Martin Casado - The Past, Present, and Future of Digital Infrastructure - [Invest Like the Best, EP.280]
What this covers
The host, Patrick, talks with Martin Casado, an a16z partner who pioneered software-defined networking and co-founded Nicira. Casado first divides cloud adoption into eras: early experimentation from 2005 to 2010, mainstream enterprise migration from 2012 to 2016, and a period from 2018 onward when cloud became a recurring third-party expense. The Dropbox story anchors the discussion. Its many small objects were a poor fit for AWS storage, so it built its own data centers for that one workload. Casado turns this into a case for selective workload repatriation and puts a number on it. In his analysis of 50 public software companies, halving cloud spend would add roughly $200 billion in aggregate market value, because COGS savings flow through the market's multiple.
The conversation then ranges widely. Casado uses Ford's Rouge River factory to argue that market size sets how finely production divides, and he applies that to applications splitting into components that become companies. He describes API-first businesses as judged by usage graphs, and he regrets passing on such deals. Hugging Face is his example. He also argues that differentiation now comes from data use rather than software ability, and he describes the AI/ML prediction pipeline as an emerging primitive.
Several positions are contested. He expects the big three clouds to keep growing and holding margins. He thinks open source matters less once infrastructure is consumed as a service. On crypto he favors a middle view, in which it layers onto traditional systems, and he leaves open whether distributed systems can support general compute. He also worries that infrastructure innovation has gone dark through consolidation.
Casado argues digital infrastructure is entering a new innovation cycle: cloud economics compress SaaS margins, pushing companies to unbundle workloads and creating room for specialized infrastructure providers, while APIs and data turn applications into composable businesses.
- Cloud costs sit inside SaaS cost of goods, so workload optimization can move valuations materially
- Markets growing large enough let each application component become its own company, as the Ford factory example illustrates
- Value is shifting from software-only differentiation to data use, and usage-based pricing fits API infrastructure
This asset isn't compiled yet
You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.
In traditional software vendoring, infrastructure costs were borne by customers, so software had roughly 80% margins; with SaaS, hosting infrastructure becomes part of the vendor's cost of goods, which fundamentally changes the cost structure and often compresses margins.
“if your product is software as a service, then part of your cost of goods is actually the infrastructure.”
Modern AI and ML can take unstructured data such as vision, natural language, and speech and make it usable in typical program logic, which previously was impossible, pushing software into the physical world and driving different hardware and software designs including ASICs.
“what modern AI and ML does, which we've never really been able to do before in systems, is take unstructured data, and digitize it, and add it to the typical logic of the program.”
Distributed systems complexity cannot be papered over with a thin software layer or hidden under an API, so the distributed paradigms developers actually use will determine what systems can be built, and whether crypto networks become general compute or only embarrassingly parallel applications remains unresolved.
“distributed systems is one of those things that you just can't paper over with a thin software layer. You can't hide it under an API.”
The core crypto innovation is enabling anonymous parties with no prior trust relationship to have strong guarantees, first through a ledger and then through more general compute, which unlocks new business use cases.
“allowing basically an anonymous set of people with no prior trust relationship to have strong guarantees on something”
Ford's Rouge River factory turned raw inputs into cars because the market was too small to support suppliers; as the car market grew, nuts, bolts, and multiple tiers of OEMs and integrators emerged, showing that market size determines how granular the unit of production becomes.
“the reason is there wasn't a sufficiently large market for cars to actually have suppliers.”
Cloud adoption moved through eras: 2005-2010 early experimentation by startups, hobbyists, and tech-forward firms like Netflix; 2012-2016 mainstream enterprise migration; and 2018-2019 onward, when cloud shifted from capital expense to a recurring third-party cost that affects finances and prompts optimization.
“the move to the cloud has implications on your finances, because now instead of you being able to buy a physical asset and internalize that, you're basically paying a portion of your income to a third party.”
Technologies like Kubernetes have matured to make it practical to run workloads across clouds or on owned infrastructure, so companies can now honestly debate which workloads belong on third-party clouds, the big three, or their own infrastructure, and optimization will follow.
“technologies that enable you to run on clouds, in between clouds, that's also matured.”
Cloud data warehouses made it cheap enough to store all of an organization's data, so data becomes a resource firms repeatedly return to, rather than something processed once because storage was expensive, replacing costly on-premise systems like Teradata.
“it's cheap enough that you can throw it all there and keep it in there.”
Dropbox was paying a tremendous amount for S3 because it stores many small objects, a workload AWS was not optimized for; Dropbox built its own data centers and software for that specific workload, probably saving the company, though it moved only specific workloads rather than the whole company.
“S3, which is the storage layer on Amazon is not optimized for this use case of many small objects.”
As infrastructure is consumed as SaaS, the value of open source decreases for functional services, because the operational code that makes a service reliable is specific to the running service and would never be open sourced; developers will simply consume an API operated by someone else.
“as we move to SaaS and that's the primary way of consuming infrastructure, which it seems to be, that open source matters a lot less.”
Crypto's promise for users is being realized and innovation has been rapid, but people conflate crypto's user-facing innovation with the mechanisms used to deliver it, so traditional infrastructure combined with distributed innovations is still a win even if the crypto-native transformation is slower.
“sometimes we conflate the innovation of crypto, what it means to the users and innovation with the actual mechanisms to get there.”
Moving workloads off the public cloud makes little sense for simple webpages or non-compute-heavy vertical SaaS, but for data-intensive or compute-heavy workloads (such as music serving or large websites) optimization can justify building owned infrastructure for a subset of workloads.
“if you're just doing a webpage, it doesn't make sense to do your own infrastructure.”
Three views of crypto exist: it is fake or a Ponzi scheme, it fully reforms technology and organizations, or it is an innovation layered on top of traditional systems; Casado leans to the middle view, where real primitives for networks, applications, and consumer monetization add to rather than replace traditional networking, storage, and clouds.
“I tend to be in the middle where I think that there's a real innovation there. I think there's real value.”
The big three cloud providers will continue to grow and preserve margins because cloud adoption is still early, many workloads should still move to the cloud, and the companies that benefit most from leaving have sweetheart pricing that gives the clouds pricing power.
“The three big cloud providers are going to continue to grow and continue to preserve margins.”
For API-first companies, the unit of consumption is a function or API call, so investors can evaluate them on usage graphs, number of users, and monetization path, making them resemble consumer or bottoms-up SaaS companies rather than enterprise server companies evaluated by buyer and go-to-market.
“you literally just can look at what the usage graphs are, how many users, how do they monetize them”
Casado has regretted rejecting API deals on the theory that developers won't pay or the go-to-market won't work, because the usage graph is smarter than investors' theorizing; the market knows what it wants, as the Hugging Face example shows.
“the graph in almost every case is just smarter than our theorying. The market actually knows what it wants.”
In the 1970s one company built the chip, motherboard, sheet metal, OS, and apps; then the OS separated from hardware and apps separated from the OS, and now the application itself is being disaggregated into independent components that can become companies.
“So now what's happening is the application itself is being disaggregated.”
Casado's analysis of 50 public software companies found that halving their cloud spend and benchmarking margins against public peers would increase aggregate stock prices by about $200 billion (the blog post conservatively stated $100 billion).
“It turned out that it would increase in aggregate the stock price by $200 billion.”
A core operational AI/ML primitive is a model prediction that makes application decisions based on analyzing data rather than rules hand-coded by people, used for applications like fraud detection and pricing, and the full pipeline from data sources to served models is becoming a standard primitive.
“those decisions, rather than being something somebody coded, it's the product of analyzing data.”
Casado is hyper bullish on digital infrastructure, believing it is the beginning and end of technology, that vertical SaaS will become horizontal thin layers atop core infrastructure, that most value will accrue to infrastructure, and that decades of growth lie ahead.
“I believe that the future like vertical SaaS ends up becoming horizontal thin layers on top of core infrastructure components.”
Functions every application needs, such as authentication, access control, email, and payments, can each be pulled out and sold as a company because the market is large enough, so investors should assume any sub-component of an application has the potential to become a business.
“If you're in tech at all, or you're an investor at all, I definitely think you should look at an application and assume that any sub-component does have the potential to now become a company”
Infrastructure innovation has entered a dark age because consolidation into three companies that keep their work under wraps leaves the rest of the industry with little access, and a few companies drive perhaps 50% of the infrastructure market.
“I think we've entered somewhat of a dark ages for infrastructure innovation.”
Because higher margins lift valuation through the market's multiple, cutting COGS can have an outsized effect: a company worth $10 billion could become worth $14 billion, giving it cash and debt capacity for hiring and other investment.
“if you're a company that's say, worth $10 billion, and you can reduce your COGS by a bit, you could now become worth $14 billion”
Pricing is not systematic like hiring or org design; it is dictated by the shape of the market and product, takes months to get right, and is often fixed by the ecosystem, so companies building on a platform like Snowflake should align with how that platform charges.
“one of the things that's just not systematic is pricing. Pricing is actually dictated by the shape of the market and the shape of the product.”
Security is becoming an API and an intrinsic primitive of application building, with authentication, policy, and data integrity increasingly stitched into programs, rather than being an afterthought bolted on as a box watching everything else.
“security itself is also becoming an API, like we had talked about, and it's becoming something that's endemic to the building of applications.”
Infrastructure businesses tend to have lower margins than software businesses, and AI businesses look more like services businesses, so Casado is skeptical that infrastructure companies will look like software companies and is more conservative about their profile.
“I tend to be the most skeptical that infrastructure businesses look like software businesses. AI businesses tend to look more like services businesses.”
Startups no longer compete mainly on software ability because basic infrastructure is solved; in most cases, differentiation between companies now depends on how they use data for matching, dynamic pricing, personalization, and acquisition.
“We've moved as an industry to the businesses competing on how they use data, not how well they use software.”
Companies pass through product (finding product-market fit), sales/growth (repeatable sales, little concern for unit economics), and operations (caring about unit economics across multiple products); the cloud went through the same three phases, and it is now in the operations phase where companies rationalize spend.
“The cloud had gone through the exact same three phases, which first was trying to find product market fit”
Applications are priced per seat while infrastructure is priced per usage, and frontend-facing companies that want to succeed must move to usage pricing if they can escape seat pricing, because that is where the model naturally ends up.
“apps are for seat pricing, and infrastructure is usage pricing, and that's basically how it is.”
In early-stage investing, free-form discussion matters more than rehearsed pitches, and Casado spends much time on founders' past experiences and lessons learned because he cares less about right answers than about self-awareness and appreciation of the sophistication of the problems being solved.
“Early on, I think free form discussion is just incredibly important because anything that's rehearsed is just not an indication of reality.”
Normally people evaluate bringing workloads in-house by comparing the savings with the cost of the team and complexity, and conclude it is not worth it. Casado argues this misses the leverage that margin improvement has on stock price, which can free up substantial cash for large companies.
“Normally when people look at this problem, they say, "Well, if you bring this inside, yes, it'll save you money... but that money won't cover the team, the complexity" But if you look at the leverage that increase in margin does to the stock price, now you can free up for a large company, potentially a lot of money”
Companies like Mighty (browser as a service), Fly (compute at the CDN tier), and Mosaic (ML as a service) are led by strong repeat founders, build their own hardware, and succeed by targeting specific workloads the cloud is not optimized for.
“what's unique about all three of these companies is all of them are doing their own hardware.”
For hosted services, what matters is open standards rather than open source, because developers still want to avoid lock-in and move between providers, which is an open standards argument rather than an open source one.
“what raises importance is actually open standards, which is, I still want be able to make sure that I'm not locked in to one”
As cloud users enter the operations phase and look for benefits outside the big three, innovation in compute, network, and storage is re-emerging and being exposed to the broader industry, restarting a cycle of innovation.
“Users of the cloud are starting to evaluate, "Where can I get benefits that are outside of the cloud?"”
The market tells you the price over time rather than upfront, so founders should avoid forcing their own pricing opinion and use early sales conversations to reverse-engineer how customers think about price; per-API-call units give room to experiment.
“the market will tell you the price over time, but not initially. The less that you have public or the less that you force your opinion on, the better it is.”
Even if crypto networks succeed, traditional switches, routers, storage and databases will still be needed for a large subset of workloads, with crypto handling embarrassingly parallel portions, and crypto companies' purchases of traditional infrastructure show that adoption of traditional infrastructure continues.
“I believe that you're still going to have traditional switches and routers to make the internet. And then I believe that you're going to still have traditional storage layers”
Multiple companies that are household names have product lines with 0% margins because all revenue flows back to the cloud services hosting them, and many infrastructure companies feel they are reselling a thin layer on top of AWS or another big cloud.
“I know multiple companies that are household names, where they've got product lines that have 0% margins because all of the money goes back to the cloud services it's hosted on.”
An average application, such as a mobile app, uses about 17 external APIs, and the industry is still early in outsourcing components, with a long way to go as applications become more complex.
“on average, an application uses 17 external APIs.”
There are roughly 100 frontend developers for every backend developer, and frontend developers are consuming more of the stack, which atrophies backend work and shifts both the type of person doing frontend work and the infrastructure supporting it.
“there's probably, I don't know, 100 frontend developers for every backend developer.”
The modern data stack divides into analytics tools that support human decisions through dashboards and AI/ML tools that embed model decisions in real-time applications, and these two sides are now converging.
“There's the analytics side, which is generally... Helps with human decisions, having a lot of data and doing dashboard. And then there's the AI/ML side”
The next wave is not many Dropbox-style end customers building their own data centers; it is third-party companies providing cloud services at better price or performance, and growing cloud size makes it viable for those solvent companies to enter.
“I don't think the answer is we're going to see a lot of drop boxes, where the end customer builds their own data center.”
The primary narrative of the last decade is the cloud, with an oligopoly of three providers (Microsoft, Google, AWS) now supplying compute, storage, networking, and higher-level services that central IT teams once provided, so builders no longer need deep infrastructure knowledge.
“You have basically an oligopoly of three providers, Microsoft, Google, and AWS. They provide all of IT, which used to be provided by central IT teams.”
Ten years ago there was broad consensus that open source was great and would take over the world, but the industry now has a more nuanced view: useful for developer tools, less so for services.
“Whereas 10 years ago, there was this broad consensus that open source is great and it's going to take over the world.”
Over the last five years, higher-level services such as machine learning workflows, analytics pipelines, and specialized databases for particular query patterns have been built on cloud primitives, producing a renaissance of infrastructure that moves much faster than when infrastructure was tied to a box.
“a number of services then built on top of those that are higher level abstractions. So for example, machine learning workflows, analytic workflows, different types of databases”
Companies whose products were APIs for developers, such as PubNub, SendGrid, and Twilio, once had billboards along Highway 101, and this shows technical functions useful only to developers can now form standalone businesses, which is exciting for infrastructure investors whose founders build such tools.
“you'd have billboards where the entire company was an API. PubNub, Sendgrid, you know, Twilio.”
Traditional infrastructure components (servers from Dell, IBM, or HP, switches from Juniper or Cisco, storage arrays from EMC, and databases from Oracle) have collapsed into software layers running on merchant hardware in the cloud, letting users buy compute, flexible networking, and scalable database services without owning boxes.
“So all those have now been, basically, collapsed into a software layer over basically merchant hardware in the cloud.”
Any problem humans pursue outside software is now coming within software's reach, including agriculture and oceanography, so every industry will be touched by technology, raising the open question whether such companies remain software businesses or become something new.
“Any problem that human beings go after that's been outside of the realm of software is currently in the realm of software.”
The hardest part of being a founder is managing one's own psychology, and the key early screen is whether a founder is optimistic enough to push through hard times yet self-aware and realistic enough to make practical decisions.
“is the person tackling this optimistic enough to push you the hard times, but self-aware enough to know and realistic enough to know to make very practical decisions.”
Developers have growing budgets, but traditional sales and marketing techniques do not work on them; they are opinionated, form factions around technologies, and the best developer companies build communities that make them sensitive to developers' problems, workflows, and aesthetics.
“traditional sales and marketing technique don't work for developers.”
Casado's major life inflections, from PhD through being invested in and hired, came from people taking big bets on him, which he attributes to others' kindness rather than his own skill.
“my entire life has just been people taking big bets on me.”
Internal access policy, such as who can access what, is a specific computer science problem that could be pulled out of programs as a standalone policy language or service that lets third parties declare and enforce access rules, and several companies are just beginning to pursue it.
“How do you build a language or a policy language that kind of accesses, that allows a third party to declare a set of rules”
Hugging Face looked like a rocket ship in its usage early on, with many theories about why it could not monetize its models or sell effectively, yet the market loved it and it became a great company.
“Hugging Face is a phenomenal company. And if you looked early on at the usage, this thing is a rocket ship”
Patrick frames Dropbox as a case where cloud has real benefits in speed and outsourced reliability but can become very costly, which is the context for the owned-infrastructure reversal.
“cloud isn't just some panacea. It has incredible benefits in terms of how quickly you can get going”