# What Would it Cost to Run Claude Opus 5 Locally?

> Cost math on self-hosting a frontier LLM: a $112,000 rig can crawl at 0.1 tokens per second because memory bandwidth, not price, is the real bottleneck.

Published: 2026-08-25
URL: https://daniliants.com/insights/what-would-it-cost-to-run-claude-opus-5-locally/
Tags: local-llm, mixture-of-experts, memory-bandwidth, cost-optimization

---

## Summary

A creator runs real cost math on self-hosting a frontier-class open-weight LLM (Moonshot AI's Kimi K3, used as a stand-in since Anthropic has never published Opus 5's parameter count) versus paying for the Claude API. The hardware sticker price, about $112,000 for a usable rig, turns out not to be the deciding factor; the real killer is memory bandwidth, which can leave four data-center GPUs running at 0.1 tokens/second if the setup isn't matched correctly. At normal individual usage, the API (roughly 70 cents for an hour of agentic coding) beats local hosting on every axis except four specific scenarios.

## Key Insight

- **Mixture-of-experts doesn't shrink storage, only compute.** Kimi K3 is 2.8 trillion parameters, 896 experts per layer, only 16 (3.7%) active per token, but every expert still has to be stored and loaded, since the router picks a different subset each token. "You're not paying to run a small model. You're paying to store a massive one while running a small slice of it."
- **Real minimum footprint:** the smallest usable quantized build (~1.5 bits/parameter, 594GB across 14 shards) needs ~610GB of memory once KV cache and runtime overhead are added, not a textbook estimate, an actual published file size.
- **Three hardware tiers:**

  - *Technically runs:* 4x A100 (160GB total VRAM) + 2TB system RAM, rented at $3.87/hr, weights spill into slower system RAM.
  - *Actually usable:* 7x RTX Pro 6000 Blackwell (96GB each) to keep everything in VRAM, these cards launched at ~$8,500 and were listed at $16,000 by Nvidia in August 2026 due to a memory shortage, so $112,000 in GPUs alone before a case exists, drawing 4.2kW.
  - *No compromises:* a DGX B300 cluster (2.3TB GPU memory per box), data-center scale.

- **The real bottleneck is memory bandwidth, not VRAM capacity ("walking speed" vs "warehouse size").** Every token requires walking the full list of active experts. High-capacity machines have disproportionately low bandwidth: Mac Studio (512GB) ~819GB/s, DGX Spark ~273GB/s, Ryzen AI Max ~215GB/s, versus RTX 5090 ~1,790GB/s or H100 ~3.35TB/s. The person who actually rented 4x A100 and tried it got **0.1 tokens/second** (one token every 10 seconds), abandoning the benchmark after 17 minutes; GPU utilization sat at 0-1%. Properly bandwidth-matched hardware gets ~20 tokens/second, a 200x spread between "technically running" and "usable."
- **Electricity:** the 7-card rig at 4.2kW continuous runs ~$565/month (~$6,786/year) at the US average rate (18.5c/kWh), real money, but dwarfed by the $112k upfront hardware cost.
- **API baseline:** Anthropic's own worked example, a 1-hour Opus 5 coding session (50k input / 15k output tokens), costs about $5 uncached, ~70 cents with cache reads.
- **Breakeven depends entirely on which API you're comparing against, not the hardware.** Two analysts reached opposite conclusions: one compared local rigs to a cheap DeepSeek API (~13c/M blended) and found local essentially never breaks even (an M3 Ultra costs 2.4x the API price in electricity alone; a DGX Spark works out to 12x). The other compared against expensive frontier API pricing and found some configs pay back in 1.6-3 months at full utilization. "Break even is not a property of your rig. It's a property of which bill you're trying to eliminate."
- Inference API prices are falling 50-200x per year (Epoch AI data), which further erodes any local hardware ROI case over time. Opus 5 itself went from $15.75/M to $5.25/M input tokens within a year, a cut, but proof the number can move either direction.
- **Four legitimate reasons to self-host anyway:** (1) industrial-scale token volume every day; (2) data that legally cannot leave the building (healthcare, legal, government); (3) agents running 24/7 where the API meter never stops; (4) vendor dependency risk, since a provider can reprice or deprecate a model unilaterally.
- Even if the compute and bandwidth problems are solved, the model you'd actually be running is Kimi K3, not Opus 5. Anthropic has never released Opus 5's weights and its own launch materials concede Kimi K3 trails proprietary frontier models. "You cannot buy Opus 5 at any price."