NVIDIA: HBM Shortage Won't Hinder AI Infrastructure Expansion
0xBroomberg
Jensen Huang told a Morgan Stanley briefing that Nvidia's quarterly revenue is approaching $100 billion, and the core challenge has shifted from whether AI demand exists to converting a massive order backlog into delivered systems — memory, networking, and power are all tight, which he says proves demand is outrunning supply.
HBM costs are rising — how does Nvidia still make more money?
Bank of America cites supply-chain data: per-rack HBM cost under Blackwell is roughly $1.9 million; upgrading to Rubin adds only $200,000–$300,000.
Nvidia plans to raise rack prices from $3–4 million (Blackwell) to $6–7 million (Rubin) — the price jump far exceeds the cost increase.
This means → Nvidia "dilutes" the HBM cost pressure through higher system pricing, keeping gross margins around 70%.
How long will the memory shortage last?
Micron and SK Hynix have both issued storage-shortage warnings; Huang expects memory supply to stay tight for several years.
The root cause: token demand from AI inference is growing faster than storage capacity can expand.
Nvidia is adapting system design — some racks will carry less memory while boosting network interconnects and optimizing compute scheduling.
In plain terms = not enough memory, so let chips "talk" to each other more and "store locally" less — trading network bandwidth for storage capacity.
What lets Nvidia move from selling GPUs to selling full infrastructure?
As AI clusters scale, data transfer between GPUs becomes a bottleneck. Customers are shifting from buying standalone GPUs to buying integrated systems — GPUs, CPUs, switch chips (specialized chips that route data), NICs, optical interconnects, and software bundled together.
Nvidia reaffirmed a CPU revenue target of roughly $20 billion this fiscal year; nearly half may come from standalone CPU racks rather than management nodes inside GPU systems.
This means → the Vera CPU is breaking into the broader server market on its own, and Nvidia's revenue base is expanding from "one chip" to "an entire data center."
What are the biggest risks?
Morgan Stanley flagged three risk categories: faster-than-expected supply expansion causing a sharp slowdown in data-center growth; a steep drop in AI development costs or the emergence of disruptive competitors; and customers accelerating adoption of ASICs — application-specific integrated circuits designed for narrow tasks, cheaper than general-purpose GPUs but less flexible.
TSMC reports earnings on July 16; whether its outlook confirms the "demand far exceeds supply" thesis is the market's next key validation point.
This reflects a deeper reality: the core assumption behind today's AI investment logic is still "demand outstrips supply" — if supply catches up, the entire valuation chain needs repricing.
Content is for reference only, not financial advice.