Read the explanation
This video audits source arithmetic, not current export law or shipment authorization. The source default state uses two hundred ninety six FP8 teraflops and multiplies by eight to produce total processing performance twenty three hundred sixty eight. Dividing by die area eight hundred fourteen gives density about two point nine one after rounding. Bars compare performance and configured cap forty eight hundred at one tenth pixel per unit. The code tests strict greater-than for performance, rounded density and interconnect thresholds, so exact equality does not trigger a violation. Those rules are an authored simplification and do not independently verify applicable jurisdiction, destination, entity, license conditions or current regulation. The original input range may normalize an off-step preset display independently from the stored state used in calculation. Moving the native FP8 control to its maximum forty five hundred yields total processing performance thirty six thousand and density forty four point two three. Bars compare default and changed rounded density at eight pixels per unit. Both performance and density exceed configured caps, so the code selects its fail branch; interconnect stays nine hundred, equal to its cap. If only either performance or density or interconnect exceeds a cap, it selects warning instead. Increasing only die area can change density without changing total processing performance, illustrating the two arithmetic inputs rather than establishing lawful compliance. Independent native original and preview contexts download the actual JSON and preserve both canvas drawings through four native widths. Cluster capacity is GPU count times per-chip FP8 teraflops divided by one million. Sixteen thousand three hundred eighty four chips at forty five hundred give seventy three point seven three exaflops after rounding. Separately the interconnect heuristic clamps bandwidth over eighteen hundred between point two and one, then multiplies its missing fraction by thirty eight, assigning nineteen percent penalty at nine hundred. Bars use five pixels per number with distinct compute and penalty units, not a comparable physical quantity. Memory bandwidth is shown and affects decorative diagrams but not these estimates. Revenue and domestic substitution captions use fixed market-size and branch factors rather than measured demand, so they are not forecasts. Device specifications and regulatory assertions remain unverified; no bypass advice, public publication or exhaustive compliance approval is claimed.