The most photographed object at Nvidia’s annual developer extravaganza this year may not have been a GPU at all. It was a robot duck. Among the packed halls and keynote theatrics, a small flock of mechanical waterfowl waddled across the show floor, drawing crowds that would make a flagship smartphone launch blush. It was charming, slightly absurd, and entirely on-brand for a company that has turned its developer conference into something closer to a technology carnival than a trade show.
But the ducks are a distraction from a more consequential undercurrent. The same event that gave us Jensen Huang’s trademark leather-jacket swagger and a standing-room-only keynote also surfaced a quieter anxiety: the neoclouds, those upstart GPU rental outfits that sprouted up to meet insatiable AI demand, are showing signs of strain. The Financial Times and others reported that some of these smaller cloud providers are struggling with the brutal economics of buying the latest silicon, building data centers fast enough, and competing on price against hyperscalers who can absorb losses their smaller rivals cannot.
That tension is the real story of this AI cycle. For two years, the narrative has been simple: demand for compute outstrips supply, and anyone with GPUs can print money. The neoclouds rode that wave hard, raising capital at eye-watering valuations and signing multi-year contracts with AI labs and enterprises desperate for capacity. Now the bill is coming due. Depreciation schedules on cutting-edge accelerators are punishing, electricity costs are climbing, and the next generation of chips arrives faster than the last one can be paid off. A neocloud that bet big on H100s a couple of years ago is now watching customers ask about Blackwell and beyond.
None of this means the AI infrastructure boom is over. Far from it. Nvidia’s own messaging at the show reinforced that the company sees demand continuing to compound, with Huang painting a picture of a computing transition as significant as the move from general-purpose CPUs to accelerated computing. The open models unveiled in recent days by Nvidia, Meta, and Apple suggest the software layer is racing to keep pace with the hardware. But the distribution of winners is shifting. The hyperscalers, with their balance sheets and captive customers, can weather cycles that would sink a startup. The neoclouds cannot.
There is a parallel worth noting in the quantum computing chatter that has re-emerged around this cycle. Quantum has long been the technology that is perpetually a decade away, and the current wave of hype has a familiar feel. Investors are again asking whether quantum is “the next AI,” a comparison that flatters both fields while clarifying neither. The more useful lesson from the neocloud shakeout is that infrastructure businesses live and die on capital structure and timing, not on the elegance of their pitch decks.
On the policy side, the flow of government money into semiconductors continues. New CHIPS awards to a handful of firms, including vertical gallium nitride and fusion-adjacent players, signal that Washington wants to broaden the domestic base beyond leading-edge logic. Japan’s sizable chip bet and the European startup ecosystem defending Ukraine’s tech sector all point to a world where semiconductor supply chains are being deliberately reshaped along geopolitical lines. That is long-term tailwind for equipment makers and a long-term headache for anyone modeling costs.
Meanwhile, the cybercrime economy keeps expanding, and AI is cutting both ways. Attackers use generative tools to scale phishing and reconnaissance; defenders use the same tools to triage alerts and patch faster. The arms race is real, but the more immediate problem for most enterprises is basic hygiene, not adversarial superintelligence.
For all the spectacle at GTC, the durable takeaway is about margin and moat. Nvidia remains the undisputed center of gravity, and its conference is now a barometer for the entire industry’s mood. But the mood underneath the confetti is more cautious than it was a year ago. The companies that rented GPUs and hoped for the best are discovering that infrastructure is a grind. The robot ducks will get the headlines, and Jensen will get the applause. The neoclouds will get the hard questions. That is what a maturing boom looks like.