Compute Limits on AI Job Replacement
One of the biggest fears around AI is that it will replace jobs. The biggest constraints on AI replacing jobs are how much new compute is being built and how much compute is needed to replace a job.
Model
Token Speed
Output tok/s per 8-GPU node · Source
(See Hardware for details)
(See Hardware for details)
Tokens per Day per Job
Output tokens to replace one full-time role
AI Buildout
Global compute added/year
Fraction for Job Replacement
How much of all AI compute goes to replacing jobs
Efficiency
Catch-all for real-world losses: overhead, downtime, cooling, batching, coordination
—
| Rack | GB200 NVL72 | 36 Grace CPUs + 72 Blackwell GPUs, liquid-cooled |
| GPUs per rack | 72 × B200 | |
| Serving nodes | 9 × 8-GPU | 8-GPU tensor-parallel group |
| IT power | 120 kW | 115 kW liquid + 5 kW air |
| PUE | 1.2 | Modern liquid-cooled facility |
| Facility power | 144 kW | 120 × 1.2 |
| Racks per GW | 6,944 | 1,000,000 ÷ 144 |
- Token speeds — API-measured from Artificial Analysis (April 2026). May differ from bare-metal GB200 throughput.
- Kimi K2.6: 154.6 tok/s — AA (Clarifai). 1T/32B MoE
- Gemini 3.1 Pro: 112.2 tok/s — AA (Google API)
- GPT-5.5: 81.9 tok/s — AA (OpenAI, high)
- Claude Sonnet 4: 67.6 tok/s — AA (Anthropic API)
- Claude Opus 4: 47.7 tok/s — AA (Anthropic API)
- DeepSeek V4 Pro: 38.4 tok/s — AA (DeepSeek API)
- GB200 NVL72 — NVIDIA, The Register, DGX docs
- AI buildout — BloombergNEF: 23.1 GW under construction globally (Sept 2025). US: 15.9 GW, EMEA: 2.9 GW, APAC: 3.2 GW
- Buildout estimates — JLL: 103 GW → 200 GW by 2030 (~19 GW/yr). IEA: ~18–20 GW/yr implied. Sightline Climate: ~36% slippage rate → 10–15 GW/yr actual completions
- 300M jobs at risk — Goldman Sachs (2023): generative AI could expose ~300M full-time jobs globally to automation
