Owned-hardware TCO is computable and electricity is only 8-20% of it over three years; rent-vs-own break-even is NOT computable, because no accelerator in this corpus has both a priced rental row and a local decode benchmark

Asked for a rent-vs-own break-even, the corpus appears to have everything needed: prices for owned hardware in data/supply/, TDP for almost every accelerator, a kWh rate, and 221 priced cloud rows. It cannot produce a break-even, because the two sides of the comparison never touch. The 18 providers and 6 accelerators in the priced rental table share ZERO accelerators with the accelerators that carry a local decode benchmark. The RTX 4090, RTX 5090, RX 7900 XTX, RX 9070 XT, Arc B580, RTX PRO 6000 and both A100s are priced locally with no rental row; H100 SXM, H200 SXM, MI300X, MI355X, B200 and B300 have rental rows and no local decode benchmark.

A rent-vs-own comparison needs the same accelerator priced both ways, or a defensible per-token price for both. The corpus's cloud rows are AGGREGATE fleet throughput and its local benchmarks are batch-1 single-user decode. Those are different populations, and value-rental-ranking-by-workload-and-the-aggregate-trap already establishes the corpus holds no accelerator with both an aggregate and a single-user row. So the two available figures cannot be divided into each other: a batch-1 tok/s is what one requester experiences, while an aggregate $/Mtok assumes the device is saturated with concurrent requests.

What IS computable: the owned-hardware TCO, which is independent of the question. Model: capex from the cheapest non-cloud supply record per card, electricity at the corpus's own rate of $0.0891/kWh from supply/us-electricity-industrial-tariff-2024 (EIA industrial average, rolling 12 months to July 2026), power at card TDP plus a stated 120W host allowance (CPU, motherboard, PSU losses, idle NVMe), 24/7 operation, 3-year horizon. card capex sys W kWh/yr@30% $/yr $/mo 3yr total nvidia-rtx-4090 $ 2,000 570 1,498 $133 $11 $ 2,400 nvidia-rtx-5090 $ 1,999 695 1,826 $163 $14 $ 2,487 nvidia-rtx-3090 $ 1,100 470 1,235 $110 $9 $ 1,430 amd-radeon-rx-7900-xtx $ 1,406 475 1,248 $111 $9 $ 1,739 amd-radeon-rx-9070-xt $ 829 424 1,114 $99 $8 $ 1,127 intel-arc-b580 $ 200 310 815 $73 $6 $ 418 nvidia-rtx-pro-6000-blackwell $16,624 720 1,892 $169 $14 $17,130 nvidia-a100-80gb-pcie $12,099 420 1,104 $98 $8 $12,394 At 70% utilization the same cards cost $241 more per year in aggregate; the ordering is unchanged and the conclusion with it. WHAT THE MODEL ACTUALLY SAYS: electricity is 15-20% of three-year cost for the $2,000 cards and under 5% for the $12,000-and-up cards. Doubling utilization roughly doubles the power line but leaves it a rounding error against capex. A 450W card and a 300W card differ by about $76/yr at 30% load - not enough to change a purchase decision on its own, which is the opposite of what the joules-per-token framing implies. Power matters for thermals, noise and circuit capacity; it does not decide value. THE HOLE TO FILL: to compute a break-even the corpus needs a SAME-ACCELERATOR pair - one priced cloud row and one batch-matched local benchmark - or an explicit aggregate-to-per-user conversion factor. Neither exists. Until one does, the answer to 'should I buy or rent' is a recommendation, not a corpus-backed number.

FILED 2026-10-11 (parent, in place of a subagent that died on HTTP 429 before making a tool call). All inputs are corpus records; the only non-corpus number is the 120W host allowance, stated here as an assumption. Secondary finding on the capex inputs themselves: ebay-used-consumer-gpu-market prices the RTX 4090 at $2,000, which is the LOW end of a $2,000-$3,000 band whose bulk sits at $2,500-$3,000, so the 4090 capex line is a best case rather than a typical one. The 3-year total for the 4090 should be read as a floor.