CFO roundtable: AI growth, pricing, and forecasting (June 2025 Fintech newsletter)

The dramatic difference at Databricks is that our pricing and revenue recognition are based entirely on output, unlike input-based consumption models. If customers don’t derive value, they don’t consume, and revenue doesn’t appear on our P&L.

Our pricing is built as a function of increasing profit in absolute terms, but decreasing margin in percent terms — we decrease unit prices as customer commitments increase. We automatically discount via our pricing calculator to encourage larger customer commitments. Locking customers into higher spend helps us de-risk revenue.

I changed my pricing more than seven times in the 40 days post-launch. It was really helpful to understand the market and customer’s appetite to pay. Right now, pricing is just a slide in my deck — something I’ll keep iterating on and improving.

We started annualizing the usage-based revenue and adding it to a new metric — ARR plus annualized usage — because enterprise customers exceed their quotas far more frequently. Without this, we’d be underselling or undercounting what we actually earn.

With consumption-based models, how do you think about ARR? You might have a committed relationship, but actual usage varies month-to-month, so traditional ARR definitions become challenging.

Unlike SaaS models, where revenue tends to be linear, consumption models are inherently nonlinear — customers surge, then optimize. We manage this volatility by focusing heavily on diversification to avoid customer concentration. … We track contract values closely, and use AI to help understand and forecast our true consumption-based ARR.

We closely monitor infrastructure spend, and if costs grow faster than usage, we quickly send engineers to optimize it — there’s a continuous cycle of managing cost efficiency.

We triangulate pricing decisions through customer value proposition, competitive benchmarking, and cost and return analysis. Given how quickly AI infrastructure and software is evolving, we continually re-evaluate. … Creative pricing and packaging happens quite often based on customer need, term, and deal size — so long as we’re carefully considering the unit economics.

You have to watch costs closely — these GPU costs are significant. We track regrettable idle GPU time as utilization loss, which directly impacts margin and efficiency. Every hour we have GPUs not being used by customers impacts our margins.

We have a human-in-the-loop (HILT) team that’s part of COGS. As algorithms improve, effective adjudications per human go up and unit costs come down, but we still have to bias toward false positives to manage risk.

Research projects might not directly translate to immediate revenue, but they create significant long-term differentiation, product development, and stickiness, becoming essential in competitive markets. For example, the investment we made in research around kernels enables us to now provide a unique differentiation to our customers to reduce infrastructure costs and increase performance.

Pure text-to-speech will inevitably become commoditized. To maintain long-term defensibility, we need sophisticated product layers — workflows, data-rich features, and APIs — so customers become deeply embedded and find switching difficult.

It can be a challenge to plan even 12 months out in AI right now. There’s constant innovation, and new use cases appear quickly. You have to remain agile and factor in change as part of your risk management strategy. … The sure constant in AI is change. Models evolve rapidly, and inference use cases we haven’t even considered today will be critical in a year.

We use Databricks itself — AI, machine learning and advanced analytics — to forecast consumption patterns at the customer, workload, and product levels. This is critical not just for financial forecasting, but also for accurately setting quotas for our large sales team. … You can’t achieve the precision required in consumption forecasting using Microsoft Excel — you have to use advanced analytics, machine learning, and AI to build those predictions.

We’ve got a product, Genie, which is basically a natural language query. So you can type in natural language into your data lake. It extracts answers. Genie will understand your data. And then it learns and understands your data more and more and more, the more you use it.

I don’t think anyone has fully cracked forecasting revenue for AI. The market is booming and changing so fast that reliable forecasting feels like more of a sanity check than precise predictions.