Newer Nvidia chips don't just run AI faster, they turn hyperscalers into cash machines (Save this).
Milk Road AI Twitter · Milk Road AI (@MilkRoadAI) · 2026-07-28
Morgan Stanley analysis projects hyperscaler net margins rising from 58% on Blackwell-based data centers to 90% on Feynman chips, while custom ASIC shipments grow 44.6% versus GPU shipments' 16.1%, reshaping AI compute economics in 2026.
Appears in
Extraction
Topics: nvidia-chipsai-infrastructurehyperscalerscustom-asicsneocloud
Claims
- Morgan Stanley projects net margins of 58% on Blackwell-based data centers, rising to 78% on Rubin and 90% on Feynman-based data centers.
- Roughly $51 of every $100 spent on Nvidia-based AI compute flows to Nvidia's 85% gross margin, a cost Amazon and Google avoid entirely with custom silicon.
- Custom ASIC shipments are growing 44.6% versus GPU shipments' 16.1% for the first time, with a 40–65% cost advantage on large-scale inference.
- Nvidia is estimated to capture $1 trillion in cumulative data center revenue from 2025–2027 at approximately 75% gross margins.
- Synergy Research expects neocloud revenue to more than triple to $23 billion in 2025 and approach $180 billion by 2030.
Key quotes
Morgan Stanley's intelligence factory model projects net margins of 58% on Blackwell based data centers, climbing to 78% on Rubin and up to 90% on the newer Feynman based data centers.
Roughly $51 of every $100 spent on Nvidia based AI compute goes straight to Nvidia's 85% gross margin, a cost Amazon and Google avoid entirely by running their own silicon, AWS Trainium and Google TPU.
Custom ASIC shipments are actually growing faster than GPU shipments for the first time this year, up 44.6% versus 16.1%, with a 40-65% cost advantage on large scale inference.