Fly.io

Fly.io built the right primitive a decade early, and the agent era just arrived to claim it: Firecracker microVMs that boot in milliseconds, sold as an API (Fly Machines), with hardware-level isolation and scale-to-zero economics — precisely the substrate every AI platform now needs to run untrusted, model-generated code safely. While competitors bolt sandboxing onto container platforms, Fly's entire architecture is the sandbox, distributed across a global anycast network that took ten years and only ~$111M to build — a capital-efficiency feat in an industry that burns that monthly. The developer brand is a genuine asset: Fly's engineering culture and radically honest technical writing bought it bottoms-up loyalty money cannot, and the disciplined retreat from GPUs — publicly admitting the bet missed and deprecating the product by July 2026 — shows a team that concentrates rather than sprawls. Managed Postgres now monetizes the existing base with the highest-attach product in hosting, and the platform's full-stack scope (compute, networking, storage, databases) means revenue per customer can expand for years. At a last reported valuation around $467M from the June 2023 Series C, the entry price is a fraction of Vercel's or Railway's marks. If agent-sandbox demand inflects, Fly is the asymmetric bet in developer infrastructure.

The numbers describe a niche business, not a platform: roughly $111M raised over a decade, no priced round since June 2023, and independent trackers estimating 2024 revenue near just $11M — traction that, if even directionally right, means Fly has been out-monetized by nearly every peer in developer cloud. The GPU deprecation, whatever its candor, is a formal concession of the AI-compute market — the single largest demand wave in infrastructure history — and the agent-sandbox thesis now faces purpose-built competitors (E2B, Modal Sandboxes) plus Cloudflare, which sells isolation atop distribution Fly cannot match. The reliability reputation is a persistent tax: years of community complaints about outages and support turned Fly's own honesty into its best-known liability, and infrastructure buyers forgive slowness before they forgive flakiness. Free-tier cuts and new charges for inter-region traffic signal margin pressure on a usage-based model that was supposed to scale gracefully. Meanwhile the competitive field owns the funnels — Vercel owns the framework, Railway and Render own simplicity, Cloudflare owns the network — leaving Fly to win on architectural elegance, which history suggests is the weakest distribution strategy in developer tools. A beloved, well-engineered, structurally subscale platform is exactly the profile that gets acqui-hired, not the one that compounds.