# Why AI Compute Comes in Rectangles — Erik Bernhardsson, Modal **URL:** \ **Category:** Content Creators\ **Tags:** gpu, latent-space, machine-learning, start-ups, \-startups\ **Created:** [22 October 2026 16:31 UTC](https://www.artofsm.art/t/why-ai-compute-comes-in-rectangles-erik-bernhardsson-modal/26640 "2026-10-22T16:31:33Z")\ **Posts on this page:** 2\ **Page:** 1
**Author:** ![merefield](https://www.artofsm.art/user_avatar/www.artofsm.art/merefield/32/74_2.png) [@merefield](https://www.artofsm.art/u/merefield)\ **Post date:** [22 October 2026 16:31 UTC](https://www.artofsm.art/t/why-ai-compute-comes-in-rectangles-erik-bernhardsson-modal/26640/1 "2026-10-22T16:31:33Z")
[![](https://www.artofsm.art/uploads/default/original/3X/4/2/427573e607afd04dc585f039c1da016e21ce81f1.jpeg "Why AI Compute Comes in Rectangles — Erik Bernhardsson, Modal") ](https://www.youtube.com/watch?v=IiZ9yd2Do8I) Erik Bernhardsson, co-founder of Modal, explains how the company builds specialized AI infrastructure that simplifies deploying and scaling AI applications by managing GPU capacity across clouds and offering products like serverless compute and isolated sandboxes optimized for AI workloads. Drawing from his experience at Spotify and [Better.com](http://Better.com), Bernhardsson highlights Modal’s focus on developer experience, deep infrastructure innovation, and strategic partnerships to address the evolving demands of AI inference, training, and batch processing in a competitive GPU market. ---
**Author:** ![merefield](https://www.artofsm.art/user_avatar/www.artofsm.art/merefield/32/74_2.png) [@merefield](https://www.artofsm.art/u/merefield)\ **Post date:** [22 October 2026 18:15 UTC](https://www.artofsm.art/t/why-ai-compute-comes-in-rectangles-erik-bernhardsson-modal/26640/3 "2026-10-22T18:15:05Z")
Erik Bernhardsson, co-founder of Modal, discusses the company’s mission to build AI infrastructure that simplifies deploying and scaling AI applications. Modal offers a compute and storage platform used by various customers for inference, sandbox environments, training, and batch jobs. The company focuses on providing a seamless developer experience through its software SDK, managing capacity underneath to enable fast iteration and production deployment. Bernhardsson shares that Modal has grown significantly since its early days, driven by demand for AI workloads like stable diffusion and the rise of large language models (LLMs). Bernhardsson reflects on his background at Spotify, where he built music recommendation systems, and at [Better.com](http://Better.com), where he gained experience in team building and developer productivity. His time at Spotify inspired the creation of Modal, as he recognized the challenges of AI infrastructure firsthand. Modal initially focused on general-purpose cloud compute orchestration but pivoted towards AI inference workloads as stable diffusion and other generative AI models gained popularity. The company emphasizes building deep infrastructure components such as container runtimes, file systems, and autoscaling tailored for expensive GPU workloads, differentiating itself from traditional cloud tools like Kubernetes. A significant part of Modal’s strategy involves managing GPU capacity across multiple cloud providers and negotiating long-term deals to secure resources amid a competitive and constrained market. Bernhardsson explains the complexity of balancing renting versus owning hardware, highlighting the financial and operational challenges of owning physical infrastructure. Modal currently rents most of its capacity through multi-year contracts but remains open to hybrid models. He also discusses the evolving GPU market, noting that while Nvidia dominates, alternative accelerators are emerging and could diversify the ecosystem in the coming years. Modal offers two main products: Modal Functions, a serverless compute platform optimized for AI inference, and Modal Sandboxes, which provide isolated environments for running code generated by LLMs. Sandboxes have seen rapid growth as AI-generated code use cases expand, particularly for reinforcement learning and background agents. Bernhardsson emphasizes the importance of innovations like sandbox snapshotting and ephemeral environments to reduce overhead and enable massive parallelization, which could transform workflows such as continuous integration. He also touches on Modal’s partnership with Anthropic and the company’s proactive approach to cybersecurity, including using AI agents to detect vulnerabilities. Looking ahead, Bernhardsson is bullish on expanding Modal’s offerings into training, batch processing, and low-latency applications, while continuing to build foundational building blocks for AI software. He highlights the importance of developer experience and the rise of AI coding agents, which influence how Modal designs its platform and deprecation policies. The company is rapidly growing its engineering team, focusing on systems, GPU performance, and research to optimize models and infrastructure. Bernhardsson envisions Modal as a multi-product AI infrastructure provider that empowers customers to build innovative applications by combining composable, scalable compute and storage primitives. ## Useful Links - [Erik Bernhardsson’s Personal Website](https://erikbern.com/) — Provides primary insights and background from the speaker himself. - [Modal Blog: How to Price Serverless GPU Compute](https://modal.com/blog/how-to-price-serverless) — Explains financial and technical aspects of Modal’s GPU compute offering. - [Modal Blog: GPU Memory Snapshots](https://modal.com/blog/gpu-mem-snapshots) — Supports understanding of Modal’s infrastructure innovations for AI workloads.