DEVOURED
I am building a cloud
A developer is building exe.dev, a new cloud platform that decouples VM provisioning from resource allocation to fix what he sees as fundamental design flaws in AWS and other major cloud providers.
What: exe.dev is a cloud service that lets you buy CPU, memory, and disk resources, then run as many VMs as you want on them, rather than provisioning individual VMs tied to specific resource allocations. It includes local NVMe storage, built-in TLS/auth proxies, global regions, and anycast networking.
Why it matters: Current cloud providers treat VMs as the fundamental unit tied to resources, forcing developers to either pay for nested virtualization overhead or use limited PaaS abstractions. The author argues remote block storage made sense for HDDs but creates 10x IOPS overhead with SSDs (1ms network RTT vs 20μs seek time), egress costs are 10x data center rates, and the rise of AI agents generating more code will amplify these pain points.
Takeaway: If you're frustrated with cloud VM limitations, expensive egress, or slow remote storage, exe.dev is in early launch and addressing these specific constraints.
Deep dive
- Traditional clouds couple VMs to CPU/memory resources, but since a VM is just a Linux process in a cgroup, you should be able to run multiple VMs on purchased resources without nested virtualization penalties
- Remote block storage made sense when disks had 10ms seek times and 1ms network RTT was acceptable overhead, but SSDs have 20μs seeks making network overhead 50x worse than local
- Configuring an EC2 VM for 200k IOPS costs $10k/month while a consumer MacBook delivers 500k IOPS out of the box
- Cloud egress pricing is 10x what data centers charge, with worse multipliers at moderate volumes unless you spend millions monthly
- Kubernetes exists primarily to paper over bad cloud abstractions, making portability and usability problems slightly more bearable but fundamentally unsolvable
- PaaS solutions trade power for convenience, leaving developers stuck when they hit obscure platform limits deep into projects
- LLM agents will generate significantly more software per developer, making current cloud pain points worse and consuming more context window working around abstraction problems
- exe.dev's initial launch provides pooled CPU/memory for running arbitrary VMs, local NVMe with async replication, automatic TLS/auth proxies, and global anycast networking
- The team is rebuilding from the ground up, including racking their own servers in data centers to control the full stack
- Future roadmap includes static IPs and user-accessible automatic disk snapshots
Decoder
- Remote block devices: Storage volumes accessed over the network rather than attached directly to the server, like AWS EBS
- IOPS: Input/Output Operations Per Second, measures how many read/write operations storage can handle
- RTT: Round-Trip Time, the latency for a network packet to travel to a destination and back
- cgroup: Linux control groups, a kernel feature that isolates and limits resources for groups of processes
- gVisor: A container runtime that provides additional isolation by implementing a user-space kernel
- Anycast: A network routing method where traffic is sent to the nearest server in a group sharing the same IP address
- Egress: Data transferred out from a cloud provider to the internet, typically expensive
Original article
exe.dev addresses the VM resource isolation problem: instead of provisioning individual VMs, users get CPU and memory and run the VMs they want.
Explore all 55 articles in today's digest
Fresh dev news, AI-digested daily. Never miss a trend at devoured.fyi