AI Agents Need Computers: 74% MoM Growth, 850K/Day Runs, & New Agent Cloud — Ivan Burazin, Daytona
TL;DR
Daytona’s pivot happened because agent builders rejected “human infra” and then chased the new product down for access — Ivan Burazin says 20-30 early users told them their first sandbox product was “garbage,” but after a New Year’s Eve prototype and two-week rebuild, every demo call ran over and customers literally called asking, “Where is my API key?”
The core insight is that AI agents need stateful, resumable computers — not just disposable code runners — Daytona built on bare metal with its own scheduler so agents can pause/resume like a laptop, combining “lambda and EC2” behavior with 60 ms single-sandbox startup times and 50,000 concurrent spin-ups in about 75 seconds.
The workload mix has changed dramatically, with RL and evals now heading toward 50% of the business — Burazin says they went from zero RL workload a few months ago to nearly half, driven by customers who need huge, bursty CPU fleets to keep expensive GPUs fully utilized during training and eval runs.
The market is exploding, but the operational problem is brutal: demand is wildly spiky and utilization is low — Daytona’s biggest customer runs about 850,000 sandboxes per day, they’ve seen requests for 500,000 concurrent CPUs, and average utilization is only 15% even though peaks hit 90%, forcing them to plan for extreme bursts.
Computer use on Windows is Daytona’s next big bet because legacy enterprise software still traps trillions in white-collar labor — Burazin estimates global knowledge-worker wages at roughly $50 trillion, argues much of that work is stuck in old Windows apps, and says Daytona can spin up Windows sandboxes in about a second versus the 3-5 minutes typical on EC2 or Azure.
Burazin’s hottest take is that many SaaS companies are faking “AI reacceleration” by reselling tokens instead of exposing real API consumption — he argues the durable winners will be companies that expose their products programmatically for agents, praising Salesforce’s move to expose every product via API as the kind of shift that creates real usage growth.
Summary
From browser IDE pioneer to agent-compute founder
Burazin opens with the long arc: he and his co-founder built one of the earliest browser-based IDEs with Codeanywhere, back before VS Code, Kubernetes, or even mature Docker existed. That forced them to build deep infrastructure themselves, and those old lessons — plus a detour into the Shift conference business — eventually fed straight into Daytona.
The pivot started when “OpenDevin as a service” didn’t work
The January pivot came after they wrapped their old dev-environment automation product around OpenDevin and noticed something strange: not many users wanted the app, but lots of agent builders wanted the runtime underneath it. Burazin went on a crash course through podcasts, blogs, conferences, and white papers, then on New Year’s Eve hacked together the first version of what Daytona is now — a prototype his CTO immediately called “absolute garbage,” before rebuilding it properly in two weeks.
The product clicked when customers wouldn’t stop calling
That rebuilt version changed everything. They brought back the same people who had dismissed the earlier product, and suddenly every 15-minute call stretched to 25 or 30 because users urgently wanted access; if Daytona didn’t send API keys fast enough, customers followed up immediately. Burazin says he’d never seen that kind of pull before, which was the moment they knew this wasn’t a nice-to-have — it was a missing primitive.
Why agent sandboxes are not just VMs with a new label
Burazin’s argument is that most people misunderstand the category: agents don’t just need tiny isolated code-execution boxes, they need “composable computers.” Daytona runs directly on bare metal with its own scheduler, keeping snapshots local on NVMe so startup is fast, state persists, and agents can pause and resume like a human closing and reopening a laptop — which he describes as combining Lambda-like speed with EC2-like persistence.
The numbers are absurd: 60 ms startup, 50,000 in 75 seconds, 850,000 runs a day
He gets concrete fast: one sandbox can start in about 60 milliseconds including network latency, and 50,000 can spin up concurrently in roughly 75 seconds. Their largest customer is running about 850,000 sandboxes per day, and Daytona has fielded demand for 500,000 concurrent CPUs — exactly the kind of scale that turns “sandbox” from a toy-sounding term into serious infrastructure.
Two very different businesses are hiding inside one product
Burazin splits usage into long-running background agents — companies like Cognition, Lovable, and Harvey — and bursty RL/eval workloads from labs. The first behaves like human traffic, with weekday noon peaks and quieter nights and weekends; the second looks like a square wave, where someone suddenly wants 10,000 or 100,000 CPUs at midnight, then nothing, which leaves Daytona averaging just 15% utilization despite occasional 90% peaks.
The compute conference confirmed everyone has the same new infrastructure headache
At Daytona’s compute conference at San Francisco’s Chase Center, Burazin gathered infrastructure founders like Neon’s Nikita and Parallel’s Perog and found the same pattern everywhere: agent-native products create unprecedented spikes that traditional cloud planning didn’t prepare people for. His takeaway wasn’t just that Daytona has this issue — it’s that the whole agent infra stack is now wrestling with “follow-the-sun” workloads on one side and impossible-to-predict burst demand on the other.
Windows computer use could unlock the next giant market
Toward the end, Burazin zooms out to computer use: the idea that sophisticated agents still need actual desktops because so much valuable work lives inside ancient enterprise software. He argues that legacy Windows apps still gate enormous chunks of healthcare, government, and finance, and says Daytona’s fast Windows sandboxing could help unlock a market tied to trillions of dollars in knowledge work — while Mac remains frustrating because Apple’s licensing and VM restrictions make true elastic agent workloads nearly impossible.
Was This Useful?
Share
Keep Reading
Make Alcreon Yours
Tune your feedFive quick questions, and the feed ranks what matters to you first.Or just get notified
The weekly Echo. Signal worth keeping in your inbox.
Every new piece, announced on X.
Read Next
See all
Playbook
The Retirement Email Isn't a Warning
Model retirements now arrive every few weeks; the config-eval-rehearsal loop turns each deprecation email from a fire drill into an afternoon swap.
Playbook
The Cheapest Model That Passes
OpenRouter lists 400 models behind one API. The fix for choosing isn't a better leaderboard, it's a four-step protocol that ends in a real eval.
Playbook
Cheap Models, Hard Tasks
Most agent workflows route every step to the frontier model by default. The bill scales with how chatty the agent gets, even when most steps don't need that brain.