AI Accelerator Hardware Update

Nvidia Blackwell Delay Impact On Other Semiconductor Payers

Nvidia (NVDA) Blackwell is now expected to ship in Nvidia’s fiscal Q4 per CEO Jensen Huang’s comments during the Company’s most recent earnings call. This is somewhere between a one to two quarter slip compared to the previous guidance of: “our production shipments will start in Q2 and ramp in Q3, and customers should have data centers stood up in Q4.”

While there is no clarity on exactly what the Company will ship in Q4, management commentary suggests that the Company will likely ship GB200 modules first followed by limited quantities of B200. The guidance of “several” billion dollars in Q4 revenues suggests that the Company plans to ship at least $3B in Blackwell products during the January quarter. While $3B is a very strong number, this suggests that only about Q4 data center shipment revenues will come from Blackwell. A crossover with Hopper revenues does not seem likely until the arrival of B200A in Q2 CY2025.

The 1 to 2 quarter delays means that the following companies are likely to see some notable changes in demand: TSMC (TSM), Advanced Micro Devices (AMD), Intel (INTC), Broadcom (AVGO), Marvel (MRVL), Micron (MU), Samsung (SSNLF), and Hynix.

While much of the semiconductor supply chain, well beyond the names discussed above, is impacted, this article is limited to these few names.

TSMC Benefits As Nvidia Loses Share

One underappreciated dynamic in the AI scene is that, with about 75% margins, most of the value in the ecosystem is accruing to Nvidia. The margins of Nvidia’s competitors are 10% to 20% below this. A different way of looking at this is that, per a given dollar of customer hardware spend, the semiconductor ecosystem, including foundries like TSMC and memory suppliers like Micron, Hynix, and Samsung, will benefit more if the hardware is supplied by Nvidia’s competitors.

For each dollar of AI accelerator hardware capex deployed, TSMC gets about 10 cents with a Nvidia deployment and closer to 20 cents with AMD deployment and closer to 25 cents with Broadcom or Marvell solutions. While these are all approximations and actual numbers can vary materially, expect TSMC to do better to the extent that Blackwell problems shift AI hardware business to Nvidia alternatives.

A $100B of incremental capex to Nvidia competitors would benefit TSMC by an incremental $10B or more!

Memory vendor implications

Similarly, memory vendors are likely to benefit from Blackwell for two reasons:

• Per unit of compute, Hopper uses more memory than Blackwell. To the extent that customers switch to Hoppers, incrementally more HBM will be used.

• To the extent that AMD gains share, AMD solutions use far more HBM than Nvidia. For example, each MI325 replacing H200 would double the HBM demand. Given Nvidia’s premium pricing, memory vendor gains are likely to be even higher per each dollar of capex. On average, HBM demand could jump by about 150% with AMD than with Nvidia. This would be a major demand boost for memory vendors – even more so than for TSMC.

• Given that Broadcom and Marvell likely command only 20% to 30% margin on HBM, compared to 75% at Nvidia, HBM use per unit of capex is also likely to be much higher with Broadcom and Marvell alternatives than Nvidia Hopper or Blackwell.

As a result of these factors, memory vendors are likely to see a tailwind to their already robust memory sales. The tailwind may not be big enough to materially change the current memory narrative, but memory stocks could strengthen meaningfully as visible signs of market share shifts emerge.

All Accelerator Competitors Benefit

The following table is a bird’s eye view of the current and near-term state of the art solutions from various players.