TL;DR
Thorsten Meyer AI published an analysis arguing that the real AI cost question is not whether open models are free to download, but when operating them beats paying per token. The report says paid APIs still win for low or uneven usage, while self-hosting can win for steady, high-volume workloads with strong privacy or sovereignty needs.
Thorsten Meyer AI published a new analysis arguing that self-hosted open-weight AI models can beat paid APIs on cost for steady, high-volume workloads, but only when buyers count hardware, power, operations, software tooling, quality gaps and depreciation rather than treating a free model download as a free system.
Confirmed Details
The report starts from a question raised after an earlier piece on Mistral and European AI sovereignty: why would a company pay a vendor to run models on-premises if it can download an open model such as Qwen at no charge? The analysis answers that the model weights may be free, but the running system is not.
Thorsten Meyer AI lists the main cost items as hardware, electricity, operations time, model updates, quantization work, queue health, throughput tuning, context handling, persistence, retries, tool routing and depreciation. The source says the model is only part of the system and that a working production setup needs a harness around it.
The report says the economic crossover depends on usage. In its illustrative scenario, self-hosted hardware breaks even near about 80 million tokens per month, while paid APIs remain better for low or spiky usage. The analysis labels that figure illustrative, not a vendor quote, and says the result moves with task difficulty, sovereignty needs and operator skill.
Local LLM Inference Optimization: A Comprehensive Guide to Quantization, Hardware Acceleration, and Efficient Private AI Deployment
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Why It Matters
Personal AI Servers: A Guide to Building Private AI Infrastructure for Secure, Offline and Self-Hosted Local LLMs for Data Privacy
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Why It Matters
The analysis reframes a common AI purchasing choice for developers, publishers and companies building production workflows. A paid API can be simpler and cheaper when usage is low, demand is uneven, or the task needs the strongest closed frontier model. Owned inference can become cheaper when traffic is steady, utilization is high and the team can keep the stack running.
The privacy and sovereignty issue is separate from raw cost. The source says self-hosting makes data control structural because prompts and outputs do not need to leave the owner’s environment. That may matter for organizations handling sensitive internal material, although the report does not test any specific legal or compliance claim.
Machine Learning Flashcards — 280+ Cards Covering ML Fundamentals, Stats, Algorithms, & Model Deployment | Study Tool for Beginners, Students, Data Science and AI Professionals
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Background
GPU-Powered Deep Learning: Mastering Parallel Computing for High-Performance AI: A Practical Guide to CUDA, Optimization, and Scalable Model Deployment
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Context
The report places the question inside a wider mid-2026 AI market split between closed Western frontier APIs and lower-cost open or open-plus-hosted models, many from Chinese labs. It says open-weight systems have narrowed the capability gap on some tasks while still lagging the hardest closed frontier workloads by six to twelve months.
Thorsten Meyer AI cites examples including DeepSeek V4 Pro, Kimi K2.6, GLM-5.1 and Qwen 3.6, and compares them with closed systems from Anthropic, OpenAI and Google. Those comparisons are presented as the source’s market snapshot, not independently verified benchmark reporting in the provided material.
The report also says Apple Silicon unified memory and mixture-of-experts designs have changed what smaller operators can run locally, including large models on desk-sized machines. The central claim is that the decision has moved from ideology to workload math: token volume, utilization, quality needs and operating capacity decide the winner.
“why would a company pay Mistral to run models on-prem when it could download Qwen and run it for free?”
— Thorsten Meyer AI, posing the core question
“The weights are free to download. Running them well is not.”
— Thorsten Meyer AI
“The honest comparison is total cost of ownership vs. per-token API.”
— Thorsten Meyer AI
“Data never leaves”
— Thorsten Meyer AI
What Remains Unclear
What Remains Unclear
The source does not provide an audited cost sheet, a publication date, or independent benchmark confirmation. Exact break-even points remain uncertain because they vary by hardware price, electricity rates, staff time, batch shape, uptime needs, task difficulty and the quality gap between open and closed systems.
It is not yet clear how quickly closed frontier models will improve, whether hosted API prices will fall, or how future open-weight licenses and release patterns may affect self-hosting plans.
What’s Next
What Happens Next
Buyers comparing APIs, vendor on-prem contracts and self-hosted open models will need workload-specific cost models rather than download-price comparisons. The next step is practical: measure monthly token volume, utilization, latency needs, privacy requirements, operator capacity and error rates, then compare those figures with API pricing and hardware depreciation before choosing a path.
Key Questions
Is a free open model actually free to run?
No. The report says the weights may cost nothing to download, but production use still carries hardware, power, operations, software and depreciation costs.
When does self-hosting beat a paid API?
According to the analysis, self-hosting can win when usage is steady and high enough to keep hardware busy. The source’s sample scenario places break-even near about 80 million tokens per month, but that number is illustrative.
Do open models now match the best closed systems?
The report says open models have narrowed the gap and can match closed systems on some tasks, but it also says closed frontier models remain ahead on the hardest long-horizon agentic work.
Why would privacy change the decision?
If a model runs locally, prompts and outputs can stay inside the owner’s environment. The report treats that as a structural advantage for sensitive data, separate from cost.
What should teams measure before buying hardware?
Teams should measure token volume, traffic steadiness, latency needs, power costs, staffing, uptime needs, model quality, privacy requirements and how quickly the hardware is likely to depreciate.
Source: Thorsten Meyer AI
