Yasu

This week I'm joining Tokenomicon + FinOps X in Amsterdam: two intense days at the FinOps Foundation's event dedicated to the cost of AI, with the people who think about this as much as we do. A few weeks ago we shipped Cost of AI to Yasu customers. Every AI dollar, by model, by team, by token type, and for the first time a peer benchmark: how your token efficiency compares to other organizations, built on the methodology we published in the Token Efficiency Index. Customers tell us it takes the guesswork out of attributing AI cost for them. What I'm looking forward to most: hearing how practitioners are measuring token cost against value, what has worked for them, what hasn't, and where they see agentic engineering in this space going in the next two years. If you are joining as well, drop me a DM or come and say hello when I am there. FinOps Foundation

The industry has spent a decade building better dashboards. Every one of them can tell you what you spent. Almost none can tell you whether you should have. Those are different questions. The first is attribution. The second is efficiency. And a dashboard, however good, was never built to answer it. That's why we did this in a specific order: first the benchmark, published as an open methodology anyone can check (39 organizations, on arXiv, link in comments). Then the product. A benchmark tells you you're 6x above your peer frontier. It doesn't get the money back. Catching it before it ships and fixing what's live, that does. That full loop, in 50 seconds, below. If you own an AI budget and haven't read the paper yet: 10 minutes, link in the comments. Yasu

Some people you interview. Some people just blow you away. Anthony Goossens joins Yasu as our GTM Lead, and by the time we sat down with him he had already dissected the cloud cost problem better than most people inside the industry. The energy he brought to that first call is the energy he brings every day. He built a Sustainability business unit as the first hire and grew it to 35 people and €5M+ in revenue. He founded a cleantech consultancy, ran a €1.3M book of business, and exited it. Before that: Technology & Project staffing across Brussels, Chicago and London. What convinced us wasn't the track record on its own. It was that he had clearly thought about the problem before we ever spoke to him and immediately understood our vision. Cloud & AI pricing is built so that the hyperscalers earn more when you spend more, and most companies overpay by up to 30% as a result. Anthony will be tackling this with our new and existing clients. He's based in Amsterdam, speaks Dutch, English and French, and has been part of Leaders for Climate Action since 2023. If you want to talk about what your cloud bill is actually doing, he's the person to message. Day one and he's already cooking🔥 Great to have you with us, Anthony 🚀

I spent years watching companies overspend on cloud. The dashboards always arrived after the money was gone. Now it's happening again, faster. 2026 is the first year companies spend more running AI than training it. Agentic workloads burn 5 to 30x more tokens per task than chatbots did. Teams that budgeted for a pilot are getting production bills nobody modelled. There is no playbook for this yet. I'm hiring the person who writes it. At Yasu, our agents catch expensive infrastructure decisions at the pull request and fix waste in production autonomously. Cloud and AI spend both. I'm looking for a founding cloud engineer to be the judgment inside those agents: half your time in the console with customers finding real savings, half turning what you learn into patterns every future customer gets automatically. Sounds interesting? Apply and we'll grab a coffee. Know someone who fits? Tag them. Coffee on me if you tag someone we end up talking to :)

A typical organization sends 4 out of 5 AI tokens to premium-tier models. Not because every request needs one. Mostly because premium is the default and nobody checks. We know this because we spent the summer analyzing AI usage data from 39 different organizations. And the uncomfortable part: every single one had at least one efficiency gap. That's right, not most but all of them. The question we kept hearing from engineering leaders is "are we spending too much on AI?" Honest answer: in isolation, that question has no answer. AI spend is usage-driven. Two companies can spend the same and one runs twice as efficiently. You can't answer "is this too much?" alone. But you can answer "how do I compare to organizations like mine?" That's peer benchmarking. So we built it. The Token Efficiency Index (TEI) condenses three levers into one 0 to 100 score, benchmarked against your peers: → Cache hit rate → Cache amortization → Premium model share And we saw some real gaps during our research: one organization at 90.8% premium usage against a peer frontier near 14%. Estimated saving: $72,000. Proud to release the full methodology paper today, built by Caden Wong Himanshu D.ng Himanshu D. and the Yasu team. It started as an intern project. It's now on arXiv, open under CC-BY. Most teams are guessing. Now there's a benchmark. Full whitepaper in the comments

Trust is the first feature we ever shipped. Not because a checklist told us to. Because of what we ask of our customers. Our ACE platform reads your cloud billing data. Every account, every workload, every line item. It is read-only access, but we know that decision still weighs heavy on the people making it. Before any engineering leader connects that, their security team asks one question: what happens to our data? We wanted the answer to be an external audit, not a slide. So we did the work. Yasu is now ISO 27001 certified reviewed and signed off by an external auditor, and GDPR compliant. For a pre-seed company, this is unusual timing. Most wait until a big deal forces it. We think that has it backwards. Security was a day-one decision, not a deal-stage scramble. The certificate is a milestone. The principle behind it is permanent.

Tokenmaxxing assumed every token was productive. But now that the big bills have started landing, most tech leaders are wondering: did all that spend really improve productivity? AI is the first infrastructure category in 15 years where the cost curve points up. For two decades, the cloud era trained every CIO and CFO on one assumption. Per-unit costs decline. S3 got cheaper. EC2 got cheaper. Reserved instances got cheaper. Optimize and the bill follows. That logic is hardwired into every IT budget. AI is trending the other way, and the consensus is it will keep rising. Frontier model providers are still subsidizing today's prices. Uber put 5,000 engineers on Claude Code in December and burned through its full-year budget by April. Microsoft pulled Claude Code from thousands of its own engineers, with per-head costs running $500 to $2,000 a month. Meta capped internal AI spend after the bill approached billions. Per-token prices have fallen 67% year on year. Enterprise bills have gone up. Three forces are pushing costs up at once. The $20-a-month chatbot was venture-subsidized. Capacity is constrained. The better the agents get, the more tokens they consume per task. Subsidies ending, true cost emerging, consumption outrunning both. The good news is there is now real focus on this. The Tokenomics Foundationis forming under The Linux Foundation, and FinOps Foundation is extending its open standards to cover token spend. Most of these efforts stop at visibility. Visibility is necessary. It is not enough. The breakthrough happens when visibility connects to outcomes, and that is what we are building @ Yasu: a more intuitive way to understand cost of AI and its benefit. The cloud era let us innovate fast. Used well, AI turbocharges that. So rather than worrying about tokenmaxxing, let's focus on output-maxxing. We have a strong view here. Always interested in different ones.

🚨🚨HIRING ALERT - FOUNDING FINOPS/DEVOPS ENGINEER 🚨🚨 We started Yasu to kill cloud waste before it happens. Not to build another dashboard that reports it after the money is already gone. Every tool on the market does that. We catch the expensive infrastructure decision in the pull request, before it ever reaches the invoice, where it is 10x cheaper to fix. The way we do it: a multi-agent system we call ACE that encodes real FinOps judgment into agents running across every customer, with guardrails for what is safe to fix autonomously in production. Now we are hiring the Founding FinOps Engineer to do exactly that, and to turn every find into a pattern the agents run for the next 100 customers. Based in the Netherlands. This is for you if you have personally cut six or seven figures off a cloud bill, you live in the console and the IaC repo, and you write code instead of clicking through 80 accounts. It is not for you if you are a FinOps consultant who lives in slides, or if you need a playbook before you start. This is an engineering role. We build software. If this is the kind of problem you want to wake up to, apply via the link in the comments. And if you know exactly who this is, tag them ❤️