FinOps & Tokenomics Bulgaria Meetup #4 - AI FinOps Loop
AI spend is now the fastest-growing line on a lot of cloud bills, and most teams still can’t say what they’re getting for it. For meetup #4 we’re giving the whole evening to that problem. We’ll look at what 472 companies said about it, take apart a real agentic FinOps platform on AWS and its bill, and then see a tool that shows engineers what their own AI usage costs.
📍 Ring Tower, Sofia
🗓 27.10.2026 18:45
🗣 Mostly Bulgarian, some English (the slides are in English)
Agenda
18:45 Doors open, registration
19:00 State of Tokenomics 2026: what enterprises are really struggling with (~15 min)
Yordan Dabov & Ivaylo Vrabchev
The Tokenomics Foundation (a Linux Foundation project) has just published its first State of Tokenomics report, based on 472 organisations across 11 industries. The surprise is that price isn’t the problem. Only 7% named cost or pricing complexity as their top challenge, and only 4% asked providers for cheaper tokens. The biggest challenge is proving ROI (43%), and what companies want most is more granular data, FOCUS and better attribution. I’ll go through the findings that matter for FinOps teams: who owns AI spend, why companies are moving off frontier models, and why the 86% who use or are evaluating a model router are so much better at showing value to the CFO.
19:15 The AI FinOps Loop: cutting AI costs with FinOps, and using AI to do FinOps (~40 min)
Ivaylo Vrabchev, Phi Partners
The idea is simple. We built a multi-agent FinOps assistant on Amazon Bedrock and AgentCore: one supervisor and four specialists for anomalies, allocation, rates and usage. In its first month it found $87k a year in savings. Then we opened its own bill. A 10-token question turned into 121k tokens, and the fleet cost more to run than the savings it found.
We’ll go through:
where the token amplification and “architecture tax” come from in multi-agent setups
a real $1.7M Bedrock bill, broken down by model
attribution per agent: IAM principals, inference profiles, request metadata, and the gateway trap
why model config in application code can’t be governed, and how Prompt Management fixes that
the levers that took the same fleet from $104k to $19k a year with the same answers: scheduling, caching, model tiering, routing, batch and guardrails
build vs. buy now that AWS FinOps Agent is in preview
five things you can do on Monday morning
Bring your hardest questions about agent costs.
*This is the extended version of my AWS Community Day talk, with more time for the parts people asked about afterwards.
20:00 PracticeGraph: see, understand and improve how you work with AI coding tools (20-30 min)
Nikolay Penchev, Team Landi
PracticeGraph is a free, open-source desktop app that turns Claude Code and Codex session logs into a daily view of your AI practice. It works for pay-per-use and subscriptions alike, and it computes everything locally: no prompts are stored, and nothing is uploaded.
We'll cover:
Cost: spend by model and day, or what a subscription is worth at API prices and how close you are to your limits.
Tips: one evidence-backed change at a time, such as a costlier default model, heavy context, poor cache reuse, unused MCP servers or missing test runs.
Prices: cheaper hosts, batch and off-peak pricing, each with its source.
Staying current: hand-curated editions refreshed through the day, covering AI news, model guidance, what other developers measured, ideas worth building, and official guides and training for each tool.
Team view: opt-in anonymous aggregates without collecting anyone's prompts.
20:30 Networking, drinks and open discussion
Sponsors: Phi Partners & Sketch Coworking
See you there!
Dancho & Ivo