Databricks’ Median ARR is Falling. It’s Great News for Databricks

Last weekend, I was reading Databricks’ latest funding announcement when a few numbers caught my attention:

• More than 1,000 customers are spending over $1 million a year. 
• More than 100 are spending over $10 million.

I kept looking at those numbers and wondering: what does the typical Databricks customer actually look like? How does that compare to Snowflake?

When people hear “typical customer”, they often immediately think ‘average’. But this can be skewed by a small number of very large accounts. A better measurement would be the median revenue per customer.  The only problem is neither Databricks nor Snowflake release this number. It was time to dust off my statistical modeling muscles and try to infer the underlying distribution from available public data.

Where I got the numbers

Snowflake is relatively straightforward to analyze because it is public and regularly reports financial and customer metrics:

Databricks is more difficult to analyze because it does not publish a consistent quarterly customer-count series. Instead, it periodically discloses customer milestones alongside revenue announcements. That meant I could not simply line up revenue and customer figures quarter by quarter. I first needed a reasonable way to estimate the customer base between those disclosed milestones.

To do that, I collected the available customer-count observations and fit a single exponential growth curve across them:

Customers(t) = α × exp(βt)

The purpose was not to claim an exact customer count for every quarter, but to create a consistent estimate of Databricks’ customer base on the specific dates when the company disclosed revenue and large-customer metrics. This gave me the denominator I needed to estimate average revenue per customer over time.

The next step was to move beyond the average and try to understand the shape of the customer base itself. Both Snowflake and Databricks disclose the number of customers generating more than $1 million and $10 million in annual revenue.

Those disclosures are useful because they reveal something about the upper tail of the customer distribution. I therefore modeled revenue per customer using a lognormal distribution. This is a useful approximation for enterprise software, where a large population of relatively modest customers can coexist with a much smaller number of accounts spending millions or tens of millions of dollars annually.

For each period, I combined the estimated average revenue per customer with the proportion of customers above $1 million to estimate the spread of the distribution and its implied median. For Databricks’ more recent disclosures, the availability of both the $1 million and $10 million thresholds provided two observations from the upper tail, allowing the distribution to be constrained more tightly.

I then repeated the same calculation across each available reporting period. The result was an estimated median revenue-per-customer series that could be compared over time and, importantly, compared with Snowflake using the same general framework.

What immediately jumps out is that Snowflake’s median is nearly double Databricks’. Great news for Snowflake, right? Not quite. 

Databricks widens the base while raising the ceiling

The most important Databricks signal is not the declining median by itself. It is the declining median combined with continued growth in the company’s largest customer cohorts.

If Databricks were simply moving toward smaller customers, the enterprise tail would not be strengthening at the same time. Instead, Databricks continues to produce more customers above $1 million and $10 million, increasing their average deal size,  while adding enough customers below the midpoint to move the estimated median lower.

Snowflake’s rising median and Databricks’ falling median describe fundamentally different customer-distribution dynamics. 

• Snowflake is moving the center of its customer distribution upward. 
• Databricks is pushing the enterprise end of the distribution higher while simultaneously expanding the number of customers below the midpoint.

The strategic advantage is the size of the future expansion pool. Every new customer that enters Databricks becomes another potential large customer later. As some customers mature into million-dollar accounts, another cohort enters behind them. The continued growth of the $1 million and $10 million customers demonstrates that Databricks is already building increasingly large relationships at the top while the declining median indicates continued expansion at the other end.

Databricks is capturing more of the market

The Snowflake versus Databricks comparison is therefore about more than who has the larger customers or the higher headline growth rate. It is about how broadly each platform is capturing the market.

Databricks is not simply winning bigger customers. It is creating more customers that can become bigger customers.

That is the stronger growth model.