The Real Cost-Per-Token Question
Cloud APIs charge per token. Local hardware charges upfront but produces tokens at near-zero marginal cost. The breakeven crossover depends on model class, hardware tier, utilisation, and electricity price. This page works the math for the four hardware tiers (consumer GPU, prosumer workstation, DGX Spark, multi-GPU server) versus three cloud cost classes (frontier, mid-tier open-weight, GPT-4o-class) at five utilisation profiles.
Key Findings
- For 7B-class models at 30 percent workstation utilisation, local breakeven against equivalent cloud APIs is reached in 4-9 months depending on hardware tier.
- For dense 70B-class models, DGX Spark generates only about 4.4 tokens per second, so breakeven against frontier-class cloud APIs takes about 8 months at moderate utilisation and is never reached against cheap open-weight APIs. A sparse 120B model at about 60 tokens per second breaks even in under a month.
- For sporadic developer use (under 10 percent utilisation), cloud APIs are dramatically cheaper, the breakeven horizon stretches to 2-4 years.
- Power and cooling add 15-25 percent to local TCO at typical US electricity rates, more in Europe.
- The cost arbitrage on cloud APIs has compressed in 2026 as inference prices fell, but the gap on data-residency-sensitive workloads still favours local heavily.
Hardware TCO (3-year amortisation)
| Hardware | Up-front $ | Annual power $ (24/7) | 3-yr TCO | Annualised TCO |
|---|---|---|---|---|
| RTX 5090 (card at September 2026 street price, host extra) | $4,300 | $540 | $5,920 | $1,973 |
| Mac Studio M5 Max 128GB | $4,799 | $310 | $5,729 | $1,910 |
| NVIDIA DGX Spark | $4,699 | $184 | $5,251 | $1,750 |
| 2x H100 80GB server | $60,000 | $6,300 | $78,900 | $26,300 |
Power costs assume 24/7 operation at $0.15/kWh US-blended. Realistic utilisation reduces power by 60-80 percent. Hardware prices were corrected on October 2, 2026: DGX Spark is $4,699, the 128GB Mac Studio uses the low end of the $4,799 to $5,400 reported range, and the RTX 5090 uses its street price. DGX Spark power assumes its 140W TDP.
Cloud API Reference Pricing (per 1M output tokens, 2026 rates)
| Class | Representative API | Output $/1M tokens |
|---|---|---|
| Frontier closed | Claude Opus 4.7, GPT-5 Pro | $60-75 |
| Mid-tier closed | Claude Sonnet 4.6, GPT-5 mini | $10-15 |
| Frontier open-weight (cloud-served) | Llama 4 70B, Qwen 3 235B | $0.50-2.00 |
| Small open-weight (cloud-served) | Llama 4 8B, Qwen 3 32B | $0.10-0.50 |
Pricing aggregated from Artificial Analysis, OpenAI pricing, Anthropic pricing, and Google AI pricing.
Breakeven Analysis: 7B Model on Mac Studio M5 Max vs Cloud API
The llama.cpp scoreboard records the M5 Max at about 120 tps on Llama 2 7B Q4. At 30 percent utilisation (7.2 hours/day), the device produces approximately 3.1M tokens/day = 93M tokens/month = ~1.14B tokens/year.
| Cloud cost class | $/1M tokens | Annual cloud cost equivalent | Months to breakeven (vs $1,910/yr Mac TCO) |
|---|---|---|---|
| Frontier ($65) | $65 | $73,800 | ~0.3 months |
| Mid-tier ($12) | $12 | $13,620 | ~1.7 months |
| Frontier open-weight ($1.25) | $1.25 | $1,419 | ~16 months |
| Small open-weight ($0.30) | $0.30 | $341 | ~67 months (5.6 years) |
Decisive: against frontier-class cloud APIs, local hardware pays back in weeks. Against cheap small-model APIs, the math reverses entirely.
Breakeven Analysis: Large Models on DGX Spark vs Cloud API
DGX Spark's output depends heavily on the model. Ollama measured a dense Llama 3.1 70B at 4.4 tps. The llama.cpp community measured the sparse gpt-oss 120B at 60.6 tps. At 30 percent utilisation that is about 42M tokens per year for the dense 70B and about 573M tokens per year for gpt-oss 120B.
| Cloud cost class | Dense 70B: annual cloud equivalent | Dense 70B: months to breakeven | gpt-oss 120B: annual cloud equivalent | gpt-oss 120B: months to breakeven |
|---|---|---|---|---|
| Frontier ($65) | $2,706 | ~8 months | $37,260 | ~0.6 months |
| Mid-tier ($12) | $499 | ~42 months | $6,879 | ~3 months |
| Frontier open-weight ($1.25) | $52 | Never | $717 | ~29 months |
Breakeven is measured against a $1,750 per year DGX Spark TCO. An earlier version of this page assumed about 40 tps on a dense 70B model, which measured results do not support.
Hidden Costs Worth Naming
- Engineering time: 20-80 hours one-time setup (model serving, monitoring, fallback paths). At loaded engineering rates, this can add $5K-20K to local TCO.
- Scaling friction: cloud handles 10x load spikes for free; local hardware caps at its rated tps.
- Model upgrades: cloud auto-upgrades to better models; local requires manual model swap and re-validation.
- Reliability and uptime: single-machine SLA is materially worse than cloud; redundancy adds 2x hardware cost.
- Compliance benefit (the other direction): data residency, HIPAA, defence regulations often make local mandatory regardless of cost arithmetic.
When Local Wins Decisively
- Workloads using frontier-class models (Claude Opus, GPT-5) at moderate-to-high utilisation
- Data-residency-mandatory workloads (defence, healthcare, EU regulated finance)
- Continuous fine-tuning workflows where iteration speed matters more than cost
- Privacy-sensitive personal AI (developer assistants, on-device productivity)
When Cloud Wins Decisively
- Sporadic or bursty workloads (under 10 percent utilisation)
- Workloads on cheap open-weight cloud APIs (Llama 4 8B at $0.30/1M tokens is hard to beat)
- Need for SOTA frontier models that change every quarter
- Teams without engineering bandwidth for self-hosting infrastructure
Brand Visibility Implications
The cost dynamics push enterprise AI workloads toward local hosting in two specific directions: frontier-class models for power users (where cost savings compound), and small open-weight models for high-volume internal automations (where data-residency and embedding-vector privacy matter). Both directions remove brand-relevant queries from cloud-API observability. For brands tracking AI visibility, this is the structural reason the open-source LLM blind spot is widening, not narrowing.
Methodology
Hardware costs from current vendor list pricing (NVIDIA, Apple, NewEgg). Cloud API costs from Artificial Analysis 2026 snapshots and vendor pricing pages. Power costs use $0.15/kWh blended US rate. Throughput from our companion tokens-per-second benchmarks page. Real workloads diverge from utilisation assumptions; the analysis is sensitive to those, sanity-check with your actual usage.
How Presenc AI Helps
Presenc AI tracks brand visibility across both cloud-served and locally-served LLM deployments, the only AI-visibility platform that does both. For enterprises modelling the local-vs-cloud decision, our cross-deployment data informs the brand-visibility risk side of the equation.