AMD Q1 25– Party Pooper

USA / China spoils the party.

• AMD’s excellent results show that it is taking share from Intel in PCs as well as gaining share in the data centre as inference becomes a much larger part of demand, but unlike Intel, Nvidia has a fix for this.
• AMD Q1 25 revenues / EPS were $7.4bn / $0.96 comfortably ahead of estimates at $7.1bn / $0.94.
• Guidance was also good with revenues / gross margins of $7.1bn – $7.1bn ($7.4bn) / 54% which was ahead of the consensus of $7.2bn.
• This is a reflection of taking share in both PCs and in the data centre, where revenues grew by 56%, YoY which I suspect was faster than Nvidia supporting AMD’s claims of share gains.
• Unfortunately, here, the good news stopped and these results were spoiled by the forecast that China restrictions could cost $1.5bn in 2025 revenues, leading to a 6% gain after the results were published turning into a no-score draw once the conference call had concluded.
• There is no doubt in my mind that AMD’s strong quarter was a result of two factors.

• • First, AI inference: where it is clear that demand for data centre compute to be used for running inference on trained models to deliver services to customers is growing far faster than training.
  • The latest trick to keep performance going is to run inference on “reasoning models” multiple times and then have another model pick the best response.
  • There are also good gains in performance to be had by allowing the model to compute for longer before coming up with an answer to the enquiry.
  • This means that compared to a regular model like GPT-4o, a “reasoning model” will use many orders of magnitude more compute.
  • Hence, as Nvidia pointed out at GTC, demand for inference compute is beginning to outstrip training.
  • AMD is doing far better in inference than it is in training because it is much easier to run inference without CUDA meaning that the Nvidia lock-in is much weaker in this segment.
  • For example, many people will train on Nvidia and then run inference on something else like AMD or xAI’s LPUs.
  • However, Nvidia has a fix for this which was launched at GTC but everyone ignored because it does put up big numbers or is a lump of silicon sitting in a big box for everyone to stare at.
  • I am referring to Dynamo which I have written on before (see here) and will summarise briefly below.
  • Second, Intel is on its knees: which is almost like taking candy from a baby.
  • Intel is in existential trouble and is currently floundering with no direction of travel and no real strategy for recovery.
  • Against this backdrop, some market participants are losing faith in Intel, which is putting wind in AMD’s sails.
  • Combine this with good decision-making by AMD and TSMC’s manufacturing prowess and Intel is being beaten at its own game and losing market share.
  • I think that AMD and the Windows on Arm camp will continue to eat into Intel’s market share in the short to medium term.
• Unfortunately, further chip restrictions mean an $800m write-off in Q2 25 and potentially $1.5bn lower revenues as a result of lost sales.
• Here, we see one of the differences between Nvidia and AMD as Nvidia has so much demand that the capacity it had earmarked for China that will now be lost and will be quite easily allocated elsewhere.
• Hence, I am not expecting any commentary from Nvidia about how revenues will be lower as a result of export restrictions to China.
• Furthermore, AMD is going to find the going harder in inference if Nvidia’s Dynamo platform gets any traction.
• Dynamo promises large improvements in efficiency in data centres with the aim of maximising tokens produced per dollar spent.
• It only works properly on Nvidia hardware, and consequently, if it is popular, it will increase Nvidia’s stickiness in inference at AMD’s expense.
• AMD is no longer an expensive stock at 22.4x 2025 PER but the looming obsolescence of x86 makes me think twice about owning it.
• I would continue to prefer Qualcomm, where the diversification strategy is working, there is no real issue with obsolescence and where the stock is much cheaper.