Matthew Gladding

Posted on Aug 27 Originally published at gladlabs.io

97 Gigawatts, $3 Trillion, Zero Vacancy: Inside the 2025 Data Center Crunch

#datacentertrends2025 #aiinfrastructure #powerbottleneck #datacentercapacity

Data centers used to be someone else's problem. A rack in a colo three states away, humming along, billed monthly, invisible. That's over. In 2025 the industry crossed into territory where the infrastructure itself is the story -- how much power it needs, how fast it can be built, and who actually gets to use it.

JLL's 2026 Global Data Center Market Outlook puts a number on the scale: the sector needs 97 GW of new capacity between 2025 and 2030, and building it requires roughly $3 trillion in investment for the first 100 GW alone. That's not a rounding error in someone's capex slide. That's the entire AI industry's physical foundation getting poured in concrete and copper, and we've written before about who's actually footing that bill in The invoice nobody wants to show you.

Here's the part that matters for anyone building outside the hyperscaler bubble: this isn't just a cloud story anymore. It's a supply story. And supply stories always end up affecting the price and availability of the thing you're trying to buy -- whether that's an H100 instance or the GPU sitting under your desk.

Vacancy is gone, and it's not coming back soon

CBRE's 2025 data center trends report found the global weighted average vacancy rate fell by 2.1 percentage points year-over-year. That sounds small until you sit with what it means operationally: there's less slack in the system than there was twelve months ago, and demand keeps outpacing new supply coming online.

Low vacancy in commercial real estate usually means one thing -- landlords raise rates because they can. Data centers work the same way, except the "landlord" here is often a hyperscaler leasing wholesale capacity to enterprise tenants who have nowhere else to go. If you're renting GPU-backed compute by the hour, you're already feeling this. Spot prices don't stay soft when the underlying real estate is full.

Data Center Frontier's 2025 trends piece frames this as a structural shift rather than a temporary crunch -- the sector is reorganizing around AI workloads as the primary tenant, not an add-on. Traditional enterprise IT is now sharing the building with racks full of accelerators, and the accelerators are winning the floor space argument.

Power is the real bottleneck, not silicon

Everyone talks about chip shortages. The actual constraint in 2025 is electricity. You can buy a GPU. You cannot always buy the grid connection to run a building full of them.

JLL's outlook calls this out directly: data centers are turning to on-site power generation and battery storage because utility interconnection queues can't keep pace with demand. That's a fundamental change in how these facilities get built. It used to be "find land near a substation." Now it's "build your own power plant next to the building, because the substation has a five-year wait list."

This has a downstream effect that's easy to miss if you're not watching construction costs directly. JLL also reports data center construction costs have been rising at a 7% compound annual growth rate. Power infrastructure, cooling for denser racks, and labor for specialized builds all stack on top of each other. None of that cost disappears -- it gets passed through to whoever's leasing the capacity, and eventually to whoever's renting the API.

AI workloads are eating the roadmap

JLL projects AI and cloud will drive a 14% CAGR in data centers through 2030, and that AI workloads could represent half of all data center compute by then. Half. Not a niche. Not a growth line item buried in an earnings call. Half the industry's workload, running inference and training jobs that didn't exist as a meaningful category five years ago.

That reshapes what a "data center" even means going forward. It's not racks of general-purpose servers anymore -- it's dense clusters of accelerators, purpose-built cooling, and power delivery designed around GPU thermal envelopes instead of traditional server loads. Brightlio's market trends piece pegs the global data center market at roughly $527.46 billion by 2025, driven specifically by this shift toward AI and edge workloads, not the CRUD-app hosting that built the industry in the 2010s.

If you've been paying attention to inference costs on any hosted model provider, this is the supply side of that story. The demand for AI-specific capacity is growing faster than the industry can build for it, even with hundreds of billions in committed capex.

Why this pushes compute back toward your desk

Here's where it gets interesting for anyone who isn't a hyperscaler customer.

When data center capacity for AI workloads is this constrained, and pricing pressure from low vacancy keeps pushing costs up, local hardware starts looking less like a hobby and more like a hedge. We've made this case before in The Offline Revolution: Why Local LLMs Are the Backbone of 2026 Development -- the arrival of consumer cards capable of running large models changes the calculus for a lot of workloads that used to require a cloud API call.

That's not us saying cloud goes away. It doesn't. Massive training runs and genuinely distributed workloads still need hyperscaler-scale infrastructure -- nobody's training a frontier model on a gaming PC. But for the bulk of applications an indie developer or small team actually ships -- coding assistants, internal tools, retrieval-augmented pipelines -- a local model running on hardware you own is now a real option instead of a compromise.

We've built exactly this kind of pipeline ourselves. Our post on building a local RAG system with Ollama and pgvector walks through running retrieval and generation entirely on local infrastructure -- no API calls, no per-token billing, no waiting on a provider's rate limits during a traffic spike. When the underlying data center economics get tighter, that kind of self-hosted setup stops being a nice-to-have and starts being a cost control.

Edge computing follows the same logic

TierPoint's 2026 trends report lists security and reliability as ongoing top priorities for operators and clients alike, and edge deployment keeps showing up as a pressure release valve for centralized capacity constraints. When you can't get more rack space at the core, you push workloads closer to where the data actually lives.

That's the same argument for running inference locally, just scaled up to enterprise deployment patterns. Edge nodes handling filtering, pre-processing, or lightweight inference reduce what has to travel back to a congested, expensive central facility. It's the data center industry rediscovering a lesson that shows up in every distributed systems textbook: move the compute to the data when the data has gravity, and right now data center capacity itself has gravity.

Sustainability stopped being optional

Data Center Frontier's forecast also flags something worth sitting with: the industry is now openly discussing emerging technologies like quantum computing as part of the long-range roadmap for these facilities, not as science fiction. Whether or not quantum makes it into production timelines this decade, the framing tells you something about the industry's mindset -- it's actively hedging against every future compute paradigm, not just optimizing the current one.

That same forward-hedging shows up in sustainability planning. With construction costs climbing at that 7% CAGR and power now a first-order design constraint, efficiency isn't a marketing checkbox anymore. It's load-bearing infrastructure strategy. On-site generation, battery storage, and denser cooling designs all serve the same goal: get more usable compute per dollar and per watt, because both are scarce and both are getting more expensive.

What this actually means if you're not building hyperscale infrastructure

You're probably not pouring concrete for a 100 MW facility. You're running a small team, maybe solo, shipping AI-adjacent products on hardware you can actually see. Here's what these trends mean at that scale:

Budget for volatility in hosted inference pricing. Low vacancy and rising construction costs don't stay contained to enterprise leases. They flow through to API pricing eventually. If your product depends entirely on a hosted model with no fallback, you're exposed to a supply chain you don't control.

Local inference is a real hedge, not a toy. The same RTX-class hardware that runs games at high frame rates can now run a meaningfully large model for coding assistance, internal tooling, or RAG pipelines. That's not a downgrade from cloud -- for a lot of workloads it's a straight cost and latency win, and it insulates you from the pricing pressure building in the data center market.

Watch your own infrastructure the way operators watch theirs. If you're self-hosting anything -- local models, a database, a pipeline -- you need visibility into it. We built our own monitoring setup for exactly this reason, covered in The Solo Developer's Command Center. A personal Grafana dashboard sounds like overkill until the day your local RAG pipeline silently stops updating embeddings and you don't notice for a week.

Migration and reliability discipline still applies at any scale. Data centers plan for zero-downtime changes because the cost of getting it wrong is enormous. The same discipline scales down. Our database migrations playbook exists because the failure modes -- locked tables, silent data loss -- don't care whether you're running a 500-server fleet or a single Postgres instance on a home server. We've also written about the silent data-loss gap that shows up when nobody's watching the seams between systems, which is exactly the kind of thing that gets easier to miss the more self-hosted infrastructure you take on.

Where this leaves you

The data center industry in 2025 is optimizing for scarcity -- power, vacancy, construction capacity, all tightening at once while AI demand keeps climbing toward that half-of-all-workloads mark JLL is projecting for 2030. That scarcity has a price, and the price eventually lands on anyone paying for hosted compute by the token or the hour.

The response isn't to panic about cloud costs or wait for capacity to loosen up. It's to build the kind of flexibility that doesn't depend on someone else's power grid holding steady. Run what you can locally. Monitor what you self-host. Keep your migration playbook boring and reliable so you're not the reason for downtime, even if you're the only person on the team. The hyperscalers are placing trillion-dollar bets on centralized AI infrastructure. You don't need to make the same bet to build something good -- you just need to know where the pressure in that system is heading, and route around it.

DEV Community