Reflecting on NVIDIA’s recent announcements at GTC 2025, I found myself alternating between genuine surprise and familiar nods. Strategic leaks, thoughtful speculation, and careful reading of the Silicon Valley tea leaves made most developments predictable, yet there's always plenty to unpack.
Datacenter Realities and Infrastructure Challenges
Beyond exciting new products, there's an urgent, ongoing issue—particularly in the U.S.—around datacenter readiness for NVIDIA’s latest hardware. Hopper racks, for example, at 120kW (peaking over 150kW) and weighing almost 3,000 pounds, significantly exceed the capabilities of traditional raised-floor datacenters and many service elevators.
Open question: How quickly can our industry realistically adapt its infrastructure to these escalating power and weight demands?
NVIDIA’s Marketing: A Continued Frustration
Regular readers know my stance on NVIDIA’s marketing tactics. Historically, the graphics accelerator industry has a habit of overstating generational improvements through selective benchmarks. Unfortunately, NVIDIA hasn’t moved away from this approach.
Two critical clarifications are important when interpreting NVIDIA's recent performance claims:
- Product Lifecycle: Hopper is essentially end-of-life, and Blackwell's release in late 2024 has been public knowledge for some time. Comparing Hopper now with Blackwell is akin to comparing last year’s smartphone model in August, with September sales figures of the brand-new release.
- Double Counting: Blackwell chips include two dies per unit, meaning actual normalized sales growth is closer to 1.8M—still a significant 50% increase, but far from the advertised triple growth.
Silicon Photonics: Surprising Innovation
One genuinely unexpected reveal was NVIDIA developing its own Silicon Photonics (SiPho). Jensen Huang convincingly explained why copper, not optical, was preferred for Blackwell due to cost and power concerns. Optical transceivers at around $1k each, consuming about 30W, quickly add up at large scale.
To illustrate, a modest 1024 GPU cluster (~$30M) would require roughly 3MW of power just for the optical transceivers, costing nearly $1M. Yet, crucial details on SiPho’s pricing and latency—key to its viability—were notably missing.
How this competes with other solutions, like Marvell's CPO platform, will be interesting to see later this year. https://www.marvell.com/company/newsroom/marvell-co-packaged-optics-architecture-custom-ai-accelerators.html
SpectrumX and the Missing Quantum
SpectrumX was mentioned 2.5x times more than InfiniBand in Jensen's keynote, and also powers xAI’s Colossus in Memphis of 100,000 GPUs. Interestingly, Quantum technology has quietly vanished from NVIDIA's broader roadmap, reappearing only briefly in specialized SiPho discussions (see slide above)
If you're finding value in this deep dive and want to continue reading further, and stay informed about digital infrastructure and technology strategy, subscribe to unlock my complete newsletter—it's free, insightful, and tailored to industry leaders: