NVIDIA's $96B Quarter and the AI Circular Economy

NVIDIA just booked $96.2 billion in revenue in a single quarter. As The Verge noted, the company is crossing into hundred-billion-dollar-a-quarter territory, a threshold no company has ever touched. That number is up over 100% year-over-year, and TLDR AI flagged it as one of the defining financial signals of the current AI moment. The guidance for the following quarter pointed even higher.

So yeah. Wild.

The Moonshots podcast crew, Peter Diamandis, Dave Blundin, Salim Ismail, and Alexander Wissner-Gross, spent a long stretch of their latest episode turning this number over and looking at what's underneath it. And honestly, the conversation got more interesting the further they dug.

The Loop Question

The thing that stuck with me from their discussion isn't the revenue figure itself. It's the question Wissner-Gross raised about the shape of the economy producing it.

Here's the structure as Blundin described it: NVIDIA sells chips to hyperscalers, hyperscalers sell compute to AI labs and vertical-use-case companies, those companies build products, some of those products generate revenue that flows back into buying more compute. The circle tightens. "The new economy only cares about itself," Blundin said, "and cares very little about the legacy economy."

That framing is either exciting or alarming depending on your disposition. The optimist's read is that this is what a genuinely new economic substrate looks like bootstrapping itself into existence. The skeptic's read is that you don't know if the loop is self-sustaining or artificially inflated until it isn't.

Wissner-Gross put the concern precisely: "I would love more clarity on how much of this demand is Nvidia-financed demand or not." He's not calling it a bubble. He's saying the financial markets might not have full visibility into whether the demand at the infrastructure layer is being backstopped, directly or indirectly, through NVIDIA's own balance sheet. Income-statement circularity (company A sells to company B, company B sells back to company A) reads differently than balance-sheet circularity (company A finances company B's purchase of company A's goods). The distinction matters enormously, and by his account, it's not legible from the outside right now.

Nobody on the pod was calling for a crash. The consensus seemed to be that compute is genuinely substituting for other economic inputs at a civilizational scale, which means any correction would likely be a "mini winter" rather than a collapse. But the transparency gap is a real thing worth watching.

The Other Shoe: TSMC

Sitting underneath all of NVIDIA's nosebleed margins is a single manufacturer. TSMC handles essentially everything NVIDIA sells, and the podcast group flagged this as the company's clearest structural vulnerability. Blundin noted that roughly a third of TSMC's output goes to NVIDIA, another third to Apple, and the rest of the world shares what remains. You don't casually disrupt that relationship. You definitely don't announce you're building competing fabs. The dependency is so deep that even moving to change it requires, as Blundin put it, doing it "very sneakily."

Meanwhile, the competitive landscape is shifting at the inference layer (inference = running an already-trained model to generate outputs, as opposed to training, which is the expensive, GPU-cluster-intensive process of building the model in the first place). The podcast discussed OpenAI's move into custom chip design as a signal that the biggest AI labs are no longer content to be NVIDIA customers. As the transcript framed it: "OpenAI is no longer just a customer of Nvidia. They're becoming a chip designer." NVIDIA dominates training, where you need massive coherent GPU clusters. Inference is a different market, more distributed, more heterogeneous, and increasingly contested.

China's Video Bet and the World Model Race

The episode's most geopolitically interesting segment was about the divergence between how American and Chinese AI labs are allocating compute. The Moonshots crew cited reporting from Runway's co-founder that video generation now accounts for roughly 70% of all AI token consumption in China, driven by short-form content and robotics applications.

Wissner-Gross offered what he called a "unified field theory": American labs are revenue-maximizing per token, which means they gravitate toward code generation and enterprise applications where the economics are strongest. Chinese labs, many of which give away model weights for free, aren't optimizing the same way. So they end up deploying more tokens on video, which is globally marketable and trust-agnostic in a way that enterprise software isn't. (Nobody worries about CCP access to their AI-generated short video the way they'd worry about a Chinese LLM processing their legal contracts.)

The deeper structural difference is between LLMs (large language models, which predict the next word or token) and world models (which try to predict the next state of physical reality). Video generation is a world-model application. So are robotics and autonomous driving. The podcast framed this as the two superpowers optimizing for different futures: the US toward software and knowledge work, China toward physical-world modeling and manufacturing.

Whether world models or language models advance the capability frontier faster is, as Wissner-Gross noted somewhat dryly, probably a trick question. The answer is likely: both, combined.

200,000 Fake Accounts and the Narrative War

One of the sharpest moments in the episode involved a disclosure from X's safety team: an investigation into suspected Chinese inauthentic accounts turned up a bot farm of approximately 200,000 accounts. The accounts were specifically pushing claims that AI data centers are driving up household electricity prices and straining the power grid.

The Moonshots crew's reaction was theatrical sarcasm ("Shocked. Shocked."), but the underlying point is genuinely unsettling. As one participant noted: "The difficult part is we have really no easy defense against it." Free speech norms that protect political discourse also protect coordinated inauthentic campaigns about infrastructure policy. The legal and technical tools for distinguishing one from the other at scale don't really exist yet.

What makes this particularly sharp is the specificity of the target. It's not generic anti-AI messaging. It's messaging aimed at slowing data center permitting and public acceptance at exactly the moment when AI infrastructure buildout is the central constraint on the industry's growth. That's not accidental.

Flock Safety and the Surveillance Petri Dish

The Flock Safety segment hit differently for me than the finance stuff. Flock, founded in 2017, has built a network of AI-powered license plate readers now operating across more than 6,000 communities in 49 states, generating what the company says is over 20 billion vehicle scans per month. The platform has expanded beyond plates to include audio detection, video cameras, drones, and software that federates all those feeds.

Ismail put the structural problem clearly: "We've transitioned from 'find this suspect' to 'find me people who behave like suspects,' and that's a very different animal." Pattern matching at essentially zero marginal cost means mass surveillance is no longer constrained by manpower. Civil liberties that were historically protected by the friction of following people, it was expensive, so it didn't scale, no longer have that protection. The constitution doesn't auto-update when transaction costs go to zero.

The proposed counter isn't to tear down the cameras. Wissner-Gross and Ismail both gestured at the argument, associated with writer David Brin, that the answer is symmetrical access: if surveillance data is captured in public spaces, make it a public resource. Point the lens both ways. Whether that's actually implementable as policy is a very different question.

The Job Numbers Nobody's Talking About

Against all the apocalyptic AI-and-jobs discourse, Ismail presented data from a Principal Financial Group survey of small and medium-sized businesses (SMBs). The finding: only 4% of the companies surveyed anticipated AI would reduce their staffing and wages. Thirty-one percent expected staffing and wages to increase. Of companies that had reduced staff, only 1.4% attributed those reductions to AI or automation.

The caveat worth holding: this is SMB data, and SMBs are not where most AI automation anxiety is actually concentrated. The Goldman Sachs warning the podcast also covered, about professional services firms facing an existential skills gap between AI-fluent junior staff and AI-resistant senior partners, is a different population. The two findings aren't contradictory. They're describing different layers of the economy experiencing different pressures at different speeds.

The part that did land for me was Wissner-Gross's framing: we're in a regime where "high agency is one of the few human traits that is actively rewarded in an era of super intelligence." That's a real insight, and it's not particularly comforting if you're someone whose job has historically been structured around executing defined tasks rather than directing outcomes.

The Companion Question

The episode closes out the AI companions thread with China's decision to regulate them, banning services for minors and restricting adult use, amid concerns about emotional dependence and demographic pressure. Wissner-Gross's counterintuitive forecast: the CCP will probably reverse course once it figures out how to make AI companions a vector for ideological control rather than a distraction from it. Neil Stephenson's "Young Lady's Illustrated Primer" as CCP product. That's a genuinely dark image.

For everyone else, the regulatory question remains wide open. Ismail's framing is probably the most useful one: "How do you extract the promise without the peril?" The legitimate benefits, companionship for the isolated, therapeutic support, social skills coaching, don't disappear just because the risks are real. But the 20-year lag between technology rollout and regulatory correction that characterized tobacco and social media isn't available here. The timeline is too compressed.

NVIDIA's $96.2 billion quarter is the headline. But the more durable story is what's being built underneath it, and who gets to see inside the loop.

---

Yuki Okonkwo covers AI and machine learning for Buzzrag.