The big clouds were designed for people clicking around web apps. Blaxel, a Y Combinator X25 startup, is betting the next wave of software - autonomous agents - needs its own plumbing. And it built plumbing that wakes in 25 milliseconds.
Try this thought experiment. You hire a worker who shows up, does eleven seconds of intense work, then disappears - and repeats that a few million times a day, sometimes from Frankfurt, sometimes from Oregon, sometimes from a data center you've never heard of. Now go rent that worker an office at a company that only leases by the month. That mismatch, roughly, is what running an AI agent on a traditional cloud feels like. Blaxel was built to fix it.
Blaxel is a San Francisco startup building cloud infrastructure designed specifically for autonomous AI agents. Its founders describe the product in four words that fit on a T-shirt: "AWS for AI agents." Underneath the slogan is a real argument. Amazon Web Services, Google Cloud and Microsoft Azure were engineered for the web-app and software-as-a-service era, when a server stayed on because a human might visit. Agents don't behave that way. They wake up, burst across regions and APIs, finish, and go quiet - and they do it constantly.
The gap shows up in three places at once. Cold starts: a new container spinning up in seconds is fine for a page load and painful for an agent that has to answer now. Cost: paying for idle capacity is tolerable when a server hosts a busy website, wasteful when an agent works in bursts. And control: an autonomous process that can reach the open internet and call arbitrary tools is a security surface a general-purpose cloud never had to think about. Blaxel's pitch is that you can't patch those three problems onto old infrastructure - you have to design for them from the metal up.
The technical heart of Blaxel is the sandbox: an isolated micro virtual machine that boots in milliseconds and resumes from standby in about 25 milliseconds. When an agent has nothing to do, the sandbox hibernates and the meter stops. Because startup and shutdown are so fast, Blaxel can charge only for the seconds an agent is genuinely working. Idle time - the thing that quietly drains a monthly server bill - is free.
That single design choice ripples outward. A customer that would otherwise keep containers warm just in case can let them sleep. In one publicly cited case, a company running more than a billion seconds of agent runtime a month - processing millions of videos - reported roughly half the cost of a traditional serverless setup. Another, handling terabytes of session-replay data monthly, tells a similar story.
Blaxel is less a single tool than a small stack, each piece shaped around how agents actually run:
Isolated microVMs with millisecond boots, ~25ms resume, auto-hibernation and container-escape protection.
A distributed filesystem so teams of agents share state and keep persistent volumes for years.
One endpoint to many model providers, with caching, rate limiting, routing and automatic failover.
Point it at a GitHub repo; get a deployed, auto-scaling agent API billed by the second.
Around those sit custom MCP (Model Context Protocol) server hosting for wiring agents to external tools, and programmable networking - per-sandbox static IPs, egress allow-lists, secret injection through a managed proxy, and geographic rules. The networking piece matters more than it sounds. It lets a customer set a policy like "this agent may not run outside the United States" and have the platform enforce it, which is the kind of sentence that unlocks regulated budgets.
The abstract version of Blaxel is "infrastructure." The concrete version is a list of things a developer can ship without owning a data center. A coding agent that clones a repo, runs untrusted code in a disposable sandbox, and throws the sandbox away when it's done. A research agent that fans out across dozens of parallel microVMs, each hitting different sources, then collapses their findings into shared storage. A voice or support agent that needs a fresh, low-latency environment the instant a user speaks. A background pipeline that grinds through millions of videos or session recordings on a schedule, paying only for the seconds it runs.
The common thread is isolation plus speed plus a bill that tracks real work. For teams building agent products, that removes the least interesting part of the job - standing up secure, scalable compute - and leaves them to focus on the agent's behavior. Given a GitHub repository, Blaxel returns a deployed, auto-scaling API. Given a burst of demand, it spins up thousands of sandboxes and then lets them sleep. It is, deliberately, framework-agnostic: it does not care whether the agent was written with one orchestration library or another.
Blaxel didn't start with usage-based pricing. It began with subscriptions, watched customer behavior during Y Combinator, and switched. The reasoning follows directly from the technology: if sandboxes can sleep for free, billing for idle time is just leaving money on the table for competitors to undercut. A free tier hands new developers up to $200 in credits; after that, it's pay-as-you-go across compute, storage and gateway usage.
Blaxel's six co-founders are not first-timers, and they are not strangers. Several have worked together for years - one pairing reportedly spans three decades. Before Blaxel, the group built ForePaaS, a multicloud data platform, and sold it to OVHcloud, Europe's largest cloud provider, where it became the analytics product suite. That history is the whole pitch behind the team: they lived inside a general-purpose cloud, felt exactly where it strained under new workloads, and left to build the version they wished existed.
The roster: Paul Sinai (CEO), Charles Drappier, Christophe Ploujoux, Nicolas Lecomte, Thomas Crochet and Mathis Joffre. Six co-founders usually raises eyebrows. A six-founder team that has already shipped infrastructure and exited together raises a different question - why would you bet against them?
In late 2025, months after graduating from Y Combinator's Spring 2025 (X25) batch, Blaxel announced a $7.3 million seed round led by First Round Capital, with Y Combinator, Liquid2, Transpose, Multimodal and a group of angels. What's notable is what Sinai said the money is not for. "Seven millions is not enough to build data centers, obviously, but I think it's important to go step by step." The plan sequences deliberately: sharpen the software and developer experience first, and only later explore custom hardware and data-center optimization.
The obvious comparison is the hyperscalers, and Blaxel is happy to make it - not to replace them wholesale, but to own the agent-shaped slice they weren't built for. The closer rivals are the compute-and-inference platforms developers already reach for: Modal, Replicate, RunPod, and sandbox players like E2B and Daytona. Blaxel's line in the sand is narrow on purpose. It does not train models and it does not sell inference. It runs the place agents live - compute, storage, networking, gateway - and leaves the models to everyone else.
Its customers reflect that focus: AI-native startups and developer teams deploying agents into production, with logos that include Shortwave, Webflow, Strapi and a roster of YC companies. The platform already handles millions of agent requests a day across 16 regions, and it arrives with SOC 2 Type II, HIPAA and ISO 27001 in hand - the unglamorous credentials that turn a promising demo into a signed enterprise contract.
There is a strategic logic to staying in the plumbing. Model providers rise and fall, agent frameworks come in and out of fashion, and application-layer startups fight bloody feature wars. The layer underneath - where the compute runs, where the state lives, where the network rules are enforced - changes more slowly and gets stickier the more an application depends on it. If Blaxel is right that agents become a durable category rather than a passing enthusiasm, the company sitting beneath all of them occupies a quietly valuable position. That is the same instinct that made owning the cloud, rather than the apps on it, one of the defining businesses of the last two decades.
Whether "hundreds of billions of agents" turns out to be prophecy or hyperbole, the underlying bet is simpler and harder to dismiss: if autonomous software becomes a large category, it will run somewhere purpose-built. Blaxel is trying to be that somewhere.