Browser infrastructure is resource-intensive. Running headful (8 GB) Chromium instances for AI browser agents requires either expensive always-on capacity or slow cold starts that frustrate users. Kernel, a browser infrastructure provider, previously relied on Docker containers with warm pooling to maintain acceptable latency, but the costs were unsustainable.
By switching to Unikraft’s snapshot-based microVM architecture, Kernel eliminated warm pooling entirely, reduced idle costs to zero, and achieved 2x faster end-to-end latency than competitors. The efficiency gains also enabled a new $30/month hobby tier to support their product-led growth strategy without burning through enormous amounts of cash.
The Opportunity
Kernel provides cloud browser infrastructure: Headful Chromium instances that developers use to create browser agents that perform tasks on the internet on a user’s behalf. Traffic patterns are spiky and unpredictable (similar to CI/CD workloads) with bursts driven by use cases like agentic commerce where purchase attempts cluster at specific times of day.
Before Unikraft, Kernel ran Docker containers with warm pooling. This created several problems:
-
High idle costs: Warm pools meant paying for capacity whether it was being used or not. For a workload with significant variance, this overhead was substantial.
-
Human-in-the-loop limitations: Many use cases involve a browser performing an action, then waiting for human input or an LLM to generate the next step. With Unikraft, Kernel’s browsers can be paused for up to 72 hours and resumed immediately from the same state.
-
Co-located code execution: Customers wanted to run their own code alongside browsers to minimize network round-trips. Useful browser automation requires many CDP (Chrome DevTools Protocol) calls, and latency between user code and the browser directly impacts performance.
-
Pricing constraints: The infrastructure costs made it difficult to offer accessible entry-level pricing. Unikraft drastically reduced this cost.
We got a lot of feedback that our entry level was a bit steep. We wanted to make it as easy as possible to just try it out and fall in love with the product.
— Gabriel Guerra, GTM Lead, Kernel
The Solution
Partnering with Unikraft, Kernel re-architected their browser infrastructure around three key innovations:
Snapshot-Based Browser Sessions
Rather than maintaining warm pools, Kernel now boots a VM with Chromium once, snapshots it, and stores it on disk. When a user requests a browser session, the system resumes from that snapshot. A proxy layer sits between users and the Unikraft infrastructure, keeping the CDP connection alive while the underlying VM can be snapshotted again during idle periods.
The result: Kernel only pays for disk storage during idle time, but not compute!
Hardware-Level Isolation
Each browser instance runs as a lightweight microVM, providing stronger isolation than containers. This is important for multi-tenant browser infrastructure handling untrusted workloads from different customers.
App Platform: Co-located Code Sandboxes
Using the same Unikraft technology, Kernel built an App Platform that runs customer code in isolated sandboxes alongside browsers. This eliminates the network latency that would otherwise degrade performance for automation scripts making many CDP calls.
To do anything useful with a Chromium browser, you need a lot of CDP calls, which requires a lot of network calls. If the network latency is high or it’s a little flaky, performance can suffer drastically.
— Catherine Jue, Co-Founder & CEO, Kernel
The Outcome
The migration to Unikraft delivered measurable improvements:
-
10x faster cold start latency: Customers who have benchmarked Kernel against competitors report an order of magnitude better performance. Cold start speed is consistently the top feedback point in customer conversations.
-
Zero idle costs: By snapshotting VMs during inactive periods, Kernel eliminated the warm pooling overhead. Unlike competitors, they don’t need to charge customers when the CDP connection is idle.
-
Viable human-in-the-loop support: Use cases that involve waiting for human input or LLM-generated actions are now easy to support, since there’s no cost penalty for idle time. Live state is preserved for up to 72 hours now.
-
$30/month hobby tier: The cost structure enabled Kernel to launch accessible pricing for their product-led growth strategy without burning money excessively.
-
Partnership opportunities: The performance improvements have helped Kernel establish partnerships with computer use model providers who need high-quality browser infrastructure.
When I talk to customers, they all bring up the cold starts and the performance, every single one of them. Unikraft has enabled us to outcompete other providers in performance and quality, which has become a key differentiator for us.
— Gabriel Guerra, GTM Lead, Kernel
| Metric | Result |
|---|---|
| Cold-start latency vs competitors | 10x faster |
| Cost during idle | $0 (disk only) |
| Business model impact | $30/month hobby tier now possible |
| Isolation level | Hardware-level (microVM) |
We couldn’t have done that without Unikraft. They’ve been our secret weapon.
— Gabriel Guerra, GTM Lead, Kernel
Learn how Unikraft can help you build faster infrastructure.