pricing.json

{
  "_meta": {
    "description": "Real GCP accelerator pricing used throughout this dataroom. Per chip-hour for TPU, per GPU-hour / per instance-hour for L4. USD.",
    "captured": "2026-05-28",
    "sources": [
      "https://cloud.google.com/tpu/pricing",
      "https://cloud.google.com/run/pricing",
      "https://cloud.google.com/compute/gpus-pricing",
      "https://cloud.google.com/products/compute/pricing/accelerator-optimized"
    ],
    "notes": "TPU prices are per chip-hour and per Google's note already include the host VM. L4 Cloud Run price is the GPU portion only; vCPU and memory are billed separately. Cloud Run does NOT offer TPU - TPU is only available via GKE, Compute Engine, and Vertex AI."
  },
  "l4": {
    "cloud_run_gpu_per_second_no_zonal_redundancy": 0.0001867,
    "cloud_run_gpu_per_hour_no_zonal_redundancy": 0.67212,
    "cloud_run_gpu_per_second_zonal_redundancy": 0.0002909,
    "cloud_run_gpu_per_hour_zonal_redundancy": 1.04724,
    "cloud_run_vcpu_per_second": 0.000024,
    "cloud_run_memory_per_gib_second": 0.0000025,
    "compute_engine_g2_standard_4_per_hour_on_demand": 0.7092,
    "compute_engine_g2_standard_4_config": "1x NVIDIA L4, 4 vCPU, 16 GiB",
    "gpu_hbm_gb": 24,
    "bf16_tflops_dense": 121,
    "region_examples": ["us-central1", "us-east4", "us-east1"]
  },
  "tpu": {
    "v5e": {
      "on_demand_per_chip_hour": 1.20,
      "dws_flex_spot_per_chip_hour": 0.60,
      "cud_1yr_per_chip_hour": 0.84,
      "cud_3yr_per_chip_hour": 0.54,
      "hbm_gb": 16,
      "bf16_tflops": 197,
      "region_examples": ["us-central1", "us-east5", "us-west1", "us-west4"]
    },
    "v6e_trillium": {
      "on_demand_per_chip_hour": 2.70,
      "dws_flex_spot_per_chip_hour": 1.35,
      "cud_1yr_per_chip_hour": 1.89,
      "cud_3yr_per_chip_hour": 1.22,
      "hbm_gb": 32,
      "bf16_tflops": 918,
      "region_examples": ["us-east1", "us-east5"]
    },
    "v5p": {
      "on_demand_per_chip_hour": 4.20,
      "cud_1yr_per_chip_hour": 2.94,
      "cud_3yr_per_chip_hour": 1.89,
      "region_examples": ["us-east5", "us-east1"]
    },
    "ironwood_v7": {
      "on_demand_per_chip_hour": 12.00,
      "cud_1yr_per_chip_hour": 8.40,
      "cud_3yr_per_chip_hour": 5.40,
      "region_examples": ["us-central1"]
    }
  },
  "compute_economics": {
    "_note": "Dense bf16 TFLOPs and HBM bandwidth per chip, with per-dollar-hour ratios. Embedding is prefill-only and compute-bound at production batch, so compute_per_dollar governs ingest token economy. See 02-hardware/throughput-ceiling-analysis.md.",
    "bf16_tflops_dense": {"l4": 121, "v5e": 197, "v6e": 918, "v5p": 459},
    "hbm_bandwidth_gb_s": {"l4": 300, "v5e": 819, "v6e": 1600, "v5p": 2765},
    "compute_per_dollar_hour": {"l4": 170.6, "v5e": 164.2, "v6e": 340.0, "v5p": 109.3},
    "compute_per_dollar_vs_l4": {"l4": 1.00, "v5e": 0.96, "v6e": 1.99, "v5p": 0.64},
    "bandwidth_per_dollar_vs_l4": {"l4": 1.00, "v5e": 1.61, "v6e": 1.40},
    "verdict": "v5e is parity with L4 on compute/$ (no token-economy gain even if software-unblocked); v6e has a ~2x ceiling, realistically ~1.3-1.5x on a saturated lane, erased by loss of scale-to-zero at any idle fraction."
  },
  "cost_ratios_vs_l4": {
    "_basis_l4_per_hour": 0.7092,
    "_note": "How many times faster (per device) a TPU chip must run to match L4 tokens-per-dollar. Ratio is invariant to commitment tier because TPU and L4 discounts are comparable.",
    "v5e_on_demand": 1.69,
    "v6e_on_demand": 3.81,
    "v5p_on_demand": 5.92,
    "ironwood_on_demand": 16.92
  },
  "token_economics": {
    "jina_embeddings_api_usd_per_1m_tokens": 0.02,
    "free_tier_tokens_per_key": 10000000,
    "source": "https://jina.ai/embeddings",
    "breakeven_tokens_per_second": {
      "_formula": "cost_per_hour / (3600 * 0.02 / 1e6)  ==  cost_per_hour / 7.2e-5",
      "l4_g2_standard_4": 9850,
      "tpu_v5e_on_demand": 16667,
      "tpu_v6e_on_demand": 37500,
      "tpu_v5p_on_demand": 58333,
      "ironwood_on_demand": 166667
    }
  }
}