Intel did pretty well this last quarter but two things they said in the last quarterly call stuck out to SemiAccurate. One was technical the other business, so lets take a look.
The first thing that SemiAccurate thought was interesting was when Intel was talking about their AI sales and chances. At about 12m into the Q3/2025 earnings call, Intel basically threw in the towel on AI training sales. The most explicit tell was the commentary when they said their AI silicon was optimized for inference plus physical and agentic AI rather than training. Additionally the GPUs are also being optimized for inference so… wait for the pushback spin but don’t buy it.
The real death knell for Intel’s AI training ambitions was earlier in the call when Lip-Bu was talking about the tie up with Nvidia, specifically the use of Nvlink. Any company who signs up for this is toast. That is both a technical statement and behavioral based on 20+ years of watching Nvidia closely. Without going too much in to detail, signing up for Nvlink allows Nvidia to pull the rug out from any ‘partner’ as soon as the market they spend the time and effort to develop becomes worth it for Nvidia to play in. Intel did it to them, ironically not intentionally, during the Sandy Bridge transition, and Nvidia has done it to many others since.
You can say a lot of things about Nvidia, but they learn. Intel is out of the AI market as long as Nvidia is dominant, and they lost any chance at margins with Nvlink too. Any Xeon with Nvlink has one customer, and a customer with other options too. Good luck with that Intel. There are more technical reasons why this is horrendously bad deal for Intel but they are a bit more nuanced. Suffice it to say it doesn’t get any better from here.
More importantly than that was the bit about tight supply environments about 27m into the call This is real and like Intel said, will persist into next year, a boon for all chip suppliers. Starting about 6-8 weeks ago, a rumor started going around high levels of the industry about supply constraints. The finger was pointed at Intel, 18a, stupid SKU mixes, and crap yields being the cause. None of this is true but SemiAccurate heard it from multiple sources in the x86 world and several OEMs too.
Since it wasn’t true and many high level sources were talking about it, what was actually going on, and where did this story come from? That is where it gets interesting, and it all starts here.
Note: The following is for professional and student level subscribers.
Disclosures: Charlie Demerjian and Stone Arch Networking Services, Inc. have no consulting relationships, investment relationships, or hold any investment positions with any of the companies mentioned in this report.
