On July 8, 2026, at 00:00 UTC, Databricks quietly flipped a switch that changes how enterprises will budget for AI going forward. Genie — the natural-language query and agent layer that sits across Databricks' platform, including Genie Agents, Genie Code, and the newly introduced Genie One — moved from a largely included feature to a metered, pay-as-you-go product. Each identified user still gets 150 DBUs (Databricks Units) of free LLM usage per month, worth roughly $10.50 at US East rates, but every token consumed beyond that allowance now shows up as a line item. Service principals, notably, get no free allowance at all. This lands against a backdrop of genuinely staggering scale: Databricks says more than 100,000 agents have been built on its platform, collectively processing over a quadrillion tokens a year, and the company crossed $5.4 billion in annualized recurring revenue in February 2026, growing around 65% year-over-year. The pricing change isn't a footnote. It's the first real test of whether "agents" can be billed like compute rather than sold like seats — and it's a preview of the conversation every data platform vendor is about to have with its customers.
What actually changed in Genie's pricing
Before July, Genie's usage was bundled in ways that made it easy for teams to treat it as a soft cost — something absorbed into an existing Databricks contract rather than tracked as its own spend category. The new model unbundles that. Genie Agents, Genie Code, and Genie One now share a single pay-as-you-go pricing scheme: a per-user monthly allowance of 150 DBUs, and metered billing in DBUs for everything past that, tied directly to the underlying LLM usage each query or agent run consumes. Databricks has simultaneously shipped budget controls that let account admins track and cap Genie spend at the account, workspace, group, or individual-user level, with alerting before limits are hit. That's a telling pairing — the same release that introduces metered billing also introduces the guardrails to stop that billing from running away, because Databricks clearly anticipates that it will, for at least some customers, without those guardrails in place.
The mechanics matter because Genie isn't a single feature anymore — it's a family of agent surfaces spanning conversational analytics, autonomous coding assistance, and the general-purpose Genie One agent. A data analyst asking Genie a handful of questions a day looks nothing like an autonomous Genie Agent left running against a pipeline, retrying, reasoning, and calling tools dozens of times per task. Under a flat or seat-based price, both of those users cost the same. Under DBU metering, they don't — and that's precisely the point. Databricks is betting that usage-based pricing is the only model that scales sanely once agents, not humans, are doing most of the querying.
Why seat-based pricing was never going to survive contact with agents
Enterprise software has priced by seat for decades because the unit made sense: a human logs in, does a day's work, logs out, and the marginal cost of that human using the software one more hour is close to zero. Agentic AI breaks that assumption in a way that's easy to underestimate. An autonomous agent doesn't clock out. It can be triggered by an event, spawn sub-tasks, retry on failure, and run continuously against a queue of work, with no natural ceiling tied to a person's working hours. Multiply that by the sheer number of agents an organization deploys — Databricks' own figure of 100,000-plus agents built on its platform, and over a quadrillion tokens processed annually, is the clearest evidence yet that agent volume is compounding in a way seat counts never did.
That's exactly why token-metered pricing has already become the default for LLM API providers like OpenAI, Anthropic, and Google — you pay for what you consume, not for a named user who might consume a little or a lot. Databricks and Snowflake have essentially been insulated from that dynamic until now, because their agent features rode along inside broader platform contracts priced on compute (DBUs, credits) that customers were already used to provisioning for ETL and warehousing workloads. Genie's move to pay-as-you-go is Databricks explicitly extending that consumption logic to the agent layer itself, rather than continuing to treat Genie as a value-added feature bundled into existing capacity. It's a tacit admission that agent usage is now large enough, and unpredictable enough, that it needs its own metered line rather than living inside the noise of a broader platform bill.
Snowflake is making the identical bet, from the opposite direction
The timing isn't a coincidence. Both Databricks and Snowflake made "agents as the primary unit of compute" the headline of their 2026 flagship conferences, and both are racing toward the same end state — owning the client an enterprise's AI agents run through, and the back end those agents call into — even though they're approaching it from different starting points. Snowflake rebranded Snowflake Intelligence into "CoWork" and Cortex Code into "CoCo," repositioning both as agent-native products rather than analytics add-ons. More significantly, on May 27, 2026, Snowflake announced it would acquire Natoma, a two-year-old startup that had raised a $7 million seed round from Greylock and Index Ventures, and built a centralized Model Context Protocol (MCP) gateway with more than 100 prebuilt MCP servers that handles identity, policy enforcement, and audit logging at the individual tool-call level for non-human identities. In plain terms: Natoma governs what an agent is allowed to touch and logs every action it takes against enterprise systems. Snowflake, which reported $4.68 billion in FY2026 total revenue, up 29% year-over-year, is betting that governance and connectivity — not raw model access — is the harder problem to solve once agents are let loose on real enterprise data.
That's a meaningfully different wager than Databricks is making with Genie's pricing change. Databricks is solving the metering-and-billing problem first, on the assumption that customers will trust their own governance to keep agent spend bounded. Snowflake is solving the governance-and-access problem first, on the assumption that unmetered but ungoverned agents are the bigger near-term risk. Neither company is wrong, and in practice both problems have to be solved — an agent that's cheap to run but has unrestricted access to production systems is just as dangerous as one that's tightly governed but bills unpredictably. Expect Snowflake to introduce its own more granular consumption pricing for CoWork and CoCo agent usage in the coming quarters, and expect Databricks to keep building out governance and access controls around Genie Agents to match. The two vendors are converging on the same product from opposite ends, and the gap between them is closing fast.
The FinOps problem nobody budgeted for
This is where the story stops being a vendor pricing update and becomes a genuine operational headache for IT and finance teams. Traditional software budgeting works because seat counts are knowable in advance — you know how many analysts you're licensing next quarter, so you know roughly what you'll pay. Metered agent spend doesn't work that way. An agent's token consumption scales with the complexity of the tasks it's given, the number of retries it needs, the length of context it pulls in, and how many other agents or sub-agents it spawns to get the job done — none of which map cleanly onto a headcount plan. A single Genie Agent workflow that starts modest can balloon in cost the moment someone points it at a messier dataset or a more ambiguous prompt, with no seat limit acting as a natural brake.
This is precisely the failure mode Databricks' new budget controls are trying to get ahead of, and it echoes a pattern already well understood in raw LLM API spend — where organizations have learned the hard way that token-metered costs need active monitoring, not just quarterly review, or they show up as a shock on the invoice. The difference with Genie and CoWork is that the metering is now embedded a layer higher, inside the data platform contract itself, where finance teams are far less accustomed to treating it as a variable-cost line rather than a fixed platform fee.
Practical takeaways for data and platform teams evaluating this shift
- Inventory every Genie and Cortex/CoWork surface already in use. Before the next renewal conversation, get a concrete list of who's using Genie Agents, Genie Code, and Genie One (or Snowflake's CoWork/CoCo equivalents), and whether any of that usage runs via service principals, which get no free DBU allowance under the new Databricks model.
- Turn on budget controls before turning on adoption. Databricks' new per-workspace, per-group, and per-user budget alerts exist specifically because usage can spike unpredictably; treat them as a prerequisite for rolling out any new Genie Agent workflow, not an optional add-on.
- Separate agent spend from platform spend in your cost reporting. If Genie or CoWork usage is still buried inside a broader DBU or credit consumption line, finance can't see it moving until it's already a problem; break it out as its own tracked category now.
- Model worst-case agent behavior, not average-case. Because autonomous agents can retry, loop, and spawn sub-tasks, budget for the tail scenario — a misconfigured agent running unattended over a weekend — rather than extrapolating from typical daily usage.
- Weigh governance and metering together when comparing vendors. Databricks is currently ahead on consumption-based pricing transparency for Genie; Snowflake, via Natoma, is currently ahead on tool-call-level governance for agent actions. Evaluate which gap is riskier for your organization's actual agent deployments, since neither vendor has fully closed both yet.
- Revisit agent design for cost efficiency, not just capability. Just as teams learned to optimize prompt length and model selection to control raw inference costs, expect a similar discipline to emerge around agent design — fewer redundant tool calls, tighter context windows, and clearer stopping conditions — now that agent runs carry a visible metered price tag inside the data platform itself.
None of this means agentic features are becoming prohibitively expensive — 150 free DBUs per user per month is a genuinely generous allowance for moderate usage, and Databricks' budget tooling is a legitimate answer to the FinOps risk it's creating. What it means is that the free ride is over for treating agent usage as an assumed, invisible part of a data platform contract. Databricks and Snowflake are both signaling, through pricing and through acquisitions, that agents are now the primary unit of value they sell — and primary units of value get metered, monitored, and eventually renegotiated at every contract cycle from here on out. The vendor conversation that used to be about seats and storage tiers is turning into one about tokens, tool calls, and who's watching the meter. Teams that get ahead of that shift now, before their first surprise Genie invoice, will be in a far better negotiating position than those who wait for finance to ask why the AI line item tripled.