You are currently viewing SemiWiki as a guest which gives you limited access to the site. To view blog comments and experience other SemiWiki features you must be a registered member. Registration is fast, simple, and absolutely free so please, join our community today!
Hence, in my opinion, focusing on AI would be easier than focusing on the foundry business. As CEO, he should prioritize maximizing shareholder value. The AI path is not only easier but also offers greater potential rewards.
The AI path is certainly not easy nor easier. Intel has always been a distant No. 3 in the GPU market, and the No. 2, AMD, will still not be able to compete favorably with NVIDIA in AI training by 2026.
Jensen has excellent vision and great execution. In addition to CUDA, NVIDIA has another crucial advantage that AMD and Intel don't have: high-performance networking (Mellanox was acquired by Jensen in 2019). Grabbing shares from NVIDIA will be more difficult for Intel than grabbing foundry shares from TSMC.
What can Intel do in AI? I think the leadership team has basically given up on the AI training market, which is wise, in my opinion. Intel will likely focus on competing in AI inference or creating new products/markets.
My favorite strategy: when you can't win, create or play a new game.
Based on my experience, there was a big uplift and earlier deployment of what I would call broad DCTO with 32/28nm, with the move to FinFET. I say broad, because both TSMC and Common Platform had to begin broad ecosystem enablement via PDKs, etc. much earlier, because the ecosystem and large customers with their own IP could no longer extrapolate transistor characteristics. That lead to many more early MPWs and a much more structured, staged PDK and foundational/interface IP validation / DTCO process. TSMC, at the apex of the ecosystem, benefited the most. Realistically, others had to buy their way in to IP access and do speculative work ahead of customer demand- a very expensive undertaking, but necessary.
I think Intel is learning, but it’s a long cycle and means exposing some of the stuff they used to hold super close to the vest, far more widely, and respond to feedback that might affect their products coming out of the oven.
There's also an intrinsic conflict of interest- I remember one case where the IDM side of an IDM/Foundry combo discovered routing rule changes for 3rd party IP on a real design that would improve yield. Guess how long that improvement took to percolate back to the PDK
