Assumptions
Defaults carry their source. Move anything.
250 MW is chosen to match Nscale's financed 230 MW Narvik site rather than a round number — the guest for this session runs support for that estate.
The assumption that actually moves the answer
Annual cost impact of each input, swept across its plausible range.
Capital intensity
Two independent routes. Reported as a range, not a midpoint.
Nameplate vs. delivered
Peak dense FP4 against MLPerf-sustained throughput.
Huang's Law, re-derived with precision held fixed
Flagship SXM parts only · dense FP16/BF16 only · the one basis every generation from Pascal to Blackwell publishes.
Why this matters. Published “GPU performance over time” curves usually plot each generation's headline number. But the headline format keeps narrowing — FP16, then FP8, then FP4 — and each halving of precision roughly doubles the number without the silicon doing more work. Comparing Blackwell FP4 to Pascal FP16 measures the arrival of new number formats as much as it measures hardware. Holding precision fixed is the only honest version of this chart.
Doubling time
Fitted on the dense FP16/BF16 series above.
Power and bandwidth, like-for-like
Flagship SXM parts only — the comparison the book's Ch. 19 table does not make.
Claim checker
Paste a throughput or power figure. It reports every part, precision and variant the number is consistent with — and flags where it is ambiguous.
The four failure modes
Each is qualifier loss. The number survives; its meaning does not.
Published claims that do not survive checking
Each entry states the claim, the published counter-evidence, and what does and does not change as a result.
The corpus
Sparse and dense are separate fields. Variants are separate records. Unpublished values are empty on purpose.
Sources
Every page actually fetched, with access date.
Ten terms to walk in with
Each with the part the room won't know. Every figure verified against the corpus; tags mark where it came from. Each entry links to the view where the number lives.
Below the rack: where the megawatts go, and who pays to get them back
The other six views stop at the rack. This one follows the power from the substation to the silicon, then prices what efficiency is worth and how the people who deliver it get paid. It reads the Scenario view's live assumptions, so move a slider there and these numbers move too.
1 · Where a megawatt goes
One megawatt drawn from the grid, split by PUE (total facility power ÷ IT power). Overhead split from a published breakdown; IT split by rack power-supply efficiency.
2 · Energy vs. capacity
Same efficiency gain, two ways to value it.
3 · The PUE trap
PUE improves on its own as IT load rises, because part of the overhead is fixed. A contract that pays on PUE pays for growth.
4 · Four ways to get paid
Projected savings come from §2's energy view. Slide how much actually shows up and watch who absorbs the miss.
5 · Who buys
Uptime is the veto. Whoever owns it signs last.
6 · How the deals start: the trust ladder
Nobody hands an unknown model the chillers on day one.
Measurement & verification
Where performance contracts live or die.
Sources for this view
Every external number above, with its tag. Defaults marked T are illustrative and labelled as such.