Modal Grows 30-40% a Month, Buys 2-3x GPUs 3-4 Months Out

Erik Bernhardsson, co-founder and chief executive of Modal, said GPU “lead times right now is three to four months”, in an interview on 20VC with Harry Stebbings published on YouTube on 2 November. He said Modal forecasts its GPU buying from monthly demand growth of about 30% to 40% over the last 12 months: projected three to four months ahead, “you end up with like 2 to 3x”, and Modal goes out and buys that capacity.

By our arithmetic, growth of 30% to 40% a month compounded over a three-to-four-month wait gives between 2.2 and 3.8 times today’s demand. His 2-to-3x sits in the lower half of that range: it matches three months at 30% (2.2x) to four months at 30% (2.9x) or three months at 40% (2.7x), while four months at 40% comes to 3.8x.

Business Pill · Lead Time Explained

A short Business Engineer explainer on lead time: why the wait between ordering something and getting it is part of its price. It teaches the general idea only and says nothing about any company in this piece.

What Bernhardsson Said on 20VC

Stebbings asked how Modal forecasts its GPU buying, given that it does not want excess supply on its books. Bernhardsson said Modal does a “straight line or exponential line extrapolation” across its customers in aggregate, and that the demand can be predicted pretty well because the customer base is diversified.

He described buying GPUs at Nvidia’s GTC conference in March as “an open air drug market” for GPUs, and said Modal does “the dirty work of doing all the GPU deals under the hood” for its customers. In the same answer he said Modal works with 25 neoclouds; Modal’s own Series D post of 14 October says its capacity pool “now spans 20+ clouds”.

Asked whether that capacity can simply be bought, he said yes: “Prices are up for sure. You have to pay more to get the same amount of GPUs.” He added that “if you’re willing to pay more than other people, you can get a lot of GPUs”, and that Modal is “actually more long than we’re short GPUs”, because it locks up longer-term reservations and rents them out on shorter durations.

Later in the interview he said “We grew 50x in the last 12 months” and that monthly growth was still running at 30% to 40%. He also said Sandboxes, Modal’s product for running AI-written code, “is a 300 million run rate business”, about a third of revenue.

What Was Posted on X

Fireside Alpha posted this 29-second clip from Bernhardsson’s 20VC interview on 2 November. Its text matches YouTube’s captions of the full episode, filler words aside, except that it reads ‘about three to four months’ where the captions have ‘three to four months’. The 2.2x-3.8x range in this piece is our arithmetic on the rates he gives in the clip.

The Arithmetic Behind 2 to 3x

The rule he described is simple: order today what demand will be when the GPUs arrive. At 30% a month, demand is 2.2 times today’s after three months and 2.9 times after four; at 40% a month, it is 2.7 times after three months and 3.8 times after four, by our arithmetic.

The same rates over a full year compound to about 23 times (30% a month) and 57 times (40% a month), our arithmetic. His 50x for the last 12 months works out to about 39% a month.

Modal’s own blog post of 14 October gives a lower figure. It says that since the Series C five months earlier “usage has tripled” and that Modal has grown “more than 15x year-over-year”; 15x a year is about 25% a month, by our arithmetic. Neither the post nor the interview says which measure each growth figure uses.

How the Clip on X Compares

Fireside Alpha posted a 29-second clip of this answer on X at 19:08 UTC on 2 November. Its text matches YouTube’s captions of the full episode, with filler words removed, except that it reads “about three to four months” where the captions have “three to four months”.

Connecting the Coverage

We reported Modal’s $750M Series D at a $15.75B valuation and its crossing of $1B in annualized run rate on 14 October, and its plans for Nvidia Vera CPUs on 19 October. Bernhardsson’s comment that prices are up matches what we have tracked: B200 rent rose 30% in two weeks on the Ornn index in October, and Nebius raised on-demand GPU prices by up to 21%.

Modal is one of several buyers spreading demand across many neoclouds; Hyperbolic works with about 30. Our AI compute contracts tracker follows who rents compute from whom.

As we read it, the business lever is capacity planning: for a GPU cloud whose demand compounds monthly, the length of the GPU wait decides how much it has to order in advance, and so how much of its balance sheet sits in capacity that customers have not yet used.

The Structural Read

As we read it, the lead time is what turns monthly growth into an order size: the longer the wait between ordering GPUs and running them, the larger the multiple of today’s demand a fast-growing cloud has to commit to.

As we read it, Bernhardsson described two sides of the same position: Modal locks up longer-term reservations and rents them out on shorter durations, so it carries the gap between what it has committed to and what customers use at any moment.

As we read it, the two growth figures Modal has given in three weeks, more than 15x year-over-year in its blog post and 50x in the interview, imply very different monthly rates, about 25% and about 39% by our arithmetic, and neither says what it measures.

Three Implications

FOR GPU CLOUDS AND THEIR LENDERS Bernhardsson put GPU lead times at three to four months; at 30% to 40% monthly growth, that is an order of 2.2 to 3.8 times current demand, by our arithmetic.

FOR BUYERS OF COMPUTE Bernhardsson said Modal sells pay-as-you-go GPU time billed per second, so customers do not have to lock up a fixed block of GPUs for years.

FOR GPU PRICES Bernhardsson said prices are up and that buyers willing to pay more than others can get a lot of GPUs, which matches the rental-price rises we tracked in October.

The Business Engineer Lens

This story maps onto the Business Engineer framework The AI Supercycle.

The framework sets out a nine-layer stack and states its first rule: “Each layer is constrained by the one below it. You cannot ship a model without compute.”

As we read it, Modal sits on the Compute Capacity layer, and Bernhardsson’s three-to-four-month GPU wait is that rule in operation: the layer below sets how far ahead Modal has to buy. The framework is the owner’s argument and does not discuss Modal.

Related coverage: Modal tops $1B run rate, raises $750M at $15.75B valuation · DeepSeek says a GPU-hour earns $38 while rent runs near $8 · Google, Microsoft and 3 rivals hold 71% of world AI compute

What Is Not Established

The quotations come from YouTube’s machine captions of the 20VC episode, not from an official transcript. Modal has not published its GPU count, contract terms, the prices it pays, or how much capacity it has on order. The 30% to 40% monthly rate is Bernhardsson’s description of the last 12 months; the multiples in this piece apply that rate as stated and are not a forecast by Modal or by us.

The 50x in the interview and the 15x in Modal’s 14 October post may measure different things; neither defines its measure. The 25-neocloud figure is from the interview; the blog says 20+ clouds. We did not contact Modal.

The Bottom Line

Modal’s chief executive says GPU lead times are three to four months and that Modal sizes its orders by projecting 30% to 40% monthly demand growth across that wait, which he put at 2 to 3 times current capacity. Compounded, those rates give 2.2 to 3.8 times, by our arithmetic, so his figure sits in the lower half of the range his own numbers produce.