The previous four articles: AI saves software labor (article 1), cost moves to consumer products (article 2), the loop is funded by B2B (article 3), enterprises stratify between cloud and on-prem (article 4). This article asks: AI is in a crazy expansion — what is the real bottleneck?

The previous four articles covered “what AI saves, where the cost moves, who captures, how enterprises choose.” This article covers “the physical bottleneck of AI expansion.” This is the series’ conclusion — from “saving” to “building,” the AI economic story is complete here.

This question is worth answering because if the bottleneck is electricity (grid, transformers), the “crazy expansion” is actually a paper boom — announced capacity ≠ actually built.

The main answer: the real bottleneck is electricity. GPUs are available (NVIDIA has capacity), but grid interconnection takes 5 years and transformers take 5 years. 2026: 12 GW announced, only 5 GW actually under construction (about one-third). The crazy expansion is partly a paper boom.

1. Question: what is the real bottleneck of AI expansion?

The conclusions of the previous four articles: AI saves software labor, cost moves to consumer products, the loop is funded by B2B, enterprises stratify between cloud and on-prem.

This article asks: AI is in a crazy expansion — what is the real bottleneck? GPUs? Memory? Electricity?

This article only handles “the bottleneck of expansion”; it does not handle “whether AI is actually useful” (handled in the first article).

2. Evaluation framework: the bottleneck of expansion depends on “can buy vs can connect”

The bottleneck of expansion depends on two things: hardware can be bought (GPUs, memory) and electricity can be connected (grid, transformers).

Hardware can be bought: NVIDIA has capacity, memory has supply (though expensive, article 2). GPUs and memory are a “can buy” problem — if you have the money, you can buy them.

Electricity can be connected: grid interconnection, transformers, switchgear — these have long lead times. Grid interconnection takes 5 years, transformers take 5 years. These are a “can connect” problem — even with the money, you can’t connect, because the lead times are long.

Why are the lead times so long? Because transformers are highly specialized equipment; there are few manufacturers in the world who can make large transformers, and capacity expansion takes time. Grid interconnection also has to wait for the utility’s scheduling — grid construction is a “public good” that a single enterprise can’t accelerate.

So “if you have the money, you can build” is wrong. GPUs and memory are a “money problem”; electricity is a “time problem.” Money can speed up purchasing, but it can’t buy time.

So: the bottleneck is not “can buy,” it is “can connect.”

A concrete example: a data center needs 100 MW of electricity. GPUs and memory can be bought (NVIDIA has capacity, memory has supply), but to get 100 MW of electricity, you have to wait for grid interconnection (5 years) and transformers (5 years). So this data center takes 5 years to build. So the bottleneck of AI expansion is “time,” not “money.”

3. Comparison: announced capacity vs actually under construction

2026:

MetricValueSource
2026 announced capacity12 GWTech Fund[1]
2026 actually under construction5 GW (about one-third)Tech Fund[1]
Transformer lead time (before 2020)24–30 monthsInside Deep Tech[2]
Transformer lead time (now)5 yearsInside Deep Tech[2]
Projects with no disclosed power strategy25%Tech Fund[1]
Build time of stalled projects12–18 monthsTech-Insider[3]

Three observations. First, announced vs under construction: 12 GW announced, 5 GW actually under construction, only about one-third. Second, transformer lead time has surged: 24–30 months before 2020, now 5 years. Third, the stalled projects only need 12–18 months of construction, but they are frozen at the announcement stage (because of grid interconnection).

Note: 5 GW actually under construction does not mean “only 5 GW built.” It means projects “with a power strategy, grid interconnection, and transformers.” The other 7 GW are projects “announced but without a power strategy” — they have to wait for electricity.

These three observations point to the same thing: the bottleneck of AI expansion is not “can buy” but “can connect.”

So the reality of the “crazy expansion” is: announced, but can’t be built (because of electricity). This also explains why there are so many “announcements” — because “announcing” doesn’t need electricity, “building” does.

4. Judgment: the crazy expansion is actually a paper boom

GPUs are available, but electricity can’t be connected. NVIDIA has capacity, memory has supply, but grid interconnection takes 5 years and transformers take 5 years.

Of the 12 GW announced, only 5 GW (about one-third) is actually under construction. So the “crazy expansion” is partly a paper boom — announced, but can’t be built (because of electricity).

For the loop in article 3, this means: the electricity bottleneck slows data center construction, slows AI deployment, slows cloud revenue. So the loop’s “win” has to wait for electricity.

A deeper layer: the electricity bottleneck is a “physical bottleneck” — unlike GPUs (which can ramp production) or memory (which can raise prices), it has to wait for grid construction and transformer production. So the electricity bottleneck is the “hard constraint” on AI expansion.

For investors, this means: the “announcement” of AI expansion should be discounted — 12 GW announced, only 5 GW actually built. So to watch AI expansion, watch “projects with a power strategy,” not “announced projects.”

So the “hard constraint” of AI expansion is electricity. Until electricity is solved, announced capacity will always exceed actually built. For the loop in article 3, this means: the loop’s “win” has to wait for electricity. For the enterprise on-prem in article 4, this means: the on-prem data centers also have to wait for electricity. So whether cloud or on-prem, both have to wait for electricity.

For the whole series, this means: the AI economic story ends up stuck at the “physical bottleneck.” Article 1 saves software labor (virtual), article 2 moves cost to memory (physical components), article 5 is stuck at electricity (physical infrastructure). From virtual to physical, AI’s bottleneck gets harder and harder.

Applicable conditions: grid construction lag, long transformer lead times.

Limitations:

  • The electricity bottleneck will ease over time (grid construction, transformer capacity expansion). But easing takes time (5 years).
  • Some projects can bypass the grid with “self-generated power” (natural gas, nuclear). But that also has a lead time.

Ending

Back to the question: what is the real bottleneck of AI expansion? The answer: electricity. GPUs are available, but grid interconnection takes 5 years and transformers take 5 years. The crazy expansion is partly a paper boom — 12 GW announced, only 5 GW actually under construction.

Applicable conditions: grid construction lag, long transformer lead times. Limitations: the electricity bottleneck will ease over time; some projects can bypass with self-generated power.

This is the last article of the “Who pays for AI” series. Five-article recap: article 1 AI saves software labor (virtual layer), article 2 cost moves to consumer products (component layer), article 3 the loop is funded by B2B (capital layer), article 4 enterprises stratify between cloud and on-prem (application layer), article 5 the real bottleneck is electricity (physical layer). From virtual to physical, the AI economic story is complete. So the AI economic story ends up stuck at the physical bottleneck.

For readers, the practical conclusion of this series is: to watch AI, don’t just watch “how impressive it is”; watch “what it saves, where the cost moves, who captures, how enterprises choose, where the physical bottleneck is.”