Modal announced on October 14 that it has crossed $1 billion in annualized run rate and raised a $750 million Series D at a $15.75 billion post-money valuation, led by Accel. At roughly 16 times run rate, the revenue story is clear. The question for the whole inference market is what that revenue turns into once the GPU bill comes due.
What Modal said
In its announcement, Modal says it "recently crossed $1B in annualized run rate with no sign of slowing down." It says usage has tripled since its Series C five months ago and that it has grown more than 15x year over year. Those are Modal's own figures, and we haven't seen them audited. Kleiner Perkins joined as a new investor, and Menlo Ventures, General Catalyst, Redpoint, Lux Capital and Bain Capital Ventures took part as existing ones. Earlier coverage put the Series C at $355 million at a $4.65 billion valuation, so the valuation has more than tripled in about five months.
The customers Modal names are worth reading closely: DoorDash, Cognition, Ramp, Linear, Hebbia, Runway, and robotics and drug-discovery labs. These are mostly agents, background jobs and heavy training or generation workloads, not chatbots. That fits what we've seen elsewhere, where a small group of heavy users drives most of the spending. Modal also says it plans more products in reinforcement learning and post-training and wants to expand its inference offerings.
What's actually being bought
Modal doesn't sell a model or a chip. It sells serverless GPU compute: start GPUs when work arrives, shut them down when it doesn't, and bill for what ran. The value is in how well that layer keeps expensive hardware busy. Cold starts, scheduling and batching decide how much paid work comes out of each GPU hour. It's the same idea behind Nebius's reported $100–150 million acquisition of nine-month-old Inferize, which we covered earlier.
Why utilization is the margin
A16z recently published a chart, compiled by Citadel Securities from Ornn, Silicon Data and Bloomberg data, showing token prices falling steeply since July while H100 rental prices stayed roughly in the $2.50 to $2.75 an hour range. Those figures are read off the chart, so treat them as approximate. If tokens keep getting cheaper and GPUs don't, there's less room to waste a GPU hour. A provider that rents GPUs at that price and sells the output at falling prices makes or loses money on how many paid tokens or seconds it fits into each hour.
That's also a useful cross-check on our earlier piece about the financing web. If H100 rent holds up, the near-term earnings of the chips behind that debt look healthier than the worst case. It doesn't settle the question of what those chips are worth in three years, but it supports the idea that the work running on them is paying for them so far.
What we're watching
As more companies pile into inference, we expect the winners to be the ones whose stack keeps GPUs busiest. We're watching whether the others in this space report utilization or gross margin, not just revenue growth. We're also watching GPU rental prices against token prices. If rent starts falling with tokens, utilization matters less. If rent holds while tokens keep falling, it matters more.