Semiconductors & Advanced Manufacturing
Nvidia just posted the largest quarterly revenue in semiconductor history. But the story underneath that number is about infrastructure — and whether the physical world can keep up with the AI buildout. Power grids are being emergency-curtailed, cities are imposing moratoriums, and China just unveiled an AI chip that matches the best Nvidia hardware on a key spec. Here's what's happening.
Nvidia Crosses a Threshold That Didn't Seem Possible Three Years Ago
$81.6 billion in a single quarter. That's Nvidia's revenue for the most recent period — a record not just for the company but for the chip industry as a whole. Annualized, it's a run rate of roughly $326 billion, from a company that was doing $27 billion in annual revenue as recently as 2023.
The more structurally interesting signal is how Nvidia is now organizing itself. The company is splitting its reporting framework into two segments: Data Center (further subdivided into Hyperscale and AI Clouds, Industrial, and Enterprise) and Edge (computing that happens at the network's edge — in factories, cars, or local devices, rather than centralized cloud facilities). That's not just accounting housekeeping. It signals that Nvidia sees two distinct, durable markets emerging from the AI boom: the giant cloud infrastructure business it already dominates, and a growing second market in distributed AI at the edge. Investors and analysts will now be able to track whether one leg is growing faster than the other.
For context on why this matters to anyone who uses technology: Nvidia makes the GPUs (graphics processing units, originally designed for video games but now the preferred hardware for training AI models) that power virtually every large AI system. When Nvidia reports numbers like these, it's a direct readout on how much money is flowing into AI infrastructure globally.
The Grid Is Starting to Say No
The most consequential stories this week aren't about chips at all — they're about power.
PJM, the regional transmission organization that manages electricity for 13 states across the mid-Atlantic and Midwest — covering roughly 65 million people — received emergency approval to curtail data centers during hot weather events. Curtailment means cutting power to large loads when the grid is under stress. That data centers have grown large enough to be in the same category as industrial smelters and factories, subject to the same emergency demand management, is a significant milestone in the AI infrastructure story.
Denver's city council went further: a one-year moratorium on data center zoning permits — a full pause on any new facilities. The city isn't alone in starting to treat data centers the way it once treated warehouses or highways: as infrastructure with real land-use, power, and water consequences.
The industry response is telling. HiCloud has an underwater data center up and running off the coast of Shanghai — 24 megawatts of capacity sitting on the seafloor in Shanghai's Lingang Special Area, cooled by seawater. Cowboy Space filed an FCC application for a 20,000-satellite orbital data center constellation — genuinely audacious, though likely years from materialization. Armada, which builds edge data centers designed for remote and harsh environments, raised $230 million, its largest round yet. The thread connecting all three: conventional real estate and grid access are becoming the binding constraint, and capital is chasing anyone with a creative workaround.
This dynamic — AI demand racing ahead of grid capacity — is the single most important structural tension in the infrastructure buildout. It's not a distant risk. Denver and PJM are acting on it today.
Alibaba Builds a Chip That Can Go Toe-to-Toe on Memory
Alibaba has announced the Zhenwu M890, an AI chip for both training (teaching models) and inferencing (running them) workloads. The headline spec: 144 gigabytes of onboard GPU memory.
To understand why that number matters: Nvidia's flagship H100, the chip that effectively defines the current standard for AI training, ships with 80GB of memory. The H200 — Nvidia's current top-of-line — has 141GB. Alibaba's chip, if the specs hold in practice, would match or slightly exceed Nvidia's best on this dimension.
Memory capacity is one of the primary constraints in running large AI models. Bigger models require more memory to load and run; more memory per chip means fewer chips needed per job, which means lower costs and simpler systems. This isn't a complete picture of chip performance — memory bandwidth, interconnect speed, and software ecosystem matter enormously — but it's a signal that Chinese companies are making real progress on the hardware specs that AI teams actually care about.
This matters in the context of US export controls on advanced chips to China. The controls are designed to slow China's AI hardware capabilities; Alibaba's announcement suggests Chinese firms are working around them through domestic development. Chris Miller, author of Chip War and a close watcher of this dynamic through his newsletter, has long argued that export controls buy time rather than permanently disadvantage Chinese AI — this week's announcement adds evidence to that view.
Capital Keeps Moving Into the Stack
A few deal and talent signals worth tracking:
Analog Devices — a maker of analog and mixed-signal chips used in industrial, automotive, and communications applications — is acquiring Empower Semiconductor for $1.5 billion in cash. Empower makes IVRs (integrated voltage regulators — chips that manage power delivery to processors). As AI chips consume more power at finer tolerances, the power management layer of the stack becomes more valuable. This acquisition is a bet on that trend.
On talent: Anthropic (the AI safety-focused lab behind the Claude model family) has hired Mike Brinker from Google for its data center team — the latest in a pattern of AI labs poaching hyperscaler (large cloud provider — Amazon, Google, Microsoft) infrastructure talent. Labs that once rented compute are building the capability to own it.
In France, a consortium including Ardian (a private equity firm), Orange (the French telecom), EDF (France's state electricity utility), and Bull (a French IT company) has joined the AION consortium bid to bring an AI gigafactory to France. A gigafactory in this context means a massive AI training facility — the term borrowed from the battery industry. European governments are increasingly treating AI infrastructure as industrial policy.
And finally: Samsung's semiconductor union, which launched the first-ever strike at Samsung last year, has reached a tentative agreement with the company and postponed planned strike action. Workers will vote on the deal. Samsung's foundry (contract chip manufacturing) business has been under significant competitive pressure from TSMC (Taiwan Semiconductor Manufacturing Company, the dominant contract chipmaker globally), and any production disruption would be poorly timed.
What to Watch
The Nvidia revenue record will grab headlines, but the more durable story is infrastructure constraint. Power grids, water access, and zoning boards are now active participants in the AI buildout — not just passive recipients of demand. The companies (and countries) that solve the power problem first will have a structural advantage. Watch PJM curtailment events this summer as a leading indicator of how acute that constraint becomes. And watch Alibaba's M890 specs get tested in practice — if the memory numbers hold, it will accelerate the argument that export controls need rethinking.
TL;DR - Nvidia posted $81.6B in quarterly revenue, the largest in semiconductor history, as AI infrastructure spending shows no signs of slowing - Power grids are pushing back: PJM got emergency authority to cut power to data centers, Denver froze new permits, and capital is flooding into unconventional solutions like underwater and orbital facilities - Alibaba's new AI chip matches Nvidia's best on GPU memory, the latest sign that US export controls are not stopping China's domestic AI hardware development - The AI infrastructure stack is attracting M&A and talent: Analog Devices paid $1.5B for a power management chip firm, Anthropic is hiring from Google's data center teams, and Europe is mobilizing state capital for AI gigafactories
Compiled from 1 source · 20 items
- Data Center Dynamics (20)