Tagged AI infrastructure · show all articles
xAI says Colossus is the fastest AI supercomputer ever built. Set beside Microsoft, Google, and Amazon's own infrastructure disclosures, the documented differences are in speed, siting, and power sourcing, not in a single winner.
xAI has published some specifics about Grok's training and a great deal about Colossus's scale. Reporting and litigation filled in most of what it has not said about power. This piece keeps each kind of evidence separate.
Every briefing in this cohort eventually runs into the same wall: not a chip shortage, but a power shortage. The industry has enough silicon ambition. It doesn't have enough electricity infrastructure to plug it into.
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.
Nvidia and Marvell are both betting on optics to connect chips together. Lightmatter's bet goes one step further: using light for the connections inside a chip package as well.
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.
Not the chip. Not the training cluster. Ayar Labs focuses on one narrow, unglamorous problem — getting light on and off a chip efficiently — and is betting that narrowness is exactly what makes the bet work.