[AINews] Modal’s $1B ARR, $16B Series D makes it NYC’s newest AI Infra Decacorn, 3x usage and 3x valuation in 5 months, 15x in 1 yr

Modal has moved from infrastructure unicorn to decacorn: the company crossed $1B in annualized run rate and closed a $750M Series D at a $15.75B valuation, led by Accel. Modal reports that usage grew more than 15x year over year. The headline framing puts the step at roughly 3x usage and 3x valuation in five months and 15x over a year. The round was announced alongside a broader news roundup covering the same period.
A large share of that business traces to a product bet made three years ago. Modal launched Sandboxes in 2023 after noticing that people were using LLMs to generate code. According to Erik, writing in his fundraise post, Sandboxes has become a $300M run-rate product on its own. The company says it has spun up close to 3 billion sandboxes, and that they scale to a million concurrent. Stated use cases span background agents, reinforcement learning, vibecoding platforms, and more.
For developers, that mix is revealing. A sandbox is an isolated execution environment where untrusted code generated by a model can run without touching the host system. Running millions of them concurrently is an orchestration and scheduling problem rather than a model problem, which is why agent platforms now depend on providers for cold starts, isolation and fan-out. The cited use cases map onto that: background agents that need to act over long horizons, RL rollouts that need many parallel environments, and coding platforms that execute whatever their users' prompts produce. Modal's positioning suggests most of its demand is coming from exactly these workloads rather than from conventional web serving, though the company has not published a revenue breakdown by segment.
Modal's CTO, Akshat, spoke at AIE NYC about the infrastructure agents now require, listing work on millions of concurrent sandboxes, building a new sandbox runtime, launching RDMA clusters securely and elastically for distributed training, and training state-of-the-art speculators for inference. RDMA—remote direct memory access—is the high-bandwidth, low-latency interconnect used to link GPUs in training clusters, and making it elastic means users can get cluster-scale interconnects on demand instead of reserving dedicated hardware. Speculators are the small draft models used in speculative decoding, which proposes tokens a larger model then verifies, trading some compute for lower latency. Taken together, the list describes a platform trying to cover both ends of the stack: tiny, short-lived sandboxes and large, tightly coupled training jobs.
The announcement came alongside the company's first user conference, where Erik gave a keynote.
The valuation trajectory is the more interesting signal for the market. A jump from unicorn to decacorn in roughly a year, with usage up more than 15x, points to demand for agent execution infrastructure rather than model training as a growth driver. Comparable data points in the same roundup tell a similar story: Together reportedly hit $1.2B in annual bookings, Fireworks passed $1B ARR, and Baseten reportedly tripled annualized revenue to $600M, per DealBook. Inference and sandbox hosting appear to be expanding faster than the model layer's ancillary markets. That is my reading of the figures, not a claim from Modal.
In practice, a $1B run rate and a $15.75B valuation imply a revenue multiple in the mid-teens, which is aggressive by conventional software standards but not unusual for infrastructure companies growing at this rate. It also raises the bar for the next round: sustaining 15x growth gets harder as the base grows, and sandbox pricing is inherently competitive because the underlying compute is fungible. A likely trade-off for customers is that a vendor with a $300M product line has an incentive to push workloads toward its own runtime and GPU pool, which may reduce portability compared with assembling sandboxes from raw containers and a cloud provider. Modal's emphasis on a new sandbox runtime suggests it believes the execution layer itself, not just the GPUs underneath, is where differentiation lives.
Why it matters: Teams building agents, coding platforms and RL pipelines now have a clear signal about where the money in AI infrastructure is moving—toward execution and sandbox capacity rather than only model endpoints. A supplier with this much scale is more likely to be a durable dependency, and the disclosed numbers (3 billion sandboxes, one million concurrent, 15x usage growth) give engineering leads concrete benchmarks for planning capacity and negotiating contracts. Inference as a whole is getting cheaper quickly; what remains scarce, on this evidence, is the ability to run enormous numbers of small, isolated workloads safely and at low latency.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Sandboxes are all you need.