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

What it claims

Erik Bernhardsson, Modal's CEO, argues in this interview that fixed-term GPU contracts do not match bursty demand. Modal's business is to absorb that mismatch by pooling many customers onto rented capacity. Several operating details go beyond the summary. Modal functions (serverless functions, mostly inference) are about two thirds of revenue, and sandboxes are about a third. Modal has about 150 engineers, and Bernhardsson says recent uptime is "four or five nines." He forecasts, by his own account after years of wrong calls, that 2027 may be "the year of non-NVIDIA accelerators."

How it works

Capacity sourcing. Until late in the year before the interview, most capacity was hyperscaler spot instances (AWS, GCP, later Azure) or one-year reservations, plus early A10 and A100 rentals from Oracle. Modal treats all regions as one global pool, not per-region clusters, and calls that early multi-region decision consequential. It now buys from nearly 25 clouds, often on three-plus-year deals with meaningful upfront payment. A procurement desk of four people works full time on this. A team of about 10 engineers manages bare metal, including network faults, bad drives and "GPUs falling off the bus."

Vendors get a spec sheet with minimum IOPS, CPU benchmark scores and north-south bandwidth. Modal then usually runs a one-node acceptance test. Commitments sometimes reach tens or hundreds of millions of dollars with vendors Modal does not yet know well. Modal favours small, scattered blocks that frontier labs skip, such as "32 nodes in Armenia" or "128 nodes in Indonesia." Modal rents and owns no hardware. Bernhardsson says buying remains open but that owning hard assets dilutes equity or requires debt backed by locked-in revenue, which Modal's product-led model lacks.

Risk view. Locked-in reservations make Modal "effectively long GPU prices." So rising prices help it, and a collapse would hurt anyone holding supply. He declines to buy far ahead of demand. His reason is correlated tail risk: prices, utilization and the resale market could all fall at once. He cites the Hunt Brothers' failed silver corner as the cautionary case.

Sandboxes. Sandboxes first ran on gVisor. Modal added full VMs because some workloads need complete Linux compatibility. Demand came from vibe-coding companies (Lovable is named), background agents at Ramp, and RL runs by "neolabs." In Bernhardsson's view, the differentiators going forward are snapshotting, forking, resume and live migration. His target is sandboxes that start in about 100 ms, run a few commands, then snapshot and sleep. One example is CI: run setup once, snapshot it, then run each test in its own sandbox, about a thousand in parallel, instead of every shard repeating installs.

Early revenue. Modal was self-funded until a seed round in early 2022. Revenue went from $100 per month in November 2022 to about $2,000 the next month and about $30,000 in January 2023, driven by Stable Diffusion. One customer, Mage, was 80 to 90% of revenue in Q1 2023. Suno became a large customer in late 2023.

Next products. Training is aimed at small and mid-sized jobs, such as computational biology, speech models and recommendation systems on around 64 GPUs. Batch is another target; one customer transcribed 5,000 years of audio in a few days. Low-latency voice and post-training (SFT and RL) are also on the list. Bernhardsson says fine-tuning is now "very optional" because open models such as Kimi K3, DeepSeek 4.1 and GLM 5.3 run near frontier quality off the shelf.

Limits

The fetched text stops at about the 46:30 mark of a transcript running past 1:23:00. Several sections are missing: sandbox security and Anthropic, forward-deployed engineers, optional code review, bring-your-own-cloud and inference providers. The transcript is auto-generated and garbles names and dates. The sandbox launch and growth years both read "2023," so the timeline is unreliable. All figures are Bernhardsson's own statements, and his sandbox market-share claim is unquantified by his own admission.