Filtered to AI Industry · show all articles
xAI rarely publishes its own postmortems. What exists instead is a paper trail of deleted prompts, court filings and letters from state officials — ten of those documented incidents, each tied to a specific, named, dated source.
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.
Grok's API looks like any other OpenAI-compatible endpoint until you read what xAI actually documents — a distinct pricing cliff, a real-time X search tool no competitor's docs describe, and a versioning policy that silently reroutes old model names.
xAI's own dated announcements, set against independently verified construction and permitting records for the Colossus datacenter, with an explicit line drawn between what the company has disclosed and what remains an unconfirmed parameter rumor.
Not a prediction about which company wins — a grounded extrapolation from custom-ASIC growth rates, photonic interconnect adoption, and packaging trends already underway, toward what the rack itself might physically become.
Not a promotional claim — an arithmetic one. A roughly 14x growth forecast over a decade implies today's AI-semiconductor boom is closer to its beginning than its middle, on the numbers alone.
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.
Inference can be cheap while accepted software remains expensive; the correct denominator is valuable, reviewed change—not tokens, diffs, or agent hours.