Tech Newsround – TSMC & OpenAI

TSMC Q2 2025 – Compare and contrast

• In contrast to ASML, TSMC reported good results and raised guidance in a clear indication that the issue hampering ASML’s outlook has very little to do with end demand for semiconductors.
• Q2 25 revenues / EPS were TWD935bn / TWD15.4 nicely ahead of estimates of TWD921bn / TWD14.5 as demand and pricing have remained robust.
• TSMC has grown revenues by 40% YoY in H1 2025 vs H1 2024 putting it in a comfortable position to raise its FY2025 guidance to 30% YoY growth from mid-20s.
• This is a pretty safe bet given how 2025 has panned out so far, and so I will not be surprised to see TSMC raise its 2025 guidance yet again at its Q3 2025 results in October.
• This is in strong contrast to ASML’s commentary, where it missed expectations for Q3 2025 in its forecast and withdrew its expectation on growth in 2026.
• This further supports my view that ASML’s problems are largely being caused by China, which has stopped stockpiling equipment ahead of incoming restrictions and is now falling hard as a percentage of sales.
• Consequently, ASML’s issues remain specific to ASML meaning that the AI-driven demand growth remains very much in swing.

OpenAI – Getting Googled

• OpenAI has added Google Cloud to its Sub-processor list, cementing its diversification away from Microsoft and raising the possibility that Google has finally won a major client that will use its in-house silicon.
• The Sub-processor list (see here) is a list of 3^rd parties who provide “processing activities for customer data” and serves as a good indication of where OpenAI gets its compute.
• The addition of Google is the first major departure, as CoreWeave was originally sourced via Microsoft, and so I don’t really count this as diversification.
• Going with Google also raises the possibility that OpenAI has started using Google’s in-house TPUs as opposed to Nvidia for training and inference.
• As by far the market leader at the moment in generative AI services, this would be a significant development if confirmed.
• However, I suspect that OpenAI may be using Google mostly, if not entirely, for inference as opposed to training, as it is in training where Nvidia’s lock is by far the strongest.
• This makes sense because models these days use far more inference compute than before, and as the user base grows, the demand for inference will grow far more quickly than training.
• This is not uncommon, and I suspect that the bulk of the business Nvidia’s main rival, AMD, has managed to win is for inference as opposed to training.
• This highlights a growing competitive threat to Nvidia, but as I have written several times already (see here), Nvidia has already moved to mitigate this problem with the launch of Dynamo.
• This also cements Microsoft’s diversification away from OpenAI, which, given the precarious state of its corporate governance, represented significant risk to Microsoft as a single source supplier.
• This has been my sole concern with Microsoft’s otherwise pretty faultless strategy, which now looks like it will be quickly addressed.
• That being said, Microsoft is pretty expensive compared to many others involved in the AI boom, and I would prefer to look at Google, TSMC or Qualcomm.