Everyone’s talking about yesterday’s Nvidia’s Q2 earnings. Not surprisingly, most of the focus was on the sophisticated, powerful GPU chips that made the company a $4 trillion symbol of the AI boom.

But the company’s huge bet on AI isn’t just about chips; it’s about the massive, billion-dollar data centers being built to house them. Data center revenue accounts for nearly 88% of Nvidia’s total sales—meaning the GPU chips, networking gear, systems, platforms, software and services that run inside AI data centers.

I’ve been noodling on something CEO Jensen Huang mentioned during the earnings call with analysts and investors—something that ties directly to my obsession with those mega facilities built to house tens of thousands of GPUs, that consume staggering amounts of energy, and which are used to train the massive models behind generative AI. These are facilities like Meta’s planned, gas-fueled campus in northern Louisiana—which President Trump touted yesterday with a photo showing it will sprawl to the size of Manhattan—or OpenAI’s $100-billion-plus Stargate Project.

On the call, Huang touted an Nvidia product called Spectrum-XGS—a hardware and software package that together let separate data centers function like one. Think of it as the pipes and traffic control that moves data between data centers quickly and predictably. 

Wait—I know your eyes are already glazing over, but hear me out. One of my nagging questions has long been: What if the billions being bet on these mega AI data centers winds up going bust?

Spectrum-XGS is built for the mega-AI clusters Huang has long predicted. But it also enables those who can’t manage to build a single mega-facility because of permitting or financing issues, to stitch multiple data centers together into unified “AI factories.”

Until now, there were only two options for finding more compute: add more chips to a single GPU rack, or pack more racks into one giant facility. Spectrum-XGS introduces a third option: link multiple sites so they work together like one colossal supercomputer. AI cloud company CoreWeave, which rents out access to GPUs, is deploying the technology to connect its own data centers.

On the earnings call, Huang highlighted Spectrum-XGS, saying it would help “prepare for these AI super factories with multiple gigawatts of computing all connected together.” That’s the growth story Huang has been telling investors for years.

But what happens if it doesn’t unfold that way? There are several other scenarios: The most dire, of course, would be a true “AI winter,” in which the AI bubble pops and demand for AI from business and consumers plummets. In that case, demand for data centers optimized for AI—whether on mega campuses or in distributed, smaller data centers—would vanish and Nvidia would have to find other sources of revenue. Another possibility is that AI models become much smaller and are mostly used on laptops and mobile devices for “inference,” or outputting results. In that case, the demand for data centers could also drop and Nvidia would need to hedge against that.

However, there is yet another scenario where not many mega-campuses get built but technology like Spectrum-XGS still winds up helping Nvidia. If the mega-campus model falters—because of power shortages, financing constraints, or local pushback—Nvidia could still win if enough demand remains from customers. With technology like Spectrum-XGS, smaller or less centrally located facilities become more usable if demand shifts from mega-campuses to distributed ones.

In other words, Nvidia has positioned itself so that whether the industry keeps building massive new hubs or turns to linking together smaller, scattered sites, customers will still need Nvidia’s hardware and software—as well as its AI chips, of course.

Of course, Nvidia’s hedge doesn’t mean local communities are protected if the massive data centers being built in their backyards end up being white elephants. Towns that banked on jobs and tax revenue could still be left with ghost campuses and hulking concrete shells. And all of this depends on Spectrum-XGS working as promised and major players signing on. Customers haven’t tested it at scale in the real world, and networking—big or small—is always messy.

Still, whether the mega AI data center boom keeps roaring or fizzles, Nvidia is positioning itself to own the invisible infrastructure that underpins whatever future system emerges. Nvidia may be best known for selling the “picks and shovels” of AI—its GPUs—but its networking “plumbing” could help ensure the company wins either way.

Speaking of industry leaders, I hope you’ll check out Titans and Disrupters of Industry, a new podcast hosted by Fortune Editor-in-Chief, Alyson Shontell, that goes in depth with the powerful thought-leaders shaping both the world of business and the very way we live. In this exclusive interview, Accenture’s Julie Sweet discusses the company’s strategic shifts and the impact of AI, tariffs, and geopolitical changes on businesses.

Also: In less than a month, I will be headed to Park City, Utah, to participate in our annual Brainstorm Tech conference at the Montage Deer Valley! Space is limited, so if you’re interested in joining, register here. I highly recommend: There’s a fantastic lineup of speakers, including Ashley Kramer, chief revenue officer of OpenAI; John Furner, president and CEO of Walmart U.S.; Tony Xu, founder and CEO of DoorDash; and many, many more!

With that, here’s the rest of the AI news.

Sharon Goldman
sharon.goldman@fortune.com
@sharongoldman

FORTUNE ON AI

Lawyers for parents who claim ChatGPT encouraged their son to kill himself say they will prove OpenAI rushed its chatbot to market to pocket billions – by Muskaan Arshad

Anthropic’s settlement with authors may be the ‘first domino to fall’ in AI copyright battles – by Beatrice Nolan

Tesla self-driving cars are being tested in Boring Co. tunnels in Las Vegas, but full autonomy is still ‘a ways off,’ convention center exec says – by Jessica Matthews

AI IN THE NEWS

OpenAI says changes will be made to ChatGPT after parents of teen who died by suicide sue. According to CBS News, OpenAI said it will strengthen ChatGPT safeguards for vulnerable users—including new protections for teens—after the parents of 16-year-old Adam Raine filed a lawsuit alleging the chatbot encouraged him to take his own life. The complaint, filed in San Francisco, claims ChatGPT discussed suicide methods with Raine more than 1,200 times and urged him to keep his plans secret, while OpenAI allegedly ignored warnings that its technology’s “emotional attachment” features could harm young users. The case has reignited concerns that AI companies are prioritizing market dominance—OpenAI’s valuation surged from $86 billion to $300 billion after launching GPT-4—over safety. OpenAI expressed condolences to the family and said it is reviewing the lawsuit, while pledging to add parental controls, teen-specific protections, and the option for teens to designate an emergency contact.

Librarians helped test which AI gave the best answers without making stuff up. The Washington Post did an interesting test of AI search tools, to determine which one would be most likely to provide a correct answer. It enlisted the help of librarians who judged a competition between nine AI search tools, asking each AI to answer 30 tough questions. They scored 900 answers from the free, default versions of Bing Copilot, ChatGPT, Claude, Grok, Meta AI and Perplexity, as well as Google’s AI overviews, its newer AI mode and its traditional web search results. The questions were designed to test five categories of common AI blind spots. The surprising winner? Google’s AI Mode, which acts like a chatbot and was added in May to the top left corner of search results. Apparently Google still rules search, AI or not. 

China seeks to triple output of AI chips in race with the US. According to the Financial Times, China is moving aggressively to expand its domestic AI chip production, with new fabrication plants tied to Huawei expected to come online by early next year that could triple the country’s total output of AI processors. SMIC, China’s leading foundry and Huawei’s biggest chip supplier, also plans to double its 7nm production capacity, which would free up supply for smaller players like Cambricon, MetaX, and Biren—helping fill the gap left by Nvidia after U.S. export bans. At the same time, leading AI startup DeepSeek is pushing a new FP8 data format standard designed to align with this next generation of Chinese chips. The wave of capacity expansion has fueled a surge in Chinese semiconductor stocks and underscores how central chipmaking has become in the U.S.-China race over AI.

AI CALENDAR

Sept. 8-10: Fortune Brainstorm Tech, Park City, Utah. Apply to attend here.

Oct. 6-10: World AI Week, Amsterdam

Oct. 21-22: TedAI San Francisco. Apply to attend here.

Dec. 2-7: NeurIPS, San Diego

Dec. 8-9: Fortune Brainstorm AI San Francisco. Apply to attend here.

EYE ON AI NUMBERS

40%

According to a new Gartner Research report, that's how many enterprise applications will embed task-specific AI agents by 2026—that is, those that are designed to master a narrow set of responsibilities such as automatically triaging IT help-desk tickets,  flagging regulatory risks for compliance workflows, or reconciling accounts for a finance team. That number is up from less than 5% today. The shift, said the researchers, could free up workers from repetitive chores while introducing new operational and security risks.

"As AI agents begin acting independently and handle tasks ranging from routine development to complex incident response without human involvement, leaders must ensure strong security and governance,” said Gartner senior director and analyst Anushree Verma in a press release.