{
  "scenarios": [
    {
      "image_token_share": 0,
      "super_gpu_hours": 3343485.6175972926,
      "nano_gpu_hours": 2060998.1515711644,
      "total_gpu_hours": 5404483.769168457,
      "cost_musd_at_3_5_8": [
        16.213,
        27.022,
        43.236
      ]
    },
    {
      "image_token_share": 0.5,
      "super_gpu_hours": 7623550.03771431,
      "nano_gpu_hours": 4698262.233680319,
      "total_gpu_hours": 12321812.271394629,
      "cost_musd_at_3_5_8": [
        36.965,
        61.609,
        98.574
      ]
    },
    {
      "image_token_share": 1,
      "super_gpu_hours": 11903614.457831327,
      "nano_gpu_hours": 7335526.315789474,
      "total_gpu_hours": 19239140.7736208,
      "cost_musd_at_3_5_8": [
        57.717,
        96.196,
        153.913
      ]
    }
  ],
  "source": "https://arxiv.org/html/2606.02800v4",
  "assumptions": [
    "Rates 3/5/8 USD per GB200 GPU-hour are assumed, not actual contract prices.",
    "Image/video proportions are hypothetical; other modes and varying resolutions not modeled.",
    "Steady-state joint benchmark throughput is applied to pre- and mid-training; runtime is not observed. Image and video rates may be concurrent rather than exclusive: sensitivity is provisional.",
    "Paper prose and table caption disagree on GPU counts; no measured runtime claimed.",
    "Increment over Qwen3-VL only; Edge, Alpamayo and posttraining excluded."
  ],
  "central_estimate": null
}