Ramp's Eric Glyman says top AI spenders doubled revenue, warning of 300x frontier model cost increases by 2026

@eglyman there's an energy version of this. every frontier query is more gpu-seconds and more watts. routing a calendar update to gpt-5.5 pro burns real power to do what a 3B model handles fine. allocating spend by need is allocating energy by need.

@eglyman @JaredSleeper 🎯 the world will move to ROI per workflow

wrote more about this here last week:

https://www.whatshotit.vc/p/whats-in-enterprise-itvc-500

@eglyman Yes there is a lot we can do, especially multi-provider and OSS. Let's talk @eglyman! Been working on intelligent model routing (gateway agnostic) for 2 years: http://notdiamond.ai

@eglyman I'm surprised you aren't building a Router to help Ramp customers save on AI spend, similar to how you started on credit cards by sharing the interchange with us. A kick ass router that sends most (if not all) of the savings to the customer would be amazing @eglyman

@eglyman This tacitly implies that the software business model doesn’t really work for AI (since its outputs are non-deterministic)

@eglyman Yep - and that allocation discipline is likely to increase the market size https://www.exponentialview.co/p/does-pricing-shrink-or-expand-markets?r=f84r&utm_campaign=post&utm_medium=web

@eglyman "Tokens are just dollars, agents are just hires." No. A hire can reason. A hire exercises judgment. A hire can refuse. An agent is a stochastic process with a meter attached. The conflation is either naive or convenient.

@eglyman “This is the future of finance: not blindly rubber-stamping or rejecting AI spend, but allocating it with the same rigor companies apply to headcount.”

@eglyman The attribution gap is real, and there's a layer under it. Reasoning tokens bill at output rates; agentic runs burn ~1,000x a simple call, and neither maps cleanly to attribution tooling. That's why the bill keeps climbing as per-token costs fall. Thread:

@eglyman > re-route just 10% of a $10M AI bill from frontier to GPT-4 level intelligence you’ve saved nearly one million dollars. This sounds like a made-up stat — it’s not.

This doesn’t just sound like ai slop — it is.

@eglyman People don't realize that today's 4B models can run locally on a cell phone and have better than GPT-4o

@eglyman 这因果反了吧——不是砸钱AI让收入翻倍，是本来就在涨的公司才烧得起这个钱；优步一个季度烧完年度预算，是因为它打车订单本来就在涨，换个增长停滞的公司同样砸2810亿token，砸出来的只有一张更难看的财报。

@eglyman the 'which spend worked' question is so underrated. everyone stares at the bill not the roi lol

@eglyman A problem which Vantage is already helping with quite a bit as well :)

http://www.vantage.sh

@eglyman The frame is right but misses why companies overspend: the cost of falling behind is invisible and unbounded. You can measure what you spent. You can't measure what you lost by not spending. Until counterfactual attribution exists, rational CFOs will over-spend.

@eglyman A huge zone opportunity for routers and more

@eglyman The future of finance is not allowing blind deployment of capital. It’s optimizing deployment given exposure and an uncertain world. This is what we do.

@eglyman This is the right take

@eglyman typically finance is very scared of what we can’t measure over time. Hopefully the attribution issue resolves before it sours the innovation.

@eglyman Indeed