云厂商（AWS、GCP、Azure）的服务器集群针对多线程任务设计，导致单线程特征明显的 CI 作业普遍面临漫长的排队与执行延迟。Blacksmith 的三位联合创始人毕业于滑铁卢大学，此前在 Faire 从事数据库与搜索架构工作。作为 CI 领域的“局外人”，他们认为现有云架构在设计上并不适于处理此类工作负载。工作负载分析表明，CI 任务的运行特征与桌面软件及视频游戏相似，关键在于极高的单核心时钟频率，而非将计算分散到大量性能较弱的线程上。基于这一发现，Blacksmith 绕过标准企业级服务器芯片，转而从一家曾托管《我的世界》（Minecraft）和《使命召唤》（Call of Duty）游戏服务器的德国数据中心租赁搭载消费级游戏 CPU 的机器；随着客户需求激增，公司已从租赁单台机器演进为直接签署采购兆瓦（MW）级电力的大型数据中心合同。就近缓存与一键迁移加速企业交付 (3:46) 除 CPU 单核主频优势外，Blacksmith 还推行就近缓存策略，将缓存的构建产物直接放置在活跃计算节点旁边，以消除重复编译和无效测试。早期投资人回忆称，看到大量原本置于个人电脑桌下的消费级游戏芯片被机架化部署用于软件编译，在当时显得十分颠覆。技术团队将这种硬件架构与优化后的缓存方案结合，显著降低了开发等待时间。在获客层面，团队早期通过校友与同行人脉进行免费试用以换取客户品牌背书，随后对工程技术主管展开针对性拓展。得益于支持现有 CI 工作流“一键迁移”的极简接入体验，公司在未设立正式销售团队的情况下接连签下五位至六位数美元的企业订单。例如，AI 应用框架 LangChain 此前在 GCP 上运行测试时，工程师需等待 15 至 20 分钟才能分配到计算实例，迁移至 Blacksmith 后该启动等待时间缩短至 10 至 15 秒。配合业务扩展，公司在纽约和旧金山展开了涵盖 15 个地铁站及地铁列车整车包装的视觉营销推广。精简工程团队应对算力潮与物理故障 (6:46) Blacksmith 目前由 30 人团队支撑大规模基础设施，其中核心工程团队仅有 12 人。这一人效杠杆建立在全自动化监控软件的基础上，同时团队严格限制产品外延，坚决拒绝客户的定制化功能需求，以确保产品复杂性始终与精简的工程规模相适应。算力供给面临的严峻考验在于硬件交付周期与脉冲式需求的错配：行业每发布一次新型基础大语言模型，全球开发者提交的 CI 任务量便会出现急剧脉冲。受限于物理硬件的采购和上架周期，工程师主要依靠自研的装箱算法（bin-packing）、作业调度启发式方案以及专用网络拓扑来压榨现有硬件的算力极限。此外，物理机房的不确定性在夏季显现，其菲尼克斯数据中心曾因重大电涌导致主备两套冷却系统同时失效，造成实例启动延迟上升，促使公司迅速增加跨地域冗余部署。面向 AI 时代的自主 QA 与信任机制 (6:46) 随着 AI 辅助编码使拉取请求（PR）数量成倍增长，工程师已无法靠人工逐行审查每一处代码变更，软件交付的核心阻力逐渐演变为验证与信任成本。Blacksmith 正着手研发自主质量保证（QA）功能，通过自动截取屏幕截图与录制运行视频等行为及视觉凭证，检验代码修改后的实际渲染效果。公司创始人表示，自动化验证机制的推进旨在消除代码审查与测试中的流转卡点，让软件从编写到最终发布的节奏尽可能逼近开发人员的“思考速度”。 Blacksmith 证明了针对 CI 特殊负载重构软硬件栈的巨大商业可行性；然而，面对未来 AI 代码爆发带来的算力峰值与底层数据中心的物理运营风险，如何在快速扩容中兼顾电网冗余与高可用性，仍是其长期需要解决的现实挑战。

Source quotes

[6:46] We are ruthless about keeping the surface area of the product small and manageable and only scaling it with the size of the team rather than giving into one of bespoke customer requirements.

Original transcript

[0:00] Every time a developer writes code, [0:01] [music] they have to run a bunch of [0:04] checks to make sure that nothing is [0:06] broken. That process is called CI and [0:09] Blacksmith writes infrastructure to run [0:11] those really, really fast. We've grown [0:13] 20x in in the last [music] year. We [0:16] recently raised our series C. This was [0:18] done at a billion dollar valuation led [0:20] by Seoia and we raised about 66 million. [0:23] We were outsiders. We were not platform [0:25] engineers. We didn't work in DevOps. In [0:27] fact, none of us had written a CI [0:29] pipeline before. [0:30] >> When I think about why we were one of [0:32] the only ones to to approach the problem [0:34] this way and to think of it this way, I [0:36] think the part of the answer here is [0:39] just being naive enough in the right way [0:42] to [music] not kill a good idea before [0:45] it meets the market. [0:51] So today with um [music] with Agentic [0:54] coding tools, developers are writing [0:56] more and more code than they ever were. [0:58] But it's not very useful to write 100 [1:00] times [music] the code if you can't ship [1:02] it at that speed. So now the bottleneck [1:04] has become not writing the software but [1:06] validating it. And that usually means [1:08] waiting on your CI. So that means making [1:10] sure that the tests are passing and [1:12] [music] that you have confidence that [1:13] what you're shipping actually works. [1:15] >> I always wanted to come to Silicon [1:17] [music] Valley and start a company. When [1:18] I was in high school, my favorite thing [1:20] to do was surf Wikipedia pages and I [1:22] would read about technology companies [1:24] and how they were founded and what they [1:25] did and I wanted to do something of my [1:27] own. I actually went to Waterloo cuz [1:28] they had a pretty great startup [music] [1:30] program and a lot of founders came from [1:33] Waterloo and they were very technical [1:34] school. After I started working at fair, [1:36] I would start thinking a lot about ideas [1:39] and and and how I can move and actually [1:41] blacksmith is my second application to [1:43] YC. The three of us were friends from [1:45] undergrad. We were all in Toronto at the [1:46] time and we all wanted to start a [1:48] company and [music] [1:49] we were riffing ideas around what we [1:52] thought we should work on and all of us [1:53] were engineers. Pretty much at every [1:55] company we worked at CI universally [1:58] sucked and it sucked because it was [2:00] super slow but managing this piece of [2:04] infrastructure was a pain across the [2:06] board and [music] we felt we had a [2:09] better way. When we initially talked [2:10] about the idea for what became [2:12] blacksmith, we were initially kind of [2:14] skeptical whether this could work. But [2:16] over time, as we did more research, our [2:18] conviction continued to grow. My [2:20] co-founders were database engineers. I [2:22] was working on search infrastructure. [2:24] But from the outside, we sort of saw [2:26] that the reason managing CI info was so [2:30] hard, the shape of the workload is very [2:32] different from what hyperscalers like [2:35] AWS, GCP, and Azure are actually really [2:37] good at. And it almost felt like a [2:39] mismatch that caused the pain. And we [2:42] felt like the real way out is building a [2:47] cloud that is extremely good at running [2:51] CI workloads. It comes down to using the [2:53] best tool for the job. What we started [2:55] to realize was the characteristics of CI [2:58] workloads are a lot similar to you know [3:03] desktop applications in particular video [3:05] games that you would run on your typical [3:07] gaming PC in that they don't care as [3:11] much about multiple kind of threads of [3:13] execution but they care about that one [3:15] single thread of execution being really [3:17] really fast. When we started off, we [3:20] would rent these servers from this [3:22] German data center services company that [3:24] was known for hosting Minecraft servers [3:27] and Call of Duty game servers. Over [3:29] time, as our scale has grown, we've [3:31] started to own more and more of the [3:33] stack. So now we sort of operate at at, [3:36] you know, much larger uh levels of [3:38] agreements with data center providers [3:40] where we're effectively buying megawatts [3:43] of power rather than individual [3:45] machines. So I actually invested in [3:47] Blacksmith twice. I led their initial [3:49] round out of Y Combinator which was on a [3:52] safe and then a few months later I [3:54] actually led their series A [music] and [3:56] joined the board. I think it was super [3:57] interesting that they were using gaming [3:59] CPUs because I'd never seen anyone else [4:01] try that solution for CI. It was even [4:03] kind of a shock to me that somebody [4:05] would even put a bunch of these things [4:06] in one place. That was sort of the weird [4:08] thing. You usually would have it at home [4:10] for your like single, you know, your big [4:12] box you have under your desk and you're [4:13] like, you know, you're playing whatever [4:15] game. But yeah, being like, "No, we need [4:16] to have buy like a hundred of these [4:17] things and use them to compile." Like [4:19] that that's where I was like, "Oh, [4:20] that's that's kind of different." [4:22] >> A large part of making CI fast is [4:25] skipping redundant work that has already [4:28] been done before. At a high level, this [4:30] is what developers refer to as caching. [4:33] A big part of Blacksmith's special sauce [4:35] is moving cache artifacts close to the [4:40] compute nodes. [4:41] >> The first 200 customers were extremely [4:43] hard. Initially, we actually asked our [4:45] friends who were working at companies [4:46] [music] to give us a shot. We asked them [4:48] for logo rights and we were able to put [4:50] these logos on our site and build a bit [4:51] of credibility and using that I would [4:54] personally called outbound CTO's and VP [4:57] of engineering personas and platform [4:58] engineers and try to get them to use the [5:00] product. The earliest customers we had [5:02] were small companies like ourselves. [5:04] They cared about moving really fast and [5:07] so they wanted really fast CI and there [5:09] was also very little red tape. The only [5:10] real way to win over engineers is to [5:12] really perform and again to solve a [5:14] problem in a way that they probably [5:16] wouldn't be able to do themselves. You [5:17] know, they're very discerning. They like [5:19] to try a bunch of new tools. Sometimes [5:20] they aren't the stickiest customers. [5:22] >> Without any sales team, they were able [5:24] to get kind of large five six figure [5:26] contracts because of [music] just how [5:28] easy Blacksmith made it to optimize and [5:31] run your CI as fast as possible. We make [5:33] it really easy. In one click, you can [5:35] switch over [music] all of your CI to [5:37] us. One of the customers that onboarded [5:39] onto Blacksmith recently was [music] [5:40] Langchain. Langchain was previously [5:42] running their CI on GCP. On GCP, their [5:45] developers would routinely wait 15 to 20 [5:48] minutes to even acquire [music] compute [5:49] to run the CI job. On Blacksmith, [5:52] because we effectively expose an [5:54] infinite [music] pool of compute, this [5:55] wait time went down from 15 to 20 [5:57] minutes to anywhere from 10 to 15 [6:00] seconds, significantly improving the [6:02] velocity at which their developers could [6:04] push code. [music] [6:07] A big push for Blacksmith now is to get [6:10] the word out that when any engineering [6:12] team thinks about CI velocity or [music] [6:15] gaining trust in the code that they're [6:17] generating, they immediately think of [6:19] Blacksmith. [music] To this effect, we [6:21] are rolling out a massive marketing [6:22] campaign across both San Francisco and [6:24] New York. We're taking over 15 subway [6:26] stations in New York. We're wrapping a [6:28] subway train in our bright yellow color. [6:31] [music] [6:32] When I joined, the team was about five [6:34] people. We were running a small cluster [6:37] of a couple hundred servers and today [6:40] we've seen kind of many orders of [6:42] magnitude more scale. Now we're a team [6:44] of 30 people, 12 engineers. It boils [6:46] down to a few key decisions we've we've [6:48] made and kind of continue to make as we [6:50] as we go along. [music] We have a lot of [6:52] bespoke software. One to just manage the [6:55] growth of our fleet. We have a lot of [6:58] kind of bespoke observability at every [7:00] layer of the stack. And secondly, we are [7:02] ruthless about keeping the surface area [7:05] of the product small and manageable and [7:08] only scaling it with the size of the [7:11] team rather than giving into one of [7:14] bespoke customer requirements. And a big [7:18] payoff of that kind of focus is that [7:20] we're able to scale uh to a very large, [7:23] you know, number of of machines and a [7:25] very large number of customers with just [7:27] a very small team. [7:28] >> So, one of the largest challenges is [7:31] keeping up with the demand. One thing [7:33] that we're seeing is every [music] time [7:35] uh new LLM drops, um we see a massive [7:38] spike in CI usage as people rush to test [7:40] it out and ship more code. So because it [7:43] takes time for to order and install [7:45] machines, we have to get creative [7:46] technically to support that load. So [7:48] that means better algorithms for placing [7:51] our jobs, getting uh more clever around [7:54] how we can pack jobs more efficiently, [7:57] more efficient network topologies. So, [7:59] right after our series B announcement [8:02] early on in the summer, a power surge in [8:05] Phoenix took out both our primary and [8:09] backup cooling systems in one of the [8:11] biggest data centers that we operate in [8:13] Phoenix. Our customers were affected [8:15] with, you know, longer startup times for [8:17] their jobs during that window of outage. [8:20] That underscored the importance of being [8:23] geographically distributed for [8:25] Blacksmith. We want to help give [8:27] developers confidence in their code [8:29] changes. And as part of that, one of the [8:31] products that we're working on is to do [8:34] autonomous QA where we gather artifacts [8:38] like screenshots or video snippets that [8:40] induce confidence in a change. With the [8:42] amount of code that's being produced, [8:44] developers won't have time to go and try [8:46] to understand if things are working. [8:48] They're going to look for evidence and [8:50] Blacksmith is going to give them [8:51] confidence to merge their code faster. [8:53] We want to remove that bottleneck and [8:55] allow developers to [music] [8:58] basically bring shipping software as [9:00] close as they can to their speed of [9:02] thought.