Are Tokens The New Headcount? How CTOs Win Resource Allocation In 2027
Adam Cohen - founder, CEO at weave.
gettyFor the last decade, the hardest question for CTOs has been around human capital allocation. How many engineers do we need? How senior? Which product gets the team? We saw industries built to answer those questions: project management software with baked-in analytics such as Jira (2002), Asana (2008) and Linear (2019) as well as financial planning and analysis (FP&A) tools such as NetSuite and Anaplan.
Now, we’re moving to the token-based economy, and CTOs are scrambling to adjust. They now have to answer an even harder question: How can you allocate tokens?
Our research found that between April and September 2026, AI cost per employee rose 5.3 times from $109 a month to $575. Code output per employee (which is measured using a normalized unit of engineering work and scored for every merged pull request as the time a high-performing engineer would need to make the change without AI) grew 2.3 times from 56 to 131 units. The cost of each unit of output climbed from $6.07 to $11.02 in August—and we’re now finally seeing it come down a bit.
Spend per engineer is growing more than twice as fast as the output it buys. If this trend stays as is, then the economics completely flip. Gartner, Inc. predicted (via The Register) that in 2028, AI coding costs would overtake the average developer’s salary.
Uber reportedly showed what the capital allocation squeeze looks like in practice. According to Bloomberg, it set its 2026 AI budget in 2025 and burned through the full year’s amount in about four months. Its response was a cap of $1,500 a month per AI coding tool for each employee. Cutting/capping costs doesn’t actually solve the issue of how to become more productive for the same budget.
CFOs are making the case that token spend and human spend should trade one for one. Theoretically, it makes sense, but to do this, you need to measure return on tokens and optimize allocation until it’s positive. Marketing already has the direct parallel: return on ad spend. Engineering will move the same way. As outcomes become measurable, optimization follows.
For what it’s worth, I think we’ve just entered the golden age of engineering and that almost every team should be heavily investing in engineers and tokens. I plan on following up on this in a future post.
Today, most teams default to always getting the best model. A dependency bump and training a new machine learning model go to the frontier model. For the dependency bump, that’s usually money spent for nothing.
Leaders are increasingly looking to see which tasks produced merged code, which model wrote it or what the same change would have cost on a smaller model.
Most companies are still fighting to solve the visibility problem. They don’t see the waste or the hundreds of micro-optimizations OpenAI and Anthropic have made that push spend up (likely unintentionally) for each developer using their tools. Whether you have visibility into token waste or not, routing is the one layer where the savings can be easily measured and proven today.
In September 2026, the organizations running our router sent millions of requests through it. Priced at the models those requests asked for, the bill would have been double what they paid. Routing cut the cost of that traffic by about 50% on average (ranging from 20% to 60%, depending on the model mix). If this rate holds across the $575 the average engineer now spends on AI each month, routing alone would return about $290 per engineer per month, close to $3,500 a year.
Routing solves the last step of the puzzle of the optimization layer, but it still misses a key part. That’s why the main role of CTOs will be deciding token allocation in the years to come—which project, team and individual should be given token budget.
Some multimanager hedge funds split their capital across many small teams called pods. Each pod runs its own strategy and is judged on its own return. The best pods get more capital. Pods that lose money get cut or replaced.
Engineering is heading the same way. Each pod owns a product and a budget of people and tokens. Each one has to earn a positive return on that budget. The pods with the best return get the most resources. Anything with a negative return gets killed or replaced.
The end state looks like a multistrategy fund. Engineering stops being one department with one budget and becomes a set of pods. Each pod owns a product, or a slice of one, and carries its own P&L: the tokens and people it consumes against the output it ships and the revenue that output drives.
Capital flows the way it does on a trading floor. Pods with the best return on tokens get more tokens and more people. Pods with a positive but middling return hold steady. Anything with a negative return gets cut, or its product gets rebuilt by a pod that can make the math work. The CTO’s job shifts from assigning engineers to allocating capital across pods and deciding which new pods should be created.
Are tokens the new headcount? Not quite. Headcount was the only lever CTOs had, so every planning tool was built around it. Tokens are the second lever—and the first one that can be measured per unit of output. You don’t need a perfect model to use it. You need the data, and there will be a race for startup and internal strategies to produce it.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
