Bloomberg Tech

Erik Bernhardsson on why AI-written code must run but never escape

Erik Bernhardsson· CEO of Modal at Modal
·~5 min·English·Bloomberg
AgentsAI InfrastructureInferenceTrainingAI CompanyOpen Source
TL;DR

Modal CEO Erik Bernhardsson explains how code sandboxes — isolated places where model-written code can run but cannot escape — went from a quiet 2023 side project to a roughly $300M business once models got good enough to write whole apps, and why they are becoming a foundational primitive of AI software.

01The Thesis

AI Applications Need a New Set of Building Blocks

<strong>The new era of AI software runs on a new set of infrastructure primitives</strong> — inference, training, and sandboxes — and Modal is betting its $750M round on supplying them.

you really kinda need a new set of building blocks for all these phases

— Erik Bernhardsson, Bloomberg Tech
Key Insight
Framing inference, training, and sandboxes as interchangeable building blocks is a positioning move: it recasts Modal from one more inference vendor into the layer every AI app depends on, which is a far harder position to displace.
02Core Primitive

Execute Code, But Don't Escape

<strong>A sandbox is an isolated place to run model-generated code — it can execute, but it cannot escape the sandbox</strong>.

to control these agents in a way where they can execute code but not escape

— Erik Bernhardsson, Bloomberg Tech
Key Insight
Execute but not escape is the entire value proposition compressed into four words: the moment a model can write code you did not review, the scarce resource stops being the code and becomes a safe place to run it.
03Inflection Point

From Side Project to a $300M Business

<strong>Sandboxes sat flat from 2023 until models could write whole apps</strong>, and then demand exploded into a roughly $300M run-rate business.

until about two years later with, I think, Sonnet 4.5 or one of those models that really could output, like, full apps.

— Erik Bernhardsson, Bloomberg Tech
Key Insight
The flat first two years are the real lesson: the product was correct before the market existed. Sandboxes only mattered once models crossed the threshold of writing code worth running, which makes Modal's growth a derivative of model capability, not of its own sales effort.
04From Niche to Primitive

More Than Hacking

<strong>Sandboxes started as a way to test what agents can hack, but became a foundational primitive for building coding models</strong> — training runs and internal tools alike.

Sandbox is really becoming more than just, you know, a way to test these agents' hacking.

— Erik Bernhardsson, Bloomberg Tech
Key Insight
The cyber-testing story is the attention-grabbing use, but Bernhardsson quietly redirects to the larger one: every team training or running a coding model needs isolated execution, so the security niche was simply the first customer of a general primitive.
05Token Economics

Cheaper Tokens, More Demand

<strong>Every drop in token prices creates more demand, not less</strong> — cheaper inference pulls more work onto the whole platform.

as token prices drop, it just creates so much more demand for these things.

— Erik Bernhardsson, Bloomberg Tech
Key Insight
This is the Jevons paradox applied to tokens: efficiency gains lower the unit price and raise total consumption. For a provider that bills for execution, falling model prices are a tailwind, not a threat.
06What's Next

A Full Computer for the Agent

<strong>The next step is a full virtual computer for the agent</strong> — Modal's new VM sandboxes hand an agent a real Linux machine to run long tasks.

agents are getting more and more capable at controlling computers.

— Erik Bernhardsson, Bloomberg Tech
Key Insight
Moving from a code sandbox to a full virtual machine is a bet on where agents are headed: from answering in one shot to operating a computer over long sessions, which needs persistent, general-purpose environments rather than a one-off code runner.