Sunk Cost Calculator Shows Local LLM Rig Takes 44 Years to Break Even

Bottom‑line: Local inference on a $3,499 Mac Studio will not pay for itself for decades

The Sunk Cost calculator (https://sunkcost.ai/) predicts a 44‑year break‑even for a Mac Studio M5 Max (64 GB) running the open‑source Qwen‑3.8‑27B model. At the assumed usage of 500 k tokens per day (15 : 1 input‑to‑output ratio) the rig saves only $0.22 per day compared with the OpenRouter API, so the hardware cost is never recovered under current pricing.

How the numbers are derived

Item Value Source
Hardware cost $3,499 Apple Mac Studio M5 Max spec sheet (2026)
Power draw (load) 145 W (stand‑in from M4 Max) Apple documentation, inferred from previous generation
Electricity price $0.17 /kWh (US residential average) U.S. EIA 2025 data
Local throughput 25 tokens / s (estimated) Bandwidth‑bound estimate: 614 GB/s ÷ 16.5 GB per token × 0.75
API throughput 80 tokens / s (OpenRouter) Provider documentation
API price $0.32 in / $2.5 out per M tokens OpenRouter Qwen‑3.8‑27B endpoint (checked 2026‑09‑03)
Daily token mix 468,750 input, 31,250 output 500 k total tokens, 15 : 1 ratio
Daily API cost $0.23 Calculated from token mix and API rates
Daily electricity cost $0.01 (0.145 kW × 0.051 kWh × $0.17)
Daily net saving $0.22 $0.23 – $0.01
Break‑even days 15,942 days (≈ 44 years) $3,499 ÷ $0.22
Break‑even tokens 7.97 B tokens 15,942 days × 500 k tokens/day

The calculator assumes static API prices (no future declines) and continuous operation at the estimated token rate. It also treats the entire hardware purchase as sunk cost, ignoring resale value or other workloads the machine might handle.

Model capability vs. cost

The tool ranks Qwen‑3.8‑27B at 34 points on the Artificial Analysis Intelligence Index v4.3, placing it in the “Sonnet‑class” tier—roughly comparable to Anthropic’s Claude Sonnet 5 or OpenAI’s GPT‑5.6 Luna. While the model is competitive in reasoning benchmarks, its throughput (25 tok/s) is only 31 % of the API’s speed, translating to 15 minutes of waiting per day for a typical 1,000‑token response.

What the calculator leaves out (and why they matter)

  • Idle power: The $0.01/day electricity estimate assumes the machine is always at 145 W load. Real‑world idle consumption adds cost.
  • Prompt‑processing time: Input tokens are assumed free because they are processed 20‑100× faster than generation, but they still consume electricity.
  • Resale value: A Mac Studio retains resale value, which would reduce the effective sunk cost if the hardware is later sold.
  • Alternative workloads: The rig can also be used for rendering, gaming, or other compute tasks, which are not accounted for in the break‑even.
  • Future API price drops: The calculator can toggle a “prices keep falling” option; historically, cloud LLM prices have declined, which would further extend the pay‑back horizon.
  • Frontier models: The comparison excludes proprietary models (e.g., Claude Opus, GPT‑6) that cannot be run locally at any price point.

Community reactions on Hacker News

redox99: “The math is wrong; token‑per‑second should be at least double with MTP and Q8 KV. Even with higher throughput, the rig never pays for itself.”
txrx0000: “I value privacy over cost. Owning the hardware prevents providers from spying on my thoughts.”
ThunderSizzle: “Claude Code costs >$100/month; a $1,350 GPU gave me a 4‑month pay‑off and full data control.”
jrflo: “44 years to break even is absurd for coding work; a $100/month subscription feels far better.”
mcone: “A used HP Omen with a 3090 runs Qwen 3.8 at 57 tok/s for $2k, cutting the pay‑off time dramatically.”
ProjectArcturis: “Local LLMs aren’t about saving money; they’re about autonomy, fine‑tuning, and avoiding lock‑in.”

The comments highlight two recurring themes: privacy/autonomy outweighing pure cost considerations, and hardware choice dramatically affecting throughput (e.g., higher‑end GPUs can double token rates, shortening the pay‑back).

Practical takeaways

  1. Pure monetary break‑even is unrealistic for a consumer‑grade Mac Studio running Qwen‑3.8‑27B under today’s API pricing.
  2. Higher‑performance GPUs (e.g., RTX 3090, A6000) can improve token‑per‑second rates, but even a 2× speedup only halves the 44‑year horizon to ~22 years.
  3. Non‑financial benefits—data privacy, model ownership, and the ability to run custom fine‑tuned models—are the primary motivators for local inference.
  4. Resale and multi‑purpose use can materially improve the economics; a machine that also serves as a render node or development server recoups cost through other revenue streams.
  5. Future API price trends are critical. If cloud providers continue to cut prices, the gap widens; if they rise, the break‑even improves marginally.

Bottom line

The Sunk Cost calculator provides a transparent, data‑driven estimate: a $3,499 Mac Studio running Qwen‑3.8‑27B saves only $0.22 per day, leading to a 44‑year pay‑back. While the financial case is weak, many users pursue local LLMs for privacy, autonomy, and multi‑purpose compute, which the calculator does not quantify.

Sources

Related