Background auto

Cloud had more than one winner

5 min read

A software team can pay AWS to run its product, Snowflake to analyze its data, and Datadog to investigate why checkout has become slow. Another team can choose Azure or Google Cloud. A smaller team might prefer DigitalOcean because it wants a simpler way to get its application running.

All of those purchases can exist in the same market. I keep thinking about that when I hear that the frontier AI labs will capture all the value.

The labs have a strong position today. Menlo Ventures' 2026 enterprise AI report estimates that Anthropic, OpenAI and Google capture 69% of deployment spending, including their own applications. Their share increased from the previous year. I understand why someone building with their models would worry about how much room remains.

But I think we are too quick to project today's concentration onto a market that could become much larger and serve many more kinds of work.

Even the clouds had competitors

AWS launched S3 in 2006. Google introduced its App Engine preview in 2008. Windows Azure became commercially available in 2010. DigitalOcean started in 2012 with an offer aimed at developers looking for simple, affordable cloud computing.

AWS kept growing through that competition. In the third quarter of 2026, Synergy Research estimated its cloud infrastructure market share at 26%, compared with Microsoft's 19% and Google's 16%. Amazon still led a market that supported substantial businesses beside it.

Microsoft and Google had capital, engineering and existing customer relationships. DigitalOcean pursued a narrower offer. These were different ways to compete, with different resources and customers. A provider could build a business without matching every service AWS offered or becoming the largest cloud.

Cloud also supported businesses that bought infrastructure and sold something else. Snowflake runs on AWS, Azure and Google Cloud. A customer can pay Snowflake for its data platform while the underlying cloud earns money supplying the infrastructure. Datadog sells monitoring for the systems teams operate. The team investigating a slow checkout still has work to do after it rents servers.

I think the expansion of demand is easy to underestimate here. As more companies ran more software in the cloud, there were more systems to monitor, more data to analyze, and more customers with different requirements. The opportunity grew along with the leading providers.

How much more work becomes possible?

On November 10, CoreWeave reported $3.66 billion in third-quarter revenue, up from $1.37 billion a year earlier. It named Canva, Substack and Wayve among its customers, alongside relationships with major labs. Another cloud provider is already selling substantial amounts of AI infrastructure. The company also reported a $709 million net loss.

fal is pursuing a more specialized offer. In its November 3 announcement about investing in fal, Bessemer said Adobe, Canva, Shopify and Amazon had moved production workloads to the company. fal supplies infrastructure for image and video generation, and model developers use it for distribution and serving. That gives model developers a route to customers and enterprises another place to buy inference.

The same Menlo report offers an example among applications. It estimates that the labs' coding products reached $24 billion in annualized enterprise spending, while the rest of the coding market grew 2.9 times to $10.1 billion. The labs captured more of the market, and the independent segment still expanded. A falling share can coexist with a growing business when total demand rises fast enough.

Retell offers a concrete example of a business using a lab's models. On November 12, its CEO reported a $100 million annualized revenue run rate. He described Asbury Automotive's developers managing service-call agents for 157 dealerships. In an October customer story, OpenAI said its models generated responses for 80% of Retell calls. Retell handles the surrounding call operations, including integrations and transfers to people. Customers buy a service, and the model supplier gets usage from delivering it.

I think there is room for more suppliers too. A company processing routine documents might prefer a smaller model that does the job at a lower cost per document. Another might need a private deployment. Those requirements leave suppliers different ways to compete.

Suppose a company can afford to review supplier contracts only when they are large enough to justify the effort. If AI makes reviewing smaller contracts economical, the number of reviews it buys could increase. The model provider can earn more from that work, and a company delivering the reviews can earn revenue too.

Those reviews would be new purchases, provided someone can deliver useful work at a price the customer will pay. We could spend years debating who captures today's spending while customers start paying for work they previously left undone.

The labs will compete for that demand themselves. They already sell applications, and better models could let them complete more of the customer's work. Cloud history gives us no assurance that a particular AI product will survive. It does give us a precedent for competing providers and independent businesses growing around powerful infrastructure companies.

I don't need the frontier labs to become weak businesses to believe there is room. Before ruling out a new supplier, I would want to see which customers it can serve well, what those customers still struggle to buy, and how much more work they would purchase if someone could deliver it affordably.

:sources 12
  1. AWS: Amazon S3 launch, March 2006
  2. Google: App Engine preview launch, April 2008
  3. Microsoft: Windows Azure general availability, February 2010
  4. DigitalOcean: Company history and developer focus
  5. Synergy Research: Q3 2026 cloud infrastructure market
  6. Snowflake: Supported cloud platforms
  7. Datadog: Infrastructure and application monitoring
  8. Menlo Ventures: 2026 State of Enterprise AI, spending estimates and methodology
  9. CoreWeave: Q3 2026 financial results, November 10, 2026
  10. Bessemer: Investing in fal, production customers and specialized infrastructure, November 3, 2026
  11. Retell: $100 million annualized revenue run rate and Asbury Automotive example, November 12, 2026
  12. OpenAI: Retell's model usage and voice-agent operations, October 27, 2026
:commands

j/k scroll · gg top · G bottom

[/] previous / next post · b blog · H home

t cycle theme · ? commands · Esc clear and close