Workload Clouds and the Death of the Attach Rate

Over the last five years, a new model has emerged: companies that own a specific computing workload and abstract away the underlying infrastructure entirely. The customer pays for the service and never has to think about what's running underneath. The vendor becomes the best in the world at running that workload (and in doing so, becomes the cloud provider for it).

I've been calling these companies "workload clouds." And I think they represent the most interesting infrastructure businesses being built today. We’ve seen companies like Snowflake, Vercel*, and newer entrants like fal* carve the path ahead, and many more startups are following suit.

As LLM’s improve, value is moving down the stack, from code to compute. If your product is ultimately a file filled with lines of code that can run anywhere, the question you have to ask yourself is whether that sustains value in a world where the marginal cost of writing code is approaching zero. If your product is a service (like the ability to run a specific compute workload better than anyone else), the opportunity ahead of you is enormous.

Settling for 5%

Traditional infrastructure software companies sold code while customers managed their own compute. The cloud made that easier, but the value split stayed the same: AWS, Azure, and GCP captured the lion's share as compute moved from on-prem to cloud, while software vendors settled for a layer on top.

That created great businesses. But those companies' ambitions were to help developers build in the cloud, not to own their workloads, which ultimately capped what they could become. Database providers like pre-Atlas Mongo and Redis provided DBMS software, but the database ran on AWS and was managed by a DB admin on the customer's team. CI/CD companies like GitLab and JFrog orchestrated your CI system, but the compute ran on a cloud provider.

The value capture split is dramatic. The big three cloud providers do ~$200B of ARR. The software that helps customers run in the cloud does a fraction of that¹. Yes, it earns higher gross margins, but the difference in absolute gross profit dollars is still enormous.

And that gap is getting harder to defend. In a world where the marginal cost of code is approaching zero, pure software becomes a commodity. That 5-10% attach rate (the % of a cloud bill customers spend on infrastructure software) goes to 1-2%. The answer for infrastructure startups is to stop looking like software providers and start looking like cloud providers: own the workload.

Snowflake and Databricks Showed the Way

The first signals came from data. In 2014, Snowflake separated compute and storage in the data warehouse and offered a service that let data teams scale without managing infrastructure. For their customers, they didn't become a layer on top of data infrastructure, they became the data infrastructure. And instead of capturing a 10% attach rate, they captured the whole thing. 

In the beginning, they paid most of that back to the clouds. But as they built expertise running the workload, margins improved to ~70%. Critically, where the compute runs (still mostly AWS) is an implementation detail the customer doesn't have to think about unless they want to.

Databricks* followed. The company started selling software to help you run Spark jobs and sent customers two bills: one from Databricks, one from their cloud provider. Over the last three years, they shifted to owning the workload entirely. In 2024, CEO Ali Ghodsi announced they were going "100% serverless." They increasingly send one bill, with no infrastructure to manage. Gross margins dropped from 85% to the mid-70s, but gross profit growth accelerated meaningfully.

It's easy to dismiss "the Data and AI Cloud" as marketing. But it brings to mind the full-stack shift of the mobile era that Chris Dixon wrote about back in 2014, when companies like Tesla, Uber, and Netflix built an end-to-end product or service instead of the partial stack approach and snapped up enormous shares of their respective markets.

The New Playbook is Spreading

The best infrastructure companies of this era are taking that playbook and running it across every production workload:

• Vercel* built the frontend cloud, then the web cloud, now the AI cloud
• Supabase and Neon* (now part of Databricks) built the Postgres clouds
• fal* built the generative media cloud
• Baseten, Together, Modal, and Fireworks built the LLM clouds
• Railway built the Python cloud
• Browserbase* built the browser cloud
• Blacksmith built the CI cloud

These companies are growing faster at a larger scale than previous generations of infrastructure startups — and notably, often in markets investors had written off as too small, because attach rates looked so low. Here’s the playbook they’re running:

1. Choose a critical, compute and/or storage-heavy workload and offer a service, not software, that runs it for developers.
2. Abstract away the complexity. No provisioning, no scaling. A higher-level primitive that’s easy to start and easy to scale.
3. Compete with the cloud providers on speed and developer experience, where they are weakest. Become the default place developers deploy that workload. With agents now building on behalf of developers, this increasingly means nailing agent experience to become a default primitive in the token path.
4. Become the best in the world at running it. Optimize your stack and pool resources across customers to achieve better performance and lower TCO than any customer could achieve alone.

In the early days, owning the workload is harder than shipping software. Costs are high, margins are low (my Notable Capital partner Glenn Solomon and I wrote more on serverless margins and how they improve over time here), and real enterprise customers are out of reach until you've earned credibility with startups and hobbyists. But over the long term, these businesses are more valuable (and, critically, more defensible) than anything built on a software attach rate.

The Core IP Isn't in the Pixels

The moat in these businesses lives in both the systems they are running and the continuous learning from running the systems. You can screenshot a UI and recreate it. You can't ask an AI to spin up a globally distributed edge deployment network and then go optimize it down the stack.

The core IP of a workload cloud happens in the background — the distributed systems work, the hardware optimization, the reliability engineering — that you as a user never see but constantly benefit from. Vercel has become renowned for its extraordinary developer taste, but that taste is the surface expression of deep underlying infrastructure optimization. The two aren't separable.

This is also why these businesses hold up against the AI threat better than pure software does. Running a reliable browser cloud or building a globally sharded database requires real work that can't be done on a laptop. And the operational expertise that accumulates over years of running the workload becomes a more formidable moat.

When companies focus on the workload, they also earn the right to serve anyone running that workload. Startups no longer have to segment by company size or end user profile, they can sell to anyone from an enterprise to an individual developer to an agent. And maniacal focus on the workload forces startups to change product, pricing, and packaging as the workload changes. The best example of which is the shift from developers as customers to agents as customers.

Clouds in the Token Path

Agents are consuming compute and creating new software at rates we've never seen. That makes being in the compute business more valuable than it's ever been.

The workload clouds in the best position are the ones sitting in the token path: either running the compute that agents need directly (Fireworks, Supabase, Browserbase*), or becoming the default place agents deploy new workloads. Right now, when Claude Code builds something in JavaScript, it deploys to Vercel by default. When it builds in Python, it deploys to Railway. There’s a ridiculous advantage to being the default.

These companies start too small for the hyperscalers to care about, and then become too good in their domains for the hyperscalers to compete effectively. More and more are going further down the stack — Railway, Blacksmith, and the inference clouds are building their own data centers, with strong early results. Owning hardware improves margins and enables deeper vertical integration. As AI lowers the barrier to custom silicon development, some will build hardware tuned specifically to their workload.

There has never been a better time to build a workload cloud. The playbook is clear, the prize is enormous, and the companies doing it are building moats that compound in ways pure software never could. 

Our team at Notable has had the opportunity to work with many of the best workload cloud companies—Vercel, fal, Browserbase, Inngest, and Neon. If you’re obsessing over how to best serve a specific workload, we’d love to meet you.

--

¹Public infrastructure SaaS companies (excluding cyber) do a combined ~$22.6B ARR, even when including NET, MDB, and SNOW, which are closer to workload clouds

Thank you to Aashay Sanghvi, Divyahans Gupta, Tristan Handy, Zac Smith, and my colleague Glenn Solomon for their feedback on this post.