Why AI Compute Comes in Rectangles — Erik Bernhardsson, Modal CEO

What it claims

Modal reports more than $1 billion in annualized run rate and nearly three billion sandboxes created. Erik Bernhardsson says revenue grew at least 50x in the two years since his first Latent Space interview. Modal functions, launched in 2021 and mostly used for inference, account for about two thirds of revenue. Sandboxes account for about a third. The company has about 150 engineers. Bernhardsson states that Modal has "by far the largest market share" in sandboxes, and adds that he has no numbers to support this.

How it works

Capacity sourcing. Modal treats its capacity as one global pool rather than as separate clusters in each region, a decision it made early. Until recently, most of that capacity was hyperscaler spot instances (preemptible) or one-year reservations from AWS, GCP, Azure and Oracle, where it rented A10s and A100s. Today Modal works with nearly 25 clouds. Many of those deals run three or more years and require meaningful upfront payment. Four people work full time on procurement. About ten engineers manage bare-metal capacity, handling network faults, bad drives and "GPUs falling off the bus."

Vendor vetting. Vendors first receive a spec sheet covering IOPS, a minimum CPU benchmark score, and north-south bandwidth for both ingress and egress. Modal then often runs acceptance tests on a single node. Some commitments reach tens or hundreds of millions of dollars with vendors Modal does not yet know well. Repeat deals depend on an internal vendor score that rates reliability, support responsiveness and fix turnaround time. Modal looks for small blocks of capacity that frontier labs pass over, such as 32 nodes in Armenia or 128 nodes in Indonesia.

Sandboxes. The first sandbox prototype was built in one or two days on Modal's existing container and gVisor isolation stack. Modal now also offers full VMs, because some workloads need complete Linux compatibility. Its current work covers snapshotting, forking, resume and live migration. The target is sandboxes that start in about 100 ms, run a few commands, then snapshot and stop. One example is CI where the test setup runs once, is snapshotted, and each test then runs in its own sandbox.

Evidence

Early revenue was about $100 per month in November 2022, about $2,000 in December 2022, and about $30,000 in January 2023. One Stable Diffusion customer, Mage, made up 80 to 90% of revenue in Q1 2023. Suno became a large customer in late 2023. Sandbox demand came from vibe-coding companies (Lovable), background agents (Ramp), and RL training at "neolabs." One batch customer transcribed 5,000 years of audio in a few days. Bernhardsson says recent uptime has reached "four or five nines."

Positions held by Bernhardsson

• Renting capacity is more capital-efficient than owning it. Buying hardware is still on the table, along with joint ventures and SPVs.
• Locked-in reservations leave Modal effectively long GPU prices. A price collapse is the larger risk, because resale markets would vanish at the same time. He does not expect a collapse soon.
• 2027 may be the year non-NVIDIA accelerators matter. He notes that he has made this prediction every year and been wrong each time. He mentions disaggregated inference, with NVIDIA handling prefill and Cerebras handling decode.
• Small and mid-sized training is underserved, for example 64-GPU jobs in computational biology and speech. Strong open models such as Kimi K3, DeepSeek 4.1 and GLM 5.3 make fine-tuning optional.

Limits

This is an interview, so the figures are self-reported by Modal's CEO and are not audited. The transcript is automatic and garbles names, for example "Eric Fern-Hartson" and "Kaluigi" (the tool is Luigi). The fetched text stops at about the 46-minute mark, partway through the CI discussion. It does not cover the later segments listed in the timestamps: sandbox security and Anthropic, forward-deployed engineers, optional code review, bring-your-own-cloud, and the rectangular-contract segment at 1:18:31. Engineering details on GPU snapshots, the resource solver and the one-million-sandbox work are in separate posts linked from the show notes, not on this page.