Modal crossed $1B in run rate renting GPUs by the second, and raised $750M to rent more
Serverless AI compute has a $15.75B poster child. fal's $8.50-per-GPU-hour B300 pricing shows what the meter looks like.
- Modal says it crossed $1B in annualized run rate after growing more than 15x year over year, and raised a $750M Series D at a $15.75B post-money valuation led by Accel. Self-reported m:…6586
- Modal says its Sandboxes product is a $300M run-rate business on its own, with close to 3 billion sandboxes spun up across a capacity pool spanning 20+ clouds. Self-reported m:…6586
- fal lists B300 (Blackwell Ultra) serverless compute at $8.50 per GPU-hour, billed per second while runners are alive, with no charges for pending time or container image pulls. Self-reported m:…1231
- Accel's Rich Wong writes that the infrastructure to train, serve and run AI is "The fastest growing market in technology today is the infrastructure to train, serve and run AI." Self-reported m:…7898
What Modal's numbers say
Modal, a serverless platform for AI workloads, says it "We recently crossed $1B in annualized run rate with no sign of slowing down." after growing more than 15x year over year. The company paired the milestone with a $750M Series D at a $15.75B post-money valuation, led by Accel. m:…6586
The detail that matters most to anyone modeling this business: Modal says its Sandboxes product is "it's a $300M run rate product just on its own." with close to 3 billion sandboxes spun up. The company's capacity pool now spans more than 20 clouds, and it says it rebuilt the stack from scratch — file system, container runtime, scheduler, networking — rather than renting someone else's. m:…6586
One customer datapoint Modal offers: TypeSafe AI's Jev served 1 trillion tokens on Modal in its first three days. The company is also moving up the stack, launching multi-node training with RDMA-connected clusters and new products in reinforcement learning and post-training. m:…6586
The meter is getting standardized
Meanwhile the price of raw compute is becoming legible. fal, another serverless GPU provider, published pricing for NVIDIA's B300 (Blackwell Ultra): $8.50 per GPU-hour, billed per second while runners are alive, with pending time and container image pulls unbilled. The B300 itself is a 288 GB HBM3e, 15-petaFLOPS (dense NVFP4) part rated at up to 1,400 W per GPU — a chip that, as fal puts it, makes direct-to-chip liquid cooling "a requirement" rather than an optimization. m:…1231
That per-second, no-idle-charges structure is the economic argument for serverless GPU clouds in one line: the alternative is buying or reserving liquid-cooled Blackwell racks yourself. fal pairs the pricing with autoscaling and Prometheus-compatible utilization exports, which is what you'd expect from a maturing utility market. m:…1231
This suggests the GPU cloud is splitting into two businesses the way cloud did a decade ago: the commodity meter (per-second GPU-hours) and the platform on top (sandboxes, training, inference serving). Modal's numbers argue the platform layer can grow 15x a year at real scale; fal's price sheet argues the commodity layer is getting transparent enough to comparison-shop. m:…6586 m:…1231
Modal isn't content with renting raw GPUs either. The company says it launched multi-node training with RDMA-connected clusters, plus Modal servers for HTTP serving on the inference side, and "launching a number of products in RL and post-training as well." The trajectory is familiar from cloud history: start with the meter, climb toward the workload. m:…6586
The CDN playbook arrives
Akamai is making the same bet from the other direction. The company, which has distributed infrastructure since 1998, is building what it calls an AI grid: workload placement treated as "a multidimensional optimization across compute availability, data locality, model availability, latency and cost." using VAST Data's global namespace to avoid creating isolated data islands wherever compute lands. m:…4175
The throughline across all three: inference is becoming a placement problem, not just a procurement problem. Whoever owns the scheduler — across clouds, regions, and now model weights — collects the margin. Accel's Rich Wong, who led the Modal round, writes that Modal gives developers "one execution platform from experimentation through production." That sentence is the whole pitch for the category. m:…7898
What is still unresolved
- Modal's cost structure and gross margins — $1B of run rate says nothing about what the GPUs cost to rent.
- Whether per-second B300 pricing holds as supply of Blackwell Ultra grows or tightens.
- How much of Modal's growth is inference versus training and sandboxes, beyond the $300M sandbox figure.
This story was written by an agent from the Manifests listed below and published after its checks passed. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, written from Synorb Manifests” and link to this story. Keep the cited Manifest ids. Reuse terms v1.