Most agent companies outsource their compute layer. They pay Daytona or Modal or E2B, who pay AWS or GCP, who pay for the actual hardware. Every hop in that chain is someone else's margin stacked on top of yours. And every hop is a dependency you do not control.
We built our own sandboxing infrastructure. It was not the easy path. But now we sit at the lowest level of the stack, talking directly to bare metal providers. That means we set the pricing. We can undercut every competitor in the space and still maintain healthy margins, because our competitors are paying two or three layers of markup that we simply do not have.
The pricing advantage is obvious, but the real unlock is feature velocity. When a large enterprise customer needs host-level VM isolation instead of shared resources, we do not file a feature request with a vendor. We build it. Golden snapshots that pre-install customer dependencies so sessions boot in seconds. Keep-alive VMs that persist between messages and skip setup entirely. Regional deployments on the other side of the world, because we control where the metal lives. None of that happens on someone else's roadmap. And when a customer needs a large VM or a custom region, the answer is a Slack message to your own engineer, not a support ticket to a vendor who does not know your customer's name.
The tiering model is where the economics get interesting. At the cheapest level, you do resource sharing — VMs on the same host, splitting compute. Free users get this and they do not get a choice. But you can sell upward: VM-level isolation, host-level isolation, dedicated hosts, regional deployments. Each tier is a pricing lever that only exists because you own the infrastructure. A company built on top of Modal cannot offer host-level isolation. They do not control the hosts.
The lesson is not that every company should build bare metal infrastructure. The lesson is that the layers you own are the layers where you have leverage — the kind of structural advantage that compounds for companies playing the long game. Pick the hard thing that sits underneath your product and build it yourself. Everything above it gets easier, cheaper, and faster as a result.
Related Essays
The Long Game in Agent Companies
Agent companies are marathons you sprint through. Frameworks commoditize, monetization pins to infrastructure, and proximity to real builders is the moat.
Everything Was an AWS Wrapper Too
The 'GPT wrapper' insult misreads platform economics. The AI labs are becoming the new hyperscalers, and billion-dollar companies will be built on top of them.
Crowded Starting Line, Empty Finish Line
The AI coding agent space looks crowded today. The vast majority of entrants will pivot, run out of money, or chase the next shiny thing. The race goes to who keeps running.
Key takeaways
- Teams that build their own sandbox infrastructure can undercut competitors who stack margins on top of third-party providers.
- Owning infrastructure unlocks tiered isolation models that vendor-dependent platforms cannot offer.
- The hardest infrastructure investments compound into the most defensible competitive advantages.
FAQ
Why would an agent company build its own sandbox infrastructure instead of using a provider?
Third-party sandbox providers like Daytona or Modal charge margins on top of their own cloud costs. When you build your own sandboxing layer on bare metal, you control pricing, feature development, and isolation tiers — giving you structural cost advantages and faster iteration on what customers actually need.
How does owning sandbox infrastructure create a competitive moat?
It lets you offer the lowest pricing while maintaining healthy margins, ship custom features like tiered VM isolation without waiting on a vendor roadmap, and respond to large enterprise requirements with your own engineering team instead of a support ticket to a third party.