The PlaybookAI Infrastructure

Modal Hit $1B a Year. Sandboxes alone are $300M.

Modal crossed $1B annualized run rate in October. The sandbox product that runs model-written code is $300M of it, the size the whole company was in May.

John TanOctober 14, 20265 min read

On October 14 Modal announced it had crossed $1B in annualized run rate and raised a $750M Series D at a $15.75B post-money valuation, led by Accel.

The round is not the interesting number. This one is, and it sits four paragraphs down in the company's own post: sandboxes are a $300M run rate product on their own.

Five months earlier, at its Series C on May 21, the whole company had just passed $300M in annualized revenue. The line item that runs model-written code is now the size the entire business was in May.

$1B

Annualized run rate, growing more than 15x year over year. Company-reported.

$300M

The sandbox product alone. One third of the company, from isolated code execution.

3bn

Sandboxes spun up since the product launched in 2023.

What a sandbox is, and why it has a price

A sandbox is a machine that appears for a few seconds, runs code nobody read first, and disappears. Every agent that writes code needs one, because the alternative is running generated code next to your production data.

Erik Bernhardsson, who founded Modal, states the reason in nine words.

Erik Bernhardsson
Erik
Bernhardsson

"LLMs write code now! And it has to run somewhere, isolated from everything else."

Erik Bernhardsson  ·  Modal  ·  October 14, 2026

That sentence is the whole $300M. Somewhere is not free, and until this announcement nobody had published what somewhere costs at scale.

Figure 1. Where the agent bill lands
Three layers. Most budgets name one.
MODEL HARNESS EXECUTION Tokens in, tokens out The loop around the model A machine to run the output On every invoice Rarely a line item $300M a year, one vendor
The execution layer is the one that shows up as a surprise, because it scales with how often agents run rather than with how many people you have. Figure by nativefirst.

The million-sandbox test

In June the engineering team ran a million sandboxes at once and published the dashboard. All one million were created in under a minute, at an average of 18,518 per second.

A monitoring dashboard headed Sandboxes showing 1,000,000 concurrent sandboxes, 1,000,000 total sandboxes created and 18,518.52 average sandboxes created per second, above a bar chart of roughly forty thousand sandboxes created per fifteen-second bucket and smaller panels for CPU, memory and network.
1M concurrent · 22 Jun 2026 One million live at once, created in 54 seconds. The memory panel reserves 195 terabytes.

Modal's own dashboard, published in Scaling to 1 million concurrent sandboxes in seconds.

Scale on its own is a vanity number. The number that matters to anyone building on top of this is the one underneath it: how long a single sandbox takes to start, because an agent waits for it. They published that too, for ten thousand sandboxes sampled from the same test.

A scatterplot titled Time to started, sandbox, n equals 10000, plotting start time in seconds against sandbox creation time, with horizontal reference lines marking p50 at 0.48 seconds, p90 at 5.55 seconds and p99 at 14.66 seconds.
Start times · n=10,000 Median 0.48s. But p90 is 5.55s and p99 is 14.66s, and the spread widens as the test loads up.

The honest version of a performance claim: the median, the tail, and the shape of the degradation.

Half a second at the median is the headline. The useful fact is the tail. One run in a hundred waits almost fifteen seconds, and the scatter gets worse the harder the system is pushed. If you fan out two thousand agent runs, you are not waiting for the median. You are waiting for the slowest one.

Who is actually paying for this

The customer list in the announcement reads as a census of the agent economy. Cognition, Ramp, Linear and Hebbia are named as building background agents. DoorDash's AI platform team moves applications from experiment to production on it. Runway generates video; Physical Intelligence, Generalist AI and 1X train robots; Chai Discovery and Achira do drug discovery; Suno makes music.

The clearest picture of what the money buys comes from a case study Modal published five days later, on the legal-AI company Legora. Eirik Drage Steen, a member of its technical staff, describes an experiment as 100 cases, 4 harness variants and 5 trials each. That multiplies out to two thousand runs, and each run gets its own sandbox booting the full application pinned to a specific commit.

Figure 2. One experiment, two thousand machines
This is what an eval costs in compute.
100 4 5 2,000 CASES HARNESS VARIANTS TRIALS EACH RUNS, ONE VM EACH × × = Every run boots the whole application, pinned to a commit. Not a mock of the agent. The agent, inside the product it ships in.
Legora's standard experiment, as described in Modal's case study. The arithmetic is why a testing discipline that sounds cheap arrives as an infrastructure bill. Figure by nativefirst.

Two thousand full application boots to answer one question about one change. That is not waste. It is what measuring an agent honestly costs, and it is the reason a sandbox line exists at all.

Why this number is worth more than a benchmark

A benchmark measures whether a model can write code. A sandbox bill measures whether anyone ran it.

Every coding benchmark this year has been argued over: saturated, gamed, run on contaminated data, scored by a harness that does half the work. None of those arguments touch $300M of billed execution. Somebody wrote code with a model, and somebody paid to put it somewhere isolated and press go, three billion times.

It is the clearest outside measure we have that agent-written code is not just being generated. It is being executed, in production, at a volume large enough to fund a company.

Three things to take from it

Budget the execution layer. If your agents write and run code, you have a compute line that scales with agent activity rather than headcount, and it is probably not in your AI budget. It was not in anyone's until this announcement gave it a size.

Check where your generated code runs today. If the answer is a long-lived container with your credentials in the environment, that is the real finding, and it is a security finding rather than a cost one.

Price your evals before you commit to them. The right number of eval runs is high, and Figure 2 is what high looks like on an invoice. Decide that deliberately rather than discovering it.

What to hold loosely

Every figure here is company-reported and unaudited, and the announcement does not define how run rate is calculated. The claim that TypeSafe's Jev served a trillion tokens on Modal in its first three days is a vendor's claim about a customer, and the first infrastructure-side volume figure anyone has published for Jev. The million-sandbox test is a test, run by the company, on a date it chose.

What is not in question is the direction of the number. $300M in May for the whole company, $300M now for one product inside it, and a capacity pool spanning more than twenty clouds to keep up. Modal runs 200-plus people across New York, San Francisco and Stockholm.

Model-written code has to land somewhere. Somewhere costs money now.

Portrait of John Tan
John Tan ↗

Founder of nativefirst.ai. Runs everything on agents and writes down what works.

Sources
1Modal, "Modal crosses $1B in Annualized Run Rate and raises $750M Series D", 14 October 2026, first party: ">$1B in annualized run rate", "grown more than 15x year-over-year", usage "tripled" since the Series C, "$750M Series D at a $15.75B post-money valuation" led by Accel with Kleiner Perkins joining, Sandboxes "a $300M run rate product just on its own", "close to 3 billion sandboxes" spun up, "LLMs write code now! And it has to run somewhere, isolated from everything else", own file system, container runtime, scheduler and networking, a capacity pool spanning 20+ clouds, and a "team of 200+ across NY, SF, and Stockholm". Company-reported and unaudited; the post does not define how run rate is calculated.
2Modal, "Modal's Series C: Raising $355M at a $4.65B valuation", 21 May 2026: "surpassing $300 million in annualized revenue", "growing fivefold since September". This is the figure the $300M sandbox line is being compared against.
3Modal, "Scaling to 1 million concurrent sandboxes in seconds": a million sandboxes run concurrently with "all 1 million in under a minute", "tens of thousands of sandboxes per second", median start times "less than half a second", scheduling in "tens of milliseconds", "millions of sandboxes per day" and "up to fifty thousand concurrent sandboxes per customer". The dashboard reproduced here reads 1,000,000 concurrent, 1,000,000 created and 18,518.52 average created per second over the window 22 June 2026, 12:05 to 12:06 EDT. The scatterplot reproduced here samples 10,000 of those sandboxes: p50 0.48s, p90 5.55s, p99 14.66s.
4Modal, "How Legora runs 2,000 parallel agent evals on Modal Sandboxes", 19 October 2026. Eirik Drage Steen, Member of Technical Staff, Legora: "A typical experiment for the agent is now 100 cases, 4 harness variants, and 5 trials each" and "Our agent wraps the entire product." Each run boots the full application pinned to a commit inside its own Sandbox.
5The claim that TypeSafe AI's Jev "served 1T tokens on Modal in its first three days" appears in Modal's own announcement. It is a vendor claim about a customer and the first infrastructure-side volume figure published for Jev.

Read next

AI is moving really fast. Keep up.

Read the Playbook