AI Cost Lab 2026
Simulators you can break and fix: cost per outcome, the token-efficiency levers, evals as the gate, agents going non-linear, and pricing what AI costs you. Three tracks, twenty steps, one fictive company whose numbers add up.
Asbjørn Jørgensen
I help startups and VCs partner with AWS to build category-defining businesses, looking after AWS's partnerships with Atomico, Index Ventures, Northzone, byFounders, NEA and many others across Europe.
On the side I'm supporting Caveman (YC F26) to build the AI efficiency category.
Startups · Investing · AI · Amsterdam · LinkedIn
Record
- 2012 Cloakroom · Founder · Consumer
- 2017 Maxwell · Founder · B2B SaaS
- 2023 – Amazon Web Services · Startups & VC partnerships · Europe
- 2026 Caveman · Investor · AI cost efficiency · YC F26
Field notes
Rooms where the work happens. Hover for colour, click to open.
-
2026byFounders Summit · the whole crew -
2026TechBBQ Copenhagen · morning run crew -
2026Dutch GP, Zandvoort · paddock with founders -
2026Atomico 20-year gala · with Niklas Zennström -
2026byFounders Voyage Ashore · Faroe Islands -
2026Mesh Community, Nusfjord · Lofoten
Projects
Caveman 2026
A small side-scrolling arcade game about cost per outcome. Smash the fluff, keep the code, hit the switches, cap the whale. Seven levels, each one a lever for cutting an AI bill, each with a real-world takeaway card and a link into the lab.
Seven levers
Seven levers that cut an AI bill, one per level of Caveman, each with a step in the lab that goes deeper. You can also just read them.
- 1Cost per outcomeMeasure cost per completed task, including retries and the price of failure. It is the only number that lets you compare models honestly.In the lab: Cost per outcome, not price per token →Play →
- 2Right-size the modelRoute by difficulty: small model first, escalate on low confidence. Cascading is a gate problem, not a model problem; invest in the validator.In the lab: Your first lever: right-size the model →Clear level 1 to unlock
- 3Trim the contextKeep context under control: summarise history, window it, retrieve only the top-k that matters. Often worth 30-50%, and reversible.In the lab: Context trimming and output caps →Clear level 2 to unlock
- 4Cache and batchPrompt caching for repeated prefixes, semantic cache for repeated questions, batch API for anything that can wait. Any change to the prompt must preserve the cached prefix.In the lab: Prompt caching →Clear level 3 to unlock
- 5Cap the spendAutoscale to zero, set budget alarms and per-key quotas, and review provisioned capacity monthly. Catch the leak before the invoice does.In the lab: Catch the leak before the invoice does →Clear level 4 to unlock
- 6Tag and measureInstrument every call: tokens in/out, model, cost, latency, per feature, user and prompt version. Find the 5% of calls that cost 50%.In the lab: Tag it, or you can't fix it →Clear level 5 to unlock
- 7Cap the agentTrace cost per trajectory, not per call. Gate agent changes on outcome, cost per success and trajectory shape together, and set the thresholds before you need them.In the lab: Trajectory cost is quadratic in steps →Clear level 6 to unlock
The short version. For the full treatment, open the lab. Levels unlock as you clear them in Caveman; progress is saved in this browser.
Contact
If you are backed by a tier 1 VC or a serial entrepreneur, let's talk! I'd love to discuss your needs around fundraising, go-to-market, deal flow and, of course, technical support while building on the AWS stack.