Cisco sees AI infrastructure revenue nearly doubling to $7.5B
The AI buildout keeps printing money for the companies selling its plumbing — and its appetite for training data just met its first organized resistance at the bookstore counter.
Cisco said hyperscalers placed $9.3 billion in AI infrastructure orders in fiscal 2026, and it expects AI infrastructure revenue to nearly double to $7.5 billion this fiscal year. The networking giant beat estimates in its fiscal fourth quarter — $1.22 adjusted earnings per share on $17.25 billion in revenue — and guided the current quarter to $18 billion to $18.2 billion, well above the $16.8 billion Wall Street expected. Hyperscalers alone accounted for $4 billion of those infrastructure orders in the quarter, a sign the cloud giants are still buying switches and routers as fast as Cisco can ship them, not just accelerators.
Yet the stock fell nearly 4% in extended trading — the same "good numbers, bad reaction" pattern investors saw with Cerebras earlier today, after a run-up of more than 60% this quarter left the bar impossibly high. The durable signal is the order book: Cisco's AI business is still small next to the chipmakers, but $9.3 billion in hyperscaler orders says the buildout's next phase is the networking layer that moves data between thousands of GPUs — and that money is starting to show up in the most traditional tech company on the list.
Rare booksellers are starting to refuse the AI industry's bulk orders, fearing their stock is being scanned and then destroyed for training data. The Irish bookstore Kennys flagged a "bananas" order for 5,000 obscure titles that came with no haggling over price — a red flag dealers now associate with AI buyers — and owner Tomás Kenny says AI is "terrifying our industry," vowing to refuse orders likely destined for training runs. The pushback follows weeks of reporting that AI firms buy physical books in bulk, cut the spines for high-speed scanning, and shred the originals, with The Atlantic documenting the backlash and 404 Media exposing the sourcing middlemen.
The fear is that irreplaceable rare books get pulped for a marginal gain in training quality. Gentler paths exist — the Internet Archive has digitized fragile volumes "the hard way" for years, and OpenAI, Microsoft, and Harvard are working together on a million-book public-domain collection — which is exactly why the cheapest option is now colliding with a cultural red line.
What to watch: whether other booksellers follow Kennys' lead, and whether any major lab publicly commits to non-destructive scanning.
Should AI labs be allowed to buy and destroy physical books to train models, or is that a line the industry shouldn't cross? Tell us in the comments.
