Semiconductors & Advanced Manufacturing
The chip shortage is over. The power shortage is just beginning.
The semiconductor industry spent years worrying about whether TSMC could make enough chips. Today's data suggests the binding constraint has shifted — it's now whether the power grid can keep up with the data centers those chips live in. This week's coverage is dominated by a gold rush in AI infrastructure, a striking divergence between China's two biggest cloud players, and a growing willingness to light natural gas on fire to keep GPUs running.
When You Need 10x the Power and You Need It Now
The clearest signal this week came from Alibaba Cloud, whose CEO Eddie Wu said the company needs 10 times its 2022 compute capacity to keep pace with AI demand. That's not a rounding error — that's an order-of-magnitude gap between where they were 4 years ago and where they need to be. Alibaba is one of the world's largest cloud providers, offering rented computing power (servers, storage, AI processing) to businesses across Asia and globally.
What makes this striking is who's sitting on the other side of the ledger: Tencent Cloud — Alibaba's primary Chinese rival, the company behind WeChat — is reportedly struggling to generate a return on investment from its GPU purchases. GPUs (graphics processing units, originally designed for video games but now the workhorse of AI training and inference) are extraordinarily expensive: a single Nvidia H100 server rack can cost hundreds of thousands of dollars. If Tencent is buying them without generating commensurate revenue, that's a warning sign that not every AI bet is paying off — even inside the world's second-largest economy.
The Alibaba-vs-Tencent split is worth watching. It suggests the AI infrastructure buildout isn't uniformly rational — some players are scaling into demand, others may be scaling into a cost hole.
The Power Grid Can't Keep Up — So They're Building Their Own
The most revealing infrastructure story this week isn't about chips at all: it's about electrons.
Elon Musk's AI company xAI reportedly deployed 19 natural gas turbines at its Colossus 2 data center in Southaven, Mississippi over the past two months. Natural gas turbines at a data center aren't a long-term plan — they're an emergency workaround when you can't get enough power from the utility grid fast enough. The fact that xAI is burning gas to run GPUs tells you something important: the bottleneck in AI compute has become less about whether Nvidia can manufacture chips and more about whether there's a wall socket big enough to plug them into.
This dynamic is showing up elsewhere simultaneously. In Salt Lake City, energy company Certarus is supplying natural gas to a 60MW data center — megawatt, a measure of electrical power, equivalent to roughly 60,000 average American homes — as a bridge until permanent grid infrastructure is installed. Near Minneapolis, a city just approved a one-year moratorium on a T5 data center proposal (T5 being a tier classification for data center reliability, with T5 being among the highest), citing infrastructure concerns.
And then there's Star Catcher, which raised $65 million to build an orbital power grid — literally beaming power from space to satellites. That's an early-stage moonshot, but the fact that investors are funding space-based power solutions signals how seriously the industry takes the ground-level constraint.
The Gigawatt Era: Scale That Would Have Seemed Science Fiction Five Years Ago
A gigawatt (GW) is 1,000 megawatts — enough power for roughly a million homes. Data centers used to be measured in megawatts. Now they're being announced in gigawatts.
Nebius — the European AI cloud company that was spun out of Yandex, the Russian internet giant, and has been rebuilding itself as a Western-market AI infrastructure player — announced plans for a second US campus in Pennsylvania at 1.2GW once fully built out. That's not a data center, it's an industrial complex. Meanwhile, a company called Cloverleaf is proposing a 633-acre data center in Canadian County, Oklahoma. For reference, 633 acres is roughly equivalent to 480 football fields.
Google filed for a campus outside Richmond, Virginia (Project Loch, three buildings). Digital Realty — one of the world's largest data center operators, essentially the landlord of the internet — broke ground on a new campus in Paris. A municipal data center was announced in Reșița, Romania.
What's happening here is a geographic diversification of AI compute infrastructure. For years, data center capacity was concentrated in a handful of US metros (Northern Virginia, Silicon Valley, Dallas) and a few European hubs. The industry is now building everywhere — partly to get closer to power sources, partly to meet data sovereignty requirements (European regulations often require data to stay in Europe), and partly because the sheer volume of capacity needed can't fit in existing corridors.
The Takeaway: Chips Solved, Power Unsolved
The narrative in semiconductors for the past several years was about supply — could TSMC (Taiwan Semiconductor Manufacturing Company, the world's dominant chip manufacturer) make enough advanced chips? That constraint has loosened. The new constraint is energy. Every AI chip that ships eventually plugs into a wall, and the world's walls aren't ready.
Watch for two trends: First, how utilities and regulators respond to the data center buildout — the Minneapolis moratorium is unlikely to be the last. Second, whether Tencent's GPU ROI struggles turn out to be an isolated case or an early signal that the AI infrastructure cycle is getting ahead of actual monetization.
TL;DR - Power, not chips, is now the AI industry's binding constraint — xAI is burning natural gas to run GPUs, and a Minneapolis suburb just blocked a major data center over grid concerns - China's cloud giants are diverging: Alibaba says it needs 10x its 2022 compute capacity; Tencent is reportedly failing to generate returns on its GPU investments — a split worth watching - The gigawatt era is here: New data center campuses are being announced at scales (1.2GW, 633 acres) that would have been unthinkable five years ago, and they're spreading globally — Pennsylvania, Oklahoma, Paris, Romania - Infrastructure is becoming the AI story: The question is no longer whether chips can be made, but whether the grid, the permitting process, and the economics can support the buildout that AI companies are demanding
Compiled from 1 source · 20 items
- Data Center Dynamics (20)