Filtered to AI Hardware & Semiconductors · show all articles
TSMC is building the wafer fab. Amkor is building the packaging plant next door — the piece that would let a chip go from raw silicon to finished, tested product without ever leaving the country.
Most accelerator deals get announced as a unit count or a dollar figure. AMD and OpenAI described theirs in gigawatts of power consumption — a framing that says as much about the industry's real constraint as it does about the chips themselves.
AMD's absorption of inference-chip startup Untether didn't make headlines the way a multi-gigawatt hyperscaler deal does. It may say more about where AMD thinks the accelerator market is actually headed.
Semiconductors have grown from 2% to 18% of S&P 500 weight — more than double the dot-com peak. A portfolio that never touches a chip stock by name is still more exposed to this one industry than most investors realize.
Not the GPU. Not the memory. A power transformer — a piece of equipment that hasn't fundamentally changed in decades — now has a longer lead time than almost anything else in the entire AI supply chain.
A chip is not carved, it is printed — the same four-step loop repeated across dozens of layers. Resolution stopped being the binding constraint some time ago; placement, variation and measurement took over.
Performance now moves on the assembly rather than the transistor. Once a die is cut into pieces, the interface between them becomes the design, and the cost of a bit is set largely by how far that bit has to travel.
Once a model no longer fits on one device, the interconnect stops being plumbing and becomes the design surface. Every way of splitting a model generates its own traffic, and the fabric decides which splits are affordable.
Performance is decided by data movement, not arithmetic. The levels of the hierarchy, the locality that tiling manufactures, and the reason a fast kernel is fast because of what it never reads.
Petaflops are the advertised number and almost never the binding one. An accelerator is a memory system with arithmetic attached, and every serious design decision of the last decade follows from that.