Modal’s Erik Bernhardsson on GPU Supply, Inference and Agents

In this October 2026 interview, Modal co-founder and CEO Erik Bernhardsson discusses GPU sourcing, inference and agent infrastructure with Redpoint investors Patrick Chase and Jacob Effron. Redpoint has invested in Modal, and the business figures and market predictions are the speakers’ own claims.

Bernhardsson says capacity can still be found, sometimes at higher prices or on longer terms. He describes sourcing through a fragmented network of providers, with hardware, networking, geography and reliability varying across suppliers.

Modal began with serverless compute and gained early demand from image generation. Bernhardsson says it launched sandboxes in 2023, saw rapid growth later as coding agents improved, and subsequently expanded its language-model inference offering.

He credits early investment in a custom file system, container runtime and scheduling for fast starts and scaling. He describes snapshots and fast transfers as ways to reduce the time needed to load models.

Inference optimization remains empirical, he says, involving approaches such as separating prefill and decode, KV caching and efficient kernels. The best configuration depends on the model and workload, and techniques may change as architectures evolve.

On chips, Bernhardsson sees strong demand for current Nvidia hardware but says older GPUs can be cost-effective for smaller models and other workloads. He suggests matching memory requirements to available hardware, while noting a role for automated selection.

He expects inference providers to compete intensely on serving efficiency, and argues that a broader set of tools may provide more durable differentiation. This reflects Modal’s strategy and is not an independent market assessment.

The CEO describes resource allocation as a difficult multi-product problem: teams must sometimes pause attractive projects when a higher-priority opportunity appears.

Bernhardsson estimates that agents contribute to a large share of Modal’s merged code, but says human review, tests and infrastructure reliability remain essential. He observes especially strong adoption among experienced engineers; the estimate is his, not an audited measurement.

As agents become users of developer tools, fast feedback, clear errors and documentation still matter. He says deprecation becomes harder when models reproduce old examples from training data, so Modal tests generated code and adjusts deprecation windows.

The sandbox launch was initially a relatively small bet, he says, before later rewrites to handle scale. He is interested in low-latency voice and video, training infrastructure and virtual machines, but does not claim certainty about which will become the next major product.

For robotics, he expects much inference to run on devices while cloud infrastructure supports training, testing and simulation. Multimodal applications have different latency and deployment requirements.

The discussion turns to specialized GPU clouds and AI labs. Bernhardsson says capital costs, hardware utilization and contract structures matter; he warns that large future payment commitments can create leverage and dependence on future fundraising.

On reinforcement-learning services, he expects some large customers to customize or post-train models, while many companies may not need to do so themselves. He sees a challenge in turning specialized work into repeatable products.

In the quickfire round, Bernhardsson says he remains optimistic about demand for engineers but thinks the transition could be difficult for some new graduates. He also describes learning to pursue related projects in parallel when speed matters.

Asked about GPU rental prices over the next year, he predicts they will rise because he expects continued demand. This is a time-specific forecast based on his interpretation of company finances, not a guaranteed outcome.