UBS data: 39% of hyperscaler AI capex flows to $NVDA today. By 2028, that share climbs to 49%—on a capex base 60% larger than now.
Hyperscaler spend goes from $492B (2025) to $1.6T (2028), a 3.3x jump. NVIDIA's data center revenue scales from $194B to $792B—4.1x growth—because its slice of the pie expands even as the pie itself triples.
Bill of materials doubled from Blackwell to Vera Rubin: single rack now costs ~$7.8M, memory alone up 400% to over $2M. Input costs doubled. Gross margin? Still 75%. NVIDIA passed the entire cost increase to buyers without hesitation. That's pricing power.
But the model just changed. July 1: NVIDIA launched a revenue-share pilot with Sharon AI and Firmus across 210,000 Grace Blackwell GPUs. Now NVIDIA sells the rack at 75% margin, then collects a cut of whatever that rack earns over its 10-year lifespan. Morgan Stanley models this second revenue stream at ~$51B/year by FY2029, at near-100% gross margin—no hardware, no depreciation.
None of that recurring revenue is priced in yet.
$NVDA trades ~22x forward earnings. Take UBS's $792B revenue estimate, apply a conservative 55% net margin (below the current ~63%), keep today's multiple with zero expansion, and you land near $400/share vs. $219 today. Margin compression assumption is likely too harsh.
If you think the company with the best GPUs, the largest scale, and a royalty on its own output somehow loses pricing power from here, you've lost the plot.
Hyperscaler spend goes from $492B (2025) to $1.6T (2028), a 3.3x jump. NVIDIA's data center revenue scales from $194B to $792B—4.1x growth—because its slice of the pie expands even as the pie itself triples.
Bill of materials doubled from Blackwell to Vera Rubin: single rack now costs ~$7.8M, memory alone up 400% to over $2M. Input costs doubled. Gross margin? Still 75%. NVIDIA passed the entire cost increase to buyers without hesitation. That's pricing power.
But the model just changed. July 1: NVIDIA launched a revenue-share pilot with Sharon AI and Firmus across 210,000 Grace Blackwell GPUs. Now NVIDIA sells the rack at 75% margin, then collects a cut of whatever that rack earns over its 10-year lifespan. Morgan Stanley models this second revenue stream at ~$51B/year by FY2029, at near-100% gross margin—no hardware, no depreciation.
None of that recurring revenue is priced in yet.
$NVDA trades ~22x forward earnings. Take UBS's $792B revenue estimate, apply a conservative 55% net margin (below the current ~63%), keep today's multiple with zero expansion, and you land near $400/share vs. $219 today. Margin compression assumption is likely too harsh.
If you think the company with the best GPUs, the largest scale, and a royalty on its own output somehow loses pricing power from here, you've lost the plot.