{
  "as_of": "2026-09-10",
  "project_assumptions": {
    "gpu_count": 100000,
    "dense_bf16_pflops_per_gpu": 2.5,
    "days": 120,
    "effective_fraction": 0.4,
    "main_pretraining_fraction": 0.6,
    "effective_project_flops": 1.0368e+27,
    "project_gpu_days": 12000000,
    "main_pretraining_equivalent_full_cluster_days": 72.0
  },
  "compute_budget_ledger": [
    {
      "stage": "cluster_peak_rate",
      "value": 250000000000000000000,
      "unit": "FLOP/s",
      "expression": "100000 * 2.5e+15"
    },
    {
      "stage": "project_duration",
      "value": 10368000,
      "unit": "seconds",
      "expression": "120 * 86400"
    },
    {
      "stage": "peak_project_budget",
      "value": 2592000000000000000000000000,
      "unit": "FLOPs",
      "expression": "2.5e+20 * 10368000"
    },
    {
      "stage": "effective_project_budget",
      "value": 1.0368e+27,
      "unit": "FLOPs",
      "expression": "2.592e+27 * 0.4"
    },
    {
      "stage": "main_pretraining_budget",
      "value": 6.220799999999999e+26,
      "unit": "FLOPs",
      "expression": "1.0368e+27 * 0.6"
    }
  ],
  "definitions": {
    "N_active": "Token-weighted average active parameter count used to approximate the main forward/backward matrix FLOPs, including shared modules and selected experts.",
    "N_total": "Complete model parameter inventory, including all routed experts.",
    "D": "Cumulative token presentations processed in the specified main-pretraining phase; count repeated dataset passes, count each logical token once across tensor/pipeline parallel shards and layers.",
    "eta": "Assumed conversion from hardware peak times elapsed allocated GPU time to useful model-compute budget on the 6ND accounting basis; includes execution, communication, recomputation and downtime overheads.",
    "f": "Assumed fraction of effective project compute allocated to the chosen main-pretraining run.",
    "tau": "Global D/N_total recipe ratio; a bounded tau is an extra modeling assumption, not a per-expert sample count.",
    "r": "N_total/N_active; a whole-model ratio that includes shared modules."
  },
  "hardware_basis": {
    "source": "https://www.nvidia.com/en-us/data-center/gb200-nvl72/",
    "nvl72_gpu_count": 72,
    "nvl72_sparse_bf16_pflops": 360,
    "dense_is_half_sparse": true,
    "per_gpu_dense_bf16_pflops": 2.5,
    "scope": "Dense here is the Tensor Core arithmetic specification; MoE expert routing is a separate notion of sparsity."
  },
  "assumption_status": {
    "gpu_count": "Rounded interpretation of previously discussed public statement; constant full-window allocation is assumed.",
    "per_gpu_peak": "Derived from NVIDIA BF16 specification; chosen arithmetic baseline, not disclosed Astra training precision.",
    "days": "Author scenario, not a disclosed training duration.",
    "eta": "Author scenario, not a measured Astra MFU.",
    "f": "Author scenario, not a disclosed stage allocation.",
    "active_parameters": "Prior author estimate from conditional serving economics.",
    "r_bounds": "Transfer assumption based on Kimi K3 and DeepSeek V4 Pro whole-model ratios.",
    "tau_bounds": "Author-selected fixed tau=30, requiring more cumulative tokens per total parameter; public recipes provide context and V4-Pro is a strict lower bound.",
    "D": "Calculated from assumed budget and active size; no independent Astra token count supplied.",
    "N_total_interval": "Joint feasible scenario envelope; not a statistical confidence interval."
  },
  "token_examples": {
    "one_million_sequences_times_4000_tokens": 4000000000.0,
    "20T_corpus_repeated_8_times": 160000000000000.0
  },
  "equations": {
    "compute": "C_pre = f*C_project = 6*N_active*D_processed",
    "ratio": "r = N_total/N_active",
    "recipe_ratio": "tau = D_processed/N_total",
    "coupling": "r*tau = C_pre/(6*N_active^2)",
    "expert_data": "D_expert ~= D_processed*k/E for balanced routing; k/E differs from N_active/N_total when shared parameters matter"
  },
  "epistemic_status": "Joint conditional envelope, not a confidence interval. Active size, compute allocation, sparsity and recipe-ratio transfer are assumptions. No independent Astra dataset measurement is available.",
  "references": [
    {
      "model": "Kimi-K3",
      "total_parameters_t": 2.8,
      "active_parameters_b": 104,
      "pretraining_tokens_t": null,
      "token_count_qualifier": "not_reported_in_cited_model_card",
      "tokens_per_total_parameter": null,
      "source": "https://huggingface.co/moonshotai/Kimi-K3",
      "total_active_ratio": 26.923076923076923,
      "tokens_per_active_parameter": null
    },
    {
      "model": "DeepSeek-V3",
      "total_parameters_t": 0.671,
      "active_parameters_b": 37,
      "pretraining_tokens_t": 14.8,
      "token_count_qualifier": "reported",
      "tokens_per_total_parameter": 22.056631892697467,
      "source": "https://arxiv.org/abs/2412.19437",
      "total_active_ratio": 18.135135135135137,
      "tokens_per_active_parameter": 400.0
    },
    {
      "model": "Kimi-K2",
      "total_parameters_t": 1,
      "active_parameters_b": 32,
      "pretraining_tokens_t": 15.5,
      "token_count_qualifier": "reported",
      "tokens_per_total_parameter": 15.5,
      "source": "https://arxiv.org/abs/2507.20534",
      "total_active_ratio": 31.25,
      "tokens_per_active_parameter": 484.375
    },
    {
      "model": "DeepSeek-V4-Pro",
      "total_parameters_t": 1.6,
      "active_parameters_b": 49,
      "pretraining_tokens_t": 32,
      "token_count_qualifier": "strict_lower_bound_more_than",
      "tokens_per_total_parameter": 20,
      "source": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro",
      "total_active_ratio": 32.6530612244898,
      "tokens_per_active_parameter": 653.0612244897959
    }
  ],
  "main_envelope": {
    "pretraining_fraction": 0.6,
    "active_parameters_b": [
      300,
      400
    ],
    "total_active_ratio": [
      26.923076923076923,
      32.6530612244898
    ],
    "assumed_tokens_per_total_parameter": [
      30,
      30
    ],
    "has_joint_solution": true,
    "joint_active_parameters_b": [
      325.33060108142297,
      358.28161310024586
    ],
    "total_parameters_t": [
      9.646043429622006,
      10.62304003531177
    ]
  },
  "allocation_sensitivity": [
    {
      "pretraining_fraction": 0.4,
      "active_parameters_b": [
        300,
        400
      ],
      "total_active_ratio": [
        26.923076923076923,
        32.6530612244898
      ],
      "assumed_tokens_per_total_parameter": [
        30,
        30
      ],
      "has_joint_solution": false,
      "joint_active_parameters_b": null,
      "total_parameters_t": null
    },
    {
      "pretraining_fraction": 0.6,
      "active_parameters_b": [
        300,
        400
      ],
      "total_active_ratio": [
        26.923076923076923,
        32.6530612244898
      ],
      "assumed_tokens_per_total_parameter": [
        30,
        30
      ],
      "has_joint_solution": true,
      "joint_active_parameters_b": [
        325.33060108142297,
        358.28161310024586
      ],
      "total_parameters_t": [
        9.646043429622006,
        10.62304003531177
      ]
    },
    {
      "pretraining_fraction": 0.8,
      "active_parameters_b": [
        300,
        400
      ],
      "total_active_ratio": [
        26.923076923076923,
        32.6530612244898
      ],
      "assumed_tokens_per_total_parameter": [
        30,
        30
      ],
      "has_joint_solution": true,
      "joint_active_parameters_b": [
        375.65942021996466,
        400.0
      ],
      "total_parameters_t": [
        11.52,
        12.266430047998846
      ]
    },
    {
      "pretraining_fraction": 1,
      "active_parameters_b": [
        300,
        400
      ],
      "total_active_ratio": [
        26.923076923076923,
        32.6530612244898
      ],
      "assumed_tokens_per_total_parameter": [
        30,
        30
      ],
      "has_joint_solution": false,
      "joint_active_parameters_b": null,
      "total_parameters_t": null
    }
  ],
  "recipe_sensitivity": [
    {
      "pretraining_fraction": 0.6,
      "active_parameters_b": [
        300,
        400
      ],
      "total_active_ratio": [
        26.923076923076923,
        32.6530612244898
      ],
      "assumed_tokens_per_total_parameter": [
        20,
        20
      ],
      "has_joint_solution": true,
      "joint_active_parameters_b": [
        398.4469851812158,
        400.0
      ],
      "total_parameters_t": [
        12.959999999999997,
        13.010513801835614
      ]
    },
    {
      "pretraining_fraction": 0.6,
      "active_parameters_b": [
        300,
        400
      ],
      "total_active_ratio": [
        26.923076923076923,
        32.6530612244898
      ],
      "assumed_tokens_per_total_parameter": [
        30,
        30
      ],
      "has_joint_solution": true,
      "joint_active_parameters_b": [
        325.33060108142297,
        358.28161310024586
      ],
      "total_parameters_t": [
        9.646043429622006,
        10.62304003531177
      ]
    },
    {
      "pretraining_fraction": 0.6,
      "active_parameters_b": [
        300,
        400
      ],
      "total_active_ratio": [
        26.923076923076923,
        32.6530612244898
      ],
      "assumed_tokens_per_total_parameter": [
        40,
        40
      ],
      "has_joint_solution": true,
      "joint_active_parameters_b": [
        300.0,
        310.2809786536805
      ],
      "total_parameters_t": [
        8.353718656060629,
        8.64
      ]
    }
  ],
  "midpoint_example": {
    "active_parameters_b": 350,
    "total_parameters_t": 9.874285714285714,
    "cumulative_pretraining_tokens_t": 296.2285714285714,
    "total_active_ratio": 28.212244897959184,
    "tokens_per_total_parameter": 30,
    "tokens_per_active_parameter": 846.3673469387755
  },
  "independent_data_example": {
    "active_parameters_b": 350,
    "cumulative_tokens_t": 30,
    "implied_pretraining_flops": 6.3e+25,
    "fraction_of_project": 0.06076388888888889
  },
  "scaling_law_sources": [
    "https://arxiv.org/abs/2502.05172",
    "https://arxiv.org/abs/2509.23678",
    "https://arxiv.org/abs/2402.07871"
  ]
}