Apple Silicon costs more than OpenRouter
Cost Comparison Methodology
- Many argue the article’s cost math is biased against local compute:
- It amortizes the full price of a high‑end Mac solely to LLM inference, ignoring that it’s also a general‑purpose laptop.
- It assumes 24/7 heavy usage and often picks pessimistic numbers (high power, high residential electricity, full-time load).
- Others counter that if people are buying maxed‑out Macs or dedicated boxes “for AI,” full amortization is fair, and under those assumptions cloud is clearly cheaper per token.
Hardware Choices & Utilization
- Several suggest a Mac Mini/Studio or cheaper used Macs/GPUs would be a more appropriate comparison than a premium MacBook.
- A recurring point: data centers get much better utilization and efficiency via batching, optimized GPUs, and cheap power, so they will usually win on raw $/token.
- Some note that if you only run local models occasionally, the hardware is “free” in practice, since you’d own a laptop anyway. Others argue that low utilization makes the cost per local token worse, not better.
Electricity, Depreciation, and Resale
- Electricity cost is small relative to hardware depreciation in most scenarios.
- Disagreement on hardware lifespan: some expect 3–5 years of heavy AI use to meaningfully reduce useful life; others call that FUD and claim well-cooled hardware can last a decade+.
- Several note Apple gear’s strong resale value, which the analysis mostly omits.
Token Accounting: Input vs Output
- Multiple commenters say focusing only on output tokens understates cloud cost for agentic workflows, where input tokens can dominate by ~10×.
- Local inference can reuse prompts and caches more aggressively, making “input” effectively cheap, and Mac hardware can prefill much faster than it decodes.
Performance & Model Quality
- Broad agreement: cloud frontier models (e.g., top Anthropic/OpenAI) are still significantly smarter and faster than typical local models.
- Some claim mid‑sized open models (Gemma, Qwen, DeepSeek etc.) are “good enough” for many coding and automation tasks, especially with fine‑tuning, but they do not match frontier performance in hard reasoning.
- Speed is a major pain point for local use; others say even slow local models are fine for asynchronous or background workloads.
Privacy, Control, and Risk
- Many say they choose local not for cost but for:
- Privacy, data sovereignty, and avoiding ToS‑driven censorship.
- Predictable costs and no surprise bills or outages.
- Control over model versioning, parameters, caching, fine‑tuning, and avoiding future rug‑pulls.
Future Pricing & Subsidies
- Several argue cloud token prices are currently subsidized by VC and may rise when the “AI bubble” cools, making local more attractive long‑term.
- Others believe open‑model competition and hardware efficiency improvements will keep inference cheap and competitive.
- Overall: consensus that today cloud wins on pure economics; local wins on control and privacy, with future pricing trends labeled as uncertain.