Atlas of the Present Atlas · v0.1

As of 28 October 2026

Dossier 02

Energy and Chips

v2.2-draft · unchecked

1. As of

Date
Version
v2.2-draft
Author / model
Atlas generator
Reviewer
unchecked (Josef)

2. In one sentence

Economic value per joule (Zeinali et al.; Stanford/UC Berkeley/UT Austin arXiv:2610.30858, E2): on 220 public GDPval occupational tasks across 44 occupations, GLM 5.2 delivers the highest net value ($72.27 per task) for 0.430 MJ of measured energy — about four hours of active laptop use —, Gemma 4 31B 56% of that with 23% of the energy, roughly 2.5 times the net value per joule; the same model on one instead of eight H100s delivers almost the same value, eight are only 22% faster and use 3.24 times the energy — dollar value here is expert time × median wage, not measured GDP. What a token costs (Niu, Sencer, Zhang, Chen; NVIDIA/Texas Tech/TACC arXiv:2610.28576, E2, measurement): GPU-board energy per generated token during decoding follows, on NVIDIA GH200 and GB300 for three dense models, a simple formula of shared weight bytes, private KV-cache bytes and realized time per token — 374 held-out runs within 3.73%; the largest tested batch needs 33.2–46.8% less energy per token than the smallest, and for Qwen2.5-7B it rises with long history from 0.323 to 0.638 J/token — GPU board only, dense models only, decode time as an input. AI memory above its standard test (Olgun et al. ETH Zurich/NYU arXiv:2610.14244, E2, measurement): on 36 real HBM2 chips a deliberately constructed access pattern draws on average 23.6% (max 35.4%) more power than the most power-hungry JEDEC measurement loop; reads cost 3.1× refresh, toggling bus signals +29.4%; the open model fits 36 chips at 7.6–8.1% error and the HBM3E memory power of three H200s within 2.4–5.3% — FPGA stress test ≠ GPU workload. Carbon–SLO (Topcu et al. Penn State arXiv:2610.14571, E2, simulation): facility MILP cuts up to 60% CO₂ vs round-robin at ~3% average SLO violation — simulator ≠ fleet. Meta fbnic (Gupta et al. arXiv:2610.07644, E2, self-report): in-house network card on hundreds of thousands of hosts cuts unplanned unavailability from 9.11% to 0.74%. MIT (McCourt, Fiete, Chuang arXiv:2610.10311, E2, simulation): LLMs trained with hardware faults get more robust with size — energy savings on faulty hardware remain a conjecture. The energy limit of AI is first a task and system limit: Meta 150 MW cluster (Ardestani et al. arXiv:2605.24461v3, 2026, E2): ~83K GB200 under a fixed power budget — a 960 W (80% TDP) power limit instead of 1200 W yields ~10–11% more cluster throughput per the operator, because more GPUs fit the budget; static imbalance leaves up to 10% of power stranded; one operator, self-report; inference energy varies by orders of magnitude by task class; sample-count energy (Kashaniyan/Jannesari 2026, E1) shows that at fixed N=8, 8×1 vs 1×8 costs 4.64–4.86× gross GPU-device energy on A100 — candidate count ≠ execution schedule; phase-decoupled power control (Kim et al. 2026, E1) on 8×B200 MoE: vendor Max-Q modest (+8.6% tokens/J, +5.2% latency), calibrated phase-decoupled +20.4% tokens/J at +3.5% e2e — lab/node ≠ fleet; dense recovers ~5× less; globally 2010–2018 +6% energy at +550% compute (Masanet); US DC risen again (LBNL); IEA ~415 TWh 2024 E2; Epoch ~20 million AI chips / 3.3× capacity since 2022 (E2); TSMC CoWoS 5.5-reticle now / 14-reticle 2028 (E2); DGX B200 ~14.3 kW; HBM/packaging/lithography tight.

Established now · E1 / E2

3. What works today

Economic value per joule: the most valuable AI system is not the most efficient, eight GPUs instead of one cost 3.24 times the energy (Zeinali et al.)

  1. E2

    Zeinali, Saad-Falcon, Narayan (Stanford), Abbasi (UC Berkeley), Hartley (UT Austin), Hennessy, Ré, Mirhoseini (Stanford; arXiv:2610.30858v1, submitted 2026-10-22, 35 pp., PDF opened 2026-10-25): 220 public GDPval tasks across 44 occupations in 13 categories (weighted by sector GDP and employment); the value of an output = expert completion time × occupation median wage (US OEWS), credited only on a win or tie against the expert deliverable (judge Gemini 3.1 Flash-Lite; three judges 72.3–77.2% pairwise agreement, 23 of 32 = 71.9% with human experts); energy measured end to end (GPU, CPU, DRAM, including tool use and waits), one run per task and configuration. Metrics: net value per joule (NVJ, value minus execution cost, USD/MJ) and time saving per joule (TSJ, h/MJ). Five locally served open models (OpenHands, vLLM): GLM 5.2 (eight H200s) delivers the highest net value, $72.27 per task (output value $74.19, 1.64 expert hours) for 0.430 MJ — about four hours of active laptop use or 10.2 minutes of one H200 at 700 W; Gemma 4 31B delivers 56% of that net value with 23% of the energy, about 2.5 times the NVJ and 2.8 times the TSJ; GPT-OSS 120B uses 68% less energy than GLM 5.2 at nearly identical cost and completion time.

    Zeinali et al. Joules-to-GDP. Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?. https://arxiv.org/abs/2610.30858. As of arXiv:2610.30858v1 (submitted 2026-10-22; cs.LG; 35 pp.); PDF research/pdfs/arXiv2610.30858.pdf; code github.com/HazyResearch/joules_to_gdp and dataset not opened. Checked 2026-10-25. Type: Paper (arXiv; PDF opened; 220 GDPval tasks, measured GPU/CPU/DRAM energy, LLM judge).
  2. E2

    Same source, occupations and hardware: energy reshuffles the rankings — Architecture & Engineering (e.g. CAD) delivers about 10% more net value per task than Computer & Mathematical (e.g. website features), but per joule Computer & Mathematical delivers about 27% more; the Production category (e.g. schedules to clear backlogs) drops from 2nd in time saving per task to 6th per joule. Qwen3.6-27B on one accelerator each: NVJ 535.5 (B200), 402.2 (H100), 391.3 (H200), 124.1 USD/MJ (AMD MI350X) — a spread of 4.32×, TSJ spread 9.27×. Going from one to eight H100s cuts completion time by only 22% while energy rises 3.24-fold at nearly the same output value, because all eight GPUs draw power throughout and communication plus sequential tool calls limit the gain. Against experts who themselves use AI (about 23% less time), GLM 5.2 matches their deliverable only about half as often.

    Zeinali et al. Joules-to-GDP. Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?. https://arxiv.org/abs/2610.30858. As of arXiv:2610.30858v1 (submitted 2026-10-22; cs.LG; 35 pp.); PDF research/pdfs/arXiv2610.30858.pdf; code github.com/HazyResearch/joules_to_gdp and dataset not opened. Checked 2026-10-25. Type: Paper (arXiv; PDF opened; 220 GDPval tasks, measured GPU/CPU/DRAM energy, LLM judge).

Joules per token: weight bytes, KV bytes and time explain decode energy on GH200/GB300 within 3.73% (Niu et al.)

  1. E2

    Niu (NVIDIA), Sencer (Texas Tech), Zhang (TACC), Chen (Texas Tech; arXiv:2610.28576v1, submitted 2026-10-20, PDF opened 2026-10-21): “Token-Energy Law” — GPU-board energy per useful output token during dense decoding with full attention history (BF16, FlashAttention) is described as an affine function of three quantities: shared weight bytes per token (weights ÷ realized batch size), private KV-cache bytes per token (growing with history) and realized decode time per token; coefficients calibrated per model and hardware/runtime. Measured on Qwen2.5-7B, Qwen2.5-14B and Mistral-7B on NVIDIA GH200 (Hopper, 144 GB HBM3e, 700 W power limit, vLLM; also tensor-parallel across two GPUs) and GB300 (Blackwell, 1,300 W). Across 374 held-out runs all condition means lie within 3.73% of measurement; the same coefficients moved from GPU0 to GPU1 of the same node without refitting: max 3.72%. Without the time term (bytes only) the maximum error on GB300 rises to 28.2–35.5%, with time only to 10.97–29.40%.

    Niu et al. Token-Energy Law. The Token-Energy Law of LLM Decode: Active Bytes and Realized Time Account for Joules per Token. https://arxiv.org/abs/2610.28576. As of arXiv:2610.28576v1 (submitted 2026-10-20; cs.LG); PDF research/pdfs/arXiv2610.28576.pdf. Checked 2026-10-21. Type: Paper (arXiv; PDF opened; measurement on GH200/GB300; lead author NVIDIA).
  2. E2

    Same source: at fixed model, platform and history, the largest tested batch needs 33.2–46.8% less energy per token than the smallest (batching spreads the weight-reading work); longer history raises energy in every series — Qwen2.5-7B on GH200 at batch 12: 0.323 → 0.609 → 0.638 J/token at 6,144 / 10,240 / 14,336 tokens of history, where the jump above 8,192 tokens also reflects a runtime fallback to slower execution. The logical byte account matches measured HBM read volume at 0.03–0.07% mean error (177 profiles). The fitted time coefficient (~139 W GH200, ~199–201 W GB300) should not be read as measured idle power, per the authors; energy is gross, with no idle baseline subtracted.

    Niu et al. Token-Energy Law. The Token-Energy Law of LLM Decode: Active Bytes and Realized Time Account for Joules per Token. https://arxiv.org/abs/2610.28576. As of arXiv:2610.28576v1 (submitted 2026-10-20; cs.LG); PDF research/pdfs/arXiv2610.28576.pdf. Checked 2026-10-21. Type: Paper (arXiv; PDF opened; measurement on GH200/GB300; lead author NVIDIA).

HBM memory: stress pattern draws 23.6% more than the strongest JEDEC test loop (Olgun et al.)

  1. E2

    Olgun, Galanopoulos, Kakolyris, Yüksel, Luo, Bostancı, Mutlu (ETH Zurich; New York University; arXiv:2610.14244v1 of 2026-10-09, extended version of a MICRO 2026 paper, PDF opened 2026-10-18): by their account the first comprehensive experimental power characterization of real HBM2 chips — 36 chips (8 GB each; vendor not disclosed by the board maker, likely Samsung per the authors) on 18 AMD Alveo U55C FPGA boards at 600 MHz (1,200 MT/s), current read from the board sensors. A read draws more current than any other tested operation, 3.1× refresh; read current varies by 8.1% across bank groups and 5.3% across banks; toggling all internal bank-group bus signals raises power by 29.4% versus keeping them constant. An access pattern built from these findings consumes on average 23.6% (at most 35.4%) more power than the most power-hungry JEDEC measurement loop (IDD4R) — the standard test loops do not capture the worst case.

    Olgun et al. HBM2 power. Understanding the Power Consumption of Modern High Bandwidth Memory: Experimental Analysis and Modeling Using Real HBM2 DRAM Chips. https://arxiv.org/abs/2610.14244. As of arXiv:2610.14244v1 (2026-10-09; cs.AR; extended version of a MICRO 2026 paper); PDF research/pdfs/arXiv2610.14244.pdf; model github.com/CMU-SAFARI/HBM-Power (not run). Checked 2026-10-18. Type: Paper (arXiv; PDF opened; measurement on 36 chips + 3 H200).
  2. E2

    Same source: the open power model built from this (Ayna, on top of DRAMPower) predicts the power of the 36 chips across six microbenchmarks and a real Llama-3.1-8B decode run at 7.6–8.1% mean absolute percentage error (for comparison DRAMSim3 21.6%, FGDRAM 22.6% with the all-zeros data pattern). Extrapolated to HBM3E, it matches the memory power reported by nvidia-smi on three NVIDIA H200 systems under a read-heavy microbenchmark (4.52 TB/s, 94% of peak) at 5.3 / 2.4 / 4.5% error for minimum / nominal / maximum supply voltage; idle 12.0–20.3%. Measured HBM3E memory power reaches up to 248 W on average across the three H200s — about 35% of the GPU’s 700 W budget.

    Olgun et al. HBM2 power. Understanding the Power Consumption of Modern High Bandwidth Memory: Experimental Analysis and Modeling Using Real HBM2 DRAM Chips. https://arxiv.org/abs/2610.14244. As of arXiv:2610.14244v1 (2026-10-09; cs.AR; extended version of a MICRO 2026 paper); PDF research/pdfs/arXiv2610.14244.pdf; model github.com/CMU-SAFARI/HBM-Power (not run). Checked 2026-10-18. Type: Paper (arXiv; PDF opened; measurement on 36 chips + 3 H200).

Carbon–SLO: facility MILP cuts up to 60% CO₂ at ~3% SLO violation (Topcu et al.)

  1. E2

    Topcu, Cim, Ozturk, Kandemir (Penn State; arXiv:2610.14571v1 of 2026-10-10; PDF opened 2026-10-14): latency-centric geo-distributed LLM-serving simulator (continuous batching, queueing, inter-region RTT; calibrated A100 latency/power) with hierarchical schedulers. Facility MILP: up to 60% carbon reduction vs round-robin; average SLO violation ~3% (TTFT/TPOT under MLPerf-style budgets).

    Topcu et al. Carbon–SLO. Carbon-Aware Geo-Distributed LLM Serving under Latency SLOs. https://arxiv.org/abs/2610.14571. As of arXiv:2610.14571v1 (2026-10-10); PDF research/pdfs/arXiv2610.14571.pdf. Checked 2026-10-14. Type: Paper (arXiv; PDF opened; simulation).
  2. E2

    At low load: MILP carbon savings 53.8–59.6% across models within ≤1.3% SLO violation; at high load savings shrink to ~32–38%; savings grow with the CI coefficient of variation across DCs. Greedy α=0 (carbon-aggressive) can match savings but violates SLO for up to ~55% of requests. Simulation; node calibration on Llama-3.1-8B (TPOT MAPE <5%, TTFT ~9.1%, energy ~2.5%); Mixtral end-to-end not validated on hardware; A100-only models.

    Topcu et al. Carbon–SLO. Carbon-Aware Geo-Distributed LLM Serving under Latency SLOs. https://arxiv.org/abs/2610.14571. As of arXiv:2610.14571v1 (2026-10-10); PDF research/pdfs/arXiv2610.14571.pdf. Checked 2026-10-14. Type: Paper (arXiv; PDF opened; simulation).

In-house network card: Meta fbnic cuts unplanned unavailability 12× (Gupta et al.)

  1. E2

    Gupta et al. (Meta Platforms, 28 authors; arXiv:2610.07644v1 of 2026-10-06; PDF opened 2026-10-10): fbnic, an in-house multi-host NIC, runs on hundreds of thousands of production hosts. Vendor multi-host NICs with shared firmware and buffers produced cascading isolation failures over seven years; fbnic isolates physically. Over ten months: unplanned unavailability (T1) from 9.11% to 0.74% (12×), 37% lower mean time to repair, 2.3× fewer hardware swaps vs vendor NICs on the same platform. The price: the entire operational burden (hardware-in-the-loop CI, firmware orchestration, repair automation) now sits with the hyperscaler. Operator self-report, no external audit.

    Gupta et al. Meta fbnic. From ASIC to Fleet: Lessons from Building and Operating a Hyperscaler NIC. https://arxiv.org/abs/2610.07644. As of arXiv:2610.07644v1 (2026-10-06); PDF research/pdfs/arXiv2610.07644.pdf. Checked 2026-10-10. Type: Paper (arXiv; PDF opened; operator self-report).

Inference energy depends on the task

  1. E1

    88 models, 10 tasks, 1,000 inferences: text classification 0.002 kWh / 1,000; image generation 2.907 kWh / 1,000 (σ 3.31). Factor >1,000. Generative ≫ discriminative; multi-purpose more expensive than task-specific. BLOOMz-7B: training+FT parity about 593 million inferences. Limits: sequential inferences, open models 2023, one hardware setting (A100).

    Luccioni, Jernite, Strubell. Power Hungry Processing: Watts Driving the Cost of AI Deployment?. https://arxiv.org/html/2311.16863. As of 2023-11-28. Checked 2026-08-28. Type: Paper (FAccT 2024).

Training energy of individual models

  1. E1

    BLOOM-176B: 1,082,990 GPU-h, 433,196 kWh dynamic, grid 57 gCO₂eq/kWh → 24.7 t dynamic; plus idle 14.6 t and embodied 11.2 t → 50.5 t lifecycle training. E1 for the BLOOM measurement; the paper’s cross-model table (GPT-3/OPT) is E2.

    Luccioni, Viguier, Ligozat. Estimating the Carbon Footprint of BLOOM, a 176B Parameter Language Model. https://arxiv.org/html/2211.02001. As of 2022-11. Checked 2026-08-28. Type: Paper.
  2. E1

    Strubell 2019: measured GPU power 2018–19 (Titan-X/1080-Ti, PUE 1.58). Transformer-big ~192 lb CO₂e; BERT-base ~1,438 lb. The NAS extrapolation (626,155 lb) is corrected by Patterson — do not read it as a world footprint.

    Strubell et al. Energy and Policy Considerations for Deep Learning in NLP. https://arxiv.org/html/1906.02243. As of 2019-06. Checked 2026-08-28. Type: Paper (ACL 2019).
  3. E2

    Patterson 2022: UMass NAS estimate 284 tCO₂e vs. actually 3.2 t (88×). Google ML energy 10–15% of Google total energy 2019–2021, of which ~⅗ inference / ⅖ training. E2 (one hyperscaler measurement); not E1 for “the world’s ML footprint is shrinking”.

    Patterson et al. The Carbon Footprint of Machine Learning Training Will Plateau, Then Shrink. https://arxiv.org/abs/2204.05149. As of 2022-04-11. Checked 2026-08-28. Type: Paper (Google/Berkeley).

Datacenter power: the flat era is broken

  1. E1

    Masanet et al. Science 2020 (Policy Forum, bottom-up model): global data-center electricity around 194 TWh (2010) → 205 TWh (2018), about 1% of world electricity; +6% energy while compute instances +550%. Intensity −20%/year. IT devices 92 → 130 TWh; infrastructure (cooling/power) falls and mostly offsets IT growth. Traditional 79% of instances (2010) → cloud including hyperscale 89% (2018). Authors: “we estimated” / “around”. 2010–2018, global only — not IEA 2024/26, not LBNL US 2023/24.

    Masanet et al. Recalibrating global data center energy-use estimates. https://datacenters.lbl.gov/sites/default/files/Masanet_et_al_Science_2020.full_.pdf. As of 2020-02-28. Checked 2026-08-31. Type: Paper (Science Policy Forum; LBNL-hosted journal PDF).
  2. E1

    Masanet’s flat phase is global 2010–2018. LBNL 2024 (source 13) measures the US break afterward: ~60 TWh 2014–16 → 176 TWh 2023. Not the same series, not the same geography. 205 TWh (world 2018, Masanet) ≠ 76 TWh (US 2018, LBNL).

    Masanet et al. Recalibrating global data center energy-use estimates. https://datacenters.lbl.gov/sites/default/files/Masanet_et_al_Science_2020.full_.pdf. As of 2020-02-28. Checked 2026-08-31. Type: Paper (Science Policy Forum; LBNL-hosted journal PDF).Shehabi et al. (LBNL 2024). 2024 United States Data Center Energy Usage Report. https://eta-publications.lbl.gov/sites/default/files/2024-12/lbnl-2024-united-states-data-center-energy-usage-report.pdf. As of 2024-12. Checked 2026-08-28. Type: Official lab report (LBNL-2001637).
  3. E1

    US DC power: ~60 TWh 2014–16, 76 TWh 2018 (1.9%), 176 TWh 2023 (4.4%). GPU-accelerated servers are the break. E1 for the 2014–2023 best estimate; the 2028 range 325–580 TWh is E2.

    Shehabi et al. (LBNL 2024). 2024 United States Data Center Energy Usage Report. https://eta-publications.lbl.gov/sites/default/files/2024-12/lbnl-2024-united-states-data-center-energy-usage-report.pdf. As of 2024-12. Checked 2026-08-28. Type: Official lab report (LBNL-2001637).
  4. E2

    LBNL 2025 update: 2024 actual 192 TWh = 4.7% of US electricity (GPU shipments 2023–24 revised down). 2030 reference 649 TWh = 11.8%; compounded 521–843 TWh. One update model, June 2026.

    Smith et al. (LBNL 2025). United States Data Center Energy Usage Report: 2025 Update. https://doi.org/10.71468/P1RP4F. As of 2026-06-18. Checked 2026-08-28. Type: Official lab report (LBNL-2001758).
  5. E2

    Global 2024 ~415 TWh, ~1.5% of world electricity; +12%/year since ~2017. US 45%, China 25%, Europe 15%. IEA: uncertainty “substantial” in the demand chapter.

    IEA. Energy and AI — Executive summary. https://www.iea.org/reports/energy-and-ai/executive-summary. As of 2025-04. Checked 2026-08-28. Type: Official report.IEA Demand. Energy demand from AI. https://www.iea.org/reports/energy-and-ai/energy-demand-from-aI. As of 2025-04. Checked 2026-08-28. Type: Official report (chapter).
  6. E2

    EU-27, 2022: DCs 45–65 TWh (1.8–2.6% of EU electricity). No official statistic; literature synthesis. Not a 2026 update.

    Kamiya, Bertoldi (JRC). Energy Consumption in Data Centres and Broadband Communication Networks in the EU. https://doi.org/10.2760/706491. As of 2024-02. Checked 2026-08-28. Type: Official report (JRC135926).

Chips: specification, foundry, lithography

  1. E1

    H100 SXM: FP8 tensor 3,958 TFLOPS (with sparsity), 80 GB, 3.35 TB/s, NVLink 900 GB/s, max TDP up to 700 W (configurable). E1 for the published specification; not for field utilization (LBNL ~70% of rated).

    NVIDIA. NVIDIA H100 GPU (product specifications). https://www.nvidia.com/en-us/data-center/h100/. As of 2026-08-28. Checked 2026-08-28. Type: Manufacturer specification.
  2. E1

    DGX B200 (manufacturer page, checked 2026-08-28): 8× NVIDIA Blackwell GPUs; GPU memory 1,440 GB total, 64 TB/s HBM3e; FP4 tensor 144 PFLOPS sparse / 72 PFLOPS dense; FP8 tensor 72 PFLOPS sparse (dense = ½); NVLink 14.4 TB/s aggregate; system power about 14.3 kW max; 10 rack units. E1 for the published specification; not for field utilization, not for token cost (the page’s SemiAnalysis FAQ is not a warrant).

    NVIDIA DGX B200. NVIDIA DGX B200 (product specifications). https://www.nvidia.com/en-us/data-center/dgx-b200/. As of 2026-08-28. Checked 2026-08-28. Type: Manufacturer specification.
  3. E2

    Epoch 2023, 47 ML accelerators: FP32 FLOP/s 2× / 2.3 years; energy efficiency 2× / 3.0 years; DRAM capacity and bandwidth 2× / ~4 years (memory wall). Peak ≠ utilization. EUV reticle limit 858 mm².

    Hobbhahn, Heim, Aydos. Trends in machine learning hardware. https://epoch.ai/publications/trends-in-machine-learning-hardware. As of 2023-11-09. Checked 2026-08-28. Type: Dataset/report.
  4. E2

    Epoch Trends (as of 2026-02-05): AI chip stock 3.4×/year; frontier LM training compute 5×/year since 2020; GPU FLOP/s/W 1.34×/year since 2008. E2 (research NGO; underlying inputs not audited). The 5×/year LM-frontier dashboard figure is corroborated as of 2026-08-31 by the May 2024 Epoch report in dossier 01 (different sample/window; figures there).

    Epoch Trends. Trends in Artificial Intelligence. https://epoch.ai/trends. As of 2026-02-05. Checked 2026-08-28. Type: Dataset/dashboard.
  5. E2

    Epoch Topic Overview AI Chips (pagefind 1 May 2026): five largest designers roughly 20 million AI chips by Q4 2025; Nvidia about half of units and over 66% of computing capacity. Data insight (CSV 16 Feb 2026): sold capacity in H100-equivalents 3.3×/year since 2022 (90% CI 2.7–4.1×; doubling 7 months, CI 6–8). Not the 3.4× dashboard (source 11) and not 5.3× training FLOP (D01 Sevilla). Inputs: revenue/filings/analysts, not a chip census.

    Epoch Chips Overview. Topic Overview: AI Chips. https://epoch.ai/publications/chips-topic-overview. As of 2026-05-01. Checked 2026-09-01. Type: Topic overview (research NGO).Epoch chip production. Global AI computing capacity is doubling every 7 months. https://epoch.ai/data-insights/ai-chip-production. As of 2026-02-16. Checked 2026-09-01. Type: Data insight (CSV stamp 16 Feb 2026).
  6. E2

    Overview, ML Hardware DB: flagship release price P100 $5,700 (2016) → H100 $34,000 (2022); H100 seventeen times as much computation per dollar. Performance per dollar roughly 2× / 2.5 years since the early 2010s. Leading chips ~40% more energy-efficient per year (~2.7-year doubling); B200 (2024) roughly three times computation/watt vs A100 (2020). One chip around 1,000 W at full load; AI DC power late 2025 roughly tens of gigawatts — order of magnitude, not a point value. Jan 2026 spending analysis: chips+compute time 54–62% of spend where breakdowns exist.

    Epoch Chips Overview. Topic Overview: AI Chips. https://epoch.ai/publications/chips-topic-overview. As of 2026-05-01. Checked 2026-09-01. Type: Topic overview (research NGO).
  7. E1

    TSMC FY 2025: wafer shipments 15.0 million 12-inch equivalents; advanced (≥7 nm) 74% of wafer revenue; N3 24%. CoWoS/InFO/SoIC named, without kwpm. E1 for AR figures; E2 for the node schedule.

    TSMC. 2025 Annual Report, Letter to Shareholders (ch. 1 PDF). https://investor.tsmc.com/static/annualReports/2025/english/pdf/2025_tsmc_ar_e_ch1.pdf. As of 2025. Checked 2026-08-28. Type: Company report.
  8. E2

    TSMC 2026 North America Tech Symposium PR (22 Apr 2026, PDF opened): “The Company is now producing 5.5-reticle size CoWoS.” E2 (vendor PR) for current 5.5-reticle production. No kwpm figure in this PDF — CoWoS kwpm remains E0.

    TSMC Tech Symposium 2026. TSMC Debuts A13 Technology at 2026 North America Technology Symposium. https://pr.tsmc.com/system/files/newspdf/attachment/49337b40ff139d51d533076cf7a945b30e107e07/2026%20Tech%20Symposium%20%28E%29_Final_wmn.pdf. As of 2026-04-22. Checked 2026-09-02. Type: Company PR PDF (Tech Symposium; news HTML Cloudflare-blocked).
  9. E1

    ASML FY 2025: total net sales €32.667 billion; EUV systems recognized 48 (2024: 44). Of those, 44 NXE in 2025; High-NA EXE arithmetically 4. E1 for FY units; 2026 guidance E3.

    ASML PR. ASML reports €32.7 billion total net sales and €9.6 billion net income in 2025. https://www.asml.com/en/news/press-releases/2026/q4-2025-financial-results. As of 2026-01-28. Checked 2026-08-28. Type: Company report.ASML AR. 2025 Annual Report — Financial performance section. https://ourbrand.asml.com/m/419103cb23dfeaa4/original/asml-2025-annual-report-financial-performance-section.pdf. As of 2025. Checked 2026-08-28. Type: Company report.

Critical materials — list and concentration

  1. E1

    USGS 2025 List of Critical Minerals (60), including gallium, germanium, silicon, tungsten. The US imported 80% of rare earths used in 2024. No AI-specific tonnage in this source.

    USGS. Interior Department releases final 2025 List of Critical Minerals. https://www.usgs.gov/news/science-snippet/interior-department-releases-final-2025-list-of-critical-minerals. As of 2025-11-06. Checked 2026-08-28. Type: Official.
  2. E2

    IEA Energy and AI PDF: China ~99% of global refined gallium supply; DC gallium demand 2030 >10% of today’s supply. Energy-security HTML of the same report family: 95% — internal discrepancy 99 vs 95, do not smooth. Minerals 2026 exec: top refiner China >90% for Ga, graphite, manganese, REE.

    IEA PDF. Energy and AI (PDF, Auszug). https://iea.blob.core.windows.net/assets/40a4db21-2225-42f0-8a07-addcc2ea86b3/EnergyandAI.pdf. As of 2025-04. Checked 2026-08-28. Type: Official report PDF (excerpt).IEA Minerals. Global Critical Minerals Outlook 2026 — Executive Summary. https://www.iea.org/reports/global-critical-minerals-outlook-2026/executive-summary. As of 2026. Checked 2026-08-28. Type: Official report (exec).

Phase-decoupled power control for disaggregated LLM serving (Kim et al.)

  1. E1

    Kim, Yoo, Ryu, Ha, Lee, Ryu (Xenoscube; arXiv:2609.11133v1, PDF opened 2026-09-20): phase-decoupled, model-calibrated power control for prefill/decode-disaggregated serving. On 8×B200 with Qwen3-Coder-480B FP8 MoE (agentic): vendor Max-Q in-flight +8.6% tokens/J at +5.2% mean e2e latency; controller BAL mode +20.4% tokens/J at +3.5% e2e — Pareto on both axes.

    Kim et al. Phase-Decoupled. Phase-Decoupled, Model-Calibrated Power Control for Disaggregated LLM Serving. https://arxiv.org/abs/2609.11133. As of arXiv:2609.11133v1 (2026-09-10); PDF research/pdfs/PhaseDecoupled_Power_arXiv2609.11133.pdf. Checked 2026-09-20. Type: Paper (arXiv; PDF opened).
  2. E1

    Three-day run: 32.3% electricity savings of a lane pair. Prefill: SM-clock window with latency floor; decode: calibrated power cap (A/B: cap +24.1% vs throughput-optimal clock +17.2% vs aggressive +15.5% tokens/J). Scope: MoE; exploratory dense comparison recovers roughly 5× less efficiency.

    Kim et al. Phase-Decoupled. Phase-Decoupled, Model-Calibrated Power Control for Disaggregated LLM Serving. https://arxiv.org/abs/2609.11133. As of arXiv:2609.11133v1 (2026-09-10); PDF research/pdfs/PhaseDecoupled_Power_arXiv2609.11133.pdf. Checked 2026-09-20. Type: Paper (arXiv; PDF opened).

Test-time scaling: candidate count ≠ energy schedule

  1. E1

    Kashaniyan & Jannesari (Iowa State; arXiv:2609.19499v1, PDF opened 2026-09-22): at fixed candidate budget N=8 compare 1×8, 2×4, 4×2, 8×1 generation schedules (a×b = a calls × b candidates/call) on Phi-3-mini and Qwen2.5-1.5B; latency, throughput, GPU-hours and gross GPU-device energy (NVML, idle not subtracted).

    Kashaniyan & Jannesari. Sample Count Is Not Enough: Candidate-Generation Strategy Shapes the Energy and Performance of LLM Test-Time Scaling. https://arxiv.org/abs/2609.19499. As of arXiv:2609.19499v1 (2026-09-16); PDF research/pdfs/SampleCount_Energy_arXiv2609.19499.pdf. Checked 2026-09-22. Type: Paper (arXiv; PDF opened).
  2. E1

    A100 GSM8K: 8×1 uses 4.64× (Phi-3) or 4.86× (Qwen) as much gross GPU-device energy as 1×8; P95 latency 5.77× or 6.12×; throughput falls to ~16.7–17.9% of the batched baseline. Accuracy N=1→8: Phi-3 +8.4 pp (81.4→89.8%), Qwen +18.4 pp (51.4→69.8%) — same aggregation, different execution schedule.

    Kashaniyan & Jannesari. Sample Count Is Not Enough: Candidate-Generation Strategy Shapes the Energy and Performance of LLM Test-Time Scaling. https://arxiv.org/abs/2609.19499. As of arXiv:2609.19499v1 (2026-09-16); PDF research/pdfs/SampleCount_Energy_arXiv2609.19499.pdf. Checked 2026-09-22. Type: Paper (arXiv; PDF opened).
  3. E1

    Pattern holds across three independently scheduled A100 jobs and in SciQ/V100 short-output (8×1: 2.57× or 3.34× energy). Scope: two small models, batch-scheduled HF generation, gross GPU — not continuous batching, not fleet TWh, not frontier serving.

    Kashaniyan & Jannesari. Sample Count Is Not Enough: Candidate-Generation Strategy Shapes the Energy and Performance of LLM Test-Time Scaling. https://arxiv.org/abs/2609.19499. As of arXiv:2609.19499v1 (2026-09-16); PDF research/pdfs/SampleCount_Energy_arXiv2609.19499.pdf. Checked 2026-09-22. Type: Paper (arXiv; PDF opened).

Meta 150 MW: a power limit below TDP beats full power (Ardestani et al.)

  1. E2

    Ardestani, Piga, Stojkovic et al. (Meta Platforms; arXiv:2605.24461v3 of 2026-10-02, PDF opened 2026-10-05): operator report on planning, validation and runtime power management of a 150 MW datacenter (5 buildings of ~30 MW, air-cooled facility, part of a 1 GW buildout) with ~83K GB200 GPUs (Catalina pods, 2×36 GPUs, in-rack AALC liquid cooling). Planning goal: maximum total throughput within a fixed power budget, not electricity cost.

    Ardestani et al. Meta 150 MW. Provisioning to Runtime Optimization of a 100 MW-Scale AI Cluster. https://arxiv.org/abs/2605.24461. As of arXiv:2605.24461v3 (2026-10-02; v1 2026-05-23); PDF research/pdfs/Meta100MW_arXiv2605.24461v3.pdf. Checked 2026-10-05. Type: Operator paper (arXiv; PDF v3 opened; Meta self-report).
  2. E2

    Performance vs. power limit (operator simulation projection, Fig. 9): 1200 → 1000 W costs ~5% per-GPU performance at −16.7% power; 900 W −12% performance at −25% power; within a fixed budget more GPUs fit → cluster throughput +9% (1000 W) or +6% (900 W). HBM bandwidth flat down to 1000 W, −15% at 800 W. Chosen: 960 W (80% TDP); Table 4: 2.4× per GPU / 1.9× cluster vs H100 (700 W) in the same 150 MW budget, ~11% above GB200 at 1200 W; Fig. 21: ~10%.

    Ardestani et al. Meta 150 MW. Provisioning to Runtime Optimization of a 100 MW-Scale AI Cluster. https://arxiv.org/abs/2605.24461. As of arXiv:2605.24461v3 (2026-10-02; v1 2026-05-23); PDF research/pdfs/Meta100MW_arXiv2605.24461v3.pdf. Checked 2026-10-05. Type: Operator paper (arXiv; PDF v3 opened; Meta self-report).
  3. E2

    Measured in operation: rack PSU telemetry systematically overestimates (against oscilloscope and DCIM sensors at the RPP); P70 of per-minute samples matches DCIM best. MSB headroom averages ≈160 kW ≈ 100 W/GPU, 13% of MSBs below 50 kW — static imbalance leaves 5–10% of power stranded. After validation 960 → 1020 W (+~2%), runtime capping “Dimmer” another ~2%; software power smoother against grid-scale swings costs <3% performance.

    Ardestani et al. Meta 150 MW. Provisioning to Runtime Optimization of a 100 MW-Scale AI Cluster. https://arxiv.org/abs/2605.24461. As of arXiv:2605.24461v3 (2026-10-02; v1 2026-05-23); PDF research/pdfs/Meta100MW_arXiv2605.24461v3.pdf. Checked 2026-10-05. Type: Operator paper (arXiv; PDF v3 opened; Meta self-report).

Claimed · E3

4. What is claimed, not shown

Value and time saving per joule as an optimisation target for models, chips and routing (author reading)

  1. E3

    Zeinali et al.: capability, cost and speed alone are insufficient to identify where AI systems deliver the greatest economic return per joule; NVJ and TSJ should serve as explicit objectives for models, accelerators, software and task routing under real energy constraints, and could be tested as training signals and for routing and stopping rules (compute only when expected value warrants the energy). A programmatic proposal; whether the returns hold beyond the benchmark is, per the authors, untested.

    Zeinali et al. Joules-to-GDP. Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?. https://arxiv.org/abs/2610.30858. As of arXiv:2610.30858v1 (submitted 2026-10-22; cs.LG; 35 pp.); PDF research/pdfs/arXiv2610.30858.pdf; code github.com/HazyResearch/joules_to_gdp and dataset not opened. Checked 2026-10-25. Type: Paper (arXiv; PDF opened; 220 GDPval tasks, measured GPU/CPU/DRAM energy, LLM judge).

Token-Energy Law as a basis for energy budgets and deployment planning (author reading)

  1. E3

    Niu et al.: the law is a reusable work–time relationship whose form stays the same across configurations; combined with independently validated performance models it could support prospective energy estimation, energy budgets and comparisons of deployment options. The paper uses measured decode time as an input; a prediction without measured time is not shown. Lead author at NVIDIA, the maker of the measured GPUs.

    Niu et al. Token-Energy Law. The Token-Energy Law of LLM Decode: Active Bytes and Realized Time Account for Joules per Token. https://arxiv.org/abs/2610.28576. As of arXiv:2610.28576v1 (submitted 2026-10-20; cs.LG); PDF research/pdfs/arXiv2610.28576.pdf. Checked 2026-10-21. Type: Paper (arXiv; PDF opened; measurement on GH200/GB300; lead author NVIDIA).

JEDEC test loops as an HBM power bound risk throttling and failures (author reading)

  1. E3

    Olgun et al.: bounding HBM power solely with JEDEC or vendor test loops risks errors or system failures, thermal throttling and thermal shutdowns in real high-power AI and HPC workloads; vendor IDD values for HBM are not public and thus cannot be independently checked. Plausible inference from the measurement — no such failure in a datacenter is observed in the paper.

    Olgun et al. HBM2 power. Understanding the Power Consumption of Modern High Bandwidth Memory: Experimental Analysis and Modeling Using Real HBM2 DRAM Chips. https://arxiv.org/abs/2610.14244. As of arXiv:2610.14244v1 (2026-10-09; cs.AR; extended version of a MICRO 2026 paper); PDF research/pdfs/arXiv2610.14244.pdf; model github.com/CMU-SAFARI/HBM-Power (not run). Checked 2026-10-18. Type: Paper (arXiv; PDF opened; measurement on 36 chips + 3 H200).

Fault-tolerant language models on unreliable hardware (McCourt, Fiete, Chuang)

  1. E2

    McCourt, Fiete, Chuang (MIT; arXiv:2610.10311v1 of 2026-10-07; PDF opened 2026-10-10) claim: language models trained with the same faults they will see at inference become more error-resilient as they grow, not worse. Basis: modified scaling laws from 40,000 GPU-hours of training on simulated faulty digital hardware (components zeroed with probability p), largest model 9.3 × 10⁸ parameters. The authors conjecture that appropriately trained LLMs may be formally fault-tolerant and that inference on low-energy, faulty hardware could save substantial energy. No real low-voltage hardware, no measured energy saving; extrapolation beyond the largest runs.

    McCourt et al. fault-tolerant LLMs. Fault-tolerant foundation models. https://arxiv.org/abs/2610.10311. As of arXiv:2610.10311v1 (2026-10-07); PDF research/pdfs/arXiv2610.10311.pdf. Checked 2026-10-10. Type: Paper (arXiv; PDF opened; simulation).

Meta: power is the bottleneck, not the chip

  1. E3

    Ardestani et al. (Meta) write that power supply for AI datacenters is now the most significant bottleneck “in the race toward AGI”, surpassing accelerator availability. That is the operator’s assessment in intro/abstract, not a measurement in the paper; the supporting figures (US capacity 1,189 → 1,230 GW 2023→2024; OpenAI GW plans) are cited third-party sources, not opened here.

    Ardestani et al. Meta 150 MW. Provisioning to Runtime Optimization of a 100 MW-Scale AI Cluster. https://arxiv.org/abs/2605.24461. As of arXiv:2605.24461v3 (2026-10-02; v1 2026-05-23); PDF research/pdfs/Meta100MW_arXiv2605.24461v3.pdf. Checked 2026-10-05. Type: Operator paper (arXiv; PDF v3 opened; Meta self-report).

IEA projections to 2030

  1. E2

    IEA base: ~945 TWh global 2030 (from 415 TWh 2024) or update 485 TWh (2025) → 950 TWh (2030), ~3% of world electricity. AI-focused DCs triple; +50% in 2025 vs. +17% all DCs. 2026 data per Key Questions: estimates. Methods annex not fully read.

    IEA. Energy and AI — Executive summary. https://www.iea.org/reports/energy-and-ai/executive-summary. As of 2025-04. Checked 2026-08-28. Type: Official report.IEA Demand. Energy demand from AI. https://www.iea.org/reports/energy-and-ai/energy-demand-from-aI. As of 2025-04. Checked 2026-08-28. Type: Official report (chapter).IEA Key Questions. Key Questions on Energy and AI — Executive Summary. https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary. As of 2026. Checked 2026-08-28. Type: Official report (exec summary).
  2. E2

    IEA Electricity 2026, demand chapter: US electricity use plus more than 420 TWh over the next five years; datacenter expansion is expected to make up about 50% of US demand growth to 2030. Global demand 2026–2030 averages 3.6%/year (past decade 2.8%); about 1,100 TWh/year added versus 700; 33,600 TWh in 2030 versus 28,200 TWh in 2025. E2 (agency forecast; 2026–2030 are projected values).

    IEA Electricity 2026. Demand — Electricity 2026. https://www.iea.org/reports/electricity-2026/demand. As of 2026. Checked 2026-08-28. Type: Official report (chapter).
  3. E2

    Energy per task down “at least an order of magnitude annually”. All conventional searches as simple AI text: <4 TWh/year. Video/reasoning/agents: “hundreds or thousands of times” more energy per query. E2 for direction and the 4 TWh illustration; E3 for the exact order of magnitude per year while methods are unread.

    IEA Key Questions. Key Questions on Energy and AI — Executive Summary. https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary. As of 2026. Checked 2026-08-28. Type: Official report (exec summary).
  4. E2

    HBM shortage “over the past six months”, expected until at least end-2027. Capex of the largest tech firms > USD 400 billion in 2025, +75% expected 2026. Grid: ~20% of planned DC projects at delay risk. Satellite method not independently replicated.

    IEA Security. AI and energy security. https://www.iea.org/reports/energy-and-ai/ai-and-energy-security. As of 2025-04. Checked 2026-08-28. Type: Official report (chapter).IEA Key Questions. Key Questions on Energy and AI — Executive Summary. https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary. As of 2026. Checked 2026-08-28. Type: Official report (exec summary).

Vendor outlook and unread splits

  1. E3

    TSMC 2026 shipments 16–17 million wafers (AR). ASML 2026 sales €34–39 billion. Guidance, not actuals.

    TSMC. 2025 Annual Report, Letter to Shareholders (ch. 1 PDF). https://investor.tsmc.com/static/annualReports/2025/english/pdf/2025_tsmc_ar_e_ch1.pdf. As of 2025. Checked 2026-08-28. Type: Company report.ASML PR. ASML reports €32.7 billion total net sales and €9.6 billion net income in 2025. https://www.asml.com/en/news/press-releases/2026/q4-2025-financial-results. As of 2026-01-28. Checked 2026-08-28. Type: Company report.
  2. E2

    Same Symposium PR: “A 14-reticle size CoWoS, capable of integrating approximately 10 large compute dies and 20 HBM stacks, is slated for production in 2028.” Then expansion beyond 14 reticles in 2029; optional SoW-X 40-reticle expected 2029. E2 as vendor roadmap (claimed schedule), not running capacity. kwpm still E0.

    TSMC Tech Symposium 2026. TSMC Debuts A13 Technology at 2026 North America Technology Symposium. https://pr.tsmc.com/system/files/newspdf/attachment/49337b40ff139d51d533076cf7a945b30e107e07/2026%20Tech%20Symposium%20%28E%29_Final_wmn.pdf. As of 2026-04-22. Checked 2026-09-02. Type: Company PR PDF (Tech Symposium; news HTML Cloudflare-blocked).
  3. E3

    Same product page: DGX B200 delivers 3× training and 15× inference versus DGX H100. The performance sections are marked “Projected performance subject to change”. E3 (vendor comparison, not measured in this log).

    NVIDIA DGX B200. NVIDIA DGX B200 (product specifications). https://www.nvidia.com/en-us/data-center/dgx-b200/. As of 2026-08-28. Checked 2026-08-28. Type: Manufacturer specification.
  4. E3

    Epoch overview: estimate 290,000–1.6 million H100-equivalents into China through end-2025 (median 660,000, “roughly a third” of China’s capacity). An estimate, not a customs census — do not collapse with the 20 million shipments.

    Epoch Chips Overview. Topic Overview: AI Chips. https://epoch.ai/publications/chips-topic-overview. As of 2026-05-01. Checked 2026-09-01. Type: Topic overview (research NGO).
  5. E3

    Industry splits cited inside Luccioni, not independently opened here: Barr 2019 (inference 80–90%); Wu 2021 Meta ~⅓. The Patterson Google split ⅗/⅖ is E2 vendor (above).

    Luccioni, Jernite, Strubell. Power Hungry Processing: Watts Driving the Cost of AI Deployment?. https://arxiv.org/html/2311.16863. As of 2023-11-28. Checked 2026-08-28. Type: Paper (FAccT 2024).

Constrained · Limit

5. Bottleneck and limit

Joules-to-GDP: dollar value = expert time × wage, not GDP; LLM judge; one run per task; only server components metered

  1. E2

    Zeinali et al. 2026: the dollar value follows GDPval (expert time × median wage) and is not measured GDP — the title describes the framing, not a macroeconomic measurement; a tie earns full value, a loss zero. Success is judged by an LLM judge with 71.9% agreement with humans on only 32 comparisons; one run per task and configuration; 220 tasks illustrate a framework, not a comprehensive survey (authors). Metered are GPU, CPU and DRAM, not cooling/facility overhead; execution cost from published rates; review, coordination, liability and adoption costs are missing. GLM 5.2 ran on H200s, the other four models on H100s — the model comparison is not hardware-matched; harness (OpenHands), backend, precision and batching not varied individually. Do not net against datacentre electricity figures or token energy (source 34). Preprint.

    Zeinali et al. Joules-to-GDP. Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?. https://arxiv.org/abs/2610.30858. As of arXiv:2610.30858v1 (submitted 2026-10-22; cs.LG; 35 pp.); PDF research/pdfs/arXiv2610.30858.pdf; code github.com/HazyResearch/joules_to_gdp and dataset not opened. Checked 2026-10-25. Type: Paper (arXiv; PDF opened; 220 GDPval tasks, measured GPU/CPU/DRAM energy, LLM judge).

Token energy: GPU board, dense models, one runtime — ≠ datacenter electricity per query, ≠ MoE

  1. E2

    Niu et al. 2026: only GPU-board energy is measured, without host CPU, server, network and cooling; only dense BF16 models up to 14 B parameters with full history; for mixture-of-experts models recalibration is not enough per the authors, routing-aware accounting is needed; multi-node and pipeline operation not validated; one runtime type (vLLM, hybrid CUDA-graph/eager execution); coefficients are local per model and hardware/runtime; decode time enters as an observed quantity. Preprint. Not to be equated with datacenter electricity totals (IEA, LBNL) or the capability-per-energy measurement in Linkages (Yang & Armour).

    Niu et al. Token-Energy Law. The Token-Energy Law of LLM Decode: Active Bytes and Realized Time Account for Joules per Token. https://arxiv.org/abs/2610.28576. As of arXiv:2610.28576v1 (submitted 2026-10-20; cs.LG); PDF research/pdfs/arXiv2610.28576.pdf. Checked 2026-10-21. Type: Paper (arXiv; PDF opened; measurement on GH200/GB300; lead author NVIDIA).

HBM2 on FPGA boards ≠ HBM3E/HBM4 in GPU fleets

  1. E2

    Olgun et al. 2026: what was measured is HBM2 at 1,200 MT/s on FPGA boards, not HBM3E/HBM4 in accelerators; vendor and part number unknown; the board sensor reports VDDC and VDDQ only combined. The 23.6% pattern is a constructed stress test, not an observed workload. The HBM3E validation rests on three H200s with undisclosed supply voltage and model values extrapolated from HBM2 currents; HBM4 figures are a projection without public IDD values. Preprint (extended version).

    Olgun et al. HBM2 power. Understanding the Power Consumption of Modern High Bandwidth Memory: Experimental Analysis and Modeling Using Real HBM2 DRAM Chips. https://arxiv.org/abs/2610.14244. As of arXiv:2610.14244v1 (2026-10-09; cs.AR; extended version of a MICRO 2026 paper); PDF research/pdfs/arXiv2610.14244.pdf; model github.com/CMU-SAFARI/HBM-Power (not run). Checked 2026-10-18. Type: Paper (arXiv; PDF opened; measurement on 36 chips + 3 H200).

Carbon–SLO simulator ≠ operator fleet

  1. E2

    Topcu et al. 2026: up to 60% CO₂ and ~3% SLO hold in the calibrated geo-simulator — no warrant for production serving, no measured fleet TWh or live carbon API. A100-only; Mixtral not hardware-validated.

    Topcu et al. Carbon–SLO. Carbon-Aware Geo-Distributed LLM Serving under Latency SLOs. https://arxiv.org/abs/2610.14571. As of arXiv:2610.14571v1 (2026-10-10); PDF research/pdfs/arXiv2610.14571.pdf. Checked 2026-10-14. Type: Paper (arXiv; PDF opened; simulation).

What the sources mark as a limit

  1. E2

    Meta 150 MW: one operator, one site, pre-training as the planning load; per-GPU and cluster throughput curves are the operator’s simulation projections; several rack power values confidential (*); no independent replication. The three gain figures (+9% Fig. 9 at 1000 W, ~11% Table 4, ~10% Fig. 21) do not coincide. No annual TWh, PUE or inference figure — operator report ≠ industry standard.

    Ardestani et al. Meta 150 MW. Provisioning to Runtime Optimization of a 100 MW-Scale AI Cluster. https://arxiv.org/abs/2605.24461. As of arXiv:2605.24461v3 (2026-10-02; v1 2026-05-23); PDF research/pdfs/Meta100MW_arXiv2605.24461v3.pdf. Checked 2026-10-05. Type: Operator paper (arXiv; PDF v3 opened; Meta self-report).
  2. E1

    Sample-count energy: N alone does not describe the systems cost of multi-candidate TTS. 4.64–4.86× applies to measured A100/HF schedules on two small models — do not read as 2026 world electricity or a continuous-serving factor. Gross GPU, idle not subtracted.

    Kashaniyan & Jannesari. Sample Count Is Not Enough: Candidate-Generation Strategy Shapes the Energy and Performance of LLM Test-Time Scaling. https://arxiv.org/abs/2609.19499. As of arXiv:2609.19499v1 (2026-09-16); PDF research/pdfs/SampleCount_Energy_arXiv2609.19499.pdf. Checked 2026-09-22. Type: Paper (arXiv; PDF opened).
  3. E1

    Phase-decoupled power: single-node lane under MoE serving ≠ fleet/cluster management; vendor Max-Q ≠ calibrated controller; dense models recover roughly 5× less — claims scoped to MoE. Lab/node ≠ deploy fleet.

    Kim et al. Phase-Decoupled. Phase-Decoupled, Model-Calibrated Power Control for Disaggregated LLM Serving. https://arxiv.org/abs/2609.11133. As of arXiv:2609.11133v1 (2026-09-10); PDF research/pdfs/PhaseDecoupled_Power_arXiv2609.11133.pdf. Checked 2026-09-20. Type: Paper (arXiv; PDF opened).
  4. E2

    Jevons: efficiency per task rises (IEA; Luccioni task classes; Epoch FLOP/s/W). At the same time more expensive tasks and more chips (dashboard 3.4×/year stock; sold-capacity 3.3×/year since 2022, sources 24–25). Do not collapse either with D01 5.3× training FLOP. IEA central path: DC power doubles by 2030. LBNL: the US era of flat kWh is over. Causal share of AI vs. rest-of-cloud vs. crypto is not finely resolved.

    IEA Key Questions. Key Questions on Energy and AI — Executive Summary. https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary. As of 2026. Checked 2026-08-28. Type: Official report (exec summary).Luccioni, Jernite, Strubell. Power Hungry Processing: Watts Driving the Cost of AI Deployment?. https://arxiv.org/html/2311.16863. As of 2023-11-28. Checked 2026-08-28. Type: Paper (FAccT 2024).Epoch Trends. Trends in Artificial Intelligence. https://epoch.ai/trends. As of 2026-02-05. Checked 2026-08-28. Type: Dataset/dashboard.Shehabi et al. (LBNL 2024). 2024 United States Data Center Energy Usage Report. https://eta-publications.lbl.gov/sites/default/files/2024-12/lbnl-2024-united-states-data-center-energy-usage-report.pdf. As of 2024-12. Checked 2026-08-28. Type: Official lab report (LBNL-2001637).
  5. E1

    Masanet 2020 itself warns: next doubling of compute instances “within the next 3 to 4 years”; “sharp energy demand growth” once that efficiency resource is tapped; “demand surge later this decade”. That is a 2020 Policy Forum projection (Cisco index), not a 2026 actual. IEA 2024 ~415 TWh global remains E2 (sources 1–2).

    Masanet et al. Recalibrating global data center energy-use estimates. https://datacenters.lbl.gov/sites/default/files/Masanet_et_al_Science_2020.full_.pdf. As of 2020-02-28. Checked 2026-08-31. Type: Paper (Science Policy Forum; LBNL-hosted journal PDF).
  6. E2

    Local ≠ global. World 2024 ~1.5%, 2030 ~3%. US 2024 4.7%, 2030 reference 11.8%. JRC is 2022 only for the EU.

    IEA. Energy and AI — Executive summary. https://www.iea.org/reports/energy-and-ai/executive-summary. As of 2025-04. Checked 2026-08-28. Type: Official report.Smith et al. (LBNL 2025). United States Data Center Energy Usage Report: 2025 Update. https://doi.org/10.71468/P1RP4F. As of 2026-06-18. Checked 2026-08-28. Type: Official lab report (LBNL-2001758).Kamiya, Bertoldi (JRC). Energy Consumption in Data Centres and Broadband Communication Networks in the EU. https://doi.org/10.2760/706491. As of 2024-02. Checked 2026-08-28. Type: Official report (JRC135926).
  7. E1

    Memory wall / packaging: compute 2×/2.3 y vs. bandwidth 2×/~4 y (Epoch 2023). Overview (source 24): HBM most binding constraint as of late 2025 (vendor quote via Epoch). IEA: HBM shortage until ≥2027. TSMC Tech Symposium 2026 (source 26): 5.5-reticle CoWoS in production; 14-reticle roadmap 2028 — reticle sizes known, capacity in kwpm still E0.

    Hobbhahn, Heim, Aydos. Trends in machine learning hardware. https://epoch.ai/publications/trends-in-machine-learning-hardware. As of 2023-11-09. Checked 2026-08-28. Type: Dataset/report.TSMC. 2025 Annual Report, Letter to Shareholders (ch. 1 PDF). https://investor.tsmc.com/static/annualReports/2025/english/pdf/2025_tsmc_ar_e_ch1.pdf. As of 2025. Checked 2026-08-28. Type: Company report.IEA Key Questions. Key Questions on Energy and AI — Executive Summary. https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary. As of 2026. Checked 2026-08-28. Type: Official report (exec summary).
  8. E1

    System vs. GPU TDP: H100 SXM up to 700 W (GPU, configurable). DGX B200 system about 14.3 kW max for 8 Blackwell GPUs plus CPU, network, memory — not 8×700 W and not field load. Peak spec ≠ utilization (LBNL ~70% of rated applies to H100 nodes, not measured here for B200).

    NVIDIA. NVIDIA H100 GPU (product specifications). https://www.nvidia.com/en-us/data-center/h100/. As of 2026-08-28. Checked 2026-08-28. Type: Manufacturer specification.NVIDIA DGX B200. NVIDIA DGX B200 (product specifications). https://www.nvidia.com/en-us/data-center/dgx-b200/. As of 2026-08-28. Checked 2026-08-28. Type: Manufacturer specification.
  9. E1

    CO₂ accounting: Strubell NAS 88× too high (Patterson). Grid intensity beats model size (BLOOM 25 t vs. GPT-3 502 t at similar size). Embodied is 22% of BLOOM’s training footprint. Standardised vendor disclosures 2026: E0.

    Strubell et al. Energy and Policy Considerations for Deep Learning in NLP. https://arxiv.org/html/1906.02243. As of 2019-06. Checked 2026-08-28. Type: Paper (ACL 2019).Patterson et al. The Carbon Footprint of Machine Learning Training Will Plateau, Then Shrink. https://arxiv.org/abs/2204.05149. As of 2022-04-11. Checked 2026-08-28. Type: Paper (Google/Berkeley).Luccioni, Viguier, Ligozat. Estimating the Carbon Footprint of BLOOM, a 176B Parameter Language Model. https://arxiv.org/html/2211.02001. As of 2022-11. Checked 2026-08-28. Type: Paper.

6. Actors and incentives

Who measures, who fabricates, who projects

  1. E1

    LBNL (DOE) measures US datacenter power bottom-up; the 2025 update revises 2024 downward.

    Shehabi et al. (LBNL 2024). 2024 United States Data Center Energy Usage Report. https://eta-publications.lbl.gov/sites/default/files/2024-12/lbnl-2024-united-states-data-center-energy-usage-report.pdf. As of 2024-12. Checked 2026-08-28. Type: Official lab report (LBNL-2001637).Smith et al. (LBNL 2025). United States Data Center Energy Usage Report: 2025 Update. https://doi.org/10.71468/P1RP4F. As of 2026-06-18. Checked 2026-08-28. Type: Official lab report (LBNL-2001758).
  2. E2

    IEA projects global DC electricity and HBM/grid bottlenecks in executive summaries; methods annex incompletely read.

    IEA. Energy and AI — Executive summary. https://www.iea.org/reports/energy-and-ai/executive-summary. As of 2025-04. Checked 2026-08-28. Type: Official report.IEA Key Questions. Key Questions on Energy and AI — Executive Summary. https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary. As of 2026. Checked 2026-08-28. Type: Official report (exec summary).
  3. E1

    TSMC reports wafer mix and advanced share; ASML reports EUV units. CoWoS reticle sizes in 2026 Symposium PR (source 26); kwpm still not in the opened letter/PR.

    TSMC. 2025 Annual Report, Letter to Shareholders (ch. 1 PDF). https://investor.tsmc.com/static/annualReports/2025/english/pdf/2025_tsmc_ar_e_ch1.pdf. As of 2025. Checked 2026-08-28. Type: Company report.ASML PR. ASML reports €32.7 billion total net sales and €9.6 billion net income in 2025. https://www.asml.com/en/news/press-releases/2026/q4-2025-financial-results. As of 2026-01-28. Checked 2026-08-28. Type: Company report.
  4. E1

    NVIDIA publishes H100 and DGX B200 specifications; the 3×/15× comparison remains a vendor projection (E3).

    NVIDIA. NVIDIA H100 GPU (product specifications). https://www.nvidia.com/en-us/data-center/h100/. As of 2026-08-28. Checked 2026-08-28. Type: Manufacturer specification.NVIDIA DGX B200. NVIDIA DGX B200 (product specifications). https://www.nvidia.com/en-us/data-center/dgx-b200/. As of 2026-08-28. Checked 2026-08-28. Type: Manufacturer specification.
  5. E2

    IEA Electricity 2026 demand chapter projects the US datacenter share of electricity growth; not identical to LBNL TWh levels.

    IEA Electricity 2026. Demand — Electricity 2026. https://www.iea.org/reports/electricity-2026/demand. As of 2026. Checked 2026-08-28. Type: Official report (chapter).
  6. E2

    Epoch AI curates hardware trends, chip sales and the chips topic overview (sources 11, 12, 24, 25); underlying revenue/analyst inputs not audited.

    Epoch Trends. Trends in Artificial Intelligence. https://epoch.ai/trends. As of 2026-02-05. Checked 2026-08-28. Type: Dataset/dashboard.Hobbhahn, Heim, Aydos. Trends in machine learning hardware. https://epoch.ai/publications/trends-in-machine-learning-hardware. As of 2023-11-09. Checked 2026-08-28. Type: Dataset/report.

7. State of the dispute

Where sources contradict or bound each other

  1. E1

    Strubell NAS vs. Patterson: 88× difference on the same NAS story. Both opened.

    Strubell et al. Energy and Policy Considerations for Deep Learning in NLP. https://arxiv.org/html/1906.02243. As of 2019-06. Checked 2026-08-28. Type: Paper (ACL 2019).Patterson et al. The Carbon Footprint of Machine Learning Training Will Plateau, Then Shrink. https://arxiv.org/abs/2204.05149. As of 2022-04-11. Checked 2026-08-28. Type: Paper (Google/Berkeley).
  2. E2

    2024 consistency IEA vs. LBNL is tight (IEA US ≈187 TWh vs. LBNL 192). 2030 levels diverge more and are not harmonised here.

    IEA. Energy and AI — Executive summary. https://www.iea.org/reports/energy-and-ai/executive-summary. As of 2025-04. Checked 2026-08-28. Type: Official report.Smith et al. (LBNL 2025). United States Data Center Energy Usage Report: 2025 Update. https://doi.org/10.71468/P1RP4F. As of 2026-06-18. Checked 2026-08-28. Type: Official lab report (LBNL-2001758).
  3. E2

    IEA gallium refining: 99% (PDF) vs. 95% (security HTML). Do not smooth.

    IEA PDF. Energy and AI (PDF, Auszug). https://iea.blob.core.windows.net/assets/40a4db21-2225-42f0-8a07-addcc2ea86b3/EnergyandAI.pdf. As of 2025-04. Checked 2026-08-28. Type: Official report PDF (excerpt).IEA Security. AI and energy security. https://www.iea.org/reports/energy-and-ai/ai-and-energy-security. As of 2025-04. Checked 2026-08-28. Type: Official report (chapter).
  4. E2

    Patterson “plateau then shrink” (Google internal, 4Ms) stands against the IEA/LBNL net rise. E2 internal; E3 as a 2026 world forecast.

    Patterson et al. The Carbon Footprint of Machine Learning Training Will Plateau, Then Shrink. https://arxiv.org/abs/2204.05149. As of 2022-04-11. Checked 2026-08-28. Type: Paper (Google/Berkeley).IEA. Energy and AI — Executive summary. https://www.iea.org/reports/energy-and-ai/executive-summary. As of 2025-04. Checked 2026-08-28. Type: Official report.Shehabi et al. (LBNL 2024). 2024 United States Data Center Energy Usage Report. https://eta-publications.lbl.gov/sites/default/files/2024-12/lbnl-2024-united-states-data-center-energy-usage-report.pdf. As of 2024-12. Checked 2026-08-28. Type: Official lab report (LBNL-2001637).

8. Open questions

  1. IEA Energy and AI methods annex + supply chapter — how 415/945 are built.
  2. TSMC 20-F / CoWoS capacity in kwpm (reticle sizes opened via 2026 Symposium PR; kwpm still E0).
  3. SK hynix / Samsung HBM annual report — not TrendForce.
  4. USGS OFR 2024-1057 gallium/germanium export restrictions.
  5. NVIDIA B200 single-GPU datasheet PDF + DGX-node measurements beyond the product-page TDP.
  6. Wu et al. 2021 (Meta) + Google 2025/26 environmental reports for current splits.
  7. Epoch chips-topic-overview (retry the timeout).
  8. JRC or ENTSO-E 2025/26 update for EU DC TWh after 2022.
  9. IEA Key Questions full PDF — HBM claim and “order of magnitude/year” methods.

9. Changes

  • v2.2-draft2026-10-25: Zeinali et al. Joules-to-GDP arXiv:2610.30858 (source 35, E2, PDF) — 220 GDPval tasks, 44 occupations, five open models; GLM 5.2 $72.27 net for 0.430 MJ; Gemma 4 31B 56% value / 23% energy (NVJ ×2.5, TSJ ×2.8); CAD +10% per task, software +27% per joule; Qwen3.6-27B NVJ spread ×4.32 (B200 top, MI350X bottom); 8×H100 −22% time / ×3.24 energy; value = expert time × wage, LLM judge; preprint.
  • v2.1-draft2026-10-21: Niu et al. Token-Energy Law arXiv:2610.28576 (source 34, E2, PDF, measurement) — Qwen2.5-7B/14B, Mistral-7B on GH200 (700 W) / GB300 (1,300 W); 374 holdout runs ≤3.73%; GPU0→GPU1 3.72%; bytes only 28–35%, time only 11–29%; batch 33.2–46.8% leaner; 0.323 → 0.638 J/token with history; HBM reads 0.03–0.07%; GPU board, dense, lead author NVIDIA.
  • v2.0-draft2026-10-18: Olgun et al. HBM2 power arXiv:2610.14244 (source 33, E2, PDF, measurement) — 36 chips; stress pattern +23.6% (max 35.4%) over IDD4R; reads 3.1× refresh; bus toggling +29.4%; model 7.6–8.1% MAPE; H200 HBM3E 2.4–5.3%, up to 248 W; FPGA ≠ GPU fleet.
  • v1.9-draft2026-10-14: Topcu et al. Carbon–SLO arXiv:2610.14571 (source 32, E2, PDF, simulation) — facility MILP up to 60% CO₂ vs RR; ~3% SLO; simulator ≠ fleet.
  • v1.8-draft2026-10-10: Meta fbnic Gupta et al. arXiv:2610.07644 (source 30, E2, PDF, self-report) — unavailability 9.11% → 0.74% (12×), MTTR −37%, 2.3× fewer swaps; MIT fault-tolerant LLMs McCourt et al. arXiv:2610.10311 (source 31, E2, PDF, simulation) — 40,000 GPU-h, up to 9.3 × 10⁸ parameters, energy saving only a conjecture.
  • v1.7-draft2026-10-05: Meta 150 MW Ardestani et al. arXiv:2605.24461v3 (source 29, E2, PDF v3) — ~83K GB200; 960 W (80% TDP) instead of 1200 W → ~10–11% cluster throughput; PSU overestimates, P70≈DCIM; MSB ≈100 W/GPU headroom, 5–10% stranded; +~2% (1020 W) +~2% (Dimmer); self-report ≠ industry.
  • v1.6-draft2026-09-22: Sample-count energy Kashaniyan/Jannesari arXiv:2609.19499 (source 28, E1, PDF) — N=8: 8×1 vs 1×8 A100 4.64–4.86× gross GPU energy / P95 5.77–6.12×; accuracy +8.4/+18.4 pp; schedule ≠ candidate count; lab ≠ fleet.
  • v1.5-draft2026-09-20: Phase-decoupled power control (Kim et al., arXiv:2609.11133, source 27, E1, PDF) — Max-Q +8.6% tokens/J / +5.2% e2e vs BAL +20.4% / +3.5%; 3-day 32.3% lane electricity; MoE scope, dense ~5× less; lab ≠ fleet.
  • v1.4-draft2026-09-02: TSMC 2026 Tech Symposium PR (source 26) opened — 5.5-reticle CoWoS now in production; 14-reticle ~10 compute+20 HBM slated for 2028 (vendor roadmap). kwpm still E0.
  • v1.3-draft2026-09-01: Epoch Topic Overview AI Chips (source 24) and chip-production data insight (source 25) opened — ~20 million chips Q4 2025; 3.3×/year sold capacity since 2022 (CI 2.7–4.1). Do not mix with 5.3× training compute.
  • v1.2-draft2026-08-31: Masanet Science 2020 (source 23, LBNL-hosted journal PDF) opened — global 2010–2018 +6% energy / +550% compute instances; 194→205 TWh. IEA 415 TWh and LBNL US untouched. Epoch 5×/year dashboard figure corroborated in D01 (no new D02 source for the report).
  • v1.1-draft2026-08-28: NVIDIA DGX B200 spec (product page) and IEA Electricity 2026 demand (US datacenter share) folded in.
  • v1.0-draftFirst version from the 2026-08-28 verification log.

10. Sources

No. Source As of Checked Grade
1 International Energy Agency. Energy and AI — Executive summary. https://www.iea.org/reports/energy-and-ai/executive-summary. Type: Official report. E2
2 International Energy Agency. Energy demand from AI. https://www.iea.org/reports/energy-and-ai/energy-demand-from-aI. Type: Official report (chapter). E2
3 International Energy Agency. AI and energy security. https://www.iea.org/reports/energy-and-ai/ai-and-energy-security. Type: Official report (chapter). E2
4 International Energy Agency. Energy and AI (PDF, Auszug). https://iea.blob.core.windows.net/assets/40a4db21-2225-42f0-8a07-addcc2ea86b3/EnergyandAI.pdf. Type: Official report PDF (excerpt). E2
5 International Energy Agency. Key Questions on Energy and AI — Executive Summary. https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary. Type: Official report (exec summary). E2
6 International Energy Agency. Global Critical Minerals Outlook 2026 — Executive Summary. https://www.iea.org/reports/global-critical-minerals-outlook-2026/executive-summary. Type: Official report (exec). E2
7 Luccioni, A. S.; Jernite, Y.; Strubell, E.. Power Hungry Processing: Watts Driving the Cost of AI Deployment?. https://arxiv.org/html/2311.16863. Type: Paper (FAccT 2024). E1
8 Luccioni, A. S.; Viguier, S.; Ligozat, A.-L.. Estimating the Carbon Footprint of BLOOM, a 176B Parameter Language Model. https://arxiv.org/html/2211.02001. Type: Paper. E1
9 Strubell, E.; Ganesh, A.; McCallum, A.. Energy and Policy Considerations for Deep Learning in NLP. https://arxiv.org/html/1906.02243. Type: Paper (ACL 2019). E1
10 Patterson, D.; Gonzalez, J.; Hölzle, U.; Le, Q.; Liang, C.; Munguia, L.-M.; Rothchild, D.; So, D.; Texier, M.; Dean, J.. The Carbon Footprint of Machine Learning Training Will Plateau, Then Shrink. https://arxiv.org/abs/2204.05149. Type: Paper (Google/Berkeley). E2
11 Epoch AI. Trends in Artificial Intelligence. https://epoch.ai/trends. Type: Dataset/dashboard. E2
12 Hobbhahn, M.; Heim, L.; Aydos, G.. Trends in machine learning hardware. https://epoch.ai/publications/trends-in-machine-learning-hardware. Type: Dataset/report. E2
13 Shehabi, A.; Smith, S. J.; Hubbard, A.; Newkirk, A.; Lei, N.; Siddik, M. A. B.; Holecek, B.; Koomey, J.; Masanet, E.; Sartor, D.. 2024 United States Data Center Energy Usage Report. https://eta-publications.lbl.gov/sites/default/files/2024-12/lbnl-2024-united-states-data-center-energy-usage-report.pdf. Type: Official lab report (LBNL-2001637). E1
14 Smith, S. J.; Hubbard, A.; Newkirk, A.; Ganeshalingam, M.; Holecek, B.; Sartor, D.; Mills, M.; Shehabi, A.. United States Data Center Energy Usage Report: 2025 Update. https://doi.org/10.71468/P1RP4F. Type: Official lab report (LBNL-2001758). E2
15 U.S. Geological Survey. Interior Department releases final 2025 List of Critical Minerals. https://www.usgs.gov/news/science-snippet/interior-department-releases-final-2025-list-of-critical-minerals. Type: Official. E1
16 Taiwan Semiconductor Manufacturing Company. 2025 Annual Report, Letter to Shareholders (ch. 1 PDF). https://investor.tsmc.com/static/annualReports/2025/english/pdf/2025_tsmc_ar_e_ch1.pdf. Type: Company report. E1
17 ASML Holding N.V.. ASML reports €32.7 billion total net sales and €9.6 billion net income in 2025. https://www.asml.com/en/news/press-releases/2026/q4-2025-financial-results. Type: Company report. E1
18 ASML Holding N.V.. 2025 Annual Report — Financial performance section. https://ourbrand.asml.com/m/419103cb23dfeaa4/original/asml-2025-annual-report-financial-performance-section.pdf. Type: Company report. E1
19 NVIDIA. NVIDIA H100 GPU (product specifications). https://www.nvidia.com/en-us/data-center/h100/. Type: Manufacturer specification. E1
20 Kamiya, G.; Bertoldi, P. (European Commission, JRC). Energy Consumption in Data Centres and Broadband Communication Networks in the EU. https://doi.org/10.2760/706491. Type: Official report (JRC135926). E2
21 NVIDIA. NVIDIA DGX B200 (product specifications). https://www.nvidia.com/en-us/data-center/dgx-b200/. Type: Manufacturer specification. E1
22 International Energy Agency. Demand — Electricity 2026. https://www.iea.org/reports/electricity-2026/demand. Type: Official report (chapter). E2
23 Masanet, E.; Shehabi, A.; Lei, N.; Smith, S.; Koomey, J.. Recalibrating global data center energy-use estimates. https://datacenters.lbl.gov/sites/default/files/Masanet_et_al_Science_2020.full_.pdf. Type: Paper (Science Policy Forum; LBNL-hosted journal PDF). E1
24 Epoch AI. Topic Overview: AI Chips. https://epoch.ai/publications/chips-topic-overview. Type: Topic overview (research NGO). E2
25 Epoch AI. Global AI computing capacity is doubling every 7 months. https://epoch.ai/data-insights/ai-chip-production. Type: Data insight (CSV stamp 16 Feb 2026). E2
26 Taiwan Semiconductor Manufacturing Company. TSMC Debuts A13 Technology at 2026 North America Technology Symposium. https://pr.tsmc.com/system/files/newspdf/attachment/49337b40ff139d51d533076cf7a945b30e107e07/2026%20Tech%20Symposium%20%28E%29_Final_wmn.pdf. Type: Company PR PDF (Tech Symposium; news HTML Cloudflare-blocked). E2
27 Kim, J. G.; Yoo, D.; Ryu, H.; Ha, S.; Lee, J.; Ryu, S. (Xenoscube). Phase-Decoupled, Model-Calibrated Power Control for Disaggregated LLM Serving. https://arxiv.org/abs/2609.11133. Type: Paper (arXiv; PDF opened). E1
28 Kashaniyan, M.; Jannesari, A. (Iowa State). Sample Count Is Not Enough: Candidate-Generation Strategy Shapes the Energy and Performance of LLM Test-Time Scaling. https://arxiv.org/abs/2609.19499. Type: Paper (arXiv; PDF opened). E1
29 Ardestani, E. K.; Piga, L.; Stojkovic, J.; Balaji, P.; Ozdal, M.; et al. (21 authors; Meta Platforms). Provisioning to Runtime Optimization of a 100 MW-Scale AI Cluster. https://arxiv.org/abs/2605.24461. Type: Operator paper (arXiv; PDF v3 opened; Meta self-report). E2
30 Gupta, P.; Duyck, A.; Kicinski, J.; Provine, J.; … Trager, L. (28 authors; Meta Platforms). From ASIC to Fleet: Lessons from Building and Operating a Hyperscaler NIC. https://arxiv.org/abs/2610.07644. Type: Paper (arXiv; PDF opened; operator self-report). E2
31 McCourt, T.; Fiete, I. R.; Chuang, I. L. (MIT). Fault-tolerant foundation models. https://arxiv.org/abs/2610.10311. Type: Paper (arXiv; PDF opened; simulation). E2
32 Topcu, A.; Cim, E.; Ozturk, U.; Kandemir, M. (Pennsylvania State University). Carbon-Aware Geo-Distributed LLM Serving under Latency SLOs. https://arxiv.org/abs/2610.14571. Type: Paper (arXiv; PDF opened; simulation). E2
33 Olgun, A.; Galanopoulos, S.; Kakolyris, A. K.; Yüksel, İ. E.; Luo, H.; Bostancı, F. N.; Mutlu, O. (ETH Zürich; New York University). Understanding the Power Consumption of Modern High Bandwidth Memory: Experimental Analysis and Modeling Using Real HBM2 DRAM Chips. https://arxiv.org/abs/2610.14244. Type: Paper (arXiv; PDF opened; measurement on 36 chips + 3 H200). E2
34 Niu, C. (NVIDIA Corporation); Sencer, B. (Texas Tech University); Zhang, W. (Texas Advanced Computing Center); Chen, Y. (Texas Tech University). The Token-Energy Law of LLM Decode: Active Bytes and Realized Time Account for Joules per Token. https://arxiv.org/abs/2610.28576. Type: Paper (arXiv; PDF opened; measurement on GH200/GB300; lead author NVIDIA). E2
35 Zeinali, A.; Saad-Falcon, J.; Narayan, A. (Stanford University); Abbasi, M. (UC Berkeley); Hartley, J. S. (UT Austin); Hennessy, J.; Ré, C.; Mirhoseini, A. (Stanford University). Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?. https://arxiv.org/abs/2610.30858. Type: Paper (arXiv; PDF opened; 220 GDPval tasks, measured GPU/CPU/DRAM energy, LLM judge). E2

11. Uncertainty log

Overall uncertainty of this entry, bound to the verification log of 2026-08-28 plus Masanet 2026-08-31, Epoch chips 2026-09-01, TSMC Symposium PR 2026-09-02, phase-decoupled power 2026-09-20 (source 27) and sample-count energy 2026-09-22 (source 28) Meta 150 MW 2026-10-05 (source 29), Meta fbnic and McCourt 2026-10-10 (30, 31), Topcu Carbon–SLO 2026-10-14 (32) Olgun et al. HBM2 power 2026-10-18 (source 33), Niu et al. Token-Energy Law 2026-10-21 (source 34) and Zeinali et al. Joules-to-GDP 2026-10-25 (source 35). 35 openings. Do not mix 3.3× sold capacity with 3.4× dashboard or 5.3× training FLOP. CoWoS reticle sizes ≠ kwpm. Max-Q ≠ calibrated controller; node ≠ fleet. Not used as warrant: journalism; 205 TWh (2018) as 2026 world electricity; the smuggling band as a census.

  • Established (layer 1): value per joule: Gemma 4 31B 56% net value with 23% energy vs GLM 5.2, 8×H100 −22% time for ×3.24 energy (E2, source 35); inference energy (Luccioni); BLOOM; Masanet 2010–2018; LBNL US DC; Epoch chips overview ~20 million / 3.3× since 2022 (E2); H100/DGX B200 spec; TSMC/ASML; CoWoS 5.5-reticle now (E2, source 26); phase-decoupled power MoE 8×B200 (E1, source 27); sample-count TTS schedule A100 4.64–4.86× (E1, source 28); Meta 150 MW / ~83K GB200 power limit 960 W, 5–10% stranded (E2, source 29); HBM2 36 chips: stress pattern +23.6% over JEDEC loop, model 7.6–8.1% MAPE, H200 HBM3E up to 248 W (E2, source 33); token energy GH200/GB300 within 3.73% from bytes + time, batch 33–47% leaner (E2, source 34); USGS.
  • Claimed (layer 2): NVJ/TSJ as a target for models, chips, routing (E3, source 35); IEA central projections to 2030; IEA Electricity 2026 US DCs about 50% of electricity growth; HBM until 2027; capex 400 billion; TSMC/ASML guidance; CoWoS 14-reticle 2028 roadmap (E2 vendor); DGX B200 3×/15× vs H100; unread industry splits; Meta “power > chips as the bottleneck” (E3, source 29); JEDEC loops as HBM bound risk throttling/failure (E3, source 33); Token-Energy Law for deployment planning (E3, source 34).
  • Constrained (layer 3): Joules-to-GDP wage proxy ≠ GDP, LLM judge, no facility overhead (source 35); Jevons net; local vs. global; 2026 training/inference split; CoWoS kwpm (E0; reticle sizes known); 2026 actual vs. IEA estimate; Masanet 2020 surge warning ≠ 2026 measurement; phase-decoupled node ≠ fleet, Max-Q ≠ calibrated, MoE ≠ dense; sample-count N≠schedule, small-model HF ≠ continuous serving; Meta one operator / simulation projection ≠ industry standard; HBM2 FPGA stress test ≠ GPU workload, HBM3E/HBM4 extrapolated; token energy GPU board/dense models ≠ datacenter electricity, ≠ MoE.

Not opened (not a warrant)

  • Epoch AI Topic Overview AI Chips — opened 2026-09-01 (source 24); chip-production data insight source 25. SK hynix/Samsung/Micron annual reports still unopened.
  • NVIDIA H100 datasheet PDF — download returned HTML; spec via product page.
  • TSMC Form 20-F / CoWoS kwpm — reticle sizes via 2026 Symposium PR (source 26); kwpm still not in the letter/PR.
  • SK hynix / Samsung / Micron HBM annual reports — not opened. HBM bottleneck only via IEA.
  • USGS OFR 2024-1057 — the China-99% claim comes from the IEA PDF, not this OFR.
  • Wu et al. 2021 (Meta) arXiv:2111.00364 — citation inside Luccioni only. E3.
  • Barr 2019 AWS Inf1 — not independent. E3.
  • IEA Energy and AI methods annex / supply chapter — not complete.
  • IEA Key Questions full PDF (138 pp) — not opened.
  • JRC/ENTSO-E 2025/26 EU DC update — not opened.
  • Kaack et al. 2022 Nature Climate Change — not opened.