Modal, the AI infrastructure platform founded by Erik Bernhardsson, has accelerated sharply: the company reports it crossed a $1 billion annualized run rate, raised $750 million in a Series D at a $16 billion valuation, and spun up nearly three billion sandboxes. In a wide‑ranging interview Bernhardsson explained Modal’s infrastructure choices and the economics behind its growth, with a particular focus on GPU capacity procurement and sandbox economics.
The "rectangle" problem: contracts vs. real demand
A conventional capacity contract delivers a fixed number of GPUs for a fixed period — if you draw capacity over time the result is a rectangle. Real workloads are not rectangular: they grow, spike, decline and otherwise vary. Buying for peak capacity leads to paying for idle hardware; buying less risks unmet demand.
Modal’s bet is a duration transformation: pool demand from thousands of customers, secure capacity underneath, and let each customer consume what they need when they need it. A smoother aggregate demand profile makes that pooled capacity more valuable. This complements approaches that make capacity tradable; Modal’s emphasis is on making capacity usable in small, elastic slices.
There is a financial flip side: locking in capacity at today’s prices exposes a provider to downside risk if GPU prices later fall. Bernhardsson said he doesn’t expect a near‑term price collapse but acknowledged the risk matters when structuring long‑term deals.
The sandbox that almost didn’t survive — now a major revenue stream
Modal launched sandboxes after noticing language models could generate runnable code; the company built an environment to execute that code. Early adoption was modest and the team considered shutting the product down. Subsequent demand from coding agents, background agents and reinforcement learning changed that trajectory: sandboxes have grown to roughly one third of Modal’s revenue, a result Bernhardsson did not predict at launch.
This growth alters infrastructure requirements: agents may create environments, run code, inspect results and retry many times, so the latency for starting and restoring environments is part of the product experience. Fast startup, snapshotting and resume semantics thus became product priorities.
Concrete infrastructure techniques for speed and elasticity
Modal’s serverless GPU approach combines multiple low‑level pieces to reduce the work of bringing a new model replica online:
- a buffer of immediately available GPUs
- a custom filesystem layer
- process checkpointing
- GPU memory snapshots
Together these components reduce startup cost and latency, enabling more responsive autoscaling and more efficient use of expensive GPU resources.
How Modal gets and manages capacity
Historically Modal sourced much capacity from hyperscalers (spot and on‑demand). As demand and market conditions evolved, the company expanded to nearly 25 clouds and multiple neocloud vendors, often negotiating multi‑year (three‑plus year) deals with upfront payments and price locks. Modal currently rents rather than owns capacity and has built a procurement desk and engineering teams devoted to managing bare‑metal providers, network variability and operational problems.
Bernhardsson framed the buy vs. rent decision as a trade‑off: ownership can reduce some costs but adds operational complexity, balance‑sheet risk and financing requirements. Renting can preserve capital efficiency for a software‑focused organization, but Modal keeps open all options as it secures capacity for future years.
Market dynamics and price risk
NVIDIA is ramping shipments, but capacity is tight and prices have risen; scarcity creates rents that benefit hardware vendors and owners of capacity in the short term. Bernhardsson worries more about the tail risk of a rapid price collapse, which would hurt players who locked in expensive supply. He also expects more diversity in accelerators over time — alternatives to NVIDIA are improving and could change the supply landscape.
Product map: functions and sandboxes
Modal’s primary products are:
- Modal Functions — serverless functions optimized for inference with rapid autoscaling (the majority of revenue).
- Modal Sandboxes — isolated, short‑lived environments for running model‑generated code and agents (about one third of revenue).
Because Modal aggregates many customers, the platform smooths demand and can slice capacity finely, enabling on‑demand allocations that feel immediate to users.
Security, agents and platform practices
There was a public incident where an application deployed by a customer was exploited; Bernhardsson clarified that the platform itself was not breached. Modal is rethinking default security restrictions and increasingly uses agents internally to probe for sandbox escape routes and vulnerabilities.
On engineering practices, Modal emphasizes strong CI/CD and testing. The company treats code review as optional in many cases, preferring automated safety nets and high test coverage to gate every single change with manual review.
Organization, go‑to‑market and FDEs
Modal has grown to roughly 200 employees and is scaling further. The company expanded its forward‑deployed engineers (FDEs) team significantly: FDEs work directly with customers on performance tuning, model optimization and migrations, and they also help identify productizable feature gaps.
Bernhardsson highlighted three current priorities: (1) capacity and the financial/legal complexity of procuring it, (2) product strategy (what building blocks the market will need next: faster snapshots, lower latency, training/batch capabilities), and (3) company culture as the organization scales rapidly.
Training, on‑prem and future opportunities
Modal is building out training, batch and other product areas beyond inference and sandboxes. Bernhardsson sees an underserved market in smaller and mid‑scale training (research and domain‑specific models) that does not require hyperscaler‑scale clusters. He remains skeptical about a broad on‑prem comeback; his view favors multi‑tenant cloud platforms and BYOC/brokered approaches in cases where customers insist on particular infrastructure.
Final perspective
Modal frames much of its value on two pillars: a developer/agent experience that accelerates iteration, and the ability to aggregate global compute and slice it elastically for customers. By combining low‑level infrastructure techniques (snapshotting, checkpointing, custom filesystem) with procurement and operational work across many clouds, Modal aims to turn the rectangular supply contracts common in the GPU market into fine‑grained, usable capacity for modern, bursty AI workloads.
