WIP: Tokenmaxxing Blog Post

We’ve all cracked a joke or two in the last month about the idea “tokenmaxxing,” but behind the concept is a real trend that has almost every technical leader scratching their heads. 

How much should an engineer spend on tokens? Until we have seen a string of quarters showing the results of using AI in software development, the “token torque” (ROI on token spend) will be hard to determine. We thought a first step would be to quantify the adoption, i.e.: even if we can’t yet measure the output, we can at least measure the input.

To that end, we reached out to our portfolio companies and friends in our network. We managed to collect a modest sample of 38 companies, who answered  2 questions:

• How much do you spend on code generation? This should include seats and tokens (e.g., Claude Code, Cursor), but exclude any consumption attributable to serving the product (e.g., in-product chatbot).
• How many people work in the R&D organization?

Our goal was to determine the average spend per software engineer, interpreted broadly. Note: for a lot of organizations, it was worth including teams like data science or even product management in the denominator.

The headline numbers from the results are:

• The average spend was $1,040 per engineer per month, and the median spend was $577 (so about half the average). The difference between median and average is explained by the extreme right skew of the distribution, as you can see further below.
• After taxes, etc. cash payroll ends up being about $25,000/month per engineer. Under that assumption, average token spend represents 4.2% of payroll and median spend is 2.3%.

It is worth looking at the distribution in a bit more detail, first in percentile terms:

For the visual folks, we have a visual representation of what looks like a normal-ish distribution under a log scale:

The log scale plot doesn’t really reflect visually the distribution, and although the data gets a little thing to have a very smooth curve when you use a natural scale, you can get a sense of the aforementioned right skew in a density plot:

We did not systematically collect data on harnesses (perhaps for next time!) but we got some anecdotal commentary from companies as part of their response:

• The bulk of the spend was split just about evenly between Claude and Cursor.
• Codex (OpenAI) usage was material in some companies, but not all.
• GitHub Copilot and Gemini (Google) use was rare, though we suspect this in particular may not reflect the broader enterprise market.
• Some companies were supportive of bring-your-own-codegen, and there were a lot more vendors in use across the employee base.

Averaging over $1,000 per engineer per month is a remarkable outcome for such a short amount of time. Typical developer-facing products like Figma or GitHub have historically priced around $20/mo./seat, and this represents a 50x ARPU increase so far. At the same time, in a top down analysis, the whisper ARR figures for Anthropic and OpenAI square well with our findings and further confirm the speed at which companies have adopted this technology.

We aim to track this data closely going forward to build out a sense of what best practices are, especially as the token torque becomes more measurable. We will publish updates regularly, so stay tuned!