TL;DR
Thorsten Meyer AI published a follow-up analysis arguing that open-weight models are not truly free once hardware, electricity, operations and quality gaps are counted. The report says self-hosting can beat paid APIs at sustained, predictable volume, while APIs still make more sense for low or uneven demand.
Thorsten Meyer AI has published a follow-up analysis arguing that companies should compare open-weight AI models against paid APIs by total operating cost, not by download price, a distinction that matters as businesses weigh self-hosted systems against vendor-run frontier models.
The analysis centers on a question raised after a prior Mistral sovereignty piece: why would a company pay to run models on premises when it could download models such as Qwen at no licensing cost? The answer given by the article is that model weights may be free, but operating them is not.
According to the piece, the real ledger includes hardware, electricity, operations time, model updates, queue health, tuning, context management, retries, tool routing, quality gaps and depreciation. The article says those costs often decide whether self-hosting beats renting model access through an API.
The report argues that the crossover depends on workload. For low-volume or spiky use, paid APIs can remain cheaper and simpler. For steady, high-volume inference, owned hardware can become cheaper because costs stop rising per token once the system is bought and running.
Why It Matters
The analysis matters because many AI buyers are being pushed from two directions: closed-model vendors sell managed access to the strongest systems, while open-model advocates point to free downloads and local control. The article says neither framing is complete without operating costs and usage patterns.
For companies handling sensitive data, the choice is also about control. The source argues that self-hosting keeps data inside the operator’s environment, while API use depends on vendor contracts and trust. For teams running large, predictable workloads, that control may come with lower long-run cost, but only if they can manage the system reliably.
AI model hardware server
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Background
The source places the debate in a mid-2026 market where open-weight models have narrowed the capability gap with closed frontier systems on many tasks, while still trailing on the hardest long-horizon agent work. It cites Chinese open-weight or open-access models, including DeepSeek, Kimi, GLM and Qwen, as part of the pressure on Western closed APIs.
The analysis says open models may lag the frontier by roughly 6 to 12 months on the most difficult work, then catch up on earlier benchmark challenges. It also says hosted open-model pricing can be far below premium closed-model pricing, but warns that benchmark results and token prices do not replace testing on a company’s own workload.
“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
“The meter never restarts.”
— Thorsten Meyer AI
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.
What Remains Unclear
The source does not establish a universal break-even point. It presents an illustrative crossover, with the result changing based on token volume, task difficulty, sovereignty needs, hardware cost, power cost and team skill. It is also unclear how fast today’s open-weight models will keep closing the quality gap with closed frontier systems.
self-hosted AI model setup
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 workload-specific testing: measuring monthly token volume, latency needs, output quality, staffing capacity and data-control requirements before choosing between API access, hosted open models or self-hosted inference.
AI model deployment tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
Are open-weight AI models actually free?
The download can be free, depending on the license. The article says running the model still brings costs for hardware, electricity, maintenance, model tuning and supporting software.
When can self-hosting beat a paid API?
According to the analysis, self-hosting is more likely to win when usage is steady, high volume and predictable, and when the operator has enough technical skill to keep the system running well.
When do APIs still make more sense?
APIs can remain the better option for low-volume, uneven or highly complex workloads, especially when a team needs the best frontier performance without managing infrastructure.
What remains unresolved?
The exact break-even point is not fixed. It depends on model quality, workload type, hardware prices, power costs, staffing and how quickly open-weight models improve.
Source: Thorsten Meyer AI
