AI Boom Self-Sabotages: Power Shortage Threatens Data Centers

The Paradox of the Three-Electric Boom

The South Korean “Sanjeonnic” boom—fueled by surging demand for AI chips, memory, and power infrastructure—is now running into its own shadow. A fresh warning from Morgan Stanley paints a picture where AI-driven power demand has become so voracious that it threatens to starve the very supply chains driving the boom.

The bank projects that US data center developers will face a 32-gigawatt net power shortfall by 2028. Even after factoring in on-site generation, fuel cells, and repurposed nuclear plant sites—measures estimated to close only about 24GW of the gap—a 34 percent deficit will remain. BlockSpace, a digital infrastructure firm, cited the report to say total additional power demand could reach 97GW by 2028, up sharply from a previous forecast of 68GW.

Construction in progress accounts for just 21GW. Available or contracted grid capacity adds another 19GW. The math leaves a shortfall of 57GW—widening significantly from the earlier estimate of 38GW.

Who Gets Left Behind

Morgan Stanley does not see this bottleneck threatening Nvidia or Broadcom’s earnings outlook next year. Both companies have the leverage to coordinate placement of their chips across whatever power is available. They can shuffle demand, prioritize accounts, and negotiate infrastructure timelines.

The vulnerability falls on the companies that sit behind the GPU: memory producers, optical communication firms, power management suppliers, and analog component makers. These are the sectors where order delays or cancellations will hit first.

“If chips cannot actually be deployed, customers may delay deliveries or cancel orders,” Morgan Stanley wrote, noting that memory and optical communication components face the greatest exposure to inventory disruption. The implication is clear: the AI revenue narrative is not universal. Chip revenue growth will track power availability, not just design interest.

The Memory Exposure

For Samsung Electronics and SK Hynix, the memory giants at the heart of Korea’s three-electric surge, the message is both indirect and sobering. Morgan Stanley did not revise earnings forecasts for either company directly. But it flagged the risk that memory order pipelines could stall precisely when demand signals have never looked stronger.

Memory accounts for an estimated 16 percent of the total accelerator-inclusive facility cost for Nvidia’s Rubicon Ultra system. That number grows larger as power constraints make it harder to place complete systems. A memory chip sitting in a warehouse is revenue unrecorded. A data center rack without enough power is a promise unfulfilled.

The trend in power density tells the story. Nvidia’s Vera Rubin rack design assumes 149 kilowatts per rack, climbing to 234kW. The Rubicon Ultra jumps from 415kW to 600kW per rack. These figures include GPUs, memory, networking, power supply, and liquid cooling combined. Higher density means fewer racks per megawatt of grid capacity. Fewer racks means memory shipments arrive slower than the order book suggests.

Efficiency’s Limits

Morgan Stanley expects power-per-token productivity to improve roughly sixfold between 2025 and 2028, and chip-level productivity to rise about fifteenfold. Those are remarkable gains. But efficiency alone will not contain total consumption. The bank’s base case holds that falling unit costs drive expanded AI usage, which pushes aggregate electricity demand higher rather than lower.

This is the classic Jevons paradox playing out at data center scale: improvements in efficiency expand the market faster than they shrink consumption. The AI buildout is not a fixed-cost project with a ceiling. It is a compounding one.

Political Friction

The power crunch is not purely a technical problem. It is becoming a political one.

Big tech companies have begun offering community support packages tied to energy costs and local infrastructure investments. Analysts see timing linked to the US midterm elections, with data centers turning into a campaign-issue flashpoint. Texas, the country’s largest data center market, has paused new permits while an audit runs its course. Virginia and Northern Virginia—the historic heartland of US data center development—now face multi-year interconnection queues that stretch well past 2030 in some substations.

When residents and local governments push back, grid connection timelines stretch. The bottleneck is no longer just about generation capacity. It is about the permission to plug in.

Second-Order Effects

The ripple effects of this constraint extend far beyond delayed shipments. Power scarcity is reshaping geographic concentration: developers are abandoning traditional hubs where queues are longest and moving toward regions with surplus generation—Luisiana, Oklahoma, parts of the Pacific Northwest—where land is cheap and grid capacity exists but infrastructure investment lags. This creates a new map of data center geography that could permanently alter regional economies and tax bases.

It is also creating a bifurcation in the semiconductor ecosystem. Large system integrators with enough capital can secure power agreements early and lock in supply chain advantages. Smaller AI companies and startups cannot. The result is a structural moat around incumbents that has less to do with technology and more to do with energy procurement capability.

There is a secondary effect on commodity markets too. Copper demand for grid infrastructure is already straining global supplies. Natural gas power plants, often proposed as transitional solutions, require fuel supply chains that are themselves constrained in many regions. The AI boom is now competing with electric vehicle manufacturing and renewable energy buildouts for the same有限 pool of critical materials and grid interconnection slots.

What Happens Next

The most consequential shift may be structural. Memory and optical suppliers have been priced as pure demand plays—AI goes up, so their revenues go up. Morgan Stanley’s note reframes the assumption: those revenues now depend on power procurement timelines, grid interconnection status, and local political tolerance.

For Korean industry, the takeaway is narrower but important. Samsung and SK Hynix dominate the memory side of the AI stack. Their growth trajectory remains intact in the medium term. But near-term order cadence may prove uneven, tracking not just chip demand but whether data centers can secure the kilowatts needed to run them. Quarterly earnings calls will increasingly feature answers about power allocation rather than just volume and price.

The three-electric boom is not over. It has simply met a constraint it did not fully price in. Power is no longer a background input. It is the binding variable.

Investors and executives who treat electricity as an infinite resource are walking into a re-rating event. The companies that navigate this constraint—whether through vertical integration, geographic diversification, or early power procurement—will define the next phase of the AI infrastructure cycle. The question is no longer who builds the fastest chip. It is who can power it.