OpenAI & AMD – Power Dynamic

2^nd largest frame order in history.

• AMD’s deal with OpenAI is a game-changer for AMD, bringing it from relative obscurity to being a major AI player, but this comes at the cost of 10% dilution for shareholders and the risk that large portions of this deal never actually materialise.
• Not to be outdone by Nvidia, AMD has signed a colossal 6GW deal with OpenAI that I estimate will generate around $60bn in revenues for AMD over the next 5 years, but it is very clear who is in the driving seat.
  • First, No-show Sam: as Sam Altman did not show up for the conference call to discuss the deal, although he did show up for the call with Nvidia a couple of weeks ago.
  • This is a demonstration of who is the most important supplier and is an indication that AMD wanted this deal more than OpenAI did.
  • Second, the 5-year warrant: which AMD has given to OpenAI to get the deal done, in my opinion.
  • AMD claims that it is for strategic alignment, but it is pretty clear to me that this was to get the deal over the line and to get the chips paid for in cash.
  • A warrant is the option, but not the obligation, to purchase a new share issued by the company at the strike price, which in this case is $0.01 per share (i.e. free) within the next 5 years.
  • AMD is issuing 160m warrants, which will all be exercised, meaning that existing shareholders will be diluted by 10%.
  • In 5 years, any unexercised warrants will expire, and there will be no ability to exercise them even if the conditions are subsequently met.
  • This is what leads me to think that the deal foresees that the entire 6GW will be deployed within 5 years from now (maybe faster).
  • However, the warrants will only be exercisable upon the deployment of each GW and should the share price obtain certain levels.
  • AMD would not say what the levels are, but it did say that the last tranche is exercisable when the whole 6GW is deployed and the share price is at or above $600 per share.
  • Based on this, I would guess at 4 increments of $80 per share, with the first tranche going at something like $280 and then $360 and so on.
  • $600 is 3x where it is today, and so from a shareholder perspective, the 10% dilution does not worry me at all as this looks like a good trade to get the deal done with OpenAI.
• However, the one area where this deal beats the deal with Nvidia hands down is that it looks like OpenAI will be paying for the chips with cash, as no OpenAI shares will be given to AMD.
• This makes this deal even less circular than the Nvidia deal, which, in my opinion, was not really circular at all, as OpenAI is simply paying for its chips with shares (which I think is not a problem).
• This is not like the deals struck in the Internet bubble, where infrastructure vendors lent telecom operators the money to buy equipment, the debts from which then caused them to collapse when demand went soft.
• If demand goes soft, neither Nvidia nor AMD will go bust as there is no debt involved here, but their valuations will certainly take a hit, which is another matter entirely.
• Hence, I think all of this talk about circular deals is needlessly negative and that the articles making this case are inaccurate.
• As far as I can figure out, the deal will appear as follows.
  • First, 6m chips sold: The current generation consumes around 1400W per chip, so if I assume an improvement in the MI450, this would equate to around 1000W per chip.
  • Based on this assumption, AMD would need to sell around 6m chips to add up to the 6GW announced.
  • Second, $60bn revenues: which will occur over the next 5 years.
  • Based on the pricing of the current generation, which is about $20,000 per chip, and assuming a large volume discount of 50%, 6m chips will equate to roughly $60bn.
  • This is broadly in line with the vague guidance that AMD was prepared to give about the deal.
  • Third, 4 years: as the revenues will commence in H2 2026, with the deal lasting 5 years from now.
  • This would mean something like $5bn in 2026 and around $14bn for the next 4 years.  
  • On top of revenues from OpenAI, this is likely to have a substantial halo effect, which will drive more customers to sign up with AMD and order more chips.
• That’s why this is a game-changer as AMD goes from an also-ran to a solid No. 2 with AI data centre chip revenues of $30bn a year or more (assuming the bubble doesn’t pop).
• It is also a big vote for the software that is used to train and run inference AI on its silicon, and the credibility of this takes a large step forward with this deal.  
• Given that revenues are currently $30bn annually, the pathway to grow faster than Nvidia in the medium term is quite clear, given how big Nvidia has become.
• Furthermore, margins will also increase thanks to scale and operating leverage (fixed costs rising more slowly than revenues).
• Although AMD’s financials are about to substantially improve (even with the 10% dilution), the valuation is not attractive.
• The EPS estimate already anticipates that net income will double over the next two years, which leaves the company on 25.8x 2027 PER.
• This is compared to Nvidia, which is on 24.8x 2027 PER and remains larger, much less risky and more powerful in its market than AMD.
• This deal looks like it will bring AMD into line with Nvidia in terms of valuation, and I think it will need more large deals to keep the upward momentum going.
• Hence, I would continue to prefer Nvidia over AMD and would continue to look to the adjacencies of inference at the edge and nuclear power to find companies with low valuations and real upside.