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!
Any thoughts on how long the AI/ML evolution will continue and the directions it will take? I feel we have at least five more years for the technologies involved to play out will see far deeper and much more diversity in the applications over what we have already seen.
Any thoughts on how long the AI/ML evolution will continue and the directions it will take? I feel we have at least five more years for the technologies involved to play out will see far deeper and much more diversity in the applications over what we have already seen.
One thing to watch is the TSMC CAPEX. TSMC cannot build fabs based on bubbles. These fabs will run for many years and TSMC averages 80%+ utilization. TSMC CAPEX will probably increase 10% next year ($40B+) which means continued growth. When TSMC cuts CAPEX then you can suggest the AI bubble is shrinking. TSMC's CAPEX has more than doubled in the last 10 years. That is huge! Intel's CAPEX was $25B at its high point. Next year Intel will probably spend closer to $20B if all goes well.
And who knows more about the future of semiconductor demand than TSMC? Nobody, absolutely.
Who knows ? We’re only just beginning the journey with the second type of broadly applicable models.
* Revolution one was vision systems and recognition and categorization with convolutional neural networks.
* Revolution two has been LLMs, primarily using transformer-based approaches today. LLMs are a big enough application that were now seeing hardware systems being optimized for transformers (specialized hardware for prefill, decode and prefill storage) and LLM inference serving software plus models being optimized for faster and more efficient hardware implementations (disaggregated inference).
But I expect all sorts of growth in different dimensions
* Huge improvements in LLM efficiency via extending co-optimization even further. DeepSeek “shocked” the AI world in late 2024 with a paper that revealed a bunch of their co-optimization techniques that made their open model extremely fast an efficient. The rest of the AI world jumped on those techniques in 2025. It may or may not be as seminal, but they just dropped another paper at the end on 2025 on Hyper-Connection via Constrained Manifolds (mHC) that might power further efficiency gains for LLMs.
* Advancements in new model types for new applications - there’s been a lot of motion on creating “physical AI” that understands the physics, kinematics and other elements of the real world, for robots and other physical autonomy. NVIDIA has talked about this extensively, most recently at CES. Yann LeCun left Meta to start a next generation Physical AI startup. Presumably Physical AI will need new forms of datasets, models and underlying techniques for training and encoding how the real world works. And if and when someone discovers techniques that generalize well and give consistently good results, we’ll probably see the same kind of co-optimization as we are seeing for LLMs using transformers.
* New model types that do interesting things, that haven’t found their killer app yet. I’ve seen a lot of interesting thoughts about GNNs (graph neural networks), and possible applications from fraud detection to some forms of Physical AI), but nothing effective enough yet, to open a new market.
