Apple Silicon costs more than OpenRouter

Recent analysis shows that Apple Silicon hardware, such as the M5 Max MacBook Pro, can be more expensive than OpenRouter for running local AI models, depending on the hardware’s lifespan and token throughput.

The comparison is based on hardware costs, electricity, and token processing rates. A MacBook Pro with M5 Max priced at $4,299, when amortized over 3 to 10 years, results in an annual cost of approximately $430 to $1,433, translating to an hourly cost of roughly $0.05 to $0.16. Electricity costs for inference at 50-100 watts are around $0.02 per hour, adding minimal expense.

Token throughput tests indicate that the M5 Max can process between 10 and 40 tokens per second for models like Gemma 4 31b. At 10 tokens per second, the cost per million tokens ranges from about $1.61 to $4.79, while at 40 tokens per second, it drops to $0.40 to $1.20. In comparison, OpenRouter’s Gemma 4 31b costs approximately 38-50 cents per million tokens.

Depending on assumptions, the Apple Silicon device’s cost per million tokens can be comparable to or significantly higher than OpenRouter’s. On the optimistic side (long lifespan, high throughput), the costs are similar; on the pessimistic side (short lifespan, lower throughput), Apple Silicon can be up to 10 times more expensive.

Why It Matters

This comparison highlights that for individual users and developers, the total cost of running local AI models on Apple Silicon devices can be higher than cloud-based or dedicated inference hardware like OpenRouter. The cost factors influence decisions on whether to run models locally or in the cloud, especially for large-scale or continuous inference tasks. It also underscores the importance of hardware efficiency and speed in AI deployment strategies.

Background

Recent years have seen increased interest in local AI inference to reduce reliance on cloud services, improve privacy, and lower ongoing costs. OpenRouter and similar providers offer affordable, high-speed inference hardware, while consumer devices like MacBook Pros are capable of running models close to performance levels of specialized hardware. This analysis compares the total costs involved over typical hardware lifespans, factoring in electricity, hardware price, and token throughput.

“On the optimistic side (50 watts, 40 tokens per second, and 10 years) the pro max is as cheap as openrouter.”

— William Angel

“On the pessimistic side (100 watts and 3 years at 10 tokens per second) the pro max is 10x the cost.”

— William Angel

What Remains Unclear

It is still unclear how real-world usage patterns, hardware depreciation, and actual inference speeds will impact the long-term cost comparison. Variations in electricity prices, hardware upgrades, and model efficiency could alter the current estimates.

What’s Next

Further testing and real-world deployment data are expected to refine these cost estimates. Developers and users will need to evaluate their specific needs, including speed requirements and hardware longevity, to decide whether local inference on Apple Silicon is cost-effective.

Key Questions

Is running AI models on Apple Silicon more cost-effective than cloud services?

It depends on factors like hardware lifespan, token throughput, and electricity costs. Under certain assumptions, it can be comparable or more expensive than dedicated inference hardware like OpenRouter.

How does inference speed compare between Apple Silicon and OpenRouter?

OpenRouter models like Gemma 4 31b can process 60-70 tokens per second, whereas Apple Silicon typically processes around 10-20 tokens per second, making it slower for real-time applications.

What factors most influence the total cost of local AI inference on Apple Silicon?

The primary factors are hardware cost amortized over its lifespan and token throughput. Electricity costs are relatively minor in comparison.

Could future hardware improvements change this cost comparison?

Yes, faster inference speeds and lower hardware costs could make Apple Silicon more competitive in the future.