Models Can Be Rented, Memory Cannot

Every company can rent the same frontier models. The model itself is no longer a barrier.

Then what is?

Ben Horowitz’s answer in his latest long-form article: a company’s own memory. Definitions, permissions, evaluated cases, customer history, and hard-won exceptions. The model will not accumulate these for you.

He calls this a system of record for agents.

For an Agent to be trusted with real work, this system of record must answer four questions:

What is true: the company’s own data.

What it means: how the business defines things. Databricks’ tests show that anchoring an Agent’s reasoning in the company’s ontology (a business concept graph) raised accuracy from about 50% to 84.5%.

What is allowed: identity, budget, guardrails.

Whether it works: scoring against the company’s own definition of “good.”

Every correction feeds back into the record. Even if the model changes, what the company has learned remains. In Ben Horowitz’s words:

“If you replace your company’s AI model and the new model inherits your definitions, permissions, evaluated cases, customer history, and hard-won exceptions, that is the foundation of a sustainable competitive advantage. If the new model starts from the same prompt as everyone else’s, it is not.”

This judgment points to a more fundamental change.

Coase asked in 1937: Why do firms exist? The answer is transaction costs. When the marginal cost of organizing a transaction inside the firm equals the cost of transacting in the market, the firm stops expanding.

Herbert Simon later observed that centralized business memory allows decision-making to be decentralized. An airline booking agent does not have to call every colleague; they just check the shared record. When the record changes, everyone knows.

Databases changed companies in this way. Agents will change them even more thoroughly.

Databricks is a good sample for observing this change. The company has ridden four waves of the technology S-curve: Spark, Lakehouse, Lakebase, and Agent.

In 2009, Hadoop dominated the “big data” era, but every round of computation had to write back to disk, making iterative workloads such as machine learning absurdly slow. Berkeley PhD student Matei Zaharia wrote Spark, keeping data in memory and making iteration ten times faster. In 2013, they founded Databricks.

Then came Lakehouse, combining the cheap storage of the data lake with the reliability of the data warehouse. After that came the 2025 acquisition of Neon, moving the operational database onto the same storage layer, called Lakebase. Agents can operate databases the way they operate code: open a branch, run an experiment, throw it away.

These technological breakthroughs came at just the right time. On Databricks, Agents already issue more queries than humans. More than 17 million databases are launched by Agents every day. METR tracking shows that the length of tasks Agents can complete independently has roughly doubled every seven months since 2019. If this trend continues into the early 2030s, Agents will be able to handle tasks that would take humans years to complete.

Ali Ghodsi put it directly: AI does not have an intelligence problem; it has a context problem. No matter how smart the model is, without the company’s data, permissions, and feedback, it can do nothing.

Databricks’ Genie lets business users who do not know SQL ask questions of data directly. Scientists at Novo Nordisk no longer wait for data engineers to write queries; they directly say, “Pull up all our obesity data.” Financial analysts at Rivian used Genie to build a revenue dashboard for the CFO. Natural language is becoming the new programming language. Andrej Karpathy said: “The hottest programming language is English.”

Elon Musk put it more radically: companies with “humans in the loop” will lose. He may be right, but the intermediate state may not be everyone becoming an engineer; it may be everyone becoming a manager.

Software engineers are already experiencing this shift: from writing code by hand to managing a group of coding Agents. Break down problems, assign tasks, review PRs, send them back for rewriting. That is the work of a manager.

Two banks rent the same model to make credit decisions. The first sends a prompt, receives an answer, and records the final decision. The second records which facts produced the answer, which were overridden by humans, and how those decisions performed later, and continuously adjusts.

A year later, a better model appears. Only one bank has become a better bank.

This is the logic of the “sovereign company.” Models can be swapped; memory is your own. Nvidia’s Jensen Huang said: every company’s business data is a gold mine. Nvidia has a $3 trillion market cap and 30,000 employees, relying on a large amount of AI inside the company.

Databricks’ revenue run rate has exceeded $7 billion, up 80% year over year, with a valuation of $190 billion. But more noteworthy than the numbers is the direction it is betting on: in the Agent era, a company’s competitive advantage is not the model; it is memory.

What will people do at work every day in the future? The answer may be: do the things that have not yet been figured out.