{"meta":{"schema_version":"1.1","effective":"2026-Q2","last_audited":"2026-05-19","doc":"Per-SKU reference street price for a single GPU module (chip + carrier, not full DGX/HGX system). Use as F1 TCO acquisition-layer seed. Refreshed quarterly from GRVCI v0_benchmark + public OEM channel reporting. Not a settlement reference — partner pilots should use their own procurement quotes.","audit_notes":"2026-05-19: cross-checked against src/data_feeds/ebay/completed_listings.py SKU_QUERY_TEMPLATES min/max bounds. Adjustments: B200_SXM6 $45k → $40k (closer to 2026-Q2 OEM channel midpoint); GB200_NVL $80k → $42k (was conflating the dual-GPU module total with per-GPU; corrected to per-GPU). All other 21 SKUs verified within their respective eBay completed-listings credible bands.","sources":["GRVCI v0_benchmark (cohort-month-0 T1-new)","OEM channel reporting (Supermicro / Dell / HPE configurators, public)","eBay completed-listings min/max bounds in src/data_feeds/ebay/completed_listings.py (secondary market sanity check)"],"live_api_status":{"msrp_api":"none — datacenter GPUs are channel-priced through OEMs (Supermicro, Dell, HPE). NVIDIA / AMD / Huawei do not publish public MSRP APIs.","tier_t1_ebay_secondary":"WIRED. eBay completed-listings persisted to gpu_capex_observations (migration 012). >=5 obs in 30d → trimmed-median × 1.15 markup. tier=LIVE_EBAY_X1.15.","tier_t2_usaspending_federal":"WIRED. US federal procurement contracts (USAspending /api/v2/search/spending_by_award) persisted to gpu_capex_observations with source='usaspending_contract'. >=3 awards in 180d → median, no markup (real new prices). tier=LIVE_USASPENDING. Preferred over T1 when both meet threshold.","tier_t3_seed":"ALWAYS_AVAILABLE. gpu_capex_seed.json — quarterly-refreshed reference. Falls back when both T1 and T2 below threshold.","operator_actions":"Daily collector cadence requires flops-capex-collector Fly scheduled machine. fly.flops-capex-collector.toml + Dockerfile.capex-collector staged by Lane A (33ce66e). Manual fire: POST /v1/refdata-jobs/gpu-capex/observations/collect (admin scope)."}},"skus":{"h100_sxm5":{"price_usd":28500,"tier":"T1-new","source":"GRVCI v0_benchmark"},"h100_nvl":{"price_usd":32000,"tier":"T1-new","source":"OEM channel"},"h200_sxm5":{"price_usd":35000,"tier":"T1-new","source":"OEM channel"},"h200_pcie":{"price_usd":30000,"tier":"T1-new","source":"OEM channel"},"gh200":{"price_usd":42000,"tier":"T1-new","source":"OEM channel"},"b100_sxm6":{"price_usd":35000,"tier":"T1-new","source":"OEM channel"},"b200_sxm6":{"price_usd":40000,"tier":"T1-new","source":"OEM channel — 2026-Q2 midpoint"},"b300_sxm7":{"price_usd":55000,"tier":"T1-new","source":"OEM channel"},"gb200_nvl":{"price_usd":42000,"tier":"T1-new","source":"OEM channel; per-GPU (GB200 SUPERCHIP is dual-GPU; module total ~$84k)"},"a100_80gb":{"price_usd":11000,"tier":"T1-new","source":"OEM channel"},"l40s":{"price_usd":9500,"tier":"T1-new","source":"OEM channel"},"rtx_6000_ada":{"price_usd":7200,"tier":"T1-new","source":"OEM channel"},"rtx_4090":{"price_usd":2000,"tier":"T1-new","source":"consumer street"},"h20_sxm5":{"price_usd":18000,"tier":"T1-new","source":"China-export channel reporting"},"h800_sxm5":{"price_usd":25000,"tier":"T1-new","source":"China-export channel reporting"},"a800_sxm4":{"price_usd":14000,"tier":"T1-new","source":"China-export channel reporting"},"l20_pcie":{"price_usd":7500,"tier":"T1-new","source":"China-export channel reporting"},"mi250x":{"price_usd":9000,"tier":"T1-new","source":"OEM channel"},"mi300x":{"price_usd":14000,"tier":"T1-new","source":"OEM channel"},"mi325x":{"price_usd":18000,"tier":"T1-new","source":"OEM channel"},"mi350":{"price_usd":22000,"tier":"T1-new","source":"AMD announced pricing (preliminary)"},"ascend_910b":{"price_usd":8000,"tier":"T1-new","source":"estimated; Huawei pricing opaque"},"ascend_910c":{"price_usd":12000,"tier":"T1-new","source":"estimated; Huawei pricing opaque"}},"count":23}