How the numbers are made

A model's architecture is public the day it ships. Its procurement footprint is not public for quarters. This is the arithmetic between the two.

Every figure below is computed from the one above it. Nothing here is a forecast of a share price and nothing is a recommendation — the chain answers if this coefficient moves, which names does it move, and roughly how much. What you do with that is your job, not ours.

The cascade

Colour is the grade of the input at that step. The argument gets weaker as it descends, and that is the honest shape of it — the physics is computed, the workload is assumed, the prices are estimates.

KV cache per token

252 KB

medium

From the attention architecture in a published config. GQA against MLA is a 7.3x difference in the HBM each token of context consumes.

x 2 x active params

Compute per token

810 GFLOP

medium high

rack power / measured throughput

Energy per token

0.000219 Wh

medium

x disclosed volume

Token volume

107 tn/day

low medium

Google's disclosure only. A FLOOR on the industry, not an estimate of it.

= intensity x volume

Sustained compute

1,003 EFLOP/s

asserted

KV cache resident

4,228 TB

asserted

Continuous power

976.38 MW

medium

= max(compute-bound, memory-bound)

Fleet

1.3m accelerators

asserted

entity:gb200 class, using that entity's own published specs rather than a number typed into this tool.

COMPUTE binds, by 30.8x. The fleet has to satisfy compute and memory both, so its size is the larger of the two requirements — 1.3m accelerators to meet compute, 41,328 to hold the cache. So an MLA-style cache compression frees headroom the fleet already has: it does not reduce the accelerator count and does not reduce HBM revenue.

This is the single most consequential line on the page. It decides whether algorithmic progress destroys memory demand or leaves it untouched, and it flips if the memory requirement overtakes the other.

Where it lands in dollars

market	physical	implied spend
HBM deployed	236.9 PB	$2.8bn
Accelerators	1.3m	$102.5bn
Facility capex	976 MW	$9.8bn

Coherence note. HBM content works out to $2,232 per accelerator, 2.8% of the unit price, against teardown estimates of 10–20% of BOM. The likely cause is a scope mismatch rather than a bad price — the accelerator figure is a system spend including networking and integration, while $/GB buys the memory alone. The two are different units of account and must not be summed, and the HBM figure should be read as a floor.

Which companies that lands on

The HBM market split by the graph's own supply weights. A row marked assumed rests on the 0.5 default rather than a disclosed share — a revenue number resting on that is a guess wearing a company name.

supplier	share	implied revenue	weight
sk-hynix	55.0%	$1.6bn	from model
samsung	25.0%	$0.7bn	from model
micron	20.0%	$0.6bn	from model

Every input, and how much it is worth trusting

input	value	grade	source
intensity — Computed from a published config or a measured series. The defensible layer.
kv_kb_per_token	252	medium	coefficients.py (Llama GQA, FP8)
wh_per_token	0.000219	medium	entity:energy-per-token
gflops_per_token	810	medium high	coefficients.py (2 x active params)
tokens_per_day_tn	107	low medium	entity:disclosed-token-volume (Google only — a FLOOR)
avg_context_tokens	32,768	asserted	VOLUME ASSUMPTION — not measured
concurrent_requests_k	500	asserted	VOLUME ASSUMPTION — not measured
workload — How the fleet is actually used. No provider publishes these — they are assumptions, and they are where the chain becomes an estimate.
mfu	0.35	asserted	Typical published serving MFU. NOT measured on this fleet
weight_fraction	0.45	asserted	Depends on model size, replication and parallelism strategy — a modell
price — Unit prices. Desk estimates from reported figures, not contract prints.
hbm_usd_per_gb	12	low	Desk estimate from reported HBM stack pricing — NOT a contract print
dc_usd_per_mw	10.0m	asserted	Industry order-of-magnitude, carried from the megawatt thesis where it
accel_usd_each	80,500	low medium	Goldman build-out model Rubin baseline, cited in supplied analysis

Composite grade: ASSERTED, set by avg_context_tokens. A chain is no stronger than its weakest term, and 5 of 11 inputs are asserted: mfu, weight_fraction, dc_usd_per_mw, avg_context_tokens, concurrent_requests_k. Those are not the weak part by accident — context length, concurrency, MFU and weight fraction are precisely the quantities no provider publishes. The physics is the defensible part. The dollar conversion and the workload shape are where this becomes an estimate, and the output must not be quoted at the precision the physics has.