Trends · Infrastructure

AI Infrastructure Trends in 2026: Power, Neoclouds, and the Pull Toward Control

Every January the infrastructure press declares a theme for the year. For 2026 the theme wrote itself: the AI buildout stopped being a software story and became a power, real-estate, and supply-chain story. If you run servers for a living — as we do in Toronto — the practical version of that shift is simpler still. The questions clients ask have changed. A year ago they asked which GPU. Now they ask whether anyone can give them power, a delivery date, and a clear answer on where their data physically sits.

This guide walks through the trends that matter for people who actually have to make placement and procurement decisions, not just read forecasts. It is deliberately even-handed: some of these trends point toward the hyperscalers, some toward specialised providers, and some toward owning your own iron. We will say plainly which is which, including where a Toronto bare-metal shop like ours is the wrong tool for the job.

The shape of the 2026 buildout

The headline numbers are large enough to lose meaning, so anchor them. Data-centre electricity use was roughly 415 TWh in 2024 — about 1.5% of world consumption, per IEA figures cited by Brookings — and several 2026 estimates put it on a path toward 1,000 TWh or more. The growth rate, not the absolute level, is the point: data-centre demand has been compounding several times faster than overall electricity demand for most of a decade, and AI is the accelerant.

Underneath that demand curve, the unit of design changed. Operators at Data Center World 2026 described campuses being engineered as single integrated systems rather than rows of general-purpose racks, with power, cooling, and network fabric co-designed for two very different patterns — tightly-coupled training clusters and distributed, always-on inference. That distinction matters for buyers because the two patterns reward different infrastructure, and most organisations only ever touch one of them.

Where the 2026 capex is going

The spending is concentrated in a way that is easy to underestimate. Analysts expect the largest cloud companies to commit over $600 billion in capital expenditure in 2026, up roughly 36% year over year, with the bulk earmarked for AI infrastructure and NVIDIA capturing the lion’s share of accelerator spend. That capital is increasingly arriving earlier in the project life cycle, with GPU-collateralised financings and “powered land” deals — a sign the market now treats GPUs and grid interconnections as long-lived physical assets rather than IT line items.

For everyone outside that top tier, the more useful read is the second-order effect. When the giants vacuum up GPUs, power contracts, and construction crews, lead times stretch for everyone else. The competitive edge in 2026 rarely comes from having the newest chip; it comes from having capacity that is actually available, on a date you can plan around.

Why did power become the binding constraint?

For most of the cloud era, the limiting resource was the chip. In 2026 it is the electron. Building a new hyperscale facility takes three to five years of permitting, construction, and grid interconnection; in some North American markets the interconnection queue alone has stretched past eight years. The IEA’s 2025 analysis put roughly a fifth of planned projects at risk of significant delay from grid congestion. Capital is plentiful; deliverable megawatts are not.

Density is the other half of the story. A single rack of eight H100-class accelerators draws several kilowatts before cooling and networking; the newest GB200 NVL72 systems land near 140 kW per rack. At that level air cooling simply cannot keep up — direct-to-chip liquid cooling moves from “nice to have” to a hard requirement, with coolant distribution units, leak detection, and compatible rack designs baked into the build. The knock-on effect for buyers is that “a GPU server” is no longer a commodity you slot into any cabinet; it is a thermal and electrical commitment the facility has to be designed around.

If you want the constraint in concrete terms, do the arithmetic a facility engineer does. A back-of-envelope rack power check turns “we want eight GPU nodes” into a number a data centre can actually answer yes or no to:

# Per-GPU power draw on a live node
$ nvidia-smi —query-gpu=power.draw,power.limit —format=csv
power.draw [W], power.limit [W]
698.42 W, 700.00 W
701.10 W, 700.00 W
 
# Sanity-check a full rack before asking a DC for power
$ GPUS_PER_NODE=8; NODES=4; W_PER_GPU=700
$ OVERHEAD=1.35   # CPU, NICs, fans, PSU loss, cooling headroom
$ echo ”$(( GPUS_PER_NODE * NODES * W_PER_GPU )) W GPU-only”
22400 W GPU-only
$ python3 -c “print(round(22400*1.35/1000,1),‘kW rack draw’)”
30.2 kW rack draw
# 30 kW in one cabinet = liquid-cooling territory, not a stock rack.

That single number — tens of kilowatts in one cabinet — is why a GPU server stopped being a commodity you slot in anywhere. The facility has to be engineered around the heat and the draw, which is precisely why power and cooling gate so many 2026 deployments.

There is a quieter trend hiding inside the power story. The industry has talked about GPUs for two years, but inference now accounts for the large majority of AI compute — by IEA’s reckoning, on track for around three-quarters of AI energy demand by the end of the decade — and agentic, multi-step workloads are pushing conventional CPU, memory, and networking demand up alongside it. The energy question is no longer “how much does one model cost to train”; it is “how much always-on capacity does production AI draw, and where.”

What is a neocloud, and should you care?

A neocloud is a cloud provider built almost entirely around renting out GPU compute — GPU-as-a-service — rather than the broad service catalogue a hyperscaler offers. The category exploded because hyperscalers could not build fast enough: providers that locked in power and sites before the 2023–2024 surge could install GPUs in six to eighteen months instead of waiting years for a new campus. Synergy Research put neocloud revenue above $23 billion in 2025, more than triple the prior year, and Gartner expects the category to hold roughly a fifth of a $267 billion AI-cloud market by 2030.

The appeal for buyers is price and focus. Specialised providers have offered GPU time well below hyperscaler list rates — some synthesis pegs the saving as high as two-thirds, with competitive hourly rates falling from around $8 toward under $2. If you have a defined GPU training or fine-tuning job, a neocloud is often the rational choice.

The caution is durability. McKinsey’s analysis is blunt about the economics: bare-metal-as-a-service margins sit near 14–16% once labour, power, and depreciation are counted, and returns flatten if utilisation drops below roughly 80%. Many providers secured chips before they built mature operations, which makes some structurally fragile. A wave of enterprise contracts signed in 2024–2025 comes up for renewal in late 2026, and the renewal criteria shift toward production concerns — sovereignty, resilience, financial stability of the provider itself. The honest summary: neoclouds are excellent for bursty GPU demand and risky as the sole home for anything mission-critical.

How real is the cloud-repatriation trend?

Repatriation is real, but it is selective rather than a wholesale exodus. The Barclays CIO study reported that 86% of surveyed technology leaders plan to repatriate at least some public-cloud workloads — the operative words being “at least some.” Nobody serious is unplugging from the cloud entirely; they are moving the specific workloads where the cloud’s economics or control model stopped making sense, and keeping the rest.

Two things make a workload a repatriation candidate. The first is predictable, steady-state demand: paying elastic cloud premiums for capacity you use 24/7 is the classic case where owned or single-tenant bare metal wins on cost. The second is a need for control — over data location, over the hardware, over a sending reputation you cannot afford to share with noisy neighbours. This is the same logic that drives our email-infrastructure clients off shared platforms: when the asset is your deliverability, renting it inside someone else’s black box is the expensive option. We cover that decision in depth in why self-host email, but the principle generalises far beyond mail.

Where does data sovereignty fit in?

Data sovereignty has moved from a compliance footnote to a top-line infrastructure driver. Gartner projects sovereign-cloud IaaS spending near $80 billion in 2026 — up about 36% on the prior year — and coined “geopatriation” for the deliberate shift of roughly a fifth of workloads to local or regional providers for regulatory and geopolitical reasons. Deloitte’s 2026 survey found 83% of leaders consider sovereign AI strategically important. The driver is not just law; it is a reluctance to have critical data sit under a foreign legal regime.

For Canadian organisations and anyone serving them, this is where the map gets specific. A handful of providers run Canadian data centres, but most treat a Toronto or Montréal node as a generic point on a global map rather than a sovereignty position. The combination that stays largely unclaimed is Canadian data residency framed as a data-sovereignty guarantee — single-tenant, PIPEDA-aligned, and outside the reach of foreign data-access laws. That is the lane MCSNET deliberately occupies. It will not matter for a startup prototyping a model; it matters a great deal for a regulated business deciding where customer records and email-sending infrastructure physically live.

Bare metal and managed email in an AI-first stack

Here is the even-handed part. If your 2026 challenge is training a frontier model across tens of thousands of tightly-coupled GPUs, a Toronto bare-metal provider is the wrong answer — that work belongs with a hyperscaler or a large neocloud with the power contracts and fabric to match. We do not pretend otherwise, and any infrastructure guide that claims one venue wins every workload is selling something.

What bare metal does win is the large, unglamorous middle: steady-state production services, single-tenant GPU for modest inference and fine-tuning, regulated data that has to stay in-country, and latency-sensitive work that benefits from a fixed, known machine. You get full root, predictable monthly cost instead of metered surprises, and hardware nobody else is sharing. For teams that have outgrown a VPS but have no business running a hyperscale GPU farm, a Toronto dedicated server is frequently the right unit of compute.

The piece almost every “AI infrastructure” roundup omits is email. As organisations bring AI workloads in-house, they increasingly want to own their communication infrastructure too — the transactional and bulk sending that no AI vendor and almost no neocloud touches. Running your own PowerMTA server hosting on dedicated IPs, with disciplined IP warming, gives you control over deliverability in exactly the way repatriation gives you control over compute. It is the same trend wearing different clothes: when the asset is strategic, you stop renting it through someone else’s reputation. The 2026 sending rules from the major mailbox providers — covered in Gmail and Yahoo requirements 2026 — only sharpen the case for owning your sending stack rather than inheriting a shared one.

A planning checklist for the year ahead

If you are making infrastructure decisions in 2026, the trends above collapse into a short, practical sequence. Work it in order; most placement mistakes come from skipping straight to hardware.

1. Classify the workload before the venue. Frontier training, bursty GPU, steady-state production, regulated data, or sending infrastructure — each maps to a different tier. The placement map above is a starting point.
2. Check power and lead time, not just specs. Ask any provider for a real delivery date and the power/cooling profile of the rack. In a constrained year, availability beats benchmarks.
3. Price steady-state honestly. If a workload runs 24/7, model owned or single-tenant cost against elastic cloud over 24–36 months, not one month.
4. Make data location a requirement, not an afterthought. Decide which datasets must stay in-country, and treat sovereignty as a hard filter on the provider list.
5. Don’t single-source anything critical. Whether it is a neocloud or a hyperscaler, keep a tested failover for mission-critical inference and sending.
6. Own what’s strategic. If compute or deliverability is core to the business, default toward control — bare metal, dedicated IPs, your own MTA — and rent the rest.

The through-line of 2026 is not “the cloud is over” or “everyone is buying GPUs.” It is that the easy era — when you could put any workload anywhere and let elasticity sort it out — has ended, because power, density, and data law now impose real constraints. The teams that do well this year are the ones that place each workload deliberately. For the steady, regulated, in-country, and sending-heavy slice of that work, control-first infrastructure is having its moment — and that is the slice we build for.