The cloud has always had a human alibi. Somebody clicks the button, tolerates the spinner, decides when a machine should start and forgives it for forgetting what happened yesterday. Paul Sinaï is building for the moment that alibi disappears. His newest customers may be AI agents: tireless little programs that run code, call tools, stall for a model response, resume at odd hours and expect their workspace to be exactly where they left it.
This is a wonderfully inconvenient customer. It can create a pull request before a conventional virtual machine has finished clearing its throat. It may need a great deal of compute for five seconds, almost none for five minutes, then the same files and processes again. A cloud designed for people browsing web apps can host such behavior. It simply makes everyone pay for the privilege of pretending the behavior is normal.
Sinaï's company, Blaxel, begins at that mismatch. Its persistent sandboxes are isolated microVM environments that scale to zero after one second of inactivity and, according to the company, resume in under 25 milliseconds with their state preserved. The easy metaphor is a laptop lid. Close it, the machine sleeps. Open it, the work remains. Blaxel wants to do that for millions of agent workspaces, spread across a global network, without charging developers while the machines are idle.
An obsession with the invisible parts
Long before agents began asking for their own computers, Sinaï was already occupied by the invisible machinery behind screens. In 2009 he founded TWM Project, which specialized in connected television delivered through the cloud. He later served as partner and CTO at Mobile2You. The subjects changed, but the instinct did not: find a stack full of awkward parts and give the user one coherent surface.
His old Medium biography supplies a useful word for the habit: “simplexity,” or the pleasure of making something complicated simple to use. It is not a promise to remove complexity. Real infrastructure refuses that bargain. The work is to accept complexity on the customer's behalf, then present it with good manners.
In 2015, that instinct became ForePaaS. Sinaï co-founded the company with Vincent Michel and François Vaillant around a blunt corporate problem. Data projects required engineering, science, analytics, storage and deployment tools that often behaved like relatives forced to share a holiday table. ForePaaS bundled the journey into a multi-cloud platform so teams could move from raw information to production applications without assembling every pipe themselves.
The platform found customers across Europe, North America and Asia, including a partnership with Mitsubishi Research Institute. ForePaaS raised $10 million in 2017. By 2022, it had a 23-person team and an acquirer that understood the plumbing: OVHcloud. The European cloud provider bought ForePaaS to extend its Platform as a Service portfolio across the data value chain. The founders and staff joined the group, and ForePaaS technology became part of OVHcloud's analytics products.
An acquisition can be a full stop, a semicolon or a comfortable chair. For Sinaï, it became field research. Inside a large provider he could see both the reach of modern cloud infrastructure and the assumptions it had accumulated. Websites and software subscriptions behave predictably enough. Autonomous agents are less polite. They fan across model providers, external APIs, databases and regions. They may wait longer than they calculate. They need isolation because the code they execute is not always known in advance. They also need memory, because an agent that wakes up amnesiac is merely a demo with a recurring hangover.
The reunion startup
Blaxel began in 2024 with a founding arrangement that makes startup advice columns twitch: six co-founders. Alongside Sinaï were Thomas Crochet, Charles Drappier, Mathis Joffre, Nicolas Lecomte and Christophe Ploujoux. This was not a collection of impressive strangers assembled by spreadsheet. The group had built infrastructure together before, much of it at ForePaaS. Sinaï had known Drappier for more than 15 years.
The team took all six founders through Y Combinator's Spring 2025 batch. Most companies arrive with fewer. Blaxel barely fit on the Demo Day stage, a detail Sinaï recounted with amused candor. His explanation was equally plain: the company had a great deal to build, and the group already knew how to build together.
There is a practical lesson under the crowded group photo. Trust can be infrastructure. People with shared technical scar tissue spend less time performing certainty for one another. They know who will keep digging through a networking failure, who notices the developer experience fraying, and which disagreement needs another hour instead of another meeting. A six-person founding team creates coordination costs. Prior history can pay some of them in advance.
The agent workload, simplified
During the eleven YC weeks, the team designed network architecture, prototyped the product and tested ideas with other founders. It entered 2025 without sandboxes. Work on a faster runtime for agents and MCP servers exposed a more immediate need: secure places where agents could execute arbitrary code and humans could preview the result. Sandboxes launched as a product line in April.
The company found a pitch sharper than “cloud for agents.” An agent should be able to keep its workspace indefinitely, wake almost immediately and pay only when it is active. Blaxel also colocates agent logic beside the sandbox to cut network delay, and wraps the compute layer with model routing, MCP hosting, batch jobs, observability and policies controlling where workloads may run.
A cloud customer without a calendar
By the end of the YC batch, Blaxel was handling millions of agent requests a day across 16 regions. One customer used the infrastructure to process millions of session replays, turning video into product insights. This is not the cinematic robot future. It is more peculiar and more commercial: an immense population of software workers opening computers, examining files and closing them again.
In July 2025, one month after the batch, Blaxel announced a $7.3 million seed round led by First Round Capital, with Y Combinator, Liquid 2, Transpose and others participating. Sinaï was careful about what that money could and could not buy. Seven million dollars does not build a fleet of data centers. It can improve interfaces, services and software, then let the company move downward into more specialized infrastructure one layer at a time.
The deeper ambition is visible in the company's language. Blaxel does not merely want agents to run on its infrastructure. It expects agents to operate infrastructure too. Traditional cloud consoles assume fingers, dashboards and patient administrators. An agent consumes APIs directly and can make thousands of small decisions. If software becomes an operator, the interface, security model and economics all change.
Sinaï's public writing shows the same systems appetite. He has written about automated replication, storage choices for AI and open table formats for petabyte data lakes. The recurring move is to look past a fashionable application and ask what must exist underneath it. Connected TV needed cloud delivery. Corporate analytics needed a coherent data platform. Agents need computers that remember how to nap.
The second act remembers
Founder profiles often make reinvention sound like a bonfire: burn the map, abandon the past, emerge in a more flattering jacket. Sinaï's second act is more interesting because it remembers. He kept the appetite for difficult platforms, the preference for abstraction and, remarkably, much of the team. What changed was the customer.
That gives Blaxel a measured kind of coherence. The company is not predicting that every agent will need a tiny computer because tiny computers are charming. It is responding to observable work: coding agents need safe execution, video agents need enormous runtime, and long tasks need state that survives the pauses. The product is built around the pause, usually the least glamorous moment in computing.
There is wit in that choice. The AI industry is crowded with systems eager to demonstrate thought. Sinaï is selling them somewhere to leave their coat. If his premise holds, the valuable part of the agent economy may include an uncountable number of rooms, quiet most of the time, each ready to restore a half-finished job before a human could notice it had gone to sleep.