TL;DR
Thorsten Meyer AI published a follow-up analysis arguing that open-weight AI models should be judged by total operating cost, not download price. The piece says self-hosting can beat paid APIs for steady, high-volume workloads, but only when hardware, power, maintenance and quality gaps are counted.
Thorsten Meyer AI has published a follow-up field note arguing that companies weighing self-hosted open models against paid AI APIs should compare total operating cost, not download price, because open weights may be free while hardware, power, support work and quality tradeoffs are not.
The field note was written as a follow-up to Meyer’s earlier piece on Mistral and European AI sovereignty. Meyer says the unresolved question was why a company would pay a vendor to run models on premises if it could download Qwen or another open-weight model without paying for the file.
The answer offered in the new analysis is that free describes only the model weights. The piece lists hardware, electricity, operations time, model updates, queue management, inference harnesses, quality gaps and depreciation as costs that remain with the customer when inference is self-hosted.
Meyer also argues that self-hosting can win when usage is steady and high enough. The article’s illustrative cost model places a break-even point near 80 million tokens a month under one set of settings, while warning that the figure is not a quote and changes with workload, task difficulty, data-sovereignty needs and operator skill.
Why It Matters
The question matters because many organizations are now deciding whether to use paid APIs, host open-weight models themselves or buy managed private deployments. The wrong comparison can make local AI look cheaper than it is, or make API spending look unavoidable when a predictable workload could be cheaper on owned hardware.
For companies with sensitive data, the article says the decision is not only about price. Self-hosting may keep data inside the organization by design, but it also shifts reliability, tuning and incident response to the buyer.
AI Hardware, Software, and Architectures Powering Modern Artificial Intelligence: From GPUs and ASICs to CUDA, Accelerators, Compilers and Runtimes
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Background
The analysis grows out of a sovereignty debate around European AI providers such as Mistral. The challenge Meyer identifies is that open-weight Chinese models including Qwen, DeepSeek and GLM are available to download, weakening a sales pitch built only on local control.
The field note says the market has changed because open models have narrowed the gap with closed frontier systems on some benchmarks while remaining far cheaper in API form. It says closed Western frontier models still lead on the hardest long-horizon agentic tasks, while open models may lag by six to 12 months and then close ground on prior tests.
“The weights are free to download. Running them well is not.”
— Thorsten Meyer AI field note
“The honest comparison is total cost of ownership vs. per-token API.”
— Thorsten Meyer AI field note
“Below some usage level the API wins decisively. Above some sustained, predictable volume, owned hardware wins.”
— Thorsten Meyer AI field note
“Data never leaves.”
— Thorsten Meyer AI field note
high-performance GPU for AI
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
What Remains Unclear
The article does not establish a universal break-even point. The economics vary by hardware purchase price, power cost, utilization, staffing, latency needs, model quality and whether a team can operate inference reliably. The source also presents benchmark and price comparisons as part of its analysis; readers should treat those as claims unless checked against current vendor pages and independent benchmark data.
AI model deployment server
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
What’s Next
The next step for buyers is a workload-specific test: measure monthly token volume, peak traffic, response-time requirements, data controls and staff capacity, then compare that result with current API pricing and hardware quotes. The cost line may keep moving as open models, inference software, chips and vendor pricing change through 2026.
electricity-efficient data center hardware
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
Is an open model actually free?
No. The field note says the file can be free to download, but running it requires hardware, power, staff time, maintenance and a production harness.
When can running your own model beat an API?
According to Meyer’s analysis, the local path is more likely to win when volume is high, steady and predictable, and when the team can keep machines well used.
What is the reported break-even point?
The source’s illustrative slider shows break-even near 80 million tokens a month under one selected setup. The figure is not presented as a universal quote; it shifts with usage, model choice, task difficulty and operating costs.
What are the biggest risks?
The risks are undercounting operations work, accepting lower output quality on hard tasks, buying hardware that ages quickly, and treating the downloaded weights as the whole system.
Why does data sovereignty affect the decision?
If inference runs on owned systems, the article says data does not need to leave the organization. That can matter for privacy, regulated work and internal control, though it does not remove the need for security and governance.
Source: Thorsten Meyer AI
