Search
Close this search box.
Our May 2024 Options Trade: AMD

It was a big news day for AMD, though most of the headlines weren’t picked up in news feeds. First, Elon Musk’s AI startup – xAI – was reported to have chosen Oracle (cloud) and AMD (GPUs) for their new datacenters to be built with the $6 bln they just raised. I would guess AMD’s resulting revenue would amount to roughly $600 mln to $1,200 mln based on the likely number of GPUs to purchase. The purchase could be a vote of confidence in AMD, or it could be the result of AMD being the only available GPU for 2024 and early 2025 delivery and/or the fact that AMD is 1/3 the price of NVIDIA. If we knew the planned startup that would help (If it’s in late 2025 or 2026 then NVIDIA would have been in play). Either way, it’s a nice contract for AMD.

Second, in a major blow to NVIDIA, Samsung did not pass NVIDIA’s testing for the latest-and-greatest HBM3e memory for GPUs (link). NVIDIA is really pushing the envelope with their new ‘Blackwell’ GPUs – which I still believe is a minor upgrade to the H100 which relies on memory and a CPU to speed it up rather than an improved GPU – and I’m guessing the heat generation of Samsung’s memory was too high for NVIDIAs hardware which crams in two GPUs and a CPU on one (albeit large) card. This leaves NVIDIA with only Micron as a supplier, and this will limit production of the Blackwell GPU through 2025. It also frees up memory for AMD to use and memory (plus GPU assembly) is the primary constraint to total industry GPU production through 2025.

Third, HSBC wrote (I don’t have the full note) that NVIDIA’s new GB200 GPU (with two GPUs and a CPU on board) will cost $60,000 to $70,000 – countering the common perception that it will cost $30,000. Jim Cramer started the $30,000 fixation when he asked NVIDIA CEO Jensen Huang how much it would cost and Huang held up one single GPU processor and said ‘this costs $30,000.’

I swear there isn’t usually daily news on AMD, but Ilya Sutskever, cofounder and chief scientist at OpenAI (and arguably the brains behind the operation) as well as Jan Leike, head of the ‘Superalignement Team’ (the team dedicated to preventing rogue AI behaviors at OpenAI) have left the company. Both announced their resignations on Twitter (or X, which still sounds dumb to me):

There are more than a few posts online that a large wave of key employees are following them out the door and they are all going to… xAI – Musk’s startup. I would say announcing the departures on Twitter adds some weight to the rumor – but realistically there is no other place to go (besides starting a new firm).

With $6 bln in hand and some of the best AI minds in the business on board, it now looks like xAI will be able to raise A LOT of money (OpenAI is valued at $80 bln) – and it sounds like xAI will be an all-AMD hardware company. If xAI is truly going to compete with OpenAI, Meta and Google, they need 500,000 NVIDIA H100 equivalent GPUs, or about $6 bln worth at $12,000 per GPU (Meta will have 600,000 at year end, OpenAI may have 700,000). Keep in mind just over half of the processing power at OpenAI is utilized to run the model and serve up user requests, so any company planning equal popularity needs to plan on 200,000 to 400,000 GPUs for running the model – or inferencing – a year or so out.

Not only does xAI look like a big AMD customer, but having a major LLM developer move to all-AMD hardware means a lot of focus will come to developing new software tools for AMD, obviously making their GPUs more appealing (AI developers opensource their software tools and work on them collaboratively…) In NVIDIAs case, the output of the A100 GPU more than doubled over the two years following the release due to optimized software).

Nathan Weiss & Keith Schaefer
Co-Founders, In The Money options trading service
Questions or comments? Reach out to us by email at: help@inthemoney.capital

More
articles

Sign up to receive our next three trades.

Name(Required)