Databricks - Value $190 billion - $7 billion in revenue - if you asked 100 random people in the US what the company did, you might get one or two who have even heard of the company.
Because I am a technological illiterate, even after reading this well written article, i'm still not sure what Databricks products are, but the results seem to be effeciency and the ability of companies in all fields to create their own AI programs while keeping their own data and IP in house.
Anyway, so glad there are such smart, talented people in America and the private funds to back them.
While “transaction costs” could have been eliminated by the Internet and now AI once data is centralized, the costs are in fact increasing. The computers were taught to create even more friction in demand pricing, reduced transparency, futures contracts, etc.
The future reality is Databricks and others take your data, sometimes embellish it with more, say a medical record and require you to pay to get it back.
The proof is the massive multi trillion dollar investments. They want their money back with interest. And the interest is usurious.
Electrification is the closest precedent for the 'big if' here: factories ran electric motors on the old steam-era driveshaft layout for decades before anyone redesigned the floor plan around distributed power, and output barely moved until they did. Agentic automation is the capability; reorganising decision rights is the adoption problem, and the two move on very different timelines.
If 'everyone becomes a manager' is where this lands, that capacity doesn't distribute itself, it flows through whatever authority structure the firm already has. A flat organisation and a hierarchical one inherit the same agent capability very differently. What happens to the firms that had the fewest layers of hierarchy left to flatten?
Companies like Databricks will prove to be just transitional until humans are fully out of the loop and AI does all the work that needs doing. The question people should be considering is how long will that transition take? 2 years? 4? 10?
At some point when AI does everything, there will no longer be any need for human driven economic systems and hierarchies. Money, debt, GDP, profit, CEO's, inflation, etc., all go out the window as archaic concepts not needed by Ai and its robot workers, who can work 24 x 365 w/o the need for rest, vacations, pay, rest, sick time, drama, etc.
At that point, there will no longer be trillionaires, billionaires, millionaires. All humans will be exactly equal in status and power and will be treated equally by the AI in charge.
Moving authority closer to the work raises a practical question: does the time and ability to exercise it move too?
I would follow one real exception. Who checks the agent’s output? Who can question its assumptions? What existing work makes room for that checking?
Then compare waiting, rework and overtime across the teams involved. Through TVE, we’re examining whether the whole arrangement improves—or whether the difficulty simply moves.
The stat that more databases are now launched by agents than by people is striking. For a small team starting with automation, do you think the first step is picking the tool or cleaning up where the data lives?
The four questions are not the same record. "What does it mean" is OntoRank over who created a definition, how widely it is used, and how recently it changed, and the piece says grounding there moved answers from about 50% to 84.5%. "What is allowed" is a different store: Unity Gateway gives each agent an identity, a budget, and guardrails because a login per employee breaks when people spin up hundreds of agents a day. I would version the definition the agent used separately from the grant it held. An answer can be grounded in a revenue formula that has since changed and still be an action the current guardrail no longer permits, and a single trace that only says the agent succeeded will not show which of those two moved.
More queries from agents than people is a wild stat. It also changes who the "user" of a data platform is. We spent years designing for analysts, and now the main consumer can't lean over and ask which table is the real one. All that implicit knowledge has to become explicit metadata, and most platforms aren't close.
The funny part is that "everyone becomes a manager" only works after the company writes down everything its good managers currently know without writing down. We may automate middle management by first turning tacit knowledge into the world’s largest onboarding document.
The two-bank example is where this argument turns from history into an investment question. Two banks rent the same frontier model to make credit decisions; the first sends its data to the model, the second is a 'sovereign' firm that owns its data and its loop. The model is the same, so the model cannot be the moat — the moat has to be the ontology, the evaluations and the permission structure built on top of the data. But that raises the question the piece does not quite answer: sovereign against whom? The second bank is sovereign against the model provider, because it can swap models when prices rise. It is not sovereign against the system that holds its ontology. If the definition of what is true, what is allowed and what worked all lives inside Lakehouse and Lakebase, then the switching cost has not disappeared; it has moved from the model to the platform. That is the Coase point applied one level up: the firm exists where internal coordination is cheaper than the market, and the vendor that owns the coordination layer gets to set the price of leaving. So my question is where the rent actually gets captured in this world. If agents outnumber people a thousand to one and the model is a commodity, does the value accrue to the system of record, to the firm that owns the best loop, or to nobody — because a loop that any firm can buy from the same vendor is itself a commodity? The second bank looks sovereign until a third bank buys the same loop.
Databricks - Value $190 billion - $7 billion in revenue - if you asked 100 random people in the US what the company did, you might get one or two who have even heard of the company.
Because I am a technological illiterate, even after reading this well written article, i'm still not sure what Databricks products are, but the results seem to be effeciency and the ability of companies in all fields to create their own AI programs while keeping their own data and IP in house.
Anyway, so glad there are such smart, talented people in America and the private funds to back them.
While “transaction costs” could have been eliminated by the Internet and now AI once data is centralized, the costs are in fact increasing. The computers were taught to create even more friction in demand pricing, reduced transparency, futures contracts, etc.
The future reality is Databricks and others take your data, sometimes embellish it with more, say a medical record and require you to pay to get it back.
The proof is the massive multi trillion dollar investments. They want their money back with interest. And the interest is usurious.
Electrification is the closest precedent for the 'big if' here: factories ran electric motors on the old steam-era driveshaft layout for decades before anyone redesigned the floor plan around distributed power, and output barely moved until they did. Agentic automation is the capability; reorganising decision rights is the adoption problem, and the two move on very different timelines.
If 'everyone becomes a manager' is where this lands, that capacity doesn't distribute itself, it flows through whatever authority structure the firm already has. A flat organisation and a hierarchical one inherit the same agent capability very differently. What happens to the firms that had the fewest layers of hierarchy left to flatten?
Companies like Databricks will prove to be just transitional until humans are fully out of the loop and AI does all the work that needs doing. The question people should be considering is how long will that transition take? 2 years? 4? 10?
At some point when AI does everything, there will no longer be any need for human driven economic systems and hierarchies. Money, debt, GDP, profit, CEO's, inflation, etc., all go out the window as archaic concepts not needed by Ai and its robot workers, who can work 24 x 365 w/o the need for rest, vacations, pay, rest, sick time, drama, etc.
At that point, there will no longer be trillionaires, billionaires, millionaires. All humans will be exactly equal in status and power and will be treated equally by the AI in charge.
Ben Horowitz (https://substack.com/@bhorowitz) and Michael McGuiness (https://substack.com/@michaelmcguiness),
Moving authority closer to the work raises a practical question: does the time and ability to exercise it move too?
I would follow one real exception. Who checks the agent’s output? Who can question its assumptions? What existing work makes room for that checking?
Then compare waiting, rework and overtime across the teams involved. Through TVE, we’re examining whether the whole arrangement improves—or whether the difficulty simply moves.
The stat that more databases are now launched by agents than by people is striking. For a small team starting with automation, do you think the first step is picking the tool or cleaning up where the data lives?
The four questions are not the same record. "What does it mean" is OntoRank over who created a definition, how widely it is used, and how recently it changed, and the piece says grounding there moved answers from about 50% to 84.5%. "What is allowed" is a different store: Unity Gateway gives each agent an identity, a budget, and guardrails because a login per employee breaks when people spin up hundreds of agents a day. I would version the definition the agent used separately from the grant it held. An answer can be grounded in a revenue formula that has since changed and still be an action the current guardrail no longer permits, and a single trace that only says the agent succeeded will not show which of those two moved.
More queries from agents than people is a wild stat. It also changes who the "user" of a data platform is. We spent years designing for analysts, and now the main consumer can't lean over and ask which table is the real one. All that implicit knowledge has to become explicit metadata, and most platforms aren't close.
The funny part is that "everyone becomes a manager" only works after the company writes down everything its good managers currently know without writing down. We may automate middle management by first turning tacit knowledge into the world’s largest onboarding document.
The two-bank example is where this argument turns from history into an investment question. Two banks rent the same frontier model to make credit decisions; the first sends its data to the model, the second is a 'sovereign' firm that owns its data and its loop. The model is the same, so the model cannot be the moat — the moat has to be the ontology, the evaluations and the permission structure built on top of the data. But that raises the question the piece does not quite answer: sovereign against whom? The second bank is sovereign against the model provider, because it can swap models when prices rise. It is not sovereign against the system that holds its ontology. If the definition of what is true, what is allowed and what worked all lives inside Lakehouse and Lakebase, then the switching cost has not disappeared; it has moved from the model to the platform. That is the Coase point applied one level up: the firm exists where internal coordination is cheaper than the market, and the vendor that owns the coordination layer gets to set the price of leaving. So my question is where the rent actually gets captured in this world. If agents outnumber people a thousand to one and the model is a commodity, does the value accrue to the system of record, to the firm that owns the best loop, or to nobody — because a loop that any firm can buy from the same vendor is itself a commodity? The second bank looks sovereign until a third bank buys the same loop.