Dora NodaThe most interesting infrastructure funding news of the summer is not really about the money. When Fly.io announced a $25M Series D, a new ex-Docker CEO, and a full rebrand around "computers for agents" on July 24, 2026, it confirmed something bigger than one company's pivot: the PaaS market has officially reorganized itself around AI agents as the primary customer. And buried in that announcement is the strongest argument yet for owning the machines your agents run on.
Here is the short version for teams deciding where their agent workloads should live: renting agent-shaped compute from a managed PaaS wins while you are bursty and experimental; owning the machines wins the moment your agents are steady-state production workers touching private code and data.
The math below shows exactly where the crossover sits — and why Fly.io's own numbers make the self-hosted case better than any self-hosted vendor could.
The announcement in 30 seconds
The facts, straight from the press release:
- $25M Series D, co-led by Dell Technologies Capital and Intel Capital, with Andreessen Horowitz, EQT, Geodesic, and YC participating. Martin Casado (a16z) and Daniel Docter (Dell) join the board.
- New CEO Scott Johnston, previously CEO of Docker, with earlier leadership roles at Puppet, Loudcloud, and Netscape. Founder Kurt Mackey moves to the board and an advisory role.
- The strongest quarter in company history, driven almost entirely by agent workloads. More than 37,000 customers build on Fly.io, and over 8,000 of them are agent-native.
- Revenue from the largest agent-native customers grew nearly 12x in 12 months. Among the company's biggest customers overall, agent-native companies now represent roughly two-thirds of revenue.
- Customer proof: agent-first companies like Firecrawl, Kilocode, and Plastic Labs run production AI applications on the platform.
The strategic headline is the repositioning: Fly.io now describes itself as "connected infrastructure designed for AI agents and the applications they build" — giving agents "real computers that have durable disk drives, secure connectivity to other systems, and the ability to scale to millions of instances." A Heroku-era PaaS just told the market its future is agent infrastructure. That framing is correct. The question is who should own the computers.
What "computers for agents" actually means
Strip away the marketing and the thesis is technical, specific, and — notably — one that self-hosted platforms share. Agents need fundamentally different infrastructure than the disposable execution environments of the first AI wave:
- Long-running, stateful compute as a first-class primitive. An agent is not a function invocation; it is a process with a working directory, installed tools, credentials, and memory of what it did yesterday. As Phonic co-founder Moin Nadeem put it in the announcement: Fly was "the only one that treated long-running, stateful compute as a first-class primitive rather than something bolted on."
- Durable disk. Agents accumulate state — checkouts, caches, artifacts, conversation logs. Ephemeral sandboxes that evaporate between runs force every session to rebuild the world.
- Millions-of-instances elasticity. Agent fleets burst: one moment you need ten workers, the next you need ten thousand sandboxes for a bulk evaluation run.
None of this is controversial. Firecracker-class microVM isolation is the de facto substrate across the industry — Fly.io, AWS Lambda, E2B, and others all converge on it — and the agent-sandbox market now counts dozens of funded competitors (E2B and Daytona alone have raised tens of millions each). Everyone agrees on what agents need. The open question is the oldest one in infrastructure: rent it by the second, or own it flat?
The rent-vs-own math for agent workloads
This is the section the title owes you, so here are real numbers at list prices. Consider a typical steady-state setup for a team running production agents: 10 always-on agent workers (roughly 1 shared CPU and 1 GB RAM each, with persistent disks), plus bursty sandbox capacity for evaluations and batch jobs.
Renting on Fly.io Machines. Current list pricing puts a shared-cpu-1x machine with 1 GB RAM at about $5.70/month always-on, persistent volumes at $0.15/GB/month, and egress at $0.02/GB in North America and Europe. So:
| Line item | Math | Monthly total |
|---|---|---|
| 10 always-on workers (1 CPU / 1 GB) | 10 × ~$5.70 | ~$57 |
| 100 GB of persistent volumes | 100 × $0.15 | ~$15 |
| Egress + stopped sandbox rootfs | estimated | ~$5–15 |
| Total | ~$75–90 |
That is honest, metered, pay-for-what-you-use pricing — and it scales linearly. Double the workers and you double the bill. Run 50 workers and you are past $300/month before egress.
Owning on a Hetzner-class box. A single dedicated machine in the ~€45/month class (Ryzen 5, 64 GB RAM, NVMe) comfortably runs 10 containerized agent workers plus supporting services — and the bill does not move when you add the 11th worker, or the 30th. A 2–3 node VPS cluster lands around €25–35/month flat. There is no per-second meter, no per-gigabyte volume line, no egress arithmetic.
| 10 workers | 50 workers | |
|---|---|---|
| Managed, metered (Fly.io-class) | ~$75–90/mo | ~$300+/mo |
| Owned, flat (dedicated box) | ~€45/mo | ~€45–90/mo (1–2 boxes) |
The crossover logic falls out immediately:
- Bursty and experimental favors renting. Fly machines auto-stop to rootfs-only billing (~$0.15/GB/month) when idle, so a prototype that runs two hours a day costs nearly nothing. Nothing you own can beat "almost zero" for sporadic use.
- Steady-state and always-on favors owning. The moment your agents are production workers that run all day — coding agents, ops agents, review agents — the meter runs all day too, and a flat box undercuts it within the first month.
- Burst amplitude is the tiebreaker. If you need to go from 10 to 10,000 instances for an hour, no owned fleet matches a vendor's spare capacity. If your burst is 10 to 50, a second owned box covers it.
Notice what this implies about Fly.io's own announcement: their revenue from top agent-native customers grew 12x because those customers' agents run constantly. Every one of those customers is past the crossover. They are paying the always-on meter at scale — which is great for Fly.io's Series D narrative and worth a hard look from each customer's CFO.
Where renting wins — honestly
Steel-manning the managed case matters, because there are workload shapes where renting is genuinely the right call:
- True hyperscale bursts. "Scale to millions of instances" is not a thing you build in a colo cage. If your product needs thousand-fold elastic spikes — bulk sandbox provisioning, giant eval sweeps — a vendor's pooled capacity is the only answer.
- Zero ops headcount. A managed PaaS means no node images to patch, no control plane to upgrade, no 3 a.m. disk alerts. For a two-person startup, that is worth more than any price delta.
- Global footprint on day one. Anycast edge, 30+ regions, traffic close to users and inference endpoints — renting buys geography you cannot replicate on owned hardware without becoming a network operator.
- Prototyping velocity.
fly launchto a running agent environment in minutes, scale-to-zero when you walk away. The meter rewards exactly the usage pattern experiments have.
If that list describes you — small team, spiky traffic, global users, no ops appetite — stop reading and rent. The rest of this post is for everyone else.
Where owning wins — the durable answer
For steady-state production agents, the owned-machines case rests on four pillars, and each one got stronger this year:
1. Cost at scale is not close. The table above is the polite version. At 50 always-on workers the managed meter runs 3–5x the flat-box cost, and agent fleets only grow. Inference tokens already dominate AI budgets; there is no reason to let the compute substrate take a second cut that scales with success.
2. Data control for agents touching private code. Agents are different from web dynos: they read your repositories, hold your credentials, and process your customers' data. Running them on machines you own — in your account, under your disk encryption, behind your network policy — removes an entire vendor from the trust chain. This is not hypothetical purism: when Anthropic shipped self-hosted sandboxes in May 2026, letting teams run agents "in any environment you control," the market signal was unmistakable. The lab that builds the agents expects serious teams to host them.
3. Hardware choice, including GPUs. Owned nodes run whatever silicon you buy — including GPU-bearing machines for inference-adjacent agent work — on your timeline. Managed agent platforms deprecate instance types, reprice accelerators, and gate regions by plan tier. Your roadmap should not depend on someone else's capacity planning.
4. No vendor-roadmap risk. Fly.io just changed CEOs, took on new board members from two infrastructure giants, and repositioned the entire company. That may all work out brilliantly — but it is a reminder that a vendor's strategy is load-bearing for your architecture when you rent. Pricing pivots, product sunsets, and acquisition-driven direction changes have a long history in PaaS (ask anyone who built on Heroku's free tier). Owned machines, running open-source control planes, do not get acqui-hired.
The deepest irony of the announcement is that Fly.io's core technical insight — agents need real, stateful, connected computers — is easier to deliver on machines you own. Durable disk, private networking, persistent identity: these are the defaults of a box in your rack, not premium features. The managed vendors had to invent "computers for agents" as a product category. Self-hosters just call them computers.
Which team are you? A decision framework
A short chooser to take back to your team:
- Prototype or side project, spiky usage, no ops time? Rent. Scale-to-zero pricing is purpose-built for you, and velocity beats unit economics until you have product-market fit.
- Production agents running all day on private code and data? Own. You are past the cost crossover, your security posture improves, and your roadmap stops depending on a vendor's board meetings.
- Both — steady baseline plus wild bursts? Hybrid. Run the always-on baseline on owned machines and burst evaluations and batch sandboxes to metered capacity. This is the shape most mature agent teams converge on.
Whichever you pick, Fly.io's announcement validates the underlying bet: agents are the workload now, and infrastructure that treats them as first-class citizens wins. The only question is whether those citizens live in your house or someone else's.
Bex.co is the open-source, AI-native Render alternative — push a git repo, get a running HTTPS service on machines you own. Star the repo on GitHub or deploy your first app today.
