Nvidia Just Proved the AI Boom Isn’t a Bubble… Yet

Nvidia just dropped another monster quarter , and the stock popping 5% pre-market is the easy part of the story. The harder question is whether this print tells us we’re in the late innings of an AI bubble… or still in the middle of a durable buildout. We don’t need a final verdict, but this report and the surrounding commentary give the bulls real ammunition to argue we’re not in classic bubble territory – at least not yet.

Start with the basics: the numbers are absurd, but they’re also clean. Revenue came in at $57 billion, up 62% year-on-year and 22% sequentially, driven overwhelmingly by data center compute and networking at $51.2 billion, up 66% YoY and 25% QoQ. Networking alone grew 162% YoY to $8.2 billion and is now 16% of the data center mix. That’s not “we built it and nobody came”; that’s “we built it and everyone is standing in line.” Gross margin at 73.6% was slightly above guidance, and Nvidia guided Q4 revenue to $65 billion (+65% YoY, +14% QoQ) with gross margin around 75%, aiming to hold margins in the mid-70s into 2027 even as input costs rise.

The other key tell: there are no “dark GPUs.” Colette Kress was explicit that clouds are sold out, the installed base (Blackwell, Hopper, Ampere) is fully utilized, and supply is still the limiting factor. Inventory rose 32% QoQ and supply commitments 63% QoQ, but that’s being framed against a booked and visible pipeline rather than a speculative “if we build it, maybe someone will use it.” Management says it has visibility to roughly $500 billion of Blackwell and Rubin revenue through the end of 2026, and reiterated that number on the call – with upside potential. That’s five times the Hopper generation pace. If you’re looking for classic bubble signs, “we have half a trillion already on the books” is not usually how the script reads.

Jensen Huang the AI bubble narrative head-on, which in itself is telling. He didn’t dodge; he disagreed. His framing is that we’re living through three overlapping platform shifts: CPU to GPU accelerated computing, AI transforming existing applications and enabling new ones, and the rise of agentic AI. Nvidia, in his view, sits at the intersection of all three. Importantly, he argued that the hundreds of billions in AI capex will “quickly” be profitable for customers – a direct answer to the “are we overbuilding?” fear. You can rightly be skeptical of any CEO selling a secular story, but this is not hand-waving about eyeballs or MAUs; this is tied to visible workloads, training and inference demand, and very real cost/performance advantages per watt and per dollar.

Analysts, for their part, are not treating this as a terminal blow-off top. Evercore raised its target from $261 to $352, still using a 35x multiple on 2030 EPS discounted back – and notes that on after-hours pricing Nvidia is at roughly 25x P/E, about a 30% discount to its 9-year median of 35x. In other words, on their math, the stock is cheaper than its historical norm despite growth rates that are dramatically higher. Evercore models revenue growth accelerating from 56% in the July 2025 quarter to 79% in July 2026, supported by Blackwell Ultra (GB300) ramping and management’s $500B+ compute pipeline. Stifel, KeyBanc, and Citi all echoed the same broad message: data center demand remains ferocious, Blackwell GB300 is now two-thirds of Blackwell revenue, networking is scaling rapidly, and the $500B pipeline is a baseline, not a top.

Citi added an important structural point: with TSMC CoWoS capacity expected to grow significantly next year and Anthropic / Middle East partnerships as incremental demand, they see upside to the $500B 2025–26 data center sales bogey. That’s a far cry from analysts nervously asking, “Is this the last good quarter?” Instead, they’re probing the upper bound of what the ecosystem can absorb.

There are also important “sanity check” signals that argue against this being a pure mania. First, Nvidia assumes zero data center compute revenue from China in the forward guide. That’s a material piece of business simply crossed out due to geopolitics, yet the company is still guiding to eye-watering growth and mid-70s margins. Second, opex is rising 19% QoQ as Nvidia leans into R&D and ecosystem investment, not financial engineering. The capital is going into product cadence and CUDA moat expansion, not marketing stunts.

The more thoughtful bubble risk is not about Nvidia’s P/E multiple; it’s about infrastructure overshoot. The “dark compute” concept is the right way to frame it. In the dot-com era, you had “dark fiber” – networks built ahead of sustainable demand. Here, the worry is data centers that get built and stuffed with GPUs that can’t be fully powered or economically utilized because grid capacity and real-world use cases can’t keep up. For now, Nvidia’s story is the opposite: chips are sold out, power and supply chain are the constraints, and partners like hyperscalers and CoreWeave are racing to secure capacity. But if we start to see signs of “dark compute” at those partners – fully built facilities that aren’t meaningfully turned on because of power or demand shortfalls – that’s when the AI bubble alarm should get louder.

The margin picture is another key tell. You could reasonably nitpick that gross margins were “just” flat to slightly up, not exploding higher with revenue. But given the capital-intensive early stage of the Blackwell cycle, holding mid-70s margins while ramping a new architecture is a sign of pricing power, not exhaustion. If we start to see gross margins roll over sharply while Nvidia is still spending heavily and talking up pipeline, that’s when you have to worry that competition, customer bargaining power, or simple overcapacity is biting.

So where does this leave the AI bubble debate after this report? Nvidia’s quarter and guide strongly support the “supercycle, not a speculative sugar high” camp: full utilization, booked visibility, disciplined margins, and a customer base with deep pockets and growing use cases. At the same time, the scale of the numbers, the dependence on a relatively concentrated set of hyperscalers, and emerging power constraints mean the risk of overshoot is absolutely real – it’s just more likely to show up first in the broader ecosystem (hyperscalers, second-tier clouds, and power infrastructure) than on Nvidia’s P&L.

In other words: if this is a bubble, it’s not popping today. Nvidia just delivered another data point that the AI buildout is still demand-led and profit-backed. The exit signal isn’t a single blowout quarter like this; it’s when you start seeing dark compute, falling margins, and order downgrades from Nvidia’s biggest partners. Until then, this report gives the bulls a credible argument that the AI trade has more runway than the doomsayers want to admit.