Body of letter:
“We have tried to set conservative forecasts for AMD's Al, given historic false starts, but higher price points plus supply chain commentary is pointing to an opportunity that is multiples of our initial target. Raising Al revs to $400 mm, but with bull case >$1.2 bn for CY24.
We have spent the last few days gathering data points from industry sources on the AMD MI300 (AI) opportunity. Our original estimate of $100 mm of Al revenue in CY24 in addition to other data center GPU such as cloud gaming or supercomputer) was intended to be conservative – and AMD management commentary has also been circumspect – but the opportunity in formation is materially larger.
Our thought process is that 1) the overall inference market for GPUs is only about $400 mm trailing, and while it's clear that growth will be explosive in the near term, AMD is coming from outside the market, and customer's biggest priority is to reduce cost per query so price will be an issue. 2) We have been here before, where customers state enthusiasm for AMD data center GPUs, but NVIDIA reacts aggressively to defend their turf with very good products and dilute the opportunity, and 3) our price assumptions for MI300 was that it would come at a significant discount to the A100 points of about $9k.
Through conversations with several industry contacts – and to be clear, none of this is coming from AMD management, who has yet to size the opportunity – we are becoming more optimistic on all three fronts.
We continue to hear about more aggressive investment in both training and inference, and our initial numbers were likely too low, as we have discussed recently. While initially we had heard about LLM inference still heavily using (PUs, that appears to be only the case for select low latency applications – which are significant. If there is no urgency for results, CPUs are fine, but for LLM it's clear that hitting the latency requirements needs GPUs or specialty silicon.
In terms of AMD's share of that opportunity, it seems more tangible and durable than our initial expectations, with solid revenue forecasts from two of the four
US hyperscalers – with prospects for a third – at unit levels that materially surprised us. We would still expect NVIDIA to be aggressive in defending its turf, but AMD is seeing solid orders from these customers…”
https://twitter.com/fundam_inv/status/1656747330216005632?s=46&t=EfdUOH7OfS4lpc8-o1U6Dg
Leave a Reply