NVIDIA Corporation: Durable high-quality growth at compelling value multiple, reit Buy

研报英文原文证据摘录

NVIDIA Corporation: Durable high-quality growth at compelling value multiple, reit Buy

Memory inflation predictable, Price power un-appreciated

We believe investors simultaneously overestimate the GM impact of rising memory input

cost, while underestimating NVDA’s pricing power, long-range planning, product co-

design, scale and supply chain alignment via over $119bn in prepurchase commitments

across memory, wafers, packaging and power.

To first order, NVDA high bandwidth memory content per rack goes up by ~$0.2-0.3mn

incrementally from Blackwell to Rubin, while product pricing is likely to go up by over $2-

3mn (from $3-4mn/rack in Blackwell, to $6-7mn/rack in Vera Rubin) because of the

advancements in multiple other parts of the rack such as an upgraded Vera CPU,

upgraded NVLink and Quantum Ethernet networking (none of which need HBM memory),

and a host of other software features that speed-up (and lower cost) time to first token.

Moreover, we flag NVDA’s comments that Vera Rubin can deliver 10x more performance

per watt (~10x lower inference token cost), 3.3x more powerful on inference, and up to

5x more powerful on training than Blackwell.

Overall we expect NVDA to maintain its roughly mid-70s GM, blended across its breadth

of cloud, enterprise, neocloud, sovereign customers, and across its breadth of products

in the rack-scale architecture.

Exhibit 1: We expect HBM $/rack goes from ~$150-300k/rack in Blackwell to ~$400k/rack in

Vera Rubin

HBM cost per rack by GPU generation

Blackwell Blackwell Ultra Rubin Rubin Ultra

GPU B200 B300 V200 V300

Architecture Oberon Oberon Oberon Kyber

Ramp 2H24-2H25 2H25-2H26 2H26-2H27 2H27-2H28

GPUs/rack 72x 72x 72x 144x

HBM/GPU 192 GB 288 GB 288 GB 576 GB

HBM GB/rack 13,824 GB 20,736 GB 20,736 GB 82,944 GB