Идеи
Meta's demand for a custom cut-down MI450 (half compute, half HBM) makes the chip unattractive for GenAI, and TBD has no interest. This could nuke AMD's volume at Meta because TBD will prefer Nvidia's
Meta's demand for a custom cut-down MI450 (half compute, half HBM) makes the chip unattractive for GenAI, and TBD has no interest. This could nuke AMD's volume at Meta because TBD will prefer Nvidia's Rubin unless AMD forces the standard MI450.
Risk: If AMD convinces Meta to adopt the full MI450, the negative stance reverses. The article calls for AMD to 'put on their big boy pants' and work with TBD.
Article details how Meta's custom GB200 (Ariel) had 14% higher TCO and was inferior to the standard NVL72 that competitors purchased. This validates Nvidia's standard SKU advantage and reinforces vend
Article details how Meta's custom GB200 (Ariel) had 14% higher TCO and was inferior to the standard NVL72 that competitors purchased. This validates Nvidia's standard SKU advantage and reinforces vendor lock-in for hyperscalers.
Risk: If Meta successfully fixes its culture, future custom designs could become more competitive, but near-term Nvidia benefits from Meta's missteps.
Grand Teton's design increased reliance on Broadcom PCIe switches (four per server), adding BOM, power, and complexity. Although the design was cancelled, it shows Meta's willingness to use Broadcom c
Grand Teton's design increased reliance on Broadcom PCIe switches (four per server), adding BOM, power, and complexity. Although the design was cancelled, it shows Meta's willingness to use Broadcom components when avoiding Nvidia networking.
Risk: Meta's frequent pivots could lead to cancelled Broadcom design wins, as seen with Rivos IP cancellation.
Meta's custom AMD MI450 uses 8-Hi HBM4 instead of 12-Hi, implying lower HBM content per GPU. If Meta shifts to standard MI450 or Nvidia Rubin, HBM demand could rise. More directly, the article's criti
Meta's custom AMD MI450 uses 8-Hi HBM4 instead of 12-Hi, implying lower HBM content per GPU. If Meta shifts to standard MI450 or Nvidia Rubin, HBM demand could rise. More directly, the article's critique of cutting HBM in half suggests hyperscalers may prefer full HBM stacks, benefiting HBM suppliers like Micron and SK Hynix.
Risk: HBM oversupply or pricing pressure; Micron's exposure to Meta is not disclosed.
This newsletter, published July 22, 2026,
features Wayne Ma
discussing AMD, NVDA, AVGO, MU.
4 trade ideas extracted by AI with direction and confidence scoring.
Speakers:
Wayne Ma
· Tickers:
AMD,
NVDA,
AVGO,
MU