Filtered to AI Hardware & Semiconductors · show all articles
Wide-bandgap semiconductors — gallium nitride and silicon carbide — are a genuinely separate boom from the AI-accelerator story, growing fast on the strength of electric vehicles and grid electrification alone.
Four major forecasters, four different 2030 semiconductor-market numbers, none of them wrong. A methodology-first look at why "the" 2030 forecast doesn't exist, and how to actually use the spread.
Every other accelerator in this cohort is designed to run whatever model architecture a customer throws at it. Etched's chip runs transformers and nothing else — an enormous bet that the architecture behind ChatGPT isn't going anywhere.
Coverage of China's chip constraints focuses on logic — SMIC, EUV, export controls. The country's leading domestic DRAM maker has its own reported gap, and it compounds the one already getting all the attention.
Nvidia showed off CUDA in 2006. It cost billions, served a market that didn't yet exist, and didn't turn a profit for roughly a decade. This is the story of the software decision that quietly became AI hardware's entire foundation.
A larger single die should, in principle, be faster. It is also exponentially more likely to contain a fatal defect. This briefing works through the actual yield arithmetic behind the industry's shift to chiplets.
Every chipmaker in this cohort dices a wafer into individual dies. Cerebras builds one processor out of an entire, uncut wafer — an idea the rest of the industry tried and abandoned decades ago.
A chip company most people had never heard of raised $5.55 billion and closed its first trading day up 68%. Here's the quick version of what just happened.
Celestial AI's acquisition by Marvell was one of the largest deals in the photonic-interconnect wave. This briefing looks at what the technology actually does, and why it was worth that price to a company already established in custom silicon.
Designing a chip's circuits and predicting how it will physically behave — heat, stress, electromagnetic interference — were historically separate disciplines, separate software, separate companies. Synopsys just spent $35 billion to make them one.