Modal, the company that runs other people's code on borrowed GPUs, just became one of the most expensive plumbing businesses in software. On October 14 it announced a $750 million Series D at a $15.75 billion post-money valuation, led by Accel, with Kleiner Perkins and every existing investor from Menlo Ventures to Bain Capital Ventures tagging along for the ride.
That is roughly triple the valuation the company carried five months ago, when its Series C closed. Modal also says it recently crossed $1 billion in annualized run rate. Do the math: investors just paid about 15.75 times revenue for a business whose core job is deciding which of twenty-plus clouds should host your Python function.
The Sandbox Is the Product Now
The most interesting number in the announcement is not the valuation. It is $300 million. That is the annualized run rate of Modal Sandboxes alone, the product that spins up isolated environments so coding agents and reinforcement-learning jobs can execute untrusted code without setting the host machine on fire.
Modal has spun up close to three billion of those sandboxes and can now handle a million concurrent. Those figures explain the round better than any pitch deck. Agentic coding stopped being a demo somewhere in the last eighteen months. It is a workload now, and workloads need somewhere to live.
Founder Erik Bernhardsson wrote that usage tripled since the Series C and grew more than fifteenfold year over year. TypeSafe AI's Jev burned through a trillion tokens on Modal in its first three days, which is the infrastructure equivalent of a houseguest drinking your entire fridge and then asking about snacks.
To keep up, Modal threw out most of the standard stack. It built its own filesystem, container runtime, scheduler, and networking, then bolted them onto a capacity pool spanning more than twenty clouds. The company describes the alternative as YAML hell and GPU procurement purgatory, which remains the most accurate description of Kubernetes-era AI infrastructure anyone has published this year.
The timing of the round is the genuinely funny part. In the same week, public markets staged a small revolt over AI spending. Nvidia, AMD, and Dell slid in a global tech rout tied to AI funding worries, Broadcom's Hock Tan warned about funding constraints, and the IMF flagged an AI-driven market correction as a risk to financial stability.
So we have a split-screen economy. Public investors are nervously counting capex receipts. Private investors are writing $750 million checks at 15.75 times revenue for a sandbox company. Both groups insist they are pricing the same future.
Modal's answer is that the future looks like more agents, more rollouts, more post-training runs, and vastly more short-lived compute. It recently shipped multi-node training with RDMA-connected clusters and no capacity commitment, plus new inference tooling and generally available virtual machines inside Sandboxes. Every one of those features is a small admission that customers want to run heavier things for shorter periods.
The customer list reads like a badge scanner at an AI conference: DoorDash, Cognition, Ramp, Linear, Hebbia, Runway, Suno, Physical Intelligence, and 1X. Nothing says "infrastructure layer" quite like being quietly present inside everyone's product while appearing in none of their marketing.
None of this makes the valuation safe. A 15.75x multiple assumes the run rate keeps compounding, which assumes agents keep multiplying, which assumes enterprises keep paying for tokens that write more code. That is a lot of assumptions stacked like a Jenga tower made of GPUs, and the table underneath it is the public market that spent last week edging away.
But the money is not really betting on a single product. It is betting that the sandbox, that boring isolation layer nobody wants to build in-house, becomes the toll booth for every agent that ever runs code. Modal has more than 200 employees across New York, San Francisco, and Stockholm, and a founder who closed his announcement with the sentence every founder closes with: he cannot think of a better time to be building software.
Maybe he is right. Or maybe the simpler summary is this. In 2026 the hottest commodity in AI is not a model, a chatbot, or a chip. It is a box that runs your code and then disappears. Someone just paid $15.75 billion to own the best box.
Comments