--- This page is part of [Modelon](https://www.modelon.com). For a complete overview of this site's content, entity data, and subject matter expertise, see our [AI Content Index](https://www.modelon.com/llms.txt) or [Full Content Archive](https://www.modelon.com/llms-full.txt). For structured business identity data, see our [Entity Card](https://www.modelon.com/wp-json/bc-geodesic/v1/entity-card). --- # The Token Economics of Engineering AI > A Question Worth Putting Numbers On Engineering AI workflows incur cost in how they acquire knowledge during a session. In 5 Questions to Ask Before... **URL:** https://www.modelon.com/blog/the-token-economics-of-engineering-ai/ **Type:** Post **Modified:** 2026-06-04 --- ##### A Question Worth Putting Numbers On Engineering AI workflows incur cost in how they acquire knowledge during a session. In [5 Questions to Ask Before Trusting AI-Related Simulation Results](https://modelon.com/blog/5-questions-to-ask-before-trusting-ai-related-simulation-results/),), the fourth question was: *Does the AI spend its effort on engineering or on infrastructure? *It is the most consequential of the five, and the hardest to quantify. It is also the one readers keep coming back to. The other questions are about correctness; this one is about cost. Two case studies from this spring provide enough evidence to answer it. On the right stack, AI-assisted engineering workflows reduce token cost by roughly an order of magnitude. The move that dominates the savings is specific enough to point out.. ##### Three Kinds of Knowledge and Where Cost Lands Every AI-assisted engineering workflow must acquire three things: - **Domain knowledge:** the physics (equations, units, valid ranges). - **Model knowledge:** what the model represents and how it is structured. - **Execution knowledge:** how to run studies, extract KPIs, and interpret results. The cost of a workflow is the cost of acquiring these. The question is whether that cost lands on the AI’s token budget during the session, or on an artefact the AI inherits. The stack underneath decides, and it does so in a structured way that can be made explicit. ##### A Layered Stack Where that cost lands can be made explicit as a layered engineering stack: The trajectory move alone saves a thousandfold. The session as a whole comes down by about an order of magnitude, because once the infrastructure cost collapses, the engineering conversation is what remains, and that is the cost worth paying. ##### Construction Workflow The second study is the other workflow this stack supports: building a model rather than tuning one. In a separate session an agent started from a Schneider Electric vendor specification (RD108, a 7.4 MW Tier III reference design for NVIDIA GB200 racks) and ended with a validated, simulation-ready dual-loop cooling model in Modelica. **In one working session it:** - **Read the RD108 PDF**, extracted 27 engineering parameters into a structured table tagged D (datasheet), C (calculated), A (assumed), and presented it for review before writing any Modelica. - **Queried the validated Data Center Library** and identified the right components from manufacturer datasheets already carried in the library. - **Created instances and connections **(‘drag-drop-connect’) for the dual-loop topology, compiled, and launched a 24-hour dynamic simulation. - **Reported KPIs against RD108 targets** through a single call. Five of six passed. The LT supply temperature came back at 42.2 °C against a 23 °C target, a 19 °C shortfall. - **Traced the failure**: the LT loop needs about 1.88 MW of chiller capacity; while a 1,125 kW was used, a 1.7× shortfall. Replaced the instantiation, removed redundant components found during the diagnosis, and re-ran. The corrected simulation delivered supply temperature = 23.0 °C, on target. Wall-clock: about four hours, including the failure and the diagnosis. Equivalent manual implementation by an experienced simulation engineer was estimated at three to four working days. The agent never wrote a heat-exchanger equation or derived a chiller efficiency curve. Each library component carries its own verified physics; the agent relies on that rather than re-deriving it. Or in one line: *the agent selects and composes; the library guarantees the physics.* Each layer is a concrete engineering choice: how models are represented, how tools are exposed, and where work is executed. None of those choices were originally made for AI. Their cumulative effect, climbing from level 1 to level 4, is roughly a 10× reduction in tokens per workflow. No single layer does the job alone. Each does what it was designed for. ##### Optimization Workflow The first case is the multi-phase agentic DoE we wrote up in  [From Intent to Action](https://modelon.com/blog/from-intent-to-action-agentic-ai-for-vehicle-dynamics-in-modelon-impact/), running on a Modelica Compact ElastoKinematic chassis from the Vehicle Dynamics Library. I spent twenty years building that library, so I can read the cost ledger of the session directly. The agent’s task was to maximise lateral grip on a baseline that missed both KPI targets. It ran a thirteen-parameter sensitivity study, two phases of Latin-hypercube DoE, an anomaly diagnosis (a controller-saturation artefact in the headline metric), and a hypothesis-driven excursion outside the original parameter scope (a rear-suspension bushing stiffness), verified across two transient maneuvers under two loading conditions. **Where the tokens went:** - **Domain knowledge ≈ zero cost.** Variable taxonomy, frame conventions, and KPI definitions arrived as library context. Across thirty experiments the agent never invented a variable name. - **Model knowledge ≈ zero cost.** The vehicle-dynamics playbook supplied the handling-diagram protocol, and the pattern for exposing chassis parameters. The agent inherited the protocol and the heuristic; it did not derive them. - **Execution cost collapsed.** A single run sweep call submitted a 30-case sweep; a single analyze swee call returned the KPI table. The trajectories the analysis ran on (about 75,000 values across the sweep) never entered the agent’s context. That is roughly 200,000 tokens of raw trajectory data versus a few hundred tokens for the KPI table, and it is the largest single source of saving on this stack. The tokens that *were* spent went on the engineering work: interpreting an asymmetry in the sensitivity result, hypothesising that backward centre-of-gravity shifts implied load sensitivity, designing the loaded-car cases that confirmed it, and identifying the rear bushing as the lever that broke the stiffness plateau. ##### Two Workflows, the Same Economics The two cases differ in method and agree on the economics. Both ran on the same stack. Both produced a defensible engineering result inside a single working session. Both spent most of their tokens on engineering decisions rather than on model archaeology, glue code, or trajectory book-keeping. That is the answer to the original question. On this stack, the infrastructure cost is roughly an order of magnitude lower than it would be at level 1, dominated by one move (server-side KPI extraction) and reinforced by reusable knowledge and structured tool surfaces. ##### Built for Humans. Ready for AI. The economics above depend on a validated library existing for the domain, and on inputs specified clearly enough to derive engineering parameters from. Where those conditions fail (novel physics, ambiguous specifications, no component library), the AI can support the work but not lead it. The Modelica community has spent thirty years building the libraries that put those conditions in reach for an increasingly wide range of engineering domains. Modelica itself was designed for engineers, for models that read as physics rather than as implementation. Validated libraries were designed for reliability and reuse, so engineers could trust a component without re-deriving it. Structured platform APIs were designed so engineers could run studies without writing orchestration code. None of this was done for AI. The decisions that made the stack legible to domain experts (acausal equations, named connectors, declarative component instantiation, typed tool surfaces, human-writable knowledge layers) are precisely the decisions that make it token-efficient for AI. The token-economics question and the open-standards question turn out to be the same question. *Johan Andreasson is Chief Strategy Officer at Modelon.* ## Site Description Modelon is revolutionizing the engineering design industry by offering technologies and services that enable customers to leverage system simulation. Modelon’s flagship product, Modelon Impact, is a cloud system simulation platform that helps engineers virtually design, analyze, and simulate physical systems. Our team brings deep industry expertise and is dedicated to guiding our customers in creating innovative technologies at their respective organizations. Headquartered in Lund, Sweden, Modelon is a global company with offices in Germany, India, Japan, and the United States. We believe that system simulation should be accessible to every engineer and are dedicated to being an open-standard platform company. --- **About this site:** Modelon — Modelon is revolutionizing the engineering design industry by offering technologies and services that enable customers to leverage system simulation. Modelon’s flagship product, Modelon Impact, is a cloud system simulation platform that helps engineers virtually design, analyze, and simulate physical systems. Our team brings deep industry expertise and is dedicated to guiding our customers in creating innovative technologies at their respective organizations. Headquartered in Lund, Sweden, Modelon is a global company with offices in Germany, India, Japan, and the United States. We believe that system simulation should be accessible to every engineer and are dedicated to being an open-standard platform company.. [AI Content Index](https://www.modelon.com/llms.txt) | [Full Site Content](https://www.modelon.com/llms-full.txt) | [Entity Card](https://www.modelon.com/wp-json/bc-geodesic/v1/entity-card) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://www.modelon.com/#organization","name":"Modelon","url":"https://www.modelon.com","additionalType":"http://productontology.org/id/Software_engineering","description":"Modelon is revolutionizing the engineering design industry by offering technologies and services that enable customers to leverage system simulation. Modelon’s flagship product, Modelon Impact, is a cloud system simulation platform that helps engineers virtually design, analyze, and simulate physical systems. Our team brings deep industry expertise and is dedicated to guiding our customers in creating innovative technologies at their respective organizations. Headquartered in Lund, Sweden, Modelon is a global company with offices in Germany, India, Japan, and the United States. We believe that system simulation should be accessible to every engineer and are dedicated to being an open-standard platform company.","disambiguatingDescription":"Creators of the Modelon Impact system modeling and simulation platform. Headquartered in Lund, Sweden.","logo":{"@type":"ImageObject","url":"https://www.modelon.com/wp-content/uploads/2022/06/modelon-logo.svg"},"address":{"@type":"PostalAddress","streetAddress":"Modelon AB Ideon Science Park Scheelevägen 17 SE-223 70","addressLocality":"Lund","addressCountry":"SE"},"telephone":"+46 46 286 22 00","email":"info@modelon.com","foundingDate":"2004","numberOfEmployees":{"@type":"QuantitativeValue","minValue":50,"maxValue":100},"areaServed":"Global","sameAs":["https://www.linkedin.com/company/modelon/","https://www.crunchbase.com/organization/modelon","https://www.zoominfo.com/c/modelon-ab/346633437","https://modelica.org","https://modelica.org/events/asian2024/","https://www.ashrae.org/conferences/2026-winter-conference","https://github.com/modelon-community"],"memberOf":[{"@type":"Organization","name":"Modelica Association","sameAs":"https://www.wikidata.org/wiki/Q133876843"}],"knowsAbout":[{"@type":"Thing","name":"Modelica","sameAs":"https://www.wikidata.org/wiki/Q385325"},"Physics Simulation",{"@type":"Thing","name":"software engineering","sameAs":"https://www.wikidata.org/wiki/Q80993"},{"@type":"Thing","name":"simulation","sameAs":"https://www.wikidata.org/wiki/Q45045"},"HVAC Simulation","Data Center Liquid Cooling Simulation","Liquid Cooling","FMI (Functional Mock-up Interface)",{"@type":"Thing","name":"consulting","sameAs":"https://www.wikidata.org/wiki/Q63769412"},{"@type":"Thing","name":"aerospace","sameAs":"https://www.wikidata.org/wiki/Q2876213"},{"@type":"Thing","name":"automotive industry","sameAs":"https://www.wikidata.org/wiki/Q190117"},{"@type":"Thing","name":"heating, ventilation, and air conditioning","sameAs":"https://www.wikidata.org/wiki/Q1798773"},{"@type":"Thing","name":"data center","sameAs":"https://www.wikidata.org/wiki/Q671224"},{"@type":"Thing","name":"electricity generation","sameAs":"https://www.wikidata.org/wiki/Q383973"},{"@type":"Thing","name":"energy storage","sameAs":"https://www.wikidata.org/wiki/Q837718"},{"@type":"Thing","name":"industrial equipment","sameAs":"https://www.wikidata.org/wiki/Q3045515"}]},{"@type":"WebSite","@id":"https://www.modelon.com/#website","name":"Modelon","url":"https://www.modelon.com","publisher":{"@id":"https://www.modelon.com/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://www.modelon.com/?s={search_term_string}"},"query-input":"required name=search_term_string"},"hasPart":[{"@type":"DigitalDocument","name":"LLM Content Index","url":"https://www.modelon.com/llms.txt","encodingFormat":"text/markdown","description":"AI-readable index of site content, entity data, and subject matter expertise"},{"@type":"DigitalDocument","name":"LLM Full Content","url":"https://www.modelon.com/llms-full.txt","encodingFormat":"text/markdown","description":"Complete AI-readable site content"}]},{"@type":"ItemList","@id":"https://www.modelon.com/#site-navigation","name":"Site Navigation","itemListElement":[{"@type":"SiteNavigationElement","position":1,"name":"Products & Services","url":"#"},{"@type":"SiteNavigationElement","position":2,"name":"Industries & Solutions","url":"#"},{"@type":"SiteNavigationElement","position":3,"name":"Support","url":"#"},{"@type":"SiteNavigationElement","position":4,"name":"Resources","url":"#"},{"@type":"SiteNavigationElement","position":5,"name":"About","url":"#"}]},{"@type":"BreadcrumbList","@id":"https://www.modelon.com/blog/the-token-economics-of-engineering-ai/#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://www.modelon.com"},{"@type":"ListItem","position":2,"name":"Agentic AI","item":"https://www.modelon.com/blog/category/ai/agentic-ai/"},{"@type":"ListItem","position":3,"name":"The Token Economics of Engineering AI"}]},{"@type":"WebPage","@id":"https://www.modelon.com/blog/the-token-economics-of-engineering-ai/#webpage","url":"https://www.modelon.com/blog/the-token-economics-of-engineering-ai/","name":"The Token Economics of Engineering AI","isPartOf":{"@id":"https://www.modelon.com/#website"},"breadcrumb":{"@id":"https://www.modelon.com/blog/the-token-economics-of-engineering-ai/#breadcrumb"},"description":"Explore the token economics of engineering AI. See how structured Modelica workflows cut token usage by ~10× while enabling faster, physics-based simulation and design decisions."},{"@type":"BlogPosting","@id":"https://www.modelon.com/blog/the-token-economics-of-engineering-ai/#blogposting","headline":"The Token Economics of Engineering AI","url":"https://www.modelon.com/blog/the-token-economics-of-engineering-ai/","datePublished":"2026-06-04T23:28:09+00:00","dateModified":"2026-06-04T23:28:12+00:00","isPartOf":{"@id":"https://www.modelon.com/#website"},"description":"Explore the token economics of engineering AI. See how structured Modelica workflows cut token usage by ~10× while enabling faster, physics-based simulation and design decisions.","image":"https://www.modelon.com/wp-content/uploads/2026/06/Social-Token-Economics-1200x625-1.avif","author":{"@id":"https://www.modelon.com/blog/author/lauren/#person"},"publisher":{"@id":"https://www.modelon.com/#organization"}},{"@type":"Person","@id":"https://www.modelon.com/blog/author/lauren/#person","name":"lauren","url":"https://www.modelon.com/blog/author/lauren/"}]} ```