Why AI Workloads Drive Data Center Acquisitions

The acquisition of Polariton Technologies, announced April 22, 2026, didn't make headlines because it was large. It made headlines because of what it signals: the data center industry is in the middle of a structural renegotiation, and AI is holding the pen.

For years, data center acquisitions followed a familiar playbook — buy capacity, consolidate colocation footprint, extract operational efficiencies. The math was simple. The thesis was durable. Then generative AI arrived at scale, and suddenly the playbook looked like it was written for a different sport.

AI workloads don't just consume more power — they consume it differently, unpredictably, and at a density that most existing facilities weren't built to handle. That reality is now the primary driver behind a wave of strategic acquisitions that looks nothing like the data center M&A of five years ago.

Data Center Acquisitions Are No Longer Just About Real Estate

Strip away the press release language from most recent data center deals, and you'll find the same core logic: buyers aren't purchasing buildings; they're purchasing capability.

The legacy acquisition model valued data centers on a relatively straightforward set of metrics — critical IT load in megawatts, power utilization effectiveness (PUE), lease terms, and geographic redundancy. A 100MW campus in Northern Virginia with a PUE under 1.4 was a trophy asset. Full stop.

That framework hasn't disappeared, but it now competes with a different set of questions: Can this facility support liquid cooling at scale? What's the fiber interconnect density? How close is it to a high-voltage transmission line? Does the design accommodate the 40-60kW per rack densities that GPU clusters demand?

When those questions can't be answered satisfactorily, acquirers move on — or they acquire the company that *can* answer them, which is exactly the logic behind deals like the Polariton acquisition. Specialty capabilities in optical interconnects or silicon photonics, for example, represent the kind of infrastructure intelligence that can't be bolt-on purchased. You acquire it, or you fall behind.

The new data center acquisition target isn't the biggest campus on the market — it's the most purpose-built one.

What AI Workloads Actually Demand From Infrastructure

To understand why acquisitions are accelerating, you need to understand what training and inference workloads actually do to a data center.

A traditional enterprise server rack runs at roughly 5-10kW of power draw. A rack dense-packed with NVIDIA H100 GPUs — the current workhorse of large model training — runs at 10-14kW per GPU, with eight GPUs per server and multiple servers per rack. Do the arithmetic: you're looking at rack densities that can exceed 100kW in cutting-edge AI deployments. Air cooling, the default for virtually all legacy data centers, stops being viable somewhere around 20-30kW per rack.

This isn't a future problem. Hyperscalers are deploying liquid cooling at scale right now. Microsoft, Google, and Meta have all announced or begun construction on liquid-cooled AI training facilities. The companies that haven't retooled — or haven't acquired partners who have — are watching their competitive position erode quarter by quarter.

There's a secondary infrastructure pressure that gets less attention: networking. AI training requires moving enormous volumes of data between GPU nodes at extremely low latency. That demands specialized high-bandwidth interconnects — InfiniBand fabrics, optical networking, and increasingly, silicon photonics solutions from companies like the one Polariton represents. Data center strategies that ignore the networking layer are optimizing for yesterday's workloads.

Scalability compounds all of this. AI model sizes have grown roughly 10x every two years over the last decade. Infrastructure built for today's models may be undersized for models two years out. Acquirers are increasingly underwriting for future capacity needs, not current ones — a meaningful shift in how deals get valued.

Where the Investment Is Actually Going

The financial flows tell a clear story. Infrastructure investments in AI-oriented data centers surged dramatically through 2024 and 2025, with hyperscalers collectively committing hundreds of billions in capex. But the more interesting signal is in the middle market — the tier below the hyperscalers where most acquisitions actually happen.

Private equity has recognized that AI-capable data center assets are supply-constrained in a way that traditional data center real estate wasn't. You can build a generic colocation facility in 18-24 months. You cannot build the specialized engineering capability, the utility relationships, the pre-permitted transmission access, and the liquid cooling infrastructure stack in the same timeframe. That scarcity premium is now baked into deal multiples.

Energy adjacency has become a decisive factor. Data centers running AI workloads at scale need reliable, ideally low-carbon power at volumes that strain regional grids. The most strategically valuable data center acquisitions of the next five years will likely involve assets co-located with dedicated power generation — whether that's utility-scale solar, natural gas peakers, or long-duration storage. Buyers who secure power access as part of an acquisition are buying something that can't be replicated by a competitor with a checkbook and a willing construction crew.

The geographic calculus is shifting too. Northern Virginia and Silicon Valley remain important, but power constraints have pushed serious AI infrastructure investment toward Texas, the Mountain West, and the upper Midwest — markets with available land, transmission capacity, and in some cases, access to cheap renewable energy from wind and solar resources that the coasts simply can't match.

What Successful Acquirers Actually Got Right

The distinguishing characteristic of the acquisitions that have held up well isn't deal size — it's thesis clarity. The buyers who've won in AI-driven data center M&A knew specifically what they were buying and why it was irreplaceable.

Equinix's ongoing strategy of acquiring interconnection-dense facilities demonstrates this. The value isn't the square footage; it's the ecosystem of carriers, cloud on-ramps, and enterprise customers that have co-located there over years or decades. That interconnection fabric takes years to cultivate and can't be rebuilt from scratch. AI workloads that require ultra-low latency access to cloud APIs make those interconnection hubs more valuable, not less.

On the infrastructure investment side, the acquisitions that have struggled share a common failure mode: buying generic capacity at a premium multiple with the assumption that AI demand would fill it. Demand has been real, but so has the preference among hyperscalers to build purpose-built facilities rather than retrofit existing ones. Generic colocation is not the same as AI-ready colocation, and the market is increasingly pricing that difference explicitly.

The Polariton deal fits a more defensible pattern — acquiring specialized technical capability that addresses a specific bottleneck in AI infrastructure. Companies that solve hard problems at the intersection of AI and physical infrastructure are worth more than their revenue multiples suggest because what they're really selling is time.

The Road Ahead

Several forces will shape how data center acquisitions evolve over the next 24-36 months.

Power will be the binding constraint. Utilities in high-demand markets are already requesting multi-year queues for large interconnection requests. Acquirers who treat power access as an afterthought will find themselves with expensive assets they can't fully monetize. The smarter play — already visible in several recent deals — is acquiring land and assets that come with existing utility relationships or permitted capacity.

Regulatory scrutiny of large infrastructure transactions is increasing, particularly where foreign capital is involved or where critical data infrastructure intersects with national security considerations. Deals that would have closed in six months two years ago are taking longer, and some aren't closing at all.

On the technology side, the next wave of AI accelerators — from AMD, Intel, and a raft of startups — will have different thermal and power profiles than today's NVIDIA-dominated ecosystem. Facilities locked into a single cooling or power architecture may face expensive retrofits. The most future-proof acquisitions will prioritize flexibility in infrastructure design over optimization for any single hardware generation.

Finally, the energy transition is not separate from the data center story — it's embedded in it. Carbon commitments from the hyperscalers are real, and the power demands of AI are making them harder to keep. Infrastructure investors who can bridge AI compute needs with clean energy solutions are positioned to capture value from both trends simultaneously.

The acquisition of Polariton is a small transaction in the context of global infrastructure M&A. But it's a clear indicator of where the serious money is looking: not at size, not at geography, but at the specific technical capabilities that AI demands and that the existing infrastructure stock wasn't built to provide. That gap between what AI needs and what the current data center inventory delivers is the defining investment opportunity in infrastructure right now — and it's nowhere near closed.

Explore the InfraSale Marketplace for more insights and opportunities.

INTERNAL LINK SUGGESTIONS

• [INTERNAL LINK: data center acquisitions]
• [INTERNAL LINK: AI workloads]
• [INTERNAL LINK: infrastructure investment trends]