.png)
2 weeks ago
1
Everyone talks about the latest AI models, but few talk about the infrastructure that makes them possible. Behind every model is a layer of compute that needs to be rented, managed, and scaled. Companies that simplify that process are quietly shaping the future of AI.
Last week, I spoke with Modal cofounder Erik Bernhardsson about how his company is reshaping access to GPU infrastructure. I messaged Erik after Modal announced its latest fundraise and he responded within ten minutes, happy to hop on a quick call. It was one of those conversations that reminds you how the biggest opportunities often start with simple problems no one else has solved elegantly. And the Modal solution is indeed elegant and clearly working, as the company continues to gain traction with developers and enterprises alike.
Proof can be found here: You can read more about the round here: Modal’s $87 M Series B announcement (led by Lux Capital at a $1.1 billion valuation).
What Modal Does
At its core, Modal is an infrastructure layer that makes it simple for developers to run code on GPUs. Think of it as a way to access compute on demand without paying for full-time rentals from Amazon or Google. Developers can spin up a workload when they need it, run it as long as they want, and pay only for what they need and what they actually use.
That might sound straightforward, but in practice it removes a huge amount of friction. It allows developers to focus on building products instead of managing clusters, provisioning instances, or worrying about scaling. Modal abstracts away the pain of managing compute in the same way Stripe abstracts payments or Twilio abstracts messaging.
An Analogy: The Electric Grid for AI
GPUs are like electricity. Everyone needs power, but no one builds a private power plant in the backyard. Instead, we all draw from the same shared grid. The grid balances supply and demand so that energy flows where it is needed without every household planning for peak load.
Modal plays the same role for AI compute. Renting raw GPUs from hyperscalers is like leasing a giant generator just to toast a bagel. It is inefficient and expensive. Modal pools demand across thousands of customers. One team might be training a model, another running inference, and another testing a quick experiment. Together, those workloads smooth out spikes and dips in demand. It is statistics in action, turning chaotic usage patterns into a predictable and efficient system.
Where Modal Fits in the Stack
Modal is not competing with AWS or Azure at the hardware level. Amazon and Google still own the physical GPUs. Modal sits above them. It reserves GPU capacity across multiple clouds and manages it as a shared pool. Thousands of customers can draw from that pool and pay only for what they actually use.
This pooling model smooths demand, raises utilization, and reduces cost. Modal has effectively turned capacity management into a product. For developers, the value is speed and simplicity.
Modal’s abstraction makes GPUs feel less like a scarce resource and more like a built-in feature of the modern developer experience.
Why Modal Works Right Now
GPU rentals are still expensive, and there are countless startups, research labs, and independent developers that need access to them. Modal makes that possible without a massive upfront investment or a dedicated cloud operations team.
The rise of AI applications has made compute volatility a real constraint. Pooled infrastructure helps smooth that volatility and ensures that smaller teams can compete without the overhead of managing clusters or long-term GPU contracts. Modal is built for that world.
The Next Challenge
Modal’s early success came from individuals and small AI teams. The next test is winning over enterprises. Convincing large companies to trust shared GPU infrastructure is not easy. They worry about compliance, data security, and vendor lock-in. Modal’s bet is that these companies will eventually realize the greater risk lies in wasting time and money managing infrastructure themselves and choose instead Modal’s elegant solution.
