Key Takeaways
Six states of GPU capacity: Announced, financed, permitted, under construction, energized, and rentable describe entirely different degrees of certainty. The industry uses “online” for all of them, which makes almost every AI capacity announcement unreadable without translation.
The gap between the states is measured in orders of magnitude: Developers have announced roughly 190 gigawatts of future capacity since 2024, of which about 16 gigawatts were scheduled for 2026, and only about 5 gigawatts were under construction. Announced capacity and buildable capacity aren’t the same category.
Grid operators already publish the category differentiator: ERCOT sorts large loads into studies submitted, studies approved, approval to energize, and observed energized, and reports the numbers separately.
Only two states are worth anything to a buyer: Operational capacity tells you what exists, and rentable capacity tells you what you can actually run on this quarter.
What Announced Compute Capacity Means
Since 2024, developers have announced approximately 190 gigawatts of future data center capacity. About 16 gigawatts of it was scheduled to begin operations during 2026, and only about 5 gigawatts was actually under construction. Those three numbers describe the same industry, and each one supports a completely different story about how much compute exists.
This is why an AI capacity announcement is so hard to price. The word “online” gets applied to a signed land option, a financed project, a permitted shell, an energized building, and a cluster somebody can rent this afternoon. The states differ by years of elapsed time and by very large probabilities of never happening at all.
Forecasters have to make the same distinction, which is why revisions arrive in both directions. BloombergNEF raised its 2030 installed United States forecast to 118 gigawatts, a 52% upward revision, largely because the announced pipeline grew by a further 101 gigawatts in seven months.
What the Capacity Pipeline Actually Counts
Most entries in a capacity pipeline represent an intention that a developer is free to abandon at no cost. Reading one as a queue of delayed projects sets an expectation that was never justified in the first place.
Early-stage entries drop out routinely because they never had land control, financing, or a customer. Our analysis of why compute demand keeps outrunning supply covers what happens to buyers who plan against the wrong number.
The same site frequently appears in a developer announcement, a utility queue, and a partner press release. Aggregating across those sources inflates the total without anybody intending to mislead.
Six States Hiding Inside One AI Capacity Announcement
Announced means somebody said it. Financed means capital is committed. Permitted means the local approvals are in hand. Under construction means steel and concrete are moving. Energized means the facility can legally draw power. Rentable means a customer can put a workload on it.
Each step drops a large share of the previous one, and the drop is unevenly distributed. Permitting and grid connection do most of the killing, and equipment does the rest: lead times for generators, chillers, and switchgear have more than doubled since 2019, with some past three years, which means a financed and permitted project can still sit dark waiting on a transformer.
Analysts who model this professionally warn against this exact error, arguing that treating gigawatt-scale announcements as ground truth produces forecasts that are confidently wrong in both directions. That’s why customers should know which state a number describes.
Why Permitting Status Moves the Date Most
A project either has its approvals, or it doesn’t, and in most jurisdictions the record is searchable. That makes permitting status the easiest thing to verify in any announcement, and the one most often left out.
Objections tend to surface after a project is already announced and financed, which is exactly when it looks most certain. Aethir has argued that delivery speed rather than scale is the scarce asset precisely because of how late this risk lands.
Megawatts announced tell you about ambition and megawatts permitted tell you about the next eighteen months.
Approval to Energize Is a Published Milestone
The most useful discovery for anyone reading capacity claims is that grid operators already publish a version of this taxonomy, with numbers attached. ERCOT sorts large-load projects into studies not submitted, under review, planning studies approved, approved to energize but not operational, and observed energized, and it reports each category separately every month.
The resulting figures are stark. Against a large-load queue running into the hundreds of gigawatts, roughly 9,042 megawatts had received approval to energize. In comparison, the highest non-simultaneous peak those loads have actually been observed drawing was about 3,883 megawatts in ERCOT reporting. Approval to energize and drawing power are two different states, and even they diverge by more than half.
The same pattern shows up wherever power is the binding constraint, which is now most data center sites. Aethir has already covered why the grid became the bottleneck for AI compute and why connection timing, rather than construction, sets most delivery dates.
Energized Capacity Versus What Gets Drawn
A facility with approval to energize is cleared to consume up to a stated limit. Whether it draws anything close to that limit depends on how much of the building is fitted out and how many customers have moved in.
Large loads rarely peak together, so the sum of individual maximums always exceeds what the system has ever served at one moment. Aggregating nameplate figures across sites overstates real consumption by a wide margin.
Even a fully energized facility fills gradually as racks land and customers migrate. Treating an energization date as a service date builds a scheduling error straight into the roadmap.
Operational GPU Capacity Is the Only Honest Denominator
Almost every alarming or reassuring statistic about AI infrastructure depends on which denominator is used. Compare committed spending against announced capacity and the sector looks over-invested. Compare it against operational capacity, and it looks desperately short. Both calculations are arithmetically correct.
Financing makes the choice consequential. Research from Columbia Business School on financing the AI buildout sets out how much of the sector runs on leverage underwritten against revenue that starts on a specific date. Projects underwritten against announced capacity and delivered as operational capacity two years later face a financing problem before an engineering one.
The supply side has its own denominator problem. Meeting forecast demand requires annual grid investment to rise sharply from its current base, according to the IEA electricity analysis. That investment is sized against expected load. Aethir has written about how the resulting scarcity keeps driving up GPU cloud pricing.
Construction Start Is the First Real GPU Capacity Signal
Announcements, options, and even permits can be walked away from. A construction start means long-lead equipment has been ordered and money is being spent that nobody recovers by canceling.
Site work is visible, permits are filed, and equipment orders show up in supplier disclosures. This is the earliest point in the ladder where a third party can verify a claim without relying on what the developer says.
The ratio of capacity under construction to capacity announced is the single most informative number in the sector. When that ratio is roughly one in three, most of the headline figure describes intention.
Rentable GPU Capacity Is the Only State You Can Buy
A standard is worth nothing if the party proposing it exempts itself, so it’s worth stating the Aethir position against the same ladder. Aethir ACCELERATE has secured access to sites totaling up to 20 megawatts across the United States and Europe, with facilities engineered for NVIDIA B300 and GB300 clusters of 64 to 256 nodes and designed to go from site selection to live capacity in months.
Site control plus design, which sits between announced and under construction. It isn’t energized capacity, and it isn’t rentable capacity, and that's the point. The commercial target of up to $700 million in contracts by the end of 2026 is a contract figure, not a capacity figure, and the two belong on separate lines.
What is already rentable is the existing network. More than 430,000 GPU containers across 94 countries and 200+ locations are running today, which is a different kind of claim from a megawatt projection. With Gartner forecasting a 47% rise in AI spending in 2026, the distinction between capacity you can book and capacity somebody hopes to build is worth real money.
Nothing here requires trusting anybody. Ask which of the six states a number describes, ask what the construction ratio is behind it, and ask what you could book this week.
The vocabulary works on any AI capacity announcement, including this one.
Explore enterprise GPU compute with Aethir and start from capacity that is already rentable.
Frequently Asked Questions
What does an AI capacity announcement actually promise?
Most announcements describe an intention backed by land control or a letter of intent, with financing, permitting, grid connection, construction, and commissioning still ahead. The reliable move is to ask which of those steps is already complete rather than which megawatt figure was quoted.
What is the difference between announced capacity and operational capacity?
Announced capacity is everything a developer has said it intends to build, including projects with no financing or permits. Operational capacity is what has been built, energized, and put into service. Since 2024, roughly 190 gigawatts have been announced, while the capacity actually under construction for 2026 was closer to 5 gigawatts.
What is approval to energize?
Approval to energize is the clearance a grid operator grants once a large load has completed its interconnection studies and construction, allowing it to draw power up to a stated limit legally. It sits near the end of the ladder but still ahead of service. ERCOT publishes the figure separately from observed consumption, and the two differ by more than half.
What counts as rentable capacity?
Rentable capacity is capacity a customer can contract and run a workload on now, rather than capacity that has been announced, financed, permitted, or energized. It is the only state that affects a roadmap this quarter. When comparing providers, it is the first number worth asking for.
