Tagged Nvidia · show all articles
Not one challenger, one strategy. A synthesis of eight genuinely different approaches to competing with Nvidia — merchant GPUs, custom ASICs, wafer-scale bets, and in-house hyperscaler silicon — and which approach actually has traction in 2026.
Nvidia designs the GPU. SK hynix supplies the memory stacked directly on top of it, at a market share high enough that a single Korean company's production schedule now sets a real ceiling on AI accelerator supply.
Blackwell was a GPU with a name. Rubin is a whole platform — six distinct chips engineered to work as one system — and that shift says as much about where AI hardware is headed as any single spec.
Nvidia's market capitalization gets treated as a headline in itself. The revenue, backlog, and data-center figures underneath it are more informative than the number everyone quotes.
Forward earnings multiple, backlog-to-revenue ratio, growth rate — the actual numbers behind Nvidia's valuation, laid out plainly, without a promotional or a doom-laden frame either way.
Groq built its entire business on the argument that GPUs are the wrong architecture for AI inference. Nvidia's reported move into Groq's technology is either the smartest hedge in the industry or a quiet admission the argument was right.
Individual accelerators keep getting faster. The copper connecting thousands of them together increasingly cannot keep up — which is why Nvidia is betting heavily on replacing that copper with light.
Groq built its whole identity on the claim that GPUs are the wrong chip for AI inference. Then Nvidia got involved. Here's the short version.
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.