xAI runs eight jet engines at a Mississippi data center as emissions fight goes on

The generators have state-level permission. Tom's Hardware reported on 3 November that the eight converted jet engine turbines are running at the site. The permits allow a 25 percent higher limit for pollutants classed as volatile organic compounds.

Those compounds, paired with carbon monoxide output, have drawn objections from environmental groups, the outlet said. The New York Times reporting it cites is where the challenge first surfaced.

What the law centre argues

The Southern Environmental Law Center has already challenged Mississippi's decision to approve the turbines, according to Tom's Hardware. The group argues that xAI has not provided guarantees that the jet-powered turbines meet emissions standards for specific pollutants, including formaldehyde. It also accuses the company of undercounting projected emissions relative to the data center's power sources.

That challenge is not the group's first move against the company in the state. The same organisation previously sued xAI over its use of gas generators without the required permits in an area with more than 100,000 inhabitants, Tom's Hardware reported.

Why converted jet engines at all? The answer is speed. Grid interconnection queues are long. On-site turbines can be ordered, delivered and fired up far faster than a new substation can be built. The trade-off is that the emissions profile moves from a regulated utility stack to a cluster of small units on the edge of a community. The permitting regime that governs them is thinner than what a power plant would face.

Tom's Hardware frames the story as companies looking for new energy sources as AI compute demand outruns available power. It does not give the site's capacity in megawatts, the number of hours the turbines run, or a company response to the formaldehyde claim. Those questions remain open.

The demand side is getting more expensive

The power scramble has a balance-sheet counterpart. Alphabet lost just over $500bn in market value on Thursday, the BBC reported on 30 October. The drop came after it disclosed that its own AI summaries were hitting search advertising revenue. Chief business officer Philipp Schindler said Google's AI Mode, which answers questions conversationally, was eating into the money made from ads placed next to traditional results.

Total revenue still grew 25 percent in the three months to the end of September, to $127.6bn, according to the results published on the Wednesday. But advertising came in weaker than analysts expected, and shares fell more than 12 percent to a six-month low. By the end of Thursday, the company's market value had dropped to $3.46 trillion from $3.97 trillion the day before.

Kim Scott, who formerly led Google's advertising business, told the BBC's Today programme the company would rather eat its own lunch than let someone else do it. Tracy Woo, a principal analyst at Forrester, said the Search results highlighted the difficulty of funding large AI investments while keeping established businesses strong.

Amazon's numbers landed in the same week and told a similar story about cash. Quarterly sales grew 11 percent to $201bn, the BBC reported, alongside a negative cash flow figure of $42bn. That is a six-fold increase in spending over three months. Chief executive Andy Jassy told analysts he repeatedly raised the opportunity in Amazon's own chips and its Bedrock developer platform. He said it would be more convenient if the spending were spread over 25 years. It is not.

Meta on the other side of the trade

Not every ad business is losing to the chatbots. CNBC reported on 1 November that Meta's ad revenue rose 26 percent in the latest quarter to $63.1bn, up from $50.1bn a year earlier. Google's search division grew 12 percent to $63.4bn. That 14-point gap is the widest in Meta's favour since 2021, and it leaves the two roughly level by size, though Google took an additional $11bn from YouTube in the quarter.

Nate Elliott, an analyst covering AI at Emarketer, told CNBC by email that changing discovery habits are potentially catastrophic for Google's ad business. That happens if people find ChatGPT, Claude or Meta's Muse better at finding what they want. Meta's own stock was not spared the mood: it fell 14 percent in October, its worst month since October 2022, after a 27 percent surge in September.

The advertising shift matters to publishers. The same answer engines that dent Google's ad clicks also remove the visit that used to carry a pageview. Discord offered one small counter-example last week. BleepingComputer reported on 29 October that the platform will let search engines index messages in Discoverable Servers, the publicly listed communities with at least 1,000 members, of which it says it hosts roughly 63,000. Server Web Pages arrive on 16 November, with message indexing from mid-December. Sharing stays off by default in the EU, the UK, Canada, Brazil, Australia and South Korea, and underage users are excluded.

Search behaviour is measurable, and odd

Whether AI assistants actually go looking for fresh information is now a research question. A paper posted to arXiv on 30 October by Nipun D. Pathirage and Kwan Srijomkwan built a benchmark of 400 multiple-choice questions. The set mixed widely known inside-knowledge items with outside-knowledge items drawn from news published after the models' training cut-offs. Across five commercial models, GPT-5 mini and Grok 4 Fast searched on over 90 percent of outside-knowledge questions but only 1 percent of inside-knowledge ones. Mistral Small searched on just 20 percent of the outside-knowledge questions and often abstained instead of using its tool.

Searching was partly triggered by explicit time cues. On uncued outside-knowledge questions, the two selective models' score fell from about 0.72 to 0.57 or 0.58. The authors are careful about what this measures. The scores capture the triggering behaviour of specific products, not metacognition in a strict sense.

Wired's 2 November guide makes the practical version of the same point. Simple, widely known lookups are a safe bet for AI search. Obscure questions are where a bot may find only one or two references and pick the wrong one. Recent Stanford University research found that when AI locates an accurate source it extracts the right information well, Wired reported. It fails when it cannot find one and answers confidently anyway.

For publishers, that is the mechanism behind the traffic story. A confident answer with no click is a lost visit. A confident answer with a bad source is worse.