a16z Partner Martin Casado: Amazon Web Services Started with Machines, Storage, and Queries, Then Moved to Higher-Level Services
Andreessen Horowitz partner Martin Casado wrote in a debate about whether to build applications upwards or infrastructure downwards: Amazon Web Services started with machines, storage, and queries, then moved to higher-level services. He also listed exceptions: Microsoft, AWS, Oracle, SAP, Databricks, and Palo Alto Networks; once in an oligopoly, moving upwards becomes possible, which was nearly impossible before.
The debate was initiated by entrepreneur Ivan Burazin. He believes that OpenAI's foray into vertical industries like law, finance, and healthcare will fail, and that Anthropic's Claude Design is rarely mentioned anymore; moving upwards is extremely difficult as each vertical industry has its default incumbents, and the switching cost is social habits rather than price or functionality. He then added that almost no company has succeeded by moving upwards; successful ones have all moved downwards; when moving downwards, you are the customer, understanding constraints, primitives, and physical limits, allowing for deep exploration; when moving upwards, you are a visitor, unaware of workflows, politics, switching costs, and tastes, which leads to failure.
Casado used the history of cloud platforms as a counterexample. AWS launched its simple storage service S3 in 2006, with objects, buckets, and keys, starting at about $0.15 per GB per month; months later, it introduced Elastic Compute Cloud EC2, followed by block storage, databases, and over 200 services, evolving from "renting a machine and storing a file" to deploying agents capable of reasoning, planning, and executing tasks. The original words from Andy Jassy's team were: first do the undifferentiated hard work of setting up data centers, fixing hard drives, and networking, then let developers request servers, storage, and databases. Storage was designed as building blocks for subsequent services, rather than a monolithic solution that does everything.
Casado himself started with software-defined networking, co-founding Nicira, which was sold to VMware for about $1.3 billion in 2012, later leading its networking and security business, and then joining a16z to invest in digital infrastructure. His caveat is that moving upwards is not absolutely prohibited, but one must first become a platform oligopoly. Microsoft moved from operating systems into office and cloud, Oracle and SAP horizontally consumed modules from databases and ERP, Databricks moved upwards from data lakes to intelligence and governance, and firewall companies are moving towards platforms; the path is always to lock in the lower layer before selling the upper layer.
In market mechanisms, this is a territorial division between model laboratories and industry software. Buyers are vertical software and system integrators in law firms and hospitals who fear being replaced by generic chat windows; sellers need to prove they can also create industry suite model companies. Capital is being repriced between "platform premiums" and "industry defaults." Beneficiaries are incumbents who already have workflows and compliance licenses, as well as cloud and data platforms identified by Casado as oligopolies; those under pressure are laboratory products that rely solely on clever models without foundational primitives or industry default positions. The social switching cost is written as a moat harder than a feature list: the stack pulls down from where the user begins to express intent.
Source: Public Information
ABAB AI Insight
Casado articulates the intuition of infrastructure investors as a prohibition with exceptions. The prohibition is: visitors cannot do others' workflows well. The exceptions are: you are already the default supplier of operating systems, cloud, databases, or data lakes. The historical facts of AWS support the exceptions—first sell the simplest primitives, use the primitive network to lock customers into accounts, identities, billing, and data gravity, then layer on SageMaker, Bedrock, and agents. This is not parachuting into hospitals from a chat window, but growing from cabinets that already host the counterpart's data.
The capital path is to first become the customer's cost center, then become the customer's product center. S3 and EC2 replace capital expenditures; higher-level services replace developer hours. Once billing, logs, and objects are in the same account, the social and engineering costs of switching clouds overlap. If OpenAI and Anthropic do not have equivalent data gravity, vertical suites are merely demonstrations. Databricks is mentioned because it has grown from the data plane to governance and intelligence, repeating the upward exploration of the cloud, rather than moving backwards from models. Palo Alto is moving towards platformization of security controls. The longer the list, the more it indicates that "moving upwards" is an oligopoly privilege, not a default strategy for startups.
Analogies include IBM moving from mainframes into software, Microsoft moving from Windows into Office and Azure, and Cisco moving from boxes into software subscriptions. Failure examples include countless hardware vendors failing to build app stores and search engines failing to build social networks. The industry stage is control: the model layer is still competing for entry points, while cloud and system record keepers define the entry point as high-end SKUs within their own accounts. Burazin says moving upwards will lead to failure, while Casado says to first become an oligopoly before moving upwards. Combined, these two statements suggest that startups should continue to dig down for primitives, while platform companies write applications as self-operated shelves.
The structural judgment is about the reconstruction of the industrial chain. The mechanism is that value shifts upwards along "who owns the default account" rather than horizontally along "who owns the strongest model." Pricing power lies with those who have turned machines, storage, and queries into bills. For models to move upwards, they must first become that kind of bill.
ABAB News · Cognitive Laws
First become the customer's utility bill, then you qualify to sell the customer's workflow
Moving upwards to create applications is an oligopoly privilege; visitors will only fail on taste
Once switching costs become social habits, feature lists no longer dictate pricing