Databricks’ Co-Founder Arsalan Tavakoli: Every Software Monopoly Falls in the Next 24 Months

Databricks is now running at a stunning $6.9B revenue run-rate, growing 80%+ year over year, with AI products alone past a $1.7B run-rate and net retention above 140%. That gives co-founder Arsalan Tavakoli-Shiraji better visibility into what enterprises are actually doing with AI budget than almost anyone. On stage at SaaStr AI, he laid out where that money is going, where it isn’t, and one claim that should change how every B2B exec thinks about competition:

Any business with a monopoly today will not have a monopoly 12 to 24 months from now.

It’s in some ways a stunning claim from a company that spent 13 years … building a deep moat in data and AI. But we’re all seeing it now, and it has direct consequences for both pricing power and competitive risk.

The summary:

• Three forces broke pricing power at once: build costs, the low end, and migrations
• Everyone is token maxing, and almost no one knows their true AI ROI
• The real bottleneck on enterprise AI is context, not the model
• Traditional BI is basically dead

Our deep dive:

Everyone Is “Token Maxing.” Almost No Enterprise Knows Their True AI ROI Today.

If you live on X, you’d think many enterprises have AI figured out. They’ve all built their own LLM research agents. They’re all automating everything. The reality on the front lines is nothing like that.

Every CEO has now told their org the same thing: if we’re not using AI, we’re behind. Go use tokens. We’ll measure your performance by it. Employees are doing exactly that, and token spend is going straight up.

The problem is what comes right after. Spend is climbing, and most leaders have no idea what they’re getting back for it. Tavakoli’s framing: everyone’s token maxing, spend is going up, and there’s no clear read on the output. That’s the actual state of enterprise AI in 2026. Not “we figured it out.” More like “we’re spending heavily and trying to find the outcome.”

For founders, that’s the opening. The companies that tie AI spend to a clear business outcome win budget right now. Standing up agents earns zero points. Driving a number does.

Data Stopped Being a Warehouse Decision. It Became a Top-Line One.

A few years ago the pitch to a CIO was straightforward: bring your data into a lake, replace the warehouse, get better analytics. Databricks used to whisper the AI part, because AI made buyers think of self-driving cars and robots.

That urgency profile has changed completely. AI is now a top-line imperative, not a back-office efficiency play. And the second enterprises commit to it, they hit the same wall: data silos, no semantic layer, no context. The hard part isn’t the model. It’s getting data clean, governed, and accessible to agents rather than to humans.

This is what most people miss. The bottleneck on enterprise AI isn’t model quality. It’s context. And context is not the same thing as data.

Context Is the Real Bottleneck, and It Goes Stale Fast

Think about onboarding a new employee. How do you explain everything that happens in your org so they can actually operate? That is what an agent needs, and almost no company has it written down.

Take a simple question: show me my top spenders on the major clouds at the end of last fiscal quarter in EMEA. Sounds trivial. But what counts as a “cloud”? What’s a “top spender”? When is the fiscal quarter? Which countries are in EMEA for this business? Every one of those is a definition someone learned by asking a colleague years ago. Multiply that across a 100,000-person org. Those definitions live buried in emails, meeting transcripts, and call notes, and they change constantly.

Most companies tried to solve this statically by writing a context doc. The doc is stale the day after it’s written. Point someone to a context document from two years ago and it’s already wrong. The hard part isn’t writing context once. It’s pulling in new information and deprecating the old continuously.

This is why Databricks built Genie Ontology, a self-improving context layer that extracts and continuously updates business knowledge from files, tickets, chats, and meetings. The takeaway for anyone building agents for the enterprise: the durable value isn’t in the agent. It’s in maintaining live context.

Traditional BI Is Basically Dead

Standalone BI is a dashboard graveyard. A handful of long-query dashboards nobody looks at, built by the 5% of an org that can actually write a query, with a one-week turnaround on every new question.

Databricks’ Genie flips that. The proof point from the session: a car manufacturer just loaded 70,000 users onto it. Not the 5% who can write SQL. The 95% who run the business and know which questions matter. They ask their own questions and get answers in 30 seconds instead of waiting a week for a data analyst.

That changes behavior. Someone has a question, fires it off mid-meeting, the answer comes back, and it shifts the decision in real time. And nobody ever has just one question. An answer leads to a follow-up, which leads to going deeper. Tavakoli noted that people are actually more stressed in the AI era because utilization went up. When you can always get the next answer, you keep pushing.

The old BI tools struggled because they had no semantic understanding of the data. Bolting “talk to your data” on top just meant converting text to SQL, which doesn’t work. You need the layer that interprets what’s being asked and maps it to what the data means. Real-time visibility into every byte in the org, for every employee, is now the expectation. BI as a category disappears into dashboards and answers.

Why No Monopoly Survives: The Mechanics

Three things are happening at once, and together they break pricing power for incumbents.

**One: the cost of building software collapsed.**When everything was a monolithic stack, building one and convincing an enterprise to adopt it was brutally hard. Now a new entrant can walk into an org that already has its data ingested and governed, build on top of that, and ship something credible fast. More builders, more competitors, in every valuable category.**Two: the low end got good.**The old low-end product was cheap and crappy. It did one workflow, badly, but it technically worked. Now AI makes those same low-end products great, especially when they pull in third-party APIs. Layer Salesforce or Shopify or Databricks data on top of a lightweight app and it stops being a one-workflow toy. “You get what you pay for” is breaking down. A new entrant with nothing to lose sets price at 30-40% of the incumbent and wins on greenfield deals.**Three, and this is the one that matters most: migration cost collapsed.**Migrations used to die on the vine. A vendor would promise to save you 50%, then the migration itself cost 5x the annual savings. Nobody moved. The person who understood the legacy system retired three replacements ago.

That math is now inverted. Code is self-descriptive, so LLMs can go into a legacy environment, understand what it does, convert it, migrate the data, and write the harnesses to validate that the new output matches the old. Databricks is doing enterprise-grade migrations in 30 days or less depending on complexity. When the cost of switching drops, willingness to pilot a new vendor goes up. And once buyers will actually try and switch, no incumbent can lean on lock-in to defend its price.

The evidence was on the floor at SaaStr AI. A large share of the companies exhibiting didn’t exist a year ago and are already at non-trivial revenue. Buyers are willing to try and migrate in a way they simply weren’t before.

Lock-In Is No Longer a Viable Long Term Strategy in B2B

There’s a Cambrian explosion of AI apps happening, and it works in two directions. It’s the greatest app-creation moment in B2B history. It also means brutal competition.

If you’re attacking an incumbent, the wedge is there. Lower cost of building, a low-end product that’s actually good, and migration costs low en