Key Takeaways
Neocloud capacity gets quoted as a single headline figure, but it’s really three: contracted power, deployed megawatts, and available GPU hours. Only the last one is something you can rent this quarter.
The Gap Is Documentable: The largest listed providers now report roughly 4.2 GW of contracted power against about 1.5 GW active.
Backlog Isn’t Inventory: A capacity backlog measured in tens of billions describes future obligations, not present GPU availability. Compute procurement that treats a backlog as inventory will book capacity it can’t actually schedule against.
Delivery Slips Are Structural: As much as half the data center capacity due online in 2026 could be delayed by permitting and grid connection. Any AI capacity planning built on announced dates should carry that as a base case.
Delivered Beats Contracted: Aethir has 2+ billion compute hours delivered across 430,000 GPU containers in 94 countries. Compute hours delivered is a backward-looking number, which is precisely why it’s harder to inflate than a pipeline figure.
What Neocloud Capacity Announcements Actually Measure
Neocloud capacity is the headline metric of the AI infrastructure cycle, and it’s also the least well-defined. A press release can describe power a provider has an option on, power under contract, racks physically installed, or GPUs a customer can rent on Tuesday, and all four get called capacity.
The market itself is growing fast enough for the ambiguity to matter: AI-optimized IaaS reached $42.276 billion in 2026, and the $5.2 trillion cost of compute through 2030 sets the scale of what’s being promised. Buyers who read one number as another end up surprised, the same way they do when GPU cloud pricing moves under a signed plan.
Contracted power is megawatts a provider has secured on paper from utilities or landlords, often for delivery years out. It’s a real commitment and a genuine competitive asset, but a workload can't run on it today.
Deployed megawatts describe capacity that’s energized, racked, and operating. The gap between contracted power and deployed megawatts is where most neocloud capacity confusion lives, and it’s usually measured in years, not months.
Even deployed megawatts overstate what you can buy, because most of that capacity is already committed to existing customers. Available GPU hours, the genuinely sellable slice, is the smallest of the three numbers and the only one that answers your question.
Why Contracted Power Isn’t GPU Availability
The gap is now large enough to read off public filings. The largest listed providers report roughly 4.2 GW of contracted power against about 1.5 GW active, a ratio close to 3:1. Still, contracted power and GPU availability are different currencies. Research on colocation and the infrastructure race behind AI suggests that up to 50% of the capacity due online in 2026 could be delayed, widening the gap further.
Customers sign multi-year commitments to guarantee future access, turning demand into a capacity backlog long before the hardware exists. The backlog then gets quoted as evidence of strength, which it is, and as evidence of availability, which it isn’t.
Berkeley Lab counts more than 2,060 GW in US interconnection queues, with a median wait above five years for projects that reached operation in 2025. A capacity backlog can clear only as fast as the pipeline allows.
Hardware begins losing value the day it’s energized, not the day it’s contracted, which pushes operators to fill new deployed megawatts immediately. That’s the same clock behind what a used GPU is actually worth, and it explains why new capacity rarely stays available for long.
The Three Numbers That Drive AI Capacity Planning
With worldwide AI spending at $2.52 trillion in 2026, almost every provider has a growth story, so the differentiator is which of their numbers you’re allowed to schedule against. Ask for contracted power, deployed megawatts, and available GPU hours as three separate figures, then plan only against the third.
Turning Deployed Megawatts Into Available GPU Hours
Apply a Committed-Capacity Filter: Ask what share of deployed megawatts is already contracted to named customers. The remainder is your real pool, and in a tight market that remainder is often a small fraction of the headline neocloud capacity.
Convert Power to Hours: A megawatt supports a finite number of accelerators once cooling and overhead are counted, and each accelerator yields a finite number of available GPU hours per month. That conversion turns a power figure into something AI capacity planning can actually use.
Check Where the Capacity Sits: Available GPU hours in the wrong region are unavailable in practice for a workload with residency or latency requirements. Reporting on grid flexibility and distributed inference notes US grid operators use only about 53% of generation capacity on average, so the headroom exists but rarely where the campuses are.
How Delivered Compute Hours Change Enterprise GPU Access
There’s one figure a provider can’t book in advance: What it has already served. Delivered compute hours are retrospective, auditable against billing, and immune to the optimism that inflates a pipeline. Aethir reports 2+ billion compute hours delivered to enterprise clients across 430,000 GPU containers in more than 200 locations and 94 countries.
Coverage of the $500 billion AI infrastructure financing push shows how much capital the build-side answer now absorbs, while the decentralized GPU advantage for inference describes the aggregation-side alternative.
Aethir and GPU Availability Without a Reservation
Enterprise GPU Access on Demand: Enterprise GPU access on the Aethir network doesn’t require a reservation sized to peak demand, so capacity flexes with request volume. That removes the main reason buyers overcommit, out of fear of losing GPU availability during a crunch.
Aggregation Instead of Construction: Capacity is onboarded from Cloud Hosts, operators who already hold power and connectivity, which is why the network grows without a construction cycle. Work on expanding data center capacity to meet demand makes the same point from the utility side: the fastest megawatt is one that already exists.
Additional compute capacity is incoming: To supplement our decentralized GPU cloud network, Aethir has launched the ACCELERATE program, which plans to build multiple mid-sized data centers, specifically for high-performance AI compute, focused on B300 and GB300 clusters. We plan to add capacity in regions where Aethir’s network already has demand, meaning we’ll build data centers with customers already secured.
A Compute Procurement Checklist for Neocloud Capacity
Reporting on how AI data center deals get financed makes clear that capacity announcements serve capital markets as much as customers, and work on the race to power AI explains why delivery lags the announcement.
Three Questions Before You Sign for AI Capacity Planning
What Is Live Today?
Ask for deployed megawatts and available GPU hours as of this month. If a provider can’t split those figures apart, the split is probably unflattering.
What Is Already Committed?
Ask what percentage of current neocloud capacity is contracted to existing customers. A high number is a healthy business and a thin pool, and your compute procurement needs to know which one it’s buying into.
What Happens if Delivery Slips?
Ask what the contract does when new capacity arrives late, since delay is the base case. Credits, exit rights, and a second source turn a slip into an inconvenience instead of a stalled roadmap.
Reading neocloud capacity properly is a small discipline with a big payoff: three numbers instead of one, and a plan built only on what you can schedule. Providers that can show delivered compute hours give AI capacity planning something firm to stand on. Buyers who ask for the split get shorter lead times, fewer surprises, and a credible second source when their first choice runs tight. That’s a better position than any headline figure can buy.
Frequently Asked Questions
What is neocloud capacity, and how is it measured?
Neocloud capacity describes the GPU compute a specialist AI cloud provider can offer, usually quoted in megawatts of power or numbers of accelerators. It’s measured at three very different stages, and only the last of them, capacity that’s energized and unsold, is something a buyer can rent now.
What does contracted power mean on a neocloud earnings call?
Contracted power is electrical capacity a provider has secured under agreement, often for delivery in future years. It signals ambition and financing strength, but it isn’t deployed megawatts, and it doesn’t translate into GPU availability this quarter.
How do I check real GPU availability before signing?
Ask for the current unsold pool in the specific region you need, expressed in accelerators or GPU hours, with a start date attached. If the answer arrives only as a company-wide power figure, GPU availability hasn’t been answered.
How many available GPU hours does a megawatt produce?
It depends on accelerator power draw, cooling overhead, and utilization, so the conversion varies by provider. The useful exercise is to ask the provider to do the conversion, because available GPU hours is the unit your workload actually consumes.
Does compute procurement change if capacity is already delivered?
Yes, substantially. When capacity is already live, compute procurement stops being a bet on a construction schedule. It becomes a straightforward sourcing decision with a start date, a price, and an exit.
