I first heard about Modal while working on my last startup, when we were looking for ways to stretch our infrastructure budget and came across their startup program which offered $25K in compute credits. At the time, I mostly saw Modal as a generous source of free GPU credits. But as I started digging deeper into what they were actually building, I realized Modal is making it dramatically easier for developers to access, run, and scale the compute needed to build AI applications. That initial encounter stuck with me, and as AI infrastructure has exploded over the past couple years, I've become increasingly interested in what Modal is building and why so many AI companies are choosing to build on top of it.

Modal's Beginnings

Before you understand Modal, you need to understand who built it.

Founder Erik Bernhardsson grew up and went to college in Sweden. He joined an obscure music streaming company there called Spotify in 2009, where he spent seven years building what would become the foundation of modern music recommendation.

Along the way he built two pieces of infrastructure that became foundational to the entire data engineering industry. For example, he built Luigi, a Python workflow engine for managing complex data pipeline dependencies. Luigi was eventually adopted by Stripe, Capital One, Foursquare, and Asana before Apache Airflow superseded it. He eventually found his way to becoming CTO of Better.com where he grew the engineering organization to 300 people before leaving in January 2021 to start Modal.

His co-founder Akshat Bubna is a highly competitive programmer. He studied math and computer science at MIT, joined Scale AI as an early engineer where he designed core engineering systems and led the Natural Language and Quality teams, and then left because the deployment infrastructure he had to work with frustrated him enough that he wanted to build something better.

For the founding team, Modal was not a generative AI company in its original conception. Erik started building it on the idea that running code in the cloud should feel as easy as running it locally, and for data and ML teams it was inexplicably painful. The original pitch was something like "a better Kubernetes for data teams." The insight Erik kept returning to from his Spotify years was that the infrastructure he might have wanted didn't exist. There was a need to run expensive, specialized compute (ML workloads) and the need to do it elastically without owning the hardware. Every solution required either managing your own cluster, dealing with Docker configuration hell, or accepting slow startup times that made iteration painful. He wanted to build something that eliminated all of it.

Modal ended up rebuilding the entire stack from scratch. This approach is expensive and takes longer than wrapping existing primitives. It's also why today's Modal can do things their competitors cannot.

What They Actually Built

Modal's product is best understood not as a collection of features but as a set of composable execution primitives that share a common substrate. Define infrastructure in code. Package dependencies into images. Request compute. Run workloads elastically. Pay for actual usage. Observe everything. Avoid Kubernetes entirely.

Inference is the most obvious product. You bring open model weights or a custom fine-tuned model, select your inference engine (vLLM, SGLang, whatever), tune the runtime, and Modal handles the rest. The key differentiator is that you own the code running inside the container. You're not calling an API into a black box. You're serving your own model with full visibility into what's happening, on hardware you're renting by the second.

Sandboxes are a core part of the Modal product, and we'll spend more time here. A sandbox is a secure, ephemeral container for executing arbitrary code. This can include untrusted or model-generated code. You give it a development environment, point an AI agent at it, and the agent can write code, run tests, query databases, inspect logs, and verify output in full isolation. The sandbox makes agents useful for real engineering work rather than just demos.

Training rounds out the platform by connecting fine-tuning, experiment runs, and reinforcement learning environments to the same execution layer. The thesis is that training and inference are increasingly a loop, and a platform that handles both sides of that loop reduces coordination overhead dramatically.

Notebooks are another part of the product that pull researchers into the Modal ecosystem before they deploy anything, which is both a good product and a smart distribution strategy.

Overall, the goal seems to be that Modal manages most of the inference and application layers of the AI stack.

Modal + Lovable's Viral Moment

In late 2025, an AI application builder called Lovable experienced a viral moment. Users were creating applications at a rate the company hadn't anticipated. Over one weekend, Lovable created 250,000 applications. That required over 1 million Modal Sandbox executions. At peak capacity, 20,000 sandboxes were running simultaneously. Yet, Modal's infrastructure handled it without pre-provisioning any additional capacity.

Lovable's co-founder Anton Osika said afterward that Modal was "the only infrastructure provider that enabled us to reliably run tens of thousands of app creation sessions in an instant." Crucially, he didn't say Modal was the cheapest or the fastest or the most technically impressive on a benchmark. He said it was the only one that worked reliably at that scale under that kind of sudden demand.

That weekend accelerated something that was already happening. Coding agent platforms such as Lovable and Cursor were discovering the same compute constraint. The sandboxed execution layer was becoming the critical piece of infrastructure between the model and the production environment. Modal had built Sandboxes two years earlier, before the market existed for them. When the market arrived, Modal was ready.

From September 2025 to May 2026, Modal's revenue went from approximately $60 million to $300 million annualized. Sandboxes now represent over one-third of total revenue. A product that barely existed as a primary use case two years ago is now generating over $100 million in annualized revenue from a single product line.

The Technical Decisions That Made It Possible

The revenue acceleration wasn't entirely luck. It was the consequence of architecture decisions made years before anyone was paying for the product. I am going to briefly go through a few of those architectural advances and decisions below.

Custom filesystem with lazy loading: Modal built its own filesystem rather than relying on existing container image layers. The result is that large images don't have to fully download before a container starts. You get sub-second cold starts even for containers with heavy dependencies because the filesystem pulls data lazily as it's actually needed.

GPU Memory Snapshotting: Loading a large language model into GPU memory can take 30–60 seconds. If every container startup requires that initialization, scale-to-zero becomes economically attractive but latency-unacceptable. Modal's solution: snapshot the GPU memory state after initialization and restore it for subsequent cold starts. The startup time drops dramatically because the expensive initialization work doesn't need to be repeated. The feature is still in alpha, but it's a potential example of Modal doing something their competitors fundamentally cannot do with off-the-shelf components.

Multi-cloud scheduling across 13 providers: Modal expanded from 5 cloud providers to 13 between the Series B and Series C. This is what makes elastic GPU access possible at scale without customers needing to reserve capacity. When a single provider has a GPU shortage, Modal's scheduler routes to where capacity is available. The fragmented GPU supply market, which is a problem for every AI company, becomes a structural advantage for a platform with enough provider relationships to aggregate across the shortages. The problem is actually very similar to my former startup's LLM routing solution that solved a similar problem with LLMs instead of GPUs in the cloud.

The Rust runtime: Modal built its worker runtime, storage infrastructure, and networking stack in Rust. From my understanding, systems written in Rust don't have the garbage collection pauses that can cause latency spikes at inopportune moments in an inference request.

What SEC Filings Reveal About Modal

There was something else I found about Modal. Sometimes, a way to source information that isn't heavily reported is by looking in the legal filings surrounding these companies.

The Form D from October 2025

Form D filings are sometimes required during an institutional raise of funding. Modal published two Form Ds during their October 2025 Series B. They revealed a few interesting findings. The forms themselves were filed under Modal's legal entity name "Modal Labs, Inc." but with the curious previous legal name of "Polyester, Inc." I could not immediately find any additional information within these legal documents or in corporation entity searches about Polyester, Inc.

The total Series B offering reported in the form was $94.4 million, with $59.7 million already sold at the time of filing and $34.8 million remaining. The round wasn't closed in one transaction — it was raised in tranches across 68+ investors over several months. It also noted zero sales commissions and zero finder's fees, which means this was entirely a direct institutional raise with no placement agents. That's a signal of founder leverage and investor demand. The raise itself ended up closing around $87M at a $1.1B valuation.

Another interesting finding from this form is the board composition the Form D reveals. Erik Bernhardsson holds both Executive Officer and Director roles. Akshat Bubna is listed only as Executive Officer, not a Director. The two outside directors listed are Sarah Catanzaro, a partner at Amplify Partners (the seed investor present in every round), and Patrick Chase, a partner at Lux Capital (which led the Series B). This governance structure — founders potentially retaining board control with two outside investor representatives — is consistent with a company that raised from a position of strength rather than desperation.

WhiteFiber (WYFI) 8-K filing

Another publicly available filing provides the first publicly disclosed specific GPU procurement contract for Modal. WhiteFiber disclosed that Modal Labs is the end customer on a $17 million, 2-year H200 GPU contract that commenced revenue in June 2026. The contract is notable for two reasons beyond the dollar amount. First, Modal is characterized as both the end customer and a reference partner "supporting ongoing R&D via input on design and development," meaning Modal is actively collaborating on hardware design with their supply partners, not just renting GPUs. Second, this is one of 13 cloud providers Modal now spans, suggesting the total GPU spend across all providers is substantially larger than this single disclosed contract.

An Exascale SEC filing comparing AI infrastructure companies

This was based on 2025 data, but it valued Modal at a 64x ARR multiple, the highest multiple of any company in the comparable set. For reference: Cloudflare 34x, Together AI 54x, Snowflake 17x, CoreWeave 13x, Lambda 6x. This is an independent third-party validation of how the market priced Modal's business as of August 2025, before the Series C that followed.

October 2026 Upcoming Conference

On October 1, 2026, Modal is hosting Runtime at The Midway in San Francisco. It says the conference is one day in three tracks, covering inference, training, and agents.

This is the first time I found where Modal has hosted its own conference. The speakers include Erik, Scott Wu from Cognition, and Bryan Catanzaro from NVIDIA. The track structure tells you exactly what Modal thinks the next 18 months of AI infrastructure will be about: the systems that execute what the models decide to do.

Competitive Landscape Heading into the Fourth Quarter of 2026

Heading into the fourth quarter, Modal sits in an increasingly crowded AI infrastructure market, but I think it's important to distinguish GPU clouds from what Modal is actually trying to become. Companies like CoreWeave and Lambda compete primarily on access to GPU capacity, while RunPod and Vast.ai push aggressively on price and flexibility. Meanwhile, hyperscalers like AWS, Google Cloud, and Azure have the obvious advantage of massive infrastructure footprints and enterprise relationships.

One question going forward is whether that abstraction layer becomes more valuable as AI workloads become more complex, or gets commoditized as the major clouds and GPU providers catch up. Modal's biggest direct competition may therefore not be another GPU cloud, but the combination of hyperscalers, specialized inference platforms, and increasingly capable developer tooling. RunPod, for example, is moving into much of the same territory, while the broader GPU market is becoming increasingly competitive on price as hardware availability improves.

My general feeling is that the market ultimately separates into two layers. There will be commodity compute, where price and GPU availability dominate, and an infrastructure abstraction layer, where developer experience and workload orchestration matter much more. Modal's opportunity is to own the latter. If it can make deploying an AI workload feel as simple as deploying a function feels today, it could increasingly become the central, foundational piece of the AI stack it wants to be.

What to Watch

I believe a few things are going to be worth watching in Modal's trajectory over the next 12–24 months:

  1. Modal's revenue is growing dramatically, but revenue from a compute platform where you're essentially reselling GPU time is not the same as recurring subscription revenue. The spread between what Modal pays 13 cloud providers for GPU capacity and what customers pay Modal is an important unknown in the business. The emergence of reinforcement learning as a major Sandbox workload is the highest-upside scenario. If frontier labs start running training experiments on Modal at scale, the revenue and moat implications are enormous.

  2. Modal runs 100,000+ concurrent sandboxes executing untrusted model-generated code. One serious container escape or cross-tenant data leak would be catastrophic for the trust the business is built on.

  3. The transition from developer-led adoption to enterprise sales is where many great infrastructure companies lose their culture and slow down. Whether Modal can add commercial infrastructure without diluting the technical depth that built the product is an organizational test, not a technical one. The real moat is customers using inference, Sandboxes, training, and notebooks together. A customer running their entire AI workflow on Modal primitives would have switching costs that are extremely high. The single most influential metric to watch is what percentage of customers use three or more Modal products.

Ultimately, if AI applications increasingly require massive amounts of dynamic, specialized compute and if the winning infrastructure is the infrastructure developers don't have to think about, Modal has a shot at becoming one of the core execution layers underneath the AI ecosystem. I honestly think Modal is one of the companies I'll be watching most closely going forward.

What Runs AI is independent research on the infrastructure, systems, and companies powering artificial intelligence. If you have comments, questions, or feedback, reach out via LinkedIn.