← Frontier Inference Margins · all research reports

In today's words — This is an archived research artifact. Where it says marginal serving gross margin, unit CM or serving contribution margin, the site now says serving margin; where it states a confidence interval, the site treats it as a judgment range, not a calibrated one. What the site adopted from any artifact is stated in the main report, which is authoritative where the two differ.

Research artifact — each page states its own provenance class in its header: verbatim originals carry SHA-256 stamps; adopted-findings summaries and reconstructions say so explicitly. Conclusions are synthesized (and where needed corrected) in the main report; each page carries its own run date.

Data-center and electricity registry

Built from checked-in evidence only. Programme rows remain programme rows and do not identify a facility.

Every triple carries two orthogonal labels: a closed basis (what kind of evidence stands behind it) and an observationKind (what the three numbers mean). Only a selected span, a tariff-derived delivered span and a two-load-state average/peak may carry different endpoints; a point, a floor, a ceiling, a milestone and a region fill are single values by construction.

Regions

id	registered content	basis	observation	source	class
`us-industrial`	$0.0617 / $0.0871 / $0.1053 per kWh	analyst-set	selected-span	electricity study of 2026-08-23 (not published separately)	analyst-set
`cn-coastal`	$0.0890 / $0.0980 / $0.1100 per kWh	measured/credibly-reported	tariff-derived-delivered	electricity study of 2026-08-23 (not published separately)	measured/credibly-reported
`cn-western`	$0.0600 / $0.0710 / $0.0870 per kWh	measured/credibly-reported	tariff-derived-delivered	electricity study of 2026-08-23 (not published separately)	measured/credibly-reported

Named facilities

id	registered content	basis	observation	source	class
`xai-colossus-c1`	Colossus C1, Memphis, Tennessee; mixed installed aggregate 220,000 / 225,000 / 230,000 across h100, h200, gb200 — no public per-SKU split (allocation: unsplit); milestone 100,000 h100 (2026-06-05)	per observation	per observation	electricity study of 2026-08-23 (not published separately)	partial / purpose-built-ai
`xai-colossus-ii`	Colossus 2, South Memphis/Whitehaven, Tennessee; 110,000 gb200 (point); 110,000 gb300 (point)	per observation	per observation	named-facility study of 2026-08-23 (not published separately)	partial / purpose-built-ai

Programmes (not facilities)

id	registered content	basis	observation	source	class
`deepseek-h800-serving-2025`	Disclosed V3/R1 production serving trace; average 1,814 / peak 2,224 h800	per observation	per observation	fleet-composition study of 2026-08-23 (not published separately)	programme
`anthropic-rainier-trainium`	Project Rainier completed Trainium2 milestone; 500,000 trn2 (milestone)	per observation	per observation	fleet-composition study of 2026-08-23 (not published separately)	programme
`anthropic-trn2-in-use-2026`	Anthropic-wide Trainium2 in-use floor; 1,000,000 trn2 (floor)	per observation	per observation	fleet-composition study of 2026-08-23 (not published separately)	programme
`anthropic-tpu-commitment`	TPU v7 capacity commitment; 1,000,000 tpu7 (ceiling)	per observation	per observation	fleet-composition study of 2026-08-23 (not published separately)	programme
`anthropic-tpu7-direct-purchase-estimate`	Estimated direct-purchase component of the TPU v7 commitment; 400,000 tpu7 (point)	per observation	per observation	fleet-composition study of 2026-08-23 (not published separately)	programme
`anthropic-tpu7-gcp-rented-estimate`	Estimated GCP-rented component of the TPU v7 commitment; 600,000 tpu7 (point)	per observation	per observation	fleet-composition study of 2026-08-23 (not published separately)	programme
`anthropic-colossus-c1-capacity`	Claude capacity at Colossus C1; no public count	per observation	per observation	electricity study of 2026-08-23 (not published separately)	programme
`anthropic-spacexai-capacity-contract`	Colossus mixed-NVIDIA reserved-capacity contract; no public count	per observation	per observation	fleet-composition study of 2026-08-23 (not published separately)	programme
`google-spacexai-capacity-contract`	Future mixed-NVIDIA reserved-capacity contract; no public count	per observation	per observation	fleet-composition study of 2026-08-23 (not published separately)	programme
`google-ironwood-ga`	Ironwood / TPU v7 Gemini serving platform; no public count	per observation	per observation	fleet-composition study of 2026-08-23 (not published separately)	programme
`ant-h20-production-node`	SGLang/Ant DeepSeek-R1 production node; 16 h20 (point)	per observation	per observation	`site/engine-data-v22.js`	programme
`huawei-cloudmatrix-384`	CloudMatrix 384 DeepSeek-R1 demonstration; 384 ascend (point)	per observation	per observation	`research/provider-dives/deepseek-gptpro.md`	programme
`deepseek-v4-ascend950-serving`	DeepSeek V4 support on Ascend 950-series supernodes; no public count	per observation	per observation	fleet-composition study of 2026-08-23 (not published separately)	programme
`zai-third-party-cloud-inference`	Cloud-hosted GLM training, hosting, and inference; no public count	per observation	per observation	fleet-composition study of 2026-08-23 (not published separately)	programme
`moonshot-kimi-k2-h800-training`	Kimi K2 H800 training topology; no public count	per observation	per observation	fleet-composition study of 2026-08-23 (not published separately)	programme
`moonshot-alibaba-hopper-2026`	Reported Alibaba Hopper-generation compute arrangement; no public count	per observation	per observation	fleet-composition study of 2026-08-23 (not published separately)	programme
`alibaba-ulanqab-m890`	Kimi K3 production service on Ulanqab M890 instances; no public count	per observation	per observation	China electricity-tariff study of 2026-08-23 (not published separately)	programme

Planning-rent quotes and the selection policy

Selection policy: low-committed — the low/committed planning observation for each hardware — a selection policy, not one literal contract class.

quote id	hardware	$/hr (lo / mid / hi)	rate class	term	as of	role
`gb200-owner-adopted-scenario-2026-09`	gb200	$4.5	reserved-neocloud-estimate	unspecified	2026-09-10	planning default
`gb300-owner-adopted-scenario-2026-09`	gb300	$6	reserved-neocloud-estimate	unspecified	2026-09-10	planning default
`trn3-owner-adopted-scenario-2026-09`	trn3	$2.2	reserved-neocloud-estimate	unspecified	2026-09-10	planning default
`h100-oneyear-lowcommitted-2026-08`	h100	$2.35 / $2.4 / $3.19	one-year-low-committed	about one year	2026-08-22	planning default
`h200-reserved-neocloud-2026-08`	h200	$3 / $3.68 / $3.99	reserved-1-2yr	about one to two years	2026-08-22	planning default
`gb200-coreweave-public-slice-2026-08`	gb200	$10.5	on-demand-public-slice	on demand	2026-08-22	alternate
`gb200-aws-capacity-block-2026-08`	gb200	$10.582	capacity-block	capacity block reservation	2026-08-22	alternate
`gb200-public-slice-capacity-block-span-2026-08`	gb200	$10.5 / $10.54 / $10.582	on-demand-public-slice-and-capacity-block	on demand and capacity block	2026-08-22	alternate
`h20-china-specialist-bare-metal-2026-08`	h20	$0.71 / $0.82 / $1.02	monthly-term-bare-metal	monthly / annual term	2026-08-22	planning default
`h20-alibaba-managed-on-demand-2026-08`	h20	$7.43 / $9.04 / $10.66	managed-cloud-on-demand	on demand	2026-08-22	alternate
`tpu7-google-3yr-committed-2026-08`	tpu7	$5.4 / $5.4 / $5.94	committed-3yr-list	three-year committed	2026-08-22	planning default
`tpu7-anthropic-strategic-estimate-2025-11`	tpu7	$1.6	strategic-estimate	multi-year strategic	2025-11-28	alternate
`trn2-aws-capacity-block-2026-08`	trn2	$2.235	capacity-block	capacity block reservation	2026-08-22	planning default
`h800-china-annual-commit-2026-provisional`	h800	$1.75	one-year-low-committed	annual commit	2026-08-22	planning default
`ascend-tender-candidate-2026-provisional`	ascend	$1.95	tender-candidate	tender candidate	2026-08-22	planning default

Hardware with no admissible public planning quote

These rows resolve through the unavailable-row path and price nothing unless a reader or a caller states the replay explicitly. There is no silent fallback to a retired value.

hardware	why no default	declared replay

Coverage accounting

Denominator: share of the MODELED fleet resting on DC-specific public evidence — not how much of the real fleet is known.

The three named-site / programme-type / generic-fill parts partition the modeled blend and sum to 100. The SKU/workload count-backed share is non-additive: it says how much of the same blend rests on a disclosed count for the exact key at the relevant workload and date, and it is never added to the three. Percentages are DERIVED from the stored key-level evidence and the selected modeled blend; the registry stores none of them.

company	preset	named-site serving	programme-type	generic fill	SKU/workload count-backed (non-additive)	facility rows	programme rows
anthropic	`opus`	50.7%	33.3%	16%	0%	`xai-colossus-c1`	`anthropic-rainier-trainium`, `anthropic-trn2-in-use-2026`, `anthropic-tpu-commitment`, `anthropic-tpu7-direct-purchase-estimate`, `anthropic-tpu7-gcp-rented-estimate`, `anthropic-colossus-c1-capacity`, `anthropic-spacexai-capacity-contract`
anthropic	`sonnet`	50%	25%	25%	0%	`xai-colossus-c1`	`anthropic-rainier-trainium`, `anthropic-trn2-in-use-2026`, `anthropic-tpu-commitment`, `anthropic-tpu7-direct-purchase-estimate`, `anthropic-tpu7-gcp-rented-estimate`, `anthropic-colossus-c1-capacity`, `anthropic-spacexai-capacity-contract`
anthropic	`haiku`	50%	25%	25%	0%	`xai-colossus-c1`	`anthropic-rainier-trainium`, `anthropic-trn2-in-use-2026`, `anthropic-tpu-commitment`, `anthropic-tpu7-direct-purchase-estimate`, `anthropic-tpu7-gcp-rented-estimate`, `anthropic-colossus-c1-capacity`, `anthropic-spacexai-capacity-contract`
openai	`gpt`	0%	0%	100%	0%	none	none
openai	`terra`	0%	0%	100%	0%	none	none
openai	`luna`	0%	0%	100%	0%	none	none
google	`gemini`	0%	0%	100%	0%	none	`google-ironwood-ga`, `google-spacexai-capacity-contract`
google	`gemflash`	0%	0%	100%	0%	none	`google-ironwood-ga`, `google-spacexai-capacity-contract`
xai	`grok`	0%	0%	100%	0%	`xai-colossus-c1`, `xai-colossus-ii`	none
deepseek	`dsr1`	0%	100%	0%	100%	none	`deepseek-h800-serving-2025`, `deepseek-v4-ascend950-serving`
deepseek	`dsv4`	0%	0%	100%	0%	none	`deepseek-h800-serving-2025`, `deepseek-v4-ascend950-serving`
deepseek	`dsv4f`	0%	0%	100%	0%	none	`deepseek-h800-serving-2025`, `deepseek-v4-ascend950-serving`
zhipu	`glm`	0%	0%	100%	0%	none	`zai-third-party-cloud-inference`
zhipu	`glm47`	0%	0%	100%	0%	none	`zai-third-party-cloud-inference`
moonshot	`kimi`	0%	0%	100%	0%	none	`moonshot-kimi-k2-h800-training`, `moonshot-alibaba-hopper-2026`, `alibaba-ulanqab-m890`