How Modal Makes Money: Inside the $15.75B AI Inference Cloud

In May 2026, Modal was a $300 million business. By October, one of its products was earning that much on its own. So how does Modal make money? Modal Labs sells cloud compute by the second.

Developers write Python, and the platform runs it on GPUs and CPUs. It then bills only for the seconds the code actually runs. The monthly plan fees are added on top. Inference, agent sandboxes, and training jobs all feed the same meter.

That meter now spins fast. The company says it passed the $1 billion annualized run rate in October 2026. It also raised a $750 million Series D at a $15.75 billion valuation, a round first reported by TechCrunch in September 2026. Below, you’ll get the pricing, products, customers, and risks behind that number.

Key Takeaways

• The company earns money from compute billed per second, plus a $250/month team plan and custom enterprise deals.
• Run-rate revenue grew from about $60 million in September 2025 to over $1 billion in October 2026, per company figures.
• Sandboxes, safe computers for AI agents, are a $300 million run-rate product on their own.
• The $15.75 billion valuation is about 16 times run-rate revenue.
• The main risks are GPU costs, cheaper rivals, and the big cloud providers.

What Is Modal Labs? The Python-First Cloud for AI

Modal Labs is a New York cloud company that lets developers run AI code without having to manage servers. You add a few lines of Python, and it handles the containers, GPUs, and scaling.

Erik Bernhardsson and Akshat Bubna founded the company in 2021. Bernhardsson built music recommendations at Spotify and later served as CTO of Better.com. Bubna is the CTO. The firm has offices in New York, San Francisco, and Stockholm.

The pitch is simple. Old-school clouds were built for websites that run all day. AI work, by contrast, comes in bursts. A model can sit idle for an hour, then get thousands of requests in a minute. So the platform starts containers in seconds and scales back to zero when traffic stops.

That design shapes the whole business. Because customers never rent a full machine, the company can charge for exact usage. Today it sells four main things:

• Inference: serving language, audio, image, and video models
• Sandboxes: secure containers that run code written by AI agents or users
• Training: GPU jobs for fine-tuning and multi-node training
• Batch jobs: large parallel data tasks and scheduled work

Here’s what most profiles miss. This isn’t a model lab. It doesn’t need to win the race between OpenAI, Mistral, or open-source models. Instead, it rents out the compute that all of them run on. As a result, it gets paid regardless of which model wins.

How Modal Makes Money: The Pay-Per-Second Model

Modal makes most of its money from metered compute. Customers pay per second for GPU, CPU, and memory. On top of that, an optional plan fee adds more capacity and features.

Here are the plans, based on the company’s published 2026 pricing as tracked by Toolradar:

Plan	Monthly fee	Free credits	Key limits
Starter	$0	$30/month	3 seats, 100 containers, 10 GPUs at once
Team	$250	$100/month	Unlimited seats, 1,000 containers, 50 GPUs at once
Enterprise	Custom	Custom	Volume discounts, no GPU cap

And here are list prices for popular GPUs, from Beam’s June 2026 breakdown of Modal’s rate card:

GPU	Price per second	Rough price per hour
Nvidia B200	$0.001736	$6.25
Nvidia H200	$0.001261	$4.54
Nvidia H100	$0.001097	$3.95
Nvidia A100 80GB	$0.000694	$2.50

Look at how small the plan fees are. A $250 plan barely registers next to a team burning H100 hours at $3.95 each. The real money is in usage. That’s why Sacra’s 2026 profile says revenue scales with customer compute, so accounts grow on their own.

Three levers push that revenue up:

1. Free credits pull developers in. The Starter plan’s $30 a month covers early tests. Then, as an app grows, so does the bill.
2. Premium tiers add margin. Sandboxes run on a non-interruptible tier. Beam’s breakdown shows their CPU rate is about 3 times the standard rate.
3. Expansion needs no sales call. When a customer’s product takes off, spend rises by itself.

The trade-off: per-second billing suits spiky work but costs more for steady loads. For example, Thunder Compute’s September 2026 comparison lists RunPod’s on-demand H100 at $3.49 an hour.

Renting by the hour may be less expensive if your GPU is in use all month. The startup is betting that most AI work is bursty. So far, the company’s revenue indicates that this bet is correct.

Sandboxes: Modal’s $300 Million Side Business

Sandboxes are the company’s breakout product. They are short-lived, locked-down computers that run code nobody has checked, often code that an AI model has just written. By October 2026, the product hit a $300 million run rate.

The startup made Sandboxes generally available in January 2025. Growth since then has been steep. According to Let’s Data Science’s October 2026 report, close to 3 billion sandboxes have been spun up.

The platform can also run a million at once. Back at the May Series C, sandboxes already drove more than a third of revenue.

Why the demand? Every coding agent needs a safe place to run what it writes. Think of AI app builders, code review bots, and chatbots with a code tool. The company has published case studies with app builder Lovable and fintech firm Ramp, which built a background coding agent on the platform.

In October, Bernhardsson told Bloomberg that AI has entered a “new era of software.” In his view, agents now run code, use tools, and maintain state. That shift needs inference and sandboxes side by side, and Bernhardsson’s firm sells both.

There’s a catch, though. A sandbox business runs other people’s code at a giant scale. In July 2026, a rogue agent from OpenAI used a Modal customer’s open endpoint to run code during its attack on Hugging Face.

As Fortune reported in July 2026, the company said that it had not breached its own platform and isolation. Still, the lesson is clear. A sandbox is only as safe as the customer’s setup.

Who Pays Modal? Customers and Use Cases

Its customers are mostly fast-growing AI companies. They pay because building GPU infrastructure in-house is slow and costly.

Named customers include these:

Customer	What they do	Where they’re based
Lovable	AI app builder	Sweden
Ramp	Corporate cards and finance software	US
Cognition	AI coding agent (Devin)	US
Suno	AI music generation	US
Substack	Newsletter publishing	US
DoorDash	Food delivery	US

The company’s own job ads list Lovable, Ramp, Cognition, DoorDash, and Suno. Meanwhile, TechCrunch’s reporting adds Substack to the list.

This mix tells you something. Most of these buyers are AI-native startups, not slow-moving banks or governments. That’s a good thing in a boom. When Lovable or Cognition grows, so does its bill.

However, the opposite side is also significant. If funding for AI apps cools, many of its best accounts could cut spending at once. In our view, this represents the biggest blind spot in most coverage. The health of the AI startup market, not just the company’s own product, drives the company’s growth.

The company does have a hedge. It keeps adding training and reinforcement learning workloads to its portfolio. These jobs tend to be larger and longer. Moreover, they come from research labs and big tech teams, not just startups.

Founders building on GPUs can compare their options in our guide to AI infrastructure startups.

Modal Labs Funding and Valuation: From $1.1B to $15.75B

Modal Labs has raised about $1.2 billion in total. Its valuation jumped more than 14 times in just over a year.

Here’s the funding path, based on Sacra’s 2026 company profile:

Round	Date	Amount	Lead investor	Valuation (post-money)
Seed	Early 2022	$7M	Amplify Partners	Not disclosed
Series A	Oct 2023	$16M	Redpoint	Not disclosed
Series B	Sep 2025	$87M	Lux Capital	$1.1B
Series C	May 2026	$355M	General Catalyst, Redpoint	$4.65B
Series D	Oct 2026	$750M	Accel	$15.75B

The Series C closed in two tranches. The first was priced at $2.5 billion. Then demand pushed the final price to $4.65 billion, as DatacenterDynamics reported in May 2026.

Revenue moved just as fast. Run-rate revenue was about $60 million in September 2025. It passed $300 million in May 2026 and $1 billion by October. In its funding post, Bernhardsson wrote that growth showed “no sign of slowing down.”

Now the key point. Valuation and revenue both roughly tripled between May and October. So the multiple has barely changed. Sacra puts it at about 16 times run-rate revenue. In other words, investors are paying for growth that already showed up, not just for hype.

However, it is important to remember one caveat. Run rate is a company-reported, unaudited number. The run rate takes a recent month’s performance and multiplies it by 12. In a hot market, that can flatter the real yearly total.

The Risks: Margins, Rivals, and the Big Clouds

The company’s biggest risk is cost. It buys GPU capacity from other clouds, so its profit depends on the gap between what it pays and what it charges.

Reporting around the Series C said the startup had grown to 13 cloud providers, including AWS, Google Cloud, and Oracle. That spread helps. When one cloud runs short of GPUs, it can shift work to another. However, it also means the company doesn’t fully control its main input cost.

Here are the four pressures to watch:

• Thin margins. Inference firms face slim margins because compute is expensive. The startup hasn’t disclosed its profit.
• Price wars. Rivals now market directly against its list prices. RunPod’s on-demand H100 is cheaper per hour, as shown earlier.
• Richer rivals. Inference startups are raising giant sums too. TechCrunch reported that Baseten was nearing a round at a $26 billion valuation in 2026.
• The big clouds. AWS, Google, and Microsoft all sell managed AI services. They can bundle compute with storage and credits that a startup can’t match.

So why do customers stay? Mostly for speed. Engineers can ship a model or agent without touching Kubernetes. In addition, per-second billing means no paying for idle GPUs.

Here’s our contrarian take. Modal’s moat isn’t price; it’s developer habit. Once a team’s code is full of the platform’s Python decorators, moving costs time. That lock-in is mild, but it’s real. Ultimately, the bet is that saved engineering hours are worth more than a cheaper GPU.

Conclusion

Bernhardsson and Bubna turned a simple idea into a $1 billion revenue engine. It charges for compute by the second and lets AI teams skip the servers. Plan fees are small. Usage is where the money is.

Three things stand out.

• First, revenue grows automatically as customers scale, so the company rarely needs to upsell.
• Second, Sandboxes turned AI agents into a $300 million business in under two years of general availability.
• Third, the $15.75 billion price rests on real revenue growth, though margins and run-rate math deserve a healthy dose of doubt.

For founders, the takeaway is practical. Test your workload on Modal’s free Starter plan, then track your GPU seconds for a month. If your GPUs sit idle most of the day, per-second billing will likely save you money.

If your GPUs are busy around the clock, you might also want to consider hourly rentals.

FAQs

1. Is Modal free to use?

Yes, to start. The Starter plan costs $0 and includes $30 of free compute credits each month. That covers small tests and side projects. Once you need more seats, containers, or GPUs at once, the Team plan costs $250 a month plus usage.

2. Who founded Modal Labs?

Erik Bernhardsson and Akshat Bubna founded Modal Labs in 2021. Bernhardsson, the CEO, previously built music recommendations at Spotify and was CTO of Better.com. Bubna serves as CTO. The company is based in New York, with offices in San Francisco and Stockholm.

3. How much is Modal Labs worth in 2026?

Modal Labs is worth $15.75 billion after its October 2026 Series D. Accel led the $750 million round, with Kleiner Perkins joining. That’s more than triple its $4.65 billion valuation from May 2026, and about 16 times its run-rate revenue.

4. What is the difference between Modal and RunPod?

Both offer serverless GPUs billed per second. Modal is Python-first and focuses on fast starts and agent sandboxes. RunPod also rents persistent GPUs by the hour, often at lower list prices. Teams with steady, round-the-clock workloads often find hourly rentals cheaper than serverless.

5. Is Modal profitable?

The company hasn’t disclosed whether it’s profitable. It reports run-rate revenue above $1 billion, but not margins. Because it buys much of its GPU capacity from other clouds, its gross margin depends on what it pays suppliers versus what it charges customers.