Capex is a promise; R&D is the promise made earlier. NVIDIA's FY2026 Form 10-K, filed 2026-02-25, reports research and development expense of $18.5 billion for the year. Two years earlier the same line was $8.7 billion. The company more than doubled what it spends inventing the next chip while its revenue tripled. These are XBRL-tagged facts, surfaced through EdgarBeast.
The pattern matters because accelerators are an architecture business, not a commodity one. A vendor that lets R&D stall while demand is hot is harvesting; a vendor that pushes R&D up alongside revenue is defending a roadmap. The filing puts NVIDIA in the second camp: $8.7 billion, then $12.9 billion, then $18.5 billion in successive years.
Set against the disclosed $215.9 billion in revenue, R&D is a smaller share of the top line than it was at lower revenue levels — the denominator grew faster. That is the comfortable position to be in: spending more in absolute dollars while the ratio falls, because sales rose faster than the lab bill.
What the 10-K will not tell you is which architecture the money bought, or the return on any single program. Those are not disclosed line items, and I will not invent them. The disclosed fact is the trajectory of the R&D line itself. The 10-K on sec.gov is the primary record; EdgarBeast indexed the XBRL facts behind it.