What the data center power crunch is — why Nvidia is buying into power developers
The data center power crunch is the bottleneck created when AI compute demand grows faster than electrical grid capacity can be added. A data center building takes roughly one to two years to construct, but delivering power to it can take longer, because interconnection queues, substation construction and transmission line permitting all run on multi-year timelines. As a result the binding constraint on AI expansion has moved from chips to electricity. In August 2026 Nvidia took a minority stake in Cloverleaf Infrastructure and committed 1.5 billion dollars to SoftBank-owned SB Energy
The three lines
- Bottleneck — Buildings take one to two years. Power takes longer. The constraint moved from chips to electricity
- Cause — Interconnection queues, substations and transmission all run in years, and money cannot buy queue position
- Response — Nvidia took a stake in Cloverleaf Infrastructure and committed 1.5 billion dollars to SB Energy
Key questions
- Why is there a data center power shortage?
- **Because construction and electrification run on different clocks.** The building itself generally goes up in one to two years. The power does not. A large AI data center draws anywhere from several hundred megawatts to gigawatt scale, and connecting that much new load requires **queuing for grid interconnection**, **building a substation**, and **securing transmission capacity** — each a multi-year process. Interconnection queues are the hardest of the three, because in many jurisdictions studies proceed **in the order applications were filed**, so paying more does not move you up. Transmission adds another layer, since every parcel of land a line crosses requires permits and local consent, and one blocked segment stalls the whole route. The result is a situation where **the chips are available but there is nowhere to plug them in.**
- Why would a chip company invest in power developers?
- **Because its product cannot ship into a site that has no electricity.** On August 21, 2026, Nvidia disclosed a **minority stake in Cloverleaf Infrastructure**; the Wall Street Journal reported the investment at several hundred million dollars. Cloverleaf, founded in 2024, works with **utilities, energy providers and investors to secure power and supporting infrastructure for data center sites**, and says it has delivered multiple gigawatt-scale projects across North America. The disclosure came days after Nvidia said it would invest **1.5 billion dollars in SB Energy**, a data center developer owned by SoftBank. The logic is straightforward: accelerator demand is irrelevant if there is no power to run it, so **capital follows the bottleneck**. Cloverleaf will deploy Nvidia's DSX platform to optimize site selection, power, cooling and compute decisions.
- Can this not be solved by generating more electricity?
- **Generating and delivering are different problems.** Adding generation capacity does nothing if there is no transmission to move the electricity to the data center, and transmission often takes longer to build than generation does. So the industry's emerging answer is to **bypass transmission** by siting generation next to the load. This publication has covered examples: Naver's **55 megawatt** facility in Sejong, built at a cost of 14.6 trillion won, and Anthropic's arrangement to absorb the electricity rate increases attributable to its demand rather than building data centers itself. That second case points at a live political question — when a data center raises regional power demand, **who pays the resulting increase for everyone else**. Co-location has its own limits too: it requires land, fuel and cooling water, and it leaves grid stability and local rates unresolved.
- What does this mean for chip suppliers in Asia?
- **It puts a grid-shaped timing risk under the memory cycle.** If power constraints delay data center construction, the demand for the AI accelerators and high bandwidth memory that go inside is delayed with it. SK Hynix and Samsung Electronics shipment plans end up indirectly tied to interconnection schedules on another continent. That matters right now because Nvidia guided on August 27, 2026 to roughly **70 percent revenue growth** next fiscal year, and Korean outlets read that as the basis for a memory upcycle — but that growth has to find somewhere to plug in first. Korea faces the domestic version of the same problem: this publication reported peak demand of **95.3 gigawatts**, within **1.8 gigawatts** of the all-time record, and large data centers add to that. **When reading AI investment news, the number to look for after the chip name is megawatts.**
The data center power crunch is the bottleneck created when AI compute demand grows faster than grid capacity can be added.
In one sentence: the building takes one to two years, and getting electricity to it takes longer.
That gap has moved the binding constraint on AI expansion from chips to electricity.
1. Why the power takes longer
| Stage | Work | Rough timeline |
|---|---|---|
| Building | Site preparation, construction | 1-2 years |
| Interconnection queue | Waiting for grid connection studies | Years |
| Substation | Stepping high voltage down to usable levels | Years |
| Transmission | Permitting, land, local consent | Years |
The point is that the bottom three do not finish in parallel with the building.
Interconnection queues are the worst of them. In many jurisdictions, studies proceed strictly in filing order. Paying more does not move you up. This is how a company with abundant capital ends up unable to accelerate its own schedule.
Transmission adds a further difficulty. Every parcel a line crosses needs permits and local consent, and one blocked segment stalls the entire route.
2. Scale is what broke it
| Facility | Power draw |
|---|---|
| Conventional commercial data center | Single to double-digit megawatts |
| Large AI data center | Hundreds of megawatts to gigawatt scale |
This publication covered why the unit itself changed in "What AI data center power is — why it is counted in gigawatts, not megawatts."
When the unit shifts by three orders of magnitude, you cannot simply attach the load to the existing grid. New substations and new transmission become necessary, and at that moment the schedule stretches into years.
The result is one line: the chips are available, and there is nowhere to plug them in.
3. Nvidia's response — a chip company investing in electricity
Capital follows the bottleneck.
| Date | Target | Detail |
|---|---|---|
| August 21, 2026 | Cloverleaf Infrastructure | Minority stake (WSJ: several hundred million dollars) |
| Days earlier | SB Energy (SoftBank-owned) | 1.5 billion dollars |
Cloverleaf Infrastructure, founded in 2024, works with utilities, energy providers and investors to secure power and supporting infrastructure for data center sites. It says it has delivered multiple gigawatt-scale projects across North America since founding.
Under the partnership, Cloverleaf will deploy Nvidia's DSX platform to optimize decisions on site selection, power, cooling and compute infrastructure.
The reasoning:
| Step | Detail |
|---|---|
| 1 | Nvidia sells accelerators |
| 2 | Accelerators only sell if they go into data centers |
| 3 | Data centers only exist where there is power |
| 4 | So invest in power |
This publication covered the financing side of the same expansion in "Nvidia's $500 billion compute financing platform — the GPU became collateral" and the model side in "Nvidia's $6 billion Poolside deal." A chip company extending in both directions along its own supply chain.
4. Why more generation is not the answer
| Problem | Fixed by building generation? |
|---|---|
| Insufficient generation capacity | Yes |
| Insufficient transmission | No |
Electricity that cannot reach the data center is not useful, and transmission frequently takes longer to build than generation.
So the emerging industry answer is to bypass transmission — site the generation next to the load.
Examples this publication has covered:
| Case | Detail |
|---|---|
| Naver, Sejong | 14.6 trillion won, a 55 megawatt owned facility |
| Anthropic | Does not build data centers; absorbs the electricity rate increases attributable to its demand |
The second is the more interesting one. When a data center raises regional power demand, local rates can rise for everyone else — and who pays that has become a separate political question.
Co-location has limits too. It needs land, fuel and cooling water, and it leaves grid stability and local rates unresolved.
5. The Asian angle
Demand side — memory shipment timing
If power constraints delay data center construction, demand for the accelerators and high bandwidth memory inside them is delayed with it. SK Hynix and Samsung Electronics shipment plans become indirectly tied to interconnection schedules on another continent.
This is live right now. Nvidia guided on August 27, 2026 to roughly 70 percent revenue growth next fiscal year, and Korean outlets read it as the basis for a memory upcycle. That growth still has to find somewhere to plug in.
Supply side — Korea's own grid
Korea faces the domestic version. This publication reported peak demand of 95.3 gigawatts, within 1.8 gigawatts of the all-time record. Large data centers add to that.
The ongoing debate about concentrating data centers around Seoul versus dispersing them is, at bottom, a grid question. Covered in "What a power reserve margin is — at what percentage does a blackout start."
6. How to read AI investment announcements
| Figure quoted | What it tells you |
|---|---|
| Investment size (dollars) | Intent |
| GPU count | Plan |
| Megawatts / gigawatts | When it actually runs |
After the chip name, look for megawatts. A data center announced without secured power is an announcement with schedule risk attached.
7. Related coverage
- "What AI data center power is — why it is counted in gigawatts, not megawatts" — why the unit changed
- "What a power reserve margin is" — how much slack a grid has
- "Sejong's 55 megawatts — Naver spends 14.6 trillion won on a sovereign AI factory"
- "Anthropic does not build its own data centers" — who absorbs the rate increase
- "Nvidia's $500 billion compute financing platform" — the capital-side bottleneck
8. What we could not confirm
- Investment size — Cloverleaf terms were not disclosed; "several hundred million" is WSJ reporting relayed secondhand.
- Timelines — interconnection, substation and transmission durations are described as multi-year; these vary widely and no specific statistic is cited.
- SB Energy — timing and conditions of the 1.5 billion dollar commitment.
- Industry figures — a claim that North American data center construction rose 69 percent year on year could not be traced to an original source, so it is not used above.
- Korean data — current interconnection queue conditions and regional concentration figures were not verified.
- DSX — the platform's specific capabilities were not confirmed against Nvidia primary materials.
Sources
- TechCrunch — Nvidia partners with data center developer Cloverleaf
- PR Newswire — Cloverleaf Infrastructure Forms Strategic Partnership with NVIDIA to Accelerate Data Center Infrastructure Development
- BNN Bloomberg — Nvidia invests in data center developer Cloverleaf Infrastructure
- Yahoo Finance — Nvidia in advanced talks to invest hundreds of millions in Cloverleaf Infrastructure, WSJ reports
- AIwire — Cloverleaf Partners with NVIDIA to Accelerate Data Center Infrastructure Development
- Hankyung — Nvidia invests in a data center power company to secure AI infrastructure