Geopolitics

AI Infrastructure Obstacles 2026-2028

Bottlenecks in data centers, energy, chips, and short-term deployment capacity.

AI is already hitting physical limits before 2028

This executive panel summarizes why the real bottleneck for AI is no longer just about the models. As of April 2026, transformers, permits, water, advanced packaging, and technical talent are constraining the pace at which the computing capacity the market promises can materialize.

Wait time for transformers

3.5 years

The shortage of silicon steel and industrial capacity makes the electrical connection of new campuses critical.

DC projects currently blocked

$ 72,000 M

Infrastructure delayed in the U.S. and Europe due to moratoriums, community opposition, and grid limitations.

Projected capacity deficit in 2028

- 40 GW

AISHA compares the likely AI demand with the physically viable capacity that can actually move forward today.

AISHA's thesis is straightforward: it is not enough to design faster GPUs or more efficient models if the infrastructure that must power them cannot be built at the same pace.

Between 2026 and 2028, AI expansion will be conditioned by the actual speed of the grid, electrical manufacturing, water availability, regulatory opposition, and the extreme concentration of advanced silicon.

AI is not held back by a single piece. It is held back when transformers, permits, CoWoS, water, and talent fail simultaneously, turning promised expansion into capacity that is impossible to power on.

1. Data centers: the gap between computational demand and available concrete

The market talks about gigawatts and giant campuses, but the reality of the sector moves at the pace of electrical permits, land, financing, and civil construction. The question is no longer how much capacity hyperscalers would like to build, but how much can actually be connected before 2028.

Demand versus global data center capacity

AISHA compares observed global capacity with realistic planned capacity and the demand that the 2028 AI wave would require.

Capacity adjusted to physical limits

AISHA places global operational capacity at around 115 GW in 2026, with minimal vacancy in Tier 1 markets.
Construction timelines have gone from about 3 years to 5-6 years due to slow connections and permits.
This means that the theoretical capacity announced by the market overestimates what can actually be powered on before 2028.

The Stargate case as a signal

Abilene demonstrates that even the project with the greatest capital ambition runs into grid, financing, and site limitations.
The expansion from 1.2 GW to 2.0 GW was abandoned, a clear symptom that multi-gigawatt campuses are not trivial.
If the most well-funded case hits that wall, the rest of the sector will hit it sooner.

Moratoriums and social opposition

Projects no longer depend solely on investors and builders, but on regulators, neighbors, and local electricity rates.
Citizen pressure and new rate classes disincentivize speculative deployments and increase the cost of growth.
AI infrastructure is beginning to compete politically with grid, water, and residential costs.

Regulatory signals to watch

AISHA highlights three particularly sensitive fronts. In the United States, state bills and partial moratoriums are proliferating; in Ireland, the concentration of data centers is straining the national grid; in Virginia, rate frameworks that penalize opportunistic expansion are already appearing.

The consequence is clear: expansion no longer resembles a software race. It resembles a constant negotiation with grid, territory, financing, and social legitimacy.

4. Resources and talent: the physical and human side of the bottleneck

AI infrastructure does not depend solely on chips and electricity. It also competes for copper, water, strategic materials, and professionals capable of installing and operating increasingly dense and complex systems.

Physical and human deployment limits

The table summarizes where the tightest bottlenecks are today and why they directly affect the pace of AI expansion.

	Severity	Critical signal	Impact on AI
Copper Materials	Critical	Structural deficit projected toward 2028 due to grid, renewables, and general electrification.	Increases the cost of cabling, substations, and grid expansions needed for AI campuses.
Fresh water Materials	Severe	Regulatory and social conflicts in regions with water stress and intensive cooling.	Limits viable locations and increases the political cost of deployment.
Gallium and germanium Materials	High	Chinese quotas and restrictions strain the base of the electronics supply chain.	Adds geopolitical fragility to key components of the industrial stack.
High-voltage electricians Talent	Critical	Shortage of professionals capable of installing large-scale power and connection systems.	Delays the opening of campuses already completed in civil works by months.
Thermal engineering Talent	High	The transition to liquid cooling exceeds the available operational experience.	Complicates rack densification and the safe deployment of next-generation clusters.
Senior AI/ML researchers Talent	Moderate	Extreme salary costs concentrate talent in a few companies.	Suffocates startups and narrows the real perimeter of frontier innovation.

AISHA distinguishes between material limitations and talent limitations because both act on the same outcome: delaying or increasing the cost of effective capacity.

AISHA insists that this layer is often underestimated because it doesn't sound as spectacular as a new GPU or a multimodal model. But without copper, water, power technicians, and thermal specialists, campuses remain nothing more than a financial projection.

The shortage of technical talent also cannot be solved with capital in the short term: it is part of the same industrial friction that slows energy infrastructure and data centers.

5. Synthesis and impact: how bottlenecks combine toward 2028

The key is not to add up isolated obstacles, but to understand how they reinforce each other. A transformer problem affects the grid; a grid problem affects data centers; a data center problem reduces the actual usefulness of hardware already on order.

Transformers -> grid -> campuses

Without transformers there is no connection; without connection, an already-built campus cannot be powered on.
This turns an apparently minor industrial bottleneck into the direct brake on AI deployment.
The electrical shortfall invalidates part of the capex already committed in real estate and hardware.

Water and permits -> opposition -> relocation

Water stress and noise turn technical infrastructure into local political conflicts.
Moratoriums and litigation force projects to move to less optimal or more remote geographies.
Expansion stops responding solely to economic efficiency and becomes dependent on territorial legitimacy.

CoWoS and talent -> less useful AI

If packaging is delayed, fewer GPUs arrive; if technicians are lacking, they take longer to enter production.
The silicon bottleneck is not solved by more orders alone, because it requires complete industrial integration.
The result is AI that scales below what its financial narrative takes for granted.

Projected severity matrix for 2028

Executive summary of the degree of blockage, trend, and possibility of mitigation before the turn of the decade.

	Severity	Trend	Mitigable before 2028?	Impact on AI
Grid and transformers Electrical infrastructure	5 - Critical	Worsening	Not structurally	Delays the powering on of new gigawatts by several years.
CoWoS advanced packaging Silicon and manufacturing	4 - High	Strained but contained	Partial by 2027	Keeps prices high and maintains scarcity of top-tier GPUs.
Citizen opposition and regulation Water, noise, and grid	4 - High	Worsening	Very limited	Forces relocations and slows strategic deployments.
Power and cooling technical talent Operations and installation	3 - Moderate	Worsening	Partial and slow	Increases capex and delays openings by 6 to 12 months.

AISHA projects that AI growth between 2026 and 2028 will be less global, more expensive, and more concentrated in territories with energy sovereignty, electrical industrial capacity, and preferential access to the silicon supply chain.

Sources

Same category

AI Infrastructure Obstacles 2026-2028

AI is already hitting physical limits before 2028

1. Data centers: the gap between computational demand and available concrete

Demand versus global data center capacity

Capacity adjusted to physical limits

The Stargate case as a signal

Moratoriums and social opposition

Regulatory signals to watch

2. Hardware and chips: the bottleneck is no longer designing, but packaging and delivering

Estimated data center GPU market share in 2026

NVIDIA and the CoWoS limit

Useful alternatives, not equivalent ones

Systemic risk

3. Energy crisis: transformers, gas, and the short-term nuclear fantasy

Wait time for high-power transformers

Transformers as the hidden bottleneck

Nuclear won't arrive in time

Tactical retreat to gas

4. Resources and talent: the physical and human side of the bottleneck

Physical and human deployment limits

5. Synthesis and impact: how bottlenecks combine toward 2028

Transformers -> grid -> campuses

Water and permits -> opposition -> relocation

CoWoS and talent -> less useful AI

Projected severity matrix for 2028

Sources

Keep exploring geopolitics

AI Infrastructure Obstacles 2026-2028

AI is already hitting physical limits before 2028

1. Data centers: the gap between computational demand and available concrete

Demand versus global data center capacity

Capacity adjusted to physical limits

The Stargate case as a signal

Moratoriums and social opposition

Regulatory signals to watch

2. Hardware and chips: the bottleneck is no longer designing, but packaging and delivering

Estimated data center GPU market share in 2026

NVIDIA and the CoWoS limit

Useful alternatives, not equivalent ones

Systemic risk

3. Energy crisis: transformers, gas, and the short-term nuclear fantasy

Wait time for high-power transformers

Transformers as the hidden bottleneck

Nuclear won't arrive in time

Tactical retreat to gas

4. Resources and talent: the physical and human side of the bottleneck

Physical and human deployment limits

5. Synthesis and impact: how bottlenecks combine toward 2028

Transformers -> grid -> campuses

Water and permits -> opposition -> relocation

CoWoS and talent -> less useful AI

Projected severity matrix for 2028

Sources

Keep exploring geopolitics

Geopolitical impact on development and costs

Dependencia del diésel y vulnerabilidad logística

Critical materials and geopolitical risks