The Hyperscalers Are Just Getting Started
Microsoft, Google, Amazon, and Meta are in a full-on arms race to build AI-native cloud infrastructure. Their collective capital spending is already surging.
- Microsoft guided to "materially higher" Azure infrastructure CapEx in fiscal 2025
- Google Cloud is ramping aggressively across all regions
- Amazon is pouring billions into custom chips while still buying NVIDIA GPUs
- Meta going "all in" with hundreds of thousands of GPU equivalents
Blackwell Changes the Game — Again
NVIDIA didn't just rest on the Hopper laurels. The Blackwell architecture announced at GTC is a monster leap: roughly 4x the training performance and 30x the inference performance of H100.
The Grace-Blackwell superchip pairs NVIDIA's CPU with its GPU in a tightly integrated package that hyperscalers love for efficiency.
This isn't a one-quarter story. Blackwell will be the workhorse platform through 2026 and beyond.
The Competition? Still Playing Catch-Up
AMD's MI300X looks solid on paper. Intel's Gaudi 3 is cheaper per token on certain inference workloads. The hyperscalers' custom silicon efforts are real.
But here's the reality check: CUDA. NVIDIA's software ecosystem is a decade ahead, and the switching costs for AI developers are enormous.
The network effects are brutal — more developers → better libraries → more performance → more adoption → even stronger moat.
Valuation: Expensive, But Worth It
Yes, the forward P/E is north of 40x. But when a company is growing revenue 100%+ year-over-year with 60–70% gross margins, the multiple makes sense.
The rule of 40 (growth + margin) is blowing past 150. Compare that to any other mega-cap tech name and NVIDIA still screens reasonably.
$140 feels like a milestone, not a peak. This ride still has plenty of runway left.