Universal Remedy

Saw this older LinkedIn post about AI tool sprawl, prompted by the news that Microsoft cancelled most of its internal Claude Code licenses and Uber burned through its entire 2026 AI budget in the first four months of the year. The advice that follows is sensible and specific: don't give every employee every AI tool, pick two or three use cases where AI genuinely changes output, go deep on those rather than wide on everything.

The frame of reference for this advice is enterprise. Microsoft and Uber are not struggling because AI doesn't work but managing token costs across tens of thousands of employees, navigating procurement governance, and discovering that agentic AI, the kind that takes multi-step autonomous actions across systems, uses vastly more tokens than the demo suggested. When Nvidia's VP of Applied Deep Learning says that for his team the cost of compute is far beyond the cost of employees, he is describing a problem of scale that requires scale to have.

There are much smaller operations in the world than Microsoft and Uber, the two that made the news here. AI is not a line item these companies are trying to rationalize across a workforce. It is the reason a team of a few people can do work that used to require a team atleast ten times larger. They are getting results that matter. For a company in that position, the Microsoft and Uber token crisis is not a cautionary tale but  evidence that the tools work well enough that large organizations are now struggling to govern their own adoption of them.

The advice to go deep rather than wide still stands, but the reasoning is different at this scale. It's not about cost containment. AI amplifies whatever capability you bring to it, which means a sharp person with a clear use case and good judgment gets an extraordinary return, and a fuzzy process handed to a model gets a faster version of the same fuzzy output. The selectivity that enterprises need for budget reasons, smaller organizations need for quality reasons. Though I would argue the focus on quality should be universal

These discussions tend to obscure is that the economics of AI are unusually asymmetric across company size. A token costs the same whether you are Microsoft or a company with twelve people. The productivity return on that token does not scale the same way. For the large organization, AI is incremental improvement on an existing process. For the small one, it can be the difference between a capability you have and one you don't and in the past could not even imagine building. Jensen Huang's formulation,  that a $500,000 engineer should be consuming $250,000 in tokens,  is an enterprise framing that makes no sense applied elsewhere. The right question for a small company is not what percentage of headcount cost should go to AI. It is what would have required a hire six months ago that no longer does.