Semiconductors & Advanced Manufacturing
The data center arms race is still accelerating — but the real constraint is no longer chips. It's power, and where you can get it.
THEME: Hyperscalers Keep Building, and the Numbers Are Staggering
The scale of cloud infrastructure investment continues to defy any expectation of a slowdown. AWS — Amazon's cloud computing division and the world's largest — operated 56.84 million square feet of data center and office space in 2025, a 16% increase year-over-year. To put that in perspective, that's roughly the footprint of 40 Empire State Buildings, and it grew by a meaningful chunk in a single year. Amazon also separately acquired a data center and office site in Maryland from investment firm T. Rowe Price in a $15 million sale-leaseback deal (a structure where the seller keeps using the property as a tenant — useful when you want to lock in a location quickly without full upfront commitment).
Microsoft, meanwhile, confirmed a new data center in Tyrone, Georgia, adding to a geographic expansion pattern that's been pushing hyperscalers — the handful of giant cloud companies like AWS, Google, Meta, and Microsoft that build at planetary scale — into secondary and tertiary markets that offer cheaper land, available power, and incentive-hungry local governments.
In China, cloud infrastructure spending reached $14.7 billion in Q4 2025 alone, up 26% year-over-year, according to the Data Center Dynamics data. This is a notable number: despite U.S. export controls on advanced chips restricting what Nvidia and others can sell into China, Chinese cloud giants are clearly finding ways to keep building — whether through domestic chip alternatives or by deploying older-generation hardware at scale.
THEME: Nvidia's Grip on AI Infrastructure Isn't Loosening
One of the more telling stories this week: Nebius, Lambda, and CoreWeave — three of the companies that rent out GPU (graphics processing unit, the chips that power AI training) capacity to researchers and startups — confirmed they have no near-term plans to buy Google's TPUs (Tensor Processing Units, Google's custom AI chips designed as an alternative to Nvidia's products).
The punchline is in the subhead: "We love Nvidia chips, say Nvidia-backed firms." All three companies have received investment from Nvidia, which creates obvious alignment — but the more important point is that Nvidia's CUDA software ecosystem (the programming environment that makes Nvidia GPUs easy to use for AI) has created a lock-in that even well-funded alternatives struggle to break. Developers write code for CUDA; switching to TPUs or other alternatives means rewriting that code. Google has been pushing TPUs for years, but customer adoption outside of Google's own internal workloads remains limited. For Nvidia, this is the best possible news: the moat is behavioral, not just technical.
THEME: Power Is Becoming the Binding Constraint
Two stories this week illustrate the same underlying pressure from very different angles.
First: a proposed $30 million data center near Austin, Texas failed to get a recommendation for approval from the local Planning and Zoning Commission. Local resistance to data centers is growing across the U.S. and Europe — driven by concerns about power consumption, water use for cooling, and limited local job creation relative to the infrastructure footprint.
Second, and more structurally interesting: a piece on whether ammonia-to-power systems could become a serious energy source for data centers. The Haber-Bosch process — a century-old industrial method for synthesizing ammonia from nitrogen and hydrogen — produces a chemical that can be "cracked" back into hydrogen and burned or fed into fuel cells to generate electricity. The appeal for data centers is that ammonia is energy-dense and easier to store and ship than hydrogen alone. This is early-stage thinking, but the fact that it's being discussed seriously signals how acute the power problem has become: operators are now exploring century-old industrial chemistry to keep the lights on.
Separately, SP Energy Networks and startup Keen AI launched a tool claiming to cut UK grid connection wait times — currently measured in weeks — to seconds. Grid connection queues have become one of the primary bottlenecks for new data center construction in the UK and Europe; the average wait for a new grid connection in Britain has stretched to years in some regions.
THEME: Infrastructure Is Going Everywhere — Including Places You Wouldn't Expect
A genuinely striking data point: QGroup launched what it claims is the first data center in The Gambia, a small West African nation. Angola's government also inaugurated a national data center the same week. Separately, OVH — a French cloud provider — launched an edge location in Auckland, New Zealand, and Slovakia completed integration of its Perun national supercomputer, with compute performance expected to exceed 25 petaflops (a petaflop is a quadrillion mathematical operations per second — roughly the compute needed for serious scientific simulation or AI model training at national scale).
In Southeast Asia, Malaysia's DPS Resources and Alibaba signed a memorandum of understanding to potentially develop data centers in Melaka — part of a broader pattern of Chinese cloud companies expanding physical infrastructure across Southeast Asia, partially as a way to serve regional customers and partially to establish a presence outside of China's borders.
The through-line: AI compute demand is so large and so geographically distributed that infrastructure is now being built into markets that had essentially no data center footprint five years ago.
BONUS SIGNAL: Quantum Gets Another Infusion
QuantWare raised $178 million in a Series B round to support development of its VIO quantum processor. The investors — Intel Capital (the investment arm of chipmaker Intel), IQT (a U.S. government-linked deep-tech fund), and ETF Partners — suggest this is being treated as serious infrastructure-adjacent technology rather than pure research. Quantum computers use quantum mechanical effects to solve certain problems exponentially faster than classical chips; they remain years from widespread commercial use, but the capital flowing in continues to grow.
The takeaway: The chip industry's attention has largely moved from "can we make enough chips?" to "can we actually power and site the infrastructure those chips require?" Nvidia's dominance looks durable — the software lock-in is as important as the hardware lead — but the next constraint shaping this industry is increasingly physical and political: land, power grids, and local government approval.
TL;DR - Hyperscalers are still building aggressively — AWS grew its footprint 16% in a year; Chinese cloud spending is up 26% despite chip export controls - Nvidia's lock-in remains ironclad — major GPU cloud providers confirm they have no plans to switch to Google's alternative AI chips, citing CUDA software compatibility - Power and permitting are now the binding constraint — a Texas data center got blocked locally; operators are exploring ammonia-as-fuel and AI grid tools to solve an infrastructure bottleneck that chips can't fix - Data center buildout is going truly global — new facilities announced or launched in The Gambia, Angola, New Zealand, and Malaysia in a single week
Compiled from 1 source · 20 items
- Data Center Dynamics (20)