By Amy Barzdukas
Head of Americas Marketing
In the second blog in our “Rise of the AI Data Center” series, inspired by our latest white paper, we deal with time to power. Data center construction schedules are driven by getting an operable data center into the owner’s hands, and getting predictable power performance is one of the longest poles in that tent.
The race to build AI infrastructure is fiercely competitive, and the difference between bringing a data center online in, say, 24 months versus 42 months can mean missing multiple generations of compute—which puts AI factories at a potentially fatal disadvantage. Today, power is a data center’s gating resource, replacing land, capital, and even hardware as the primary constraint on AI infrastructure deployment. Without power, your racks of GPUs are merely expensive decor.
Case in point: In late 2024, Microsoft CEO Satya Nadella famously said Microsoft had billions of dollars' worth of AI GPUs in inventory just … waiting: "The biggest issue we're having now isn't chips—it's power,” Nadella said. The company then spent $11.1 billion in Q1 2026 to lease "warm shells" that it could activate faster than new construction. Meta made a similar announcement in April 2026, partnering with CoreWeave in a $21 billion deal to expedite its AI infrastructure.
Microsoft’s and Meta’s strategic pivots are clear signs that time to power has become a board-level variable, not just a facilities metric. Time-to-power roadblocks delay AI training runs and frontier models and, for hyperscalers, can strand billions in capital investment and foreclose the possibility of future AI dominance.
On the other hand: The strategy pivot opens huge opportunities for AI infrastructure providers.
What’s the Problem? The Grid Interconnection Queue
When it comes to delayed time to power, the grid interconnection queue takes much of the blame.
Lawrence Berkeley National Laboratory's 2025 "Queued Up" analysis noted that projects installed in 2024 took, on average, five years to complete the interconnection process—up from three years in 2015 and two years in 2008. The league leader is California Independent System Operator (CAISO), where in 2024 new projects had taken 9.2 years to work their way to the front of the line. A data center breaking ground in 2026 likely won’t be approved to tap stable grid power until at least 2030.
There are many reasons why. Principal among them: The capacity just isn’t there to meet the growing demand. For example, in Northern Virginia, currently the world's largest data center market, Dominion Energy recently reported it was processing applications for 60 GW of power—but could offer only 8 GW of available capacity. That's a 7.5x demand-supply gap.
And the grid saddles data center operators with more troubles than simply restricted supply. It was engineered for 20th century loads that drew power in relatively predictable, gradual ways. (Motors actually resisted sudden changes in current, acting to stabilize grid frequency.) In contrast, AI data centers draw massive power with near-instant variability. The grid was never meant for this; at gigawatt scale, those swings destabilize and degrade a local grid segment. It’s a core reason why utilities struggle to handle AI loads even when raw generating capacity technically exists.
It’s worth noting that lead times for gas turbines, the main onsite power source for data centers going their own way, now extend well beyond five years, while used gas turbine generators have become a very hot commodity. All because operators are trying to overcome grid restrictions by any means available.
Construction, Supply Chains, and GPU Power Issues
Traditional construction models layer on additional risks. These include supply chain delays for long-lead electrical equipment, local permitting queues, and onsite construction mishaps.
Consider transformers: A 50 MW AI data center typically requires four to six primary transformers, which require lead times of 52 to 120 weeks plus three to six months for transport. An AI factory could be fully built but have its GPU clusters in storage, unable to process a single workload while waiting on power infrastructure.
Now on to GPU power issues. As noted above, the grid wasn’t designed to accommodate data center power characteristics. But the infrastructure they’ve been built into is more than happy to pass those power demands straight up to the utility grid. So what do utilities do? Reject data center power permits upfront or, worse, shut off power to operational facilities that violate their power profiles.
You see why the key to a successful AI factory is not the capacity that operators have built in, but how long before that capacity is energized and generating value.
The Modular Response: Why Wait?
There are two key strategic ways to accelerate time to power. One is to build your own microgrid, which does an end run around the interconnection queue entirely; we’ll dive into that more deeply in a later blog.
The other strong remedy involves equipment prefabrication and modularity. This moves integration, testing, and quality control into a planned factory setting and out of chaotic construction sites; enables fast, simple expansion and replication; and streamlines the commissioning process.
The typical data center is built and commissioned in 12–24 months. Operators that use standardized, modular assets cut these times by as much as 50%, which takes on greater significance as current construction and commissioning timelines are trending even longer.
These time-to-power savings derive from:
- Parallel workstreams—Factory fabrication of electrical gear takes place alongside site prep and civil work, rather than sequentially. This alone should shave several months off a project.
- Faster commissioning—Pretested and prewired equipment requires far less onsite assembly, troubleshooting, and validation. For example, a modular substation that would take months to assemble in the field might arrive ready to energize in weeks.
- Reduced dependency on skilled labor—Moving electrical integration and related work into a controlled environment with a permanent skilled workforce reduces labor bottlenecks and scheduling risk, both shortening time to power and making it more predictable.
- Repeatability—When designs are proven, procurement is streamlined, and the process is standardized, the buildout becomes almost routine. Hyperscalers building their tenth identical modular deployment move much faster than when they built their first.
- Inside-Out Power Architecture—By carefully matching server power profiles to gray space architecture, buffering the grid from fluctuations it can’t handle, operators ensure they don’t face a “Sophie’s Choice” like this: Either throttling down the expensive, cutting-edge GPUs they just bought, or ripping out the entire power architecture and starting over, which is a sure way to add another year to their time to power.
Get a Headstart in the AI Race
The operators who win the AI era will be those who solve time to power while others … wait.
To learn more about Delta’s modular data center solutions, start with the
Delta InfraSuite
To start this blog series at the beginning, go to
Load Volatility: The Invisible Killer in AI Data Centers
To continue in the series, go to
What Your TCO Model Doesn't Know Could Cost You Millions
Follow us on LinkedIn