Martin Casado from a16z argues that cloud computing has entered an 'operational phase' where companies are cutting costs because cloud bills are eating their profits. He estimates that if SaaS firms halved their cloud spending, their total market cap could rise by $200 billion. He's bullish on API-first companies like Twilio and Stripe, which turn software features into plug-and-play services. He also highlights Snowflake (a cheap cloud data warehouse) and Databricks (an AI/ML platform) as key players in the modern data stack, where the real competitive edge now comes from how well you use data, not just how well you code.
At a Glance Martin Casado (a16z partner) reviewed the evolution of digital infrastructure in a podcast, noting that cloud computing has gone through phases from virtualization to dominance by public cloud giants (e.g., AWS, Azure). However, large cloud users are now facing pressure from declining pr
Martin Casado (a16z partner, pioneer of software-defined networking, co-founder of Nicira) reviews the evolution of digital infrastructure from virtualization to the public cloud oligopoly (AWS/Azure/GCP). His core thesis: the cloud has entered an "operational phase," where large SaaS users have lost approximately $200 billion in potential market capitalization as cloud costs erode profits, creating a window for third-party specialized infrastructure companies—though the core profit sources of the three major cloud providers will not be disrupted in the near term.
Martin Casado argues that the development of the cloud has gone through three phases: the product phase (2005-2010, finding product-market fit), the growth phase (2012-2016, enterprises migrating to the cloud at scale, disregarding unit economics), and the operations phase (2018 to present, beginning to rationalize cost structures). We are currently in the third phase, where enterprises are seriously evaluating which workloads should remain on the cloud and which should be built in-house.
Historical context: Early cloud users were primarily startups and experimental projects; from 2012 to 2016, traditional enterprises and governments began large-scale adoption; after 2018, financial impact became the focus—because the cloud transformed capital expenditure (buying servers) into operating expenditure (pay-per-use), fundamentally altering the profit structure of SaaS companies.
Mechanism breakdown: The combination of the SaaS model and the cloud model creates a "dual cost effect." Traditional software companies sell licenses, and customers run the infrastructure themselves, so COGS does not include infrastructure costs. However, for every subscription fee a SaaS company collects, it must pay a portion to AWS and others—this directly compresses gross margins. Casado notes: "I know several household-name companies with product-line gross margins of 0%, because all revenue goes back to the cloud services hosting them."
Data chain: Casado's team analyzed 50 publicly traded software companies, assuming they cut cloud costs in half (the industry generally believes in-house operations can reduce costs by 50%). After recalculating profit margins and applying valuation multiples from comparable companies, the total market capitalization of these companies would increase by approximately $200 billion. He writes: "If you are a company with a $10 billion market cap, slightly reducing COGS could turn you into a $14 billion company."
Extrapolation: This trend is particularly important during the current market correction—every company is looking for ways to save money. However, Casado emphasizes that the core profit sources of the three major cloud providers will not be impacted in the short term, because (1) cloud penetration is still low, with a large number of workloads not yet migrated; (2) large customers already have historical preferential pricing; (3) the workloads most likely to be migrated are not the primary profit drivers for cloud providers.
Martin Casado argues that as the market matures, the unit of monetization becomes increasingly granular. By analogy with the automotive industry: in 1913, Ford's Rouge plant took raw materials in and produced finished cars out, because there was no sufficiently large supplier market; today, cars have a complete supply chain from nuts and bolts to multi-tier OEMs. Software is undergoing the same decomposition process—every "helper library" within an application could become an independent API company.
Mechanism breakdown: A typical application requires functions such as user authentication, permission control, email sending, and payments. In the past, these were libraries written by developers themselves; now they have become external API calls. Casado notes: "Five years ago, driving down Highway 101 in Silicon Valley, you could see billboards for entire companies that were just an API—HubSpot, SendGrid, Twilio." Currently, a mobile application uses an average of 17 external APIs, but he believes this is still in its early stages.
Investment methodology: The evaluation approach for API-first companies is fundamentally different from that for traditional infrastructure companies. Casado says: "In the past, to evaluate a server company, you had to ask who the buyers were, what the sales process looked like, and what the ACV was. Now with these API companies, you just look at the usage curve—how many users, how they monetize—it's more like a consumer company or a SaaS company." He admits to having missed several good opportunities due to "questioning monetization ability," with the lesson being: "The usage curve is almost always smarter than our theories. The market knows what it wants."
Pricing principles: Casado offers three psychological anchors: (1) Pricing is often determined by the market and should follow the pricing model of the ecosystem (e.g., if built on Snowflake, match its billing approach); (2) The market will tell you the price over time, so do not impose your own pricing opinions early on; (3) Early sales discussions are essentially pricing exploration—because the unit of API calls is small enough, there is ample room for experimentation.
Extrapolation: The API-first trend will accelerate, as front-end developers (at a roughly 100:1 ratio to back-end developers) are taking on more application logic and naturally prefer using third-party APIs over building their own back ends. This, in turn, will drive more back-end functions to be "API-ized."
Martin Casado argues that ten years ago, investors would ask a dog-walking company, "Can you build it? Can you scale it? Can you go global?" Today, these are no longer issues—anyone can build a global-scale website using cloud technology. The real differentiator has become "how you handle data": how to match dog walkers with dogs, how to dynamically set prices, how to make personalized recommendations, and how to run customer acquisition campaigns.
Mechanism breakdown: The modern data stack is divided into two major branches—the analytics side (helping humans make decisions, e.g., dashboards) and the AI/ML side (becoming part of the application itself, providing real-time decisions). Casado points out that in the past, the analytics side and the AI/ML side were separate, but they are now converging. He lists key companies: Databricks (AI/ML cornerstone), Snowflake/BigQuery/Redshift (analytics side), and DBT and Fivetran (ETL layer).
Data chain: Snowflake's success lies in offering a cloud service that is "cheap enough for you to dump all your data into it"—in the past, companies could only use expensive data warehouses like Teradata, processing data once and then discarding it; now, data can be revisited continuously like a resource.
Implication: The entire workflow from data sources to model serving is becoming the "new primitive layer." Casado says, "We are still trying to figure out which categories will ultimately form. This is part of the active due diligence every day—understanding how this stack actually fits together." He believes this is one of the biggest "battles for dominance" in the current technology landscape.
Martin Casado argues that the core breakthrough of AI/ML lies in enabling systems to process unstructured data (vision, natural language, speech) for the first time, digitizing it with sufficient precision and integrating it into program logic. This pushes software into the physical world—agriculture, oceanography, autonomous driving, and other areas previously outside the scope of software are now becoming software problems.
Mechanism Breakdown: This workload is sufficiently distinct that it drives full-stack innovation from ASIC chips to data centers. Casado cites examples such as Mosaic (a company building data centers specifically for AI workloads) and Mighty (browser-as-a-service, offloading browser computing to the cloud)—both of which build their own hardware, select their own servers, and stack their own racks.
Key Judgment: Casado believes that infrastructure businesses are unlike software businesses—AI businesses resemble service businesses, and infrastructure businesses often have lower profit margins. He states: "I tend to take a more conservative view of the profit structure of infrastructure companies."
Implications: Casado is a "technology optimist" and a "technology maximalist"—he believes that any problem humans face now falls within the domain of software. However, he also raises an unresolved question: "Are these companies still software companies? Or have they become agricultural companies? Or something entirely different?" This question is of great personal interest to him.
Martin Casado argues that the future of cryptocurrency has three possibilities: (1) the most pessimistic – it is all a scam; (2) the most optimistic – a complete restructuring of everything from technology to organization (encrypted routers, encrypted storage, DAOs, DeFi); (3) a middle path – adding a "consumption and monetization layer" on top of traditional infrastructure, similar to what social networks did for the internet.
Mechanism Breakdown: As a distributed systems expert, Casado focuses most on the "nature of applications" – will crypto applications remain in the realm of "awkward parallelism" (e.g., DeFi), or will they move toward general-purpose computing (e.g., a "crypto version of AWS")? This determines whether traditional infrastructure (switches, routers, storage layers) will still dominate.
Data Chain: Casado notes that the traditional infrastructure companies on whose boards he serves are selling large volumes of products to crypto companies – "This indicates that traditional infrastructure has significant adoption in the crypto space."
Extrapolation: Casado leans toward the middle path – crypto has genuine innovation (enabling anonymous groups to have strong guarantees about something without trust relationships), but most models of traditional infrastructure still apply. "You still need to build databases, you still need storage, you still need to understand asset trade-offs." He believes the answer to this question "involves enormous amounts of capital, and no one knows the answer."
| Position | Guest View | Key Data |
|---|---|---|
| Snowflake | Bullish | Provides a sufficiently low-cost cloud data warehouse, allowing enterprises to "dump all data into it"; alongside BigQuery and Redshift as a cornerstone of the analytics layer |
| Databricks | Bullish | AI/ML layer cornerstone company with a complete ecosystem |
| Mighty | Bullish | Browser as a service, self-built hardware, founded by a serial entrepreneur |
| Fly.io | Bullish | Allows developers to run compute workloads at the CDN layer, self-built hardware |
| Mosaic | Bullish | Machine learning as a service, builds data centers for AI workloads, self-built hardware |
| Hugging Face | Bullish | Early usage "growing like a rocket," the market loves it |
| Orbit | Bullish | Helps companies build and visualize developer ecosystems |
| Auth0 | Bullish | A classic case of authentication as an API |
| Twilio | Bullish | Pioneer of API-first companies |
| Stripe | Bullish | Pioneer of API-first companies |
| Dropbox | Case Study | Because S3 did not optimize its small object storage scenario, building its own data center "may have saved the company" |
| Netflix | Case Study | Famous early case of full migration to the cloud |
| AWS/Azure/GCP | Neutral to Cautious | The three major cloud providers will continue to grow, but core profit sources will not be impacted in the short term |
| Redshift | Neutral | Larger than Snowflake, but Snowflake is an independent company |
| BigQuery | Neutral | Alongside Snowflake and Redshift |
| DBT | Neutral | Modern company in the ETL layer |
| Fivetran | Neutral | Modern company in the ETL layer |
1. The cloud has entered the "operations phase," with $200 billion in market capitalization "wasted" on cloud costs (Casado): An analysis of 50 publicly listed software companies shows that if cloud costs were halved, total market capitalization could increase by approximately $200 billion. However, the core profit sources of the three major cloud providers will not be impacted in the short term—because the workloads most likely to be migrated are not their primary profit drivers.
2. "The usage curve is smarter than your theory" (Casado): When evaluating API-first companies, directly examining the usage curve is more reliable than any theory about monetization capability. He admits to having missed several good projects due to doubts about monetization, and the lesson is that "the market knows what it wants."
3. Every "auxiliary library" in an application could become an independent API company (Casado): Drawing an analogy to the evolution of the automotive industry from vertical integration to multi-tier suppliers. Currently, a mobile application uses an average of 17 external APIs, and this is still in the early stages.
4. The ratio of front-end developers to back-end developers is approximately 100:1, and the front end is consuming more of the stack (Casado): Front-end developers take on more application logic and naturally prefer using third-party APIs over building their own backends, which accelerates the API-first trend.
5. The basis of competition has shifted from "who writes software better" to "who uses data better" (Casado): Ten years ago, investors asked, "Can it be built?" Today, they ask, "How to match, price, and personalize?"—data capability has become the sole differentiator.
6. AI/ML pushes software from the "information domain" into the "physical world" (Casado): For the first time, systems can process unstructured data (vision, language, speech) and integrate it into program logic, driving software into fields such as agriculture and oceanography that were previously outside its scope.
7. Infrastructure businesses are not like software businesses—AI is more like a service business with lower profit margins (Casado): He holds a conservative view on the profit structure of infrastructure companies, believing they naturally have lower profit margins.
8. Three futures for cryptocurrency—the most uncertain but most impactful variable (Casado): On a spectrum from "all a scam" to "completely restructuring everything," he leans toward the middle path—adding consumption and monetization layers on top of traditional infrastructure. The key question is: Will crypto applications remain in the "awkward parallel" category, or will they move toward general-purpose computing?