May 21, 2026 · 1h 11m · latent-space
AI Agents Need Computers: 74% MoM Growth, 850K/Day Runs, & New Agent Cloud — Ivan Burazin, Daytona
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
Daytona CEO Ivan Burazin discusses building specialized bare-metal compute infrastructure for AI agents, detailing how sub-100ms sandboxes handle massive reinforcement learning spikes and drive the emergence of the Agent Cloud.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. How this is scored →
The hosts as informed peer 5.4 Guest teaching 3.9 Guest disagreement 0.9 The hosts pushing back 1.5
speaking balance: gold is the hosts, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 1:03:04 Calling out token reselling as fake SaaS acceleration
Ivan rejects the market convention of giving premium multiples to SaaS vendors simply wrapping model tokens, arguing their margins and stickiness are inferior.
Hardest push from the hosts ▶ 26:01 Challenging the compute equivalence between Neon and GPU inferenceShawn refuses Ivan's broad grouping of spiky infra peers, arguing that S3-backed architectures like Neon do not face the same strict capacity provisioning issues as GPU inference providers.
Biggest teaching moment ▶ 38:49 Breaking down the technical and legal barriers of macOS sandboxesIvan educates Shawn on why macOS sandboxing is uniquely constrained, detailing Apple's 24-hour per-user licensing rule and prohibition of cross-machine memory snapshot migration.
The host holds their own ▶ 47:54 Citing SemiAnalysis on cascading hardware bottlenecksShawn demonstrates domain depth by citing his interview with Doug O'Laughlin, tracing the hardware shortage trajectory from GPUs to memory, networking, and now CPUs.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The hosts as informed peer | Guest teaching | Guest disagreement | The hosts pushing back | Why |
|---|---|---|---|---|---|---|
| Host Message: Sustaining Independent AI Engineering Content | 2 | 1 | 0 | 0 | Shawn opens with a housekeeping message and trades friendly anecdotes with Ivan about their early interactions and mutual history with Shift and Codeanywhere. | |
| Evolution of Daytona: From Cloud IDEs to Agent Computers | 5 | 3 | 1 | 1 | Ivan outlines Daytona's transition from early browser IDEs to composable computers for agents, while Shawn references his own early investments and 'End of Localhost' essay. | |
| The Sandbox Pivot: Rapid Prototyping and Inbound Pull | 4 | 4 | 0 | 1 | Shawn asks about the pivotal moment behind Daytona's shift to sandboxes. Ivan recounts vibe coding the MVP over New Year's Eve and witnessing immediate, aggressive customer pull. | |
| Technical Architecture: Bare Metal Performance and Custom Schedulers | 6 | 5 | 1 | 2 | Shawn probes Daytona's architectural choices compared to standard VM setups. Ivan explains why they opted for custom schedulers and bare-metal NVMe preloaded snapshots for sub-second spinup times. | |
| Workload Dynamics: Long-Running Agents vs. Spiky RL Runs | 6 | 4 | 1 | 1 | The conversation covers workload patterns, contrasting predictable follow-the-sun long-running agents with spiky, square-wave reinforcement learning runs that cause 15% average utilization. | |
| Infrastructure Challenges: Burst Handling and Compute Conference Insights | 7 | 4 | 1 | 3 | Shawn drills down into how different infrastructure peers handle bursts, questioning whether S3-native architectures like Neon face the same compute issues as GPU inference providers. | |
| Reinforcement Learning Scale: Dynamic Resizing and Hardened Containers | 5 | 5 | 1 | 1 | Ivan breaks down RL training demands, highlighting dynamic sandbox resizing to prevent OOM errors and Docker-in-Docker isolation via Sysbox. | |
| Computer Use: Enterprise Windows Sandboxing and macOS Limits | 6 | 6 | 1 | 2 | Ivan discusses the massive enterprise market for Windows-based computer use automation and explains the rigid licensing and hardware limitations that hinder macOS sandboxing. | |
| Market Positioning: B2B Agent Labs and Hardware Bottlenecks | 7 | 4 | 1 | 1 | Ivan details the economics of B2B/B2B2C agent labs versus developer-facing products, while Shawn connects hardware bottlenecks back to insights from SemiAnalysis. | |
| Open Source Philosophy, AGPL Licensing, and Rapid Procurement | 7 | 4 | 1 | 2 | They discuss open-source licensing dynamics under AGPLv3, and Shawn shares his experience from Temporal regarding enterprise procurement acceleration. | |
| Agent-First Infrastructure: Version Control and CI/CD Bottlenecks | 6 | 4 | 1 | 2 | Shawn asks about agent version control needs and CI bottlenecks, and Ivan shares bizarre customer workarounds like dumping JSON snapshots straight to S3. | |
| Company Culture: High-Touch Support and Founder Endurance | 3 | 2 | 1 | 1 | Ivan details Daytona's high-touch support ethos, tight-knit veteran team, and willingness to endure personal sacrifice to outwork competitors. | |
| Future of Compute: API Monetization and the Agent Cloud | 6 | 5 | 2 | 2 | Ivan delivers a sharp critique of SaaS companies marking up AI token reselling instead of opening raw API access, before outlining the emerging 'Agent Cloud' primitive landscape. |
Statements from this episode (28)
Burazin: AI Agents Require Diverse, Composable Computer Architectures
“And our belief is strongly that agents today and going forward will need all these different compositions of computers to do different types of tasks.”
Wang: Daytona reports massive 74% month-over-month growth
“Like you have been reporting 74% month-to-month growth, and it also, it's just been going for a while.”
Burazin: Infrastructure built for human developers fails for AI agents
“Most people thought it was the same infrastructure for humans and agents. We understood a quarter ago. It's not, we just didn't know what was the right primitive.”
Burazin: Most sandbox providers run Firecracker on preemptible VMs
“Most providers run this on top of the VMs. And also firecracker. Yeah. They run a firecracker on VMs... The common way to do it is that they, one, that the state of the machine or the hard disk is not part of the sandbox itself. And the other thing is they're …”
Burazin: AI agents need persistent, stateful computers like human laptops
“Agents will be like humans in the sense of you don't want your laptop to be shut down until you're done with work. Like, and you want to close the lid and open the lid. It's the same state. So agents would want that like pause and come back. They want those tw…”
Burazin: Daytona runs on bare metal with local NVMe preloads
“The reason why Daytona is like super, super fast and you see this on benchmarks is we essentially, we run on bare metal. We have our own scheduler. We use the underlining disk CPU and RAM. Of the underlying machine, which means your IOPS are insanely fast beca…”
Daytona spins up a single agent sandbox in 60ms
“And so our time to spin up one is 60 milliseconds with network agency. So requests, spin up, reply, 60, the whole thing, 60 milliseconds.”
Daytona can spin up 50,000 concurrent sandboxes in 75 seconds
“But if you want to spin up 50,000 at once, we are now at about 75 seconds. So it takes about 75 seconds to spin up concurrently 50,000.”
Daytona's top customer runs 850,000 daily agent instances
“But the biggest customer of ours does like about 850,000 every single day is sort of where they were there just shy of a million every single day that they're running.”
Burazin: Singapore is Daytona's number one city by user headcount
“It's interesting that our, I talked to you about this, our number one city by user Is Singapore. Which is an interesting one, right? Not by revenue, just by, just like by individual headcount”
Daytona's mean compute utilization is 15%, peaking at 90%
“And so right now, Daytona's mean utilization is 15% one five. So it's very low. But it's very spiky, but we get up to 90%.”
Burazin: CPU environments must spin up instantly to prevent costly GPU idle time
“The reason why a lot of people come to us is because GPUs are more expensive than CPUs, right? So you want your GPU running at what? A hundred percent the entire time. And so when you're running runs on CPUs, when the CPU cycle is like down and spinning up the…”
Burazin: RL workloads will reach 50% of Daytona's volume this month
“It will be this one 50%, yeah.”
Burazin: Daytona sandboxes dynamically resize memory on the fly
“We haven't got into features, but an interesting feature is that it's very hard to OOM or out of memory, our sandboxes, because we can dynamically on the flyer resize, which is like impossible on almost any other thing.”
Burazin: Daytona operates roughly 1,000 Slack Connect customer channels
“We literally have about a thousand Slack connect channels, something like that.”
Burazin: Daytona is investing heavily in unreleased Computer Use features
“And so one of the things, the newer things we were talking about earlier is we made a big bet and put a lot of investment on computer use.
That is not seen publicly in the light of day.
We haven't GA'd that yet.”
Burazin: Daytona built a Windows sandbox that spins up in one second
“And the only option right now is an EC two with Windows or on, on, on Azure.
Both of them take anywhere from three to five minutes to spin up.
We've created an actual sandbox.
So it's a second instead of milliseconds, but you have like point in time snapshots,…”
Burazin: Apple is shooting itself in the foot with macOS licensing restrictions
“And from, if anyone in Apple is listening, I very much feel that they are shooting itself in the foot of the scale of the revenue of compute or licensing they could get if they would just enable
A concurrency model similar to what you can get on a Windows and …”
Burazin: Selling to End Developers Is Hard to Scale Due to Price Sensitivity
“When your focus is the end developer, it is a very hard sell because they're very, Price sensitive, very price conscience, very, you know, around that. And there's very, it's very hard to scale. Your cap is the number of people that are willing to spin up for …”
Burazin: CPUs Will Become the Next Critical Bottleneck for AI Agents
“You will get to the point, and Dylan Patel was at the conference talking about, from Semi-Analysis that talks usually about GPUs, was also talking about how CPUs will now be a bottleneck because it will be the constraint. You won't be able to grow, or we won't…”
Burazin: Daytona open-sources everything except feature-flagged Windows and GPU support
“Everything outside of what's under a feature flag today, which is like the Windows stuff, GPU stuff, whatever it is in this open source. It is there. So everything is there, like our own scheduler, everything's there.”
Burazin: Daytona completes enterprise procurement in five days due to AI demand
“Usually when you would go through procurement to become a vendor of large companies, it would take you like two, three months. We could have done five days now.”
Burazin: GitHub introduces too much overhead for AI agent inner loops
“What we saw from our customers was that they were all trying to figure out how to do versioning. Everyone is doing it in different ways. There were some really weird ways where people were doing that. And the reason was that GitHub as is Was an overhead. Like …”
Burazin: Daytona customer bypassed Git by dumping codebase JSON snapshots to S3
“Actually the most interesting one is we had one customer that will literally Take the entire code base inside the sandbox and every, I forgot what the time sequence was. They would just dump it all into a JSON and then push that to S three. And that's it.”
Burazin: Agent-generated PR volume has made CI/CD the main engineering bottleneck
“So the amount of PRs being created is insane right now, right? In general. Everyone's creating a bunch of PRs, like everyone. And then all that has to go through CI. And then that's the bottleneck. Like everyone's bottleneck. Like not just action, like not jus…”
Burazin: SaaS revenue bumps from reselling LLM tokens will drop back down
“And I think that there will be a cold shower when people understand, like, no one's actually going to use and pay for these agents and tokens, and that wasn't actually real acceleration, but it'll drop back down.”
Burazin: Daytona is launching GPU sandboxes for non-inference agent workloads
“Oh yeah, we will, but not for inference. Like essentially what we think about is like the GPU sandbox. So if you think of like, if you have a GPU in your computer, that is what you have a GPU in the sandbox. So there are workloads that do need GPUs. Again, I a…”
Burazin: A dedicated agent cloud bundling sandboxes and primitives will emerge
“There will be a cloud built out specifically for agents. And so that cloud will have sandboxes and it will have web search and it'll have databases like SQLite or Neon or whatever, specifically for agent and other things. We are not at the end of the new infra…”