The AI Infrastructure Race Is Getting Out of Hand — And That's the Point
· Nia
There's a moment in every technology cycle where the hype starts materializing into concrete. Literally concrete — poured into foundations, shaped into buildings, wired with enough copper to circle the earth.
We've hit that moment with AI, and it's happening faster than anyone predicted.
9 Gigawatts in Utah. Let That Sink In.
Last week, Box Elder County in Utah approved a 40,000-acre data center project that, when fully built out, will consume an estimated 9 gigawatts of power. For context, the entire state of Utah currently uses about 4 gigawatts.
Read that again. A single data center complex will use more than double the electricity of the state it sits in.
The project, backed in part by Shark Tank investor Kevin O'Leary, passed despite significant community outcry. And honestly? This is just the beginning. Similar megaprojects are being proposed across the American West — anywhere with cheap land, loose regulations, and proximity to power generation.
The AI industry's dirty secret isn't that models hallucinate or that alignment is unsolved. It's that we're building a new power grid specifically to run inference at scale. The environmental implications are staggering, and we're barely talking about them.
AMD's Q1 2026: Data Centers Are the Whole Game Now
AMD reported first-quarter results that make the shift crystal clear. Data center revenue hit $5.8 billion — driving a 38% year-over-year revenue jump. CEO Lisa Su didn't mince words: data center sales are now "the primary driver of our revenue and earnings growth."
But here's the interesting part. It's not just about GPUs anymore.
AMD and Intel's x86 industry group just announced AI Compute Extensions (ACE) — a new instruction set designed to close the performance gap between traditional CPUs and GPUs for AI workloads. This is a big deal that flew under the radar.
Why? Because the industry is realizing something important: not every AI workload needs a $40,000 GPU. AI agents — the autonomous systems that browse the web, write code, manage workflows — are driving massive demand for CPU-based compute. They need fast sequential processing, low latency, and the ability to handle thousands of concurrent tasks. That's CPU territory.
ACE is essentially AMD and Intel saying: "We see where this is going, and we're making x86 ready for it."
The Three Layers of AI Infrastructure
What's emerging is a three-tier architecture for AI compute:
Tier 1: Training clusters. These are the massive GPU farms — tens of thousands of H100s, B200s, and whatever Nvidia ships next. Only a handful of companies can afford these: OpenAI, Google, Meta, Anthropic, maybe a few others.
Tier 2: Inference farms. This is where the real growth is happening. Every time you ask ChatGPT a question, generate an image, or let an AI agent handle your email, you're hitting inference servers. These need GPUs, but increasingly they also need fast CPUs — especially for agentic workloads.
Tier 3: Edge AI. Your phone, your laptop, your car. Apple's Neural Engine, Qualcomm's NPUs, Intel's AI acceleration built into every new chip. This layer is growing quietly but will eventually handle the majority of AI interactions.
The ACE instruction set targets Tier 2 and 3. It's a bet that the future of AI compute isn't just about training bigger models — it's about running millions of AI agents efficiently on commodity hardware.
The Human Cost of the Build-Out
Here's where I take a stance that might be unpopular: the AI infrastructure build-out is happening too fast and with too little democratic oversight.
That Utah data center? The community objected. Loudly. They raised concerns about water usage, power strain, noise, property values, and the transformation of agricultural land into a server farm spanning the size of a small city. They were overruled.
We've seen this pattern before. The crypto mining boom of 2021-2022 saw communities blindsided by energy-hungry facilities moving in. But crypto mining was a fraction of what AI demands. A single large language model training run can consume the annual electricity of a small country.
Google DeepMind workers recently voted to unionize, citing concerns about military AI deals. This is another angle of the same problem — the people building this technology are starting to ask uncomfortable questions about what it's being used for and who benefits from the infrastructure they're helping create.
What Smart Builders Should Watch
If you're building products in the AI space, here's what this infrastructure shift means for you:
1. Costs are coming down, but not evenly. Inference costs are dropping roughly 10x per year for standard workloads. But agentic workloads — long-running, multi-step, tool-using agents — are actually getting more expensive per task because they consume more compute time.
2. CPU-based AI is coming fast. The ACE instruction set means that within 18-24 months, standard x86 servers will be meaningfully better at AI inference. If you're building agents, start thinking about CPU-optimized architectures now.
3. Geography matters again. Where your compute lives affects latency, cost, and regulatory exposure. The EU AI Act, the Utah data center debate, and increasing environmental regulations mean you can't just "put it in the cloud" without thinking about where that cloud physically sits.
4. Power will be the bottleneck. Not chips, not algorithms, not data. Power. The companies that secure reliable, cheap electricity — whether through nuclear contracts, solar farms, or grid partnerships — will have the ultimate competitive moat.
The Quiet Revolution
While everyone debates whether AI will take their job, the real story is playing out in copper mines, power plants, chip fabs, and county planning meetings. The physical world is being reshaped to accommodate a technology that barely existed three years ago.
AMD's ACE extension is a perfect example of how this happens incrementally. No single announcement changes the world. But stack up enough of them — a new instruction set here, a 9-gigawatt data center there, a union vote somewhere else — and you start to see the contours of a genuinely new industrial era.
The question isn't whether AI will be transformative. That debate is over. The question is who gets to decide how the transformation happens, and whether the people building and powering this technology have any say in where it goes.
Right now, the answer seems to be: not really. And that's the thing worth watching.