THE INFRASTRUCTURE BEHIND AI  |  ARTICLE 2

Why the race to build AI infrastructure is becoming a race for generation, grid capacity and programme certainty

In the first edition of The Infrastructure Behind AI, I argued that AI is no longer simply a technology story.

Behind every new model, cloud service and AI application sits an increasingly complex physical infrastructure: data centres, power, transmission, substations, water, cooling, fibre, specialist equipment and the people required to bring all of it together.

Since then, one part of that infrastructure challenge has become increasingly difficult to ignore.

Power.

The race to build AI capacity is rapidly becoming a race to secure enough electricity, in the right place, at the right time.

And that changes the delivery challenge considerably.

AI is changing the electricity equation

For much of the past two decades, US electricity demand grew relatively slowly. That is changing.

The US Energy Information Administration says electricity demand has been rising steadily since 2020, after more than a decade of little change, and identifies data centres as a significant driver. It expects US electricity load to increase by 1.9% in 2026 and another 2.5% in 2027, with particularly strong growth in Texas and the PJM region. [1]

Globally, the direction is equally clear.

The International Energy Agency expects electricity generation required to supply data centres to rise from around 460 TWh in 2024 to more than 1,000 TWh by 2030. [2]

Meanwhile, the scale of individual developments is changing.

Microsoft announced a new data centre campus in Pecos, Texas in June that it says will eventually add approximately 2GW of capacity. [3]

Microsoft has also highlighted how individual AI racks have moved from consuming tens of kilowatts to hundreds of kilowatts, while entire campuses can now operate at gigawatt scale. Its conclusion is revealing: power is no longer something a data centre simply plugs into. It increasingly has to be designed around from the outset. [4]

That feels like an important shift.

Building the data centre is only part of the programme

A developer can secure the land. The design can progress. Finance can be committed. Long-lead equipment can be ordered. Construction can begin.

But none of that creates usable AI capacity until sufficient power can actually reach the facility.

That makes the true programme considerably larger than the data centre itself.

Generation  →  Transmission  →  Grid Connection  →  Substation  →  Data Centre  →  Energisation  →  Commissioning  →  Operational Readiness

Each stage has its own stakeholders, approvals, procurement constraints, construction programmes and risks. Yet ultimately, they all have to converge.

From a Planning and Project Controls perspective, that is where this becomes particularly interesting.

The critical path may no longer sit entirely inside the project schedule. It may sit hundreds of miles away in a transmission upgrade, a generating project, a substation programme, a grid study, a transformer manufacturing slot or an approval process controlled by an entirely different organisation.

Texas provides an extraordinary example

Texas is one of the world’s most active data centre markets and an important market for us at Osirian. It also illustrates the scale of the challenge.

In June, ERCOT said it was tracking more than 438,000MW of large-load connection requests, with nearly 89% coming from data centres. It introduced a new process for assessing projects of 75MW or more because the previous project-by-project system was struggling with the volume of applications. [5]

Not every one of those proposed projects will be built, of course. But that is part of the challenge.

Grid operators must distinguish credible demand from speculative demand while developers need sufficient certainty over power availability to make investment decisions.

By August, the issue had become significant enough that ERCOT paused part of its new process while additional verification of data centre projects was undertaken. The EIA subsequently reduced its Texas electricity-demand forecast following the pause. [6]

At the same time, Texas is planning major new transmission infrastructure, including its first 765kV transmission lines, to move increasing quantities of electricity across the state. [7]

This isn’t simply a data centre construction boom. It is increasingly an energy and infrastructure programme.

There probably isn’t one answer to the power problem

Much of the discussion around AI power quickly becomes a debate about which energy technology should win. I suspect that misses the practical point.

The scale and speed of demand mean the answer is likely to involve multiple technologies.

The IEA expects renewables to meet nearly half of the additional global electricity demand from data centres over the next five years, followed by natural gas and coal, with nuclear becoming increasingly important towards the end of the decade and beyond. [2]

The technology companies themselves are already pursuing similarly diverse strategies.

Meta announced agreements this year supporting up to 6.6GW of nuclear capacity by 2035, spanning existing plants and advanced nuclear technologies. [8]

Its Louisiana expansion provides an interesting contrast. Meta says the associated energy agreement will fund seven new natural-gas generating plants, three grid-scale batteries, nuclear uprates and additional purchased power. [9]

Google, meanwhile, contracted more than 12GW of new clean energy in 2025 alone and is investing across technologies including renewables, nuclear, enhanced geothermal and longer-term technologies such as fusion. [10]

The story isn’t nuclear versus gas versus renewables. It is how enough reliable, affordable and deliverable power can be brought online quickly enough to support extraordinary growth in demand.

The sustainability challenge hasn’t disappeared either

This is where another tension becomes visible.

Google’s electricity demand increased 37% in 2025, its largest annual load growth to date.

Despite contracting record quantities of clean energy and reducing its operational emissions by 2%, Google acknowledges that its AI infrastructure expansion is currently moving faster than electricity grids are decarbonising. It identifies long grid-connection waits, fragmented markets, supply-chain delays and regulatory bottlenecks among the reasons new carbon-free generation cannot always arrive quickly enough. [10]

That doesn’t mean the technology industry’s environmental ambitions have disappeared. It demonstrates just how difficult it is for physical infrastructure to keep pace with digital demand.

You can scale software extraordinarily quickly. Power stations, transmission networks and substations operate on rather different timescales.

A milestone is only as good as the evidence behind it

This brings me back to Planning and Project Controls.

It is easy to place an energisation milestone on a programme. The more important question is: what evidence supports the date?

• How mature is the grid connection?
• Have the necessary network studies been completed?
• What transmission reinforcement is required?
• Is land secured for the substation?
• Where are the transformers and switchgear in the supply chain?
• Are permits and approvals aligned with construction?
• When does temporary power become permanent power?
• What happens to commissioning if energisation moves three months?
• And what happens to the commercial case if operational capacity arrives six months later than planned?

These aren’t simply engineering questions. They are programme questions.

As AI infrastructure becomes larger and more interconnected, understanding the dependencies between different programmes becomes just as important as controlling the activities within an individual schedule.

And solving power may reveal the next constraint

There is another issue worth watching. Even where power can be secured, the infrastructure still has to be built, commissioned, operated and maintained.

In April, CBRE and Meta announced LevelUp, a multi-year programme designed to recruit and train thousands of fibre technicians to work on Meta data centre projects across the US.

Meta currently has 27 US data centres either operational or under construction, with more planned. CBRE says the programme is specifically intended to address a growing shortage of technicians capable of installing fibre, network equipment and other mission-critical infrastructure. [11]

It is a relatively small part of the wider story, but an important one.

The constraint can move: first land, then power, then equipment, then skilled construction resource, then commissioning, then the technical workforce needed to operate the facility.

Removing one bottleneck doesn’t necessarily remove the critical path. It may simply move it somewhere else.

The organisations that understand the whole programme will have an advantage

The enormous capital flowing into AI infrastructure makes speed important. But speed without programme certainty creates a different set of risks.

A data centre completed before its power infrastructure is ready is not operational capacity. Neither is a powered building whose commissioning programme has slipped, whose critical equipment hasn’t arrived or whose operational workforce isn’t ready.

That is why I believe Planning, Scheduling and Project Controls will become increasingly strategic as the AI infrastructure build-out continues.

The challenge isn’t simply to build faster. It is to understand how power, grid infrastructure, construction, equipment, people, energisation and commissioning converge on the date when an asset actually becomes productive.

AI may be advancing at extraordinary speed. The infrastructure supporting it still has to obey the realities of engineering, construction and time.

And increasingly, that may be where the real critical path lies.

About Osirian Consulting

For more than 20 years, Osirian Consulting has specialised in providing experienced Planning, Scheduling and Project Controls professionals for complex capital projects.

As our work expands into the US data centre and AI infrastructure market, we’re increasingly interested in the challenges facing the people actually delivering these programmes, from power availability and grid connections through construction, commissioning and operational readiness.

I’d be interested to hear from those working in the sector: where do you see the biggest programme constraint today? Is it still power, or is the critical path already moving somewhere else?

Sources

[1] US Energy Information Administration (EIA), electricity demand outlook: https://www.eia.gov/todayinenergy/detail.php?id=67344

[2] International Energy Agency, Energy and AI: Energy supply for AI: https://www.iea.org/reports/energy-and-ai/energy-supply-for-ai

[3] Microsoft, Pecos data centre announcement: https://blogs.microsoft.com/blog/2026/06/22/powering-the-next-wave-of-ai-expanding-capacity-with-our-new-datacenter-in-pecos/

[4] Microsoft, The yield imperative: https://blogs.microsoft.com/blog/2026/09/01/the-yield-imperative-turning-ai-infrastructure-into-useful-intelligence/

[5] ERCOT, large-load connection process: https://www.ercot.com/news/release/06182026-puct-approves-ercots

[6] ERCOT market notice, August 2026: https://www.ercot.com/services/comm/mkt_notices/M-A080326-01

[7] ERCOT 2025 Annual Report: https://www.ercot.com/files/docs/2026/03/19/2025-ERCOT-Annual-Report-Final-Single-Pages-March-19-2026.pdf

[8] Meta, nuclear energy projects: https://about.fb.com/news/2026/01/meta-nuclear-energy-projects-power-american-ai-leadership/

[9] Meta, Louisiana data centre expansion: https://about.fb.com/news/2026/07/teachers-local-businesses-win-as-meta-expands-louisiana-data-center/

[10] Google, 2026 Environmental Report: https://blog.google/company-news/outreach-and-initiatives/sustainability/2026-environmental-report/

[11] CBRE, Meta and CBRE announce LevelUp: https://www.cbre.com/press-releases/meta-and-cbre-announce-levelup