14 Stocks Targeting AI Infrastructure Bottlenecks

Author: George Kikvadze

DeepFlow Tech

Shenchao Summary: George Kikvadze, Vice Chairman of Bitfury Group, proposes a counterintuitive perspective: the most profitable opportunities in the AI sector lie not in the model layer, but in infrastructure bottlenecks such as power, cooling, memory, and networking. He outlines seven critical bottlenecks in AI systems and discloses his portfolio of 14 targeted assets, which have generated approximately 60% returns so far. This “bottleneck investing” framework is essential reading for anyone focused on AI investments.

To understand where the money is made in AI, don't look at headline news—look at where the system is under pressure.

The simplest analogy: Today’s AI is like a factory with unlimited orders, but power, wiring, and cooling can’t keep up.

This mismatch itself is an opportunity.

After conducting thorough due diligence, we bet on the following AI bottleneck combination:

$CEG $GEV $VST $WMB $PWR $ETN $VRT $MU $ANET $ALAB $ASML $LRCX $CIFR $IREN

Most investors are asking: "Who will win in AI?" But that's the wrong question.

The question is: Where will the system break? Who profits from fixing it?

In the market, dependencies are leverage.

AI's dependencies are not abstract at all—they are all physical:

• Megawatt-scale electricity
• Transformer delivery lead time
• Cooling capacity per rack
• Memory bandwidth

The economic center of gravity is shifting toward these areas.

AI expansion → Infrastructure strain → Forced investment → Bottlenecks → Pricing power → Profit upgrades

When demand is inelastic and supply is constrained: prices move first, profits follow, and stock prices are reevaluated last.

A few numbers explain everything:

Nearly 50% of data center projects across the U.S. are currently delayed—not due to lack of demand or funding, but because they cannot secure electricity. Transformer delivery timelines have stretched from 24 months before 2020 to over five years today, while data center construction takes 18 months. The math doesn’t add up.

Hyperscale vendors will spend $700 billion on AI infrastructure alone by 2026, nearly six times the 2022 level. Amazon: $200 billion; Google: $175–185 billion; Meta: $115–135 billion. None are slowing down.

Semiconductors currently account for 42% of the total market capitalization of the S&P 500 Information Technology sector, more than doubling since the 2022 bear market low and exceeding their 2013 weight by more than fourfold. Semiconductors also contribute 47% of the IT sector’s forward EPS, nearly tripling their contribution from 2023.

The market is flooding into the hashing power layer at an unprecedented scale.

But computing power is no longer a bottleneck.

Capital is flooding into chips, but the real constraints have shifted elsewhere.

This discrepancy represents a trading opportunity.

• Electricity: Foundation

AI can't expand without electricity.

The U.S. needs to add new capacity equivalent to its entire current data center power base every two years to keep up with AI demand projections before 2030. Nuclear power is the only baseload source capable of providing the scale and reliability required by hyperscalers, but even the fastest nuclear restarts would take several years.

Underlying assets: $CEG $GEV $VST $WMB

These are not utility stocks; they are AI capacity providers. The market hasn't completed this recategorization yet. This mispricing is an opportunity.

Constellation Energy ($CEG) operates the largest fleet of nuclear power plants in the U.S. and is one of the few suppliers capable of providing large-scale, reliable, zero-carbon baseload power. Hyperscalers are accelerating long-term power purchase agreements with nuclear providers, placing Constellation directly on this demand pathway.

GE Vernova ($GEV) is building the power generation backbone for the next energy cycle, covering gas turbines, renewable energy, and grid solutions. As demand for AI accelerates, the ability to rapidly deploy power at scale becomes critical—and GE Vernova’s gas turbines and electrification capabilities are at the heart of this effort.

Vistra Corp ($VST) has a diversified power generation portfolio, including nuclear, natural gas, and retail electricity, enabling it to meet both baseload and peak demand. The highly variable power requirements of AI workloads make this flexibility especially valuable.

Williams Companies ($WMB) operates one of the largest natural gas pipelines in the U.S., supplying fuel to bridge the gap between current demand and the scale of future nuclear power. In the expansion of AI infrastructure, natural gas is the fastest way to bring additional electricity online. Williams essentially serves as an energy raw material supplier for AI growth.

The Grid and Electrification: Constraints Behind the Power

Generating electricity is one thing; transmitting it is even harder.

The U.S. grid interconnection queue is now extended beyond 2030. Meeting existing commitments alone will require over $50 billion in transmission investment over the next decade, not including the addition of a new AI data center.

Underlying assets: $PWR $ETN

The schedule slips here, and profit margins expand here. Companies that solve the "last mile" delivery problem possess enduring long-term pricing power.

Quanta Services ($PWR) is a leading contractor that builds and upgrades transmission infrastructure to connect power generation with consumption. As grid congestion becomes a primary bottleneck for AI expansion, Quanta is directly positioned along a multi-year, non-discretionary capital expenditure pathway. Its backlog serves as a leading indicator of grid stress.

Eaton Corporation ($ETN) provides power distribution systems, switchgear, and power management technologies that enable the safe and efficient large-scale delivery of electricity. As data centers move toward higher power densities and more complex energy flows, Eaton’s components have evolved from standardized hardware into critical infrastructure.

Cooling: Silent Ceiling

Heat kills performance. Thermodynamics has no software patches.

The goal of the next-generation AI infrastructure is 250 kilowatts per rack, compared to the previous standard of 10–15 kilowatts in enterprise data centers a decade ago. Liquid cooling is no longer optional—it is essential infrastructure. For every GPU sold, a corresponding cooling capacity is required, and this ratio will not change.

Target: $VRT

Vertiv holds a near-monopoly in cooling for hyperscale data centers. This is one of the most underestimated components of the entire AI stack, as no one cares about cooling until the cluster goes down.

Vertiv Holdings ($VRT) designs and deploys thermal management systems that keep high-density AI clusters running under extreme power loads. As cabinets transition from air cooling to liquid cooling, Vertiv is at the heart of this structural upgrade cycle, expanding in direct alignment with AI compute deployment. This is not an optional expense—it’s a necessity for operations.

Memory: The Next Bottleneck

AI is shifting from being compute-constrained to being memory-constrained.

As models grow larger and inference demands surge, memory bandwidth and capacity have become the limiting factors, rather than raw processing power. Supply of HBM (High Bandwidth Memory) is already tight. The world’s top three AI memory suppliers control over 90% of global HBM output. Micron is the primary beneficiary in the West.

Primary asset: $MU

This is the next wave of upward earnings revisions. Most portfolios are not yet positioned for it. When the market catches on, they will be.

Micron Technology ($MU) is one of the few global manufacturers capable of mass-producing advanced HBM, a critical component for AI training and inference workloads. As memory becomes a bottleneck in system performance, Micron has transitioned from a historically cyclical supplier to a structural beneficiary of AI demand. This shift has not yet been fully reflected in its valuation, leaving room for continued earnings upgrades and multiple expansion.

Network: Throughput Layer

The speed of an AI cluster depends on the slowest connection.

A single network bottleneck can bring an entire cluster of thousands of GPUs to a halt, wasting hundreds of millions of dollars in capital per facility. As cluster scale expands to 100,000 GPUs, interconnect issues multiply exponentially—one choke point, and the entire system grinds to a stop.

Underlying assets: $ANET $ALAB

Quiet, critical, underpositioned. No one talks about the network until it breaks.

Arista Networks ($ANET) builds high-performance network infrastructure that enables seamless data flow across large-scale AI clusters. When workloads demand ultra-low latency and high throughput, Arista’s software-defined networking becomes critical to maintaining cluster efficiency. The cost of downtime or inefficiency is extremely high, and Arista captures value by ensuring systems operate at full speed.

Astera Labs ($ALAB) operates within the data path to ensure high-speed connections between GPUs, CPUs, and memory in AI systems. As cluster density increases, bottlenecks shift from the network edge to chip-to-chip communication—exactly where Astera focuses. In high-performance AI environments, if communication between components isn't fast enough, the entire system slows down.

Manufacturing: Long-cycle constraints

You cannot scale AI without chip manufacturing capabilities. You cannot manufacture advanced chips without the right manufacturing tools.

ASML’s EUV lithography machines have a production cycle of over a year and cost more than $200 million each, with no credible alternatives. Every advanced chip on Earth—from NVIDIA’s H100 to Apple’s M-series—requires their equipment. Lam Research’s etch and deposition tools are embedded in every major semiconductor fabrication line worldwide.

Underlying assets: $ASML $LRCX

Long-cycle constraints. More difficult to disrupt than any software moat. Discussion热度 is far below what it should be.

ASML Holding ($ASML) is the sole supplier of EUV lithography systems, the most advanced chip manufacturing tools in existence and a prerequisite for producing cutting-edge semiconductors. With a multi-year order backlog and no viable competitors, ASML controls a critical bottleneck in the global semiconductor supply chain.

Lam Research ($LRCX) supplies etch and deposition equipment that form the backbone of semiconductor manufacturing. Its tools are deeply embedded in all major fabs, making it a cyclical and indispensable partner in chip capacity expansion. As AI-driven demand fuels ongoing capacity growth, Lam benefits from long-cycle revenues directly tied to global semiconductor manufacturing expansion.

Misclassification: Source of Alpha

This is the part most investors overlook—and the most asymmetric opportunity on the entire map.

There is a category of companies that the market prices as A, but their operational and financial reality has already become B.

Take $CIFR (Cipher Digital) and $IREN (IREN Limited) for example.

The market is still seeing Bitcoin miners.

What they are becoming is far more valuable: AI power infrastructure and HPC data center platforms.

These companies secured low-cost electricity when no one was paying attention and built their infrastructure before demand arose. Today, these are precisely the two things that hyperscale providers are scrambling to acquire.

Cipher Digital has begun its transformation, signing 15-year leases with investment-grade hyperscale tenants (third AI/HPC campus) and securing a $200 million revolving credit facility from top global banks. These are not speculative moves—they are long-term revenue commitments.

IREN implements the same strategy across multiple sites, combining energy acquisition with scalable data center development. Its advantage is speed: it already controls the land, power, and infrastructure needed for the transition to AI workloads.

The market still sees miners. The balance sheet already looks like that of an infrastructure company.

This gap will converge, and the convergence will not be slow.

This isn't a pile of stocks; it's a system.

Each position corresponds to a specific constraint in the AI stack, and each constraint must be resolved for the system to function. This is discipline.

• Power: $CEG $GEV $VST $WMB
• Grid: $PWR $ETN
• Cooling: $VRT
• Memory: $MU
• Network: $ANET $ALAB
• Manufactured by: $ASML $LRCX
• Incorrect classification: $CIFR $IREN

We are transitioning from a scarcity of computing power to a scarcity of infrastructure.

This means:

• GPU is no longer the only narrative
• Electricity, power grids, memory, and cooling have become the primary drivers of profitability.
• Returns follow discipline, not hype.

Most portfolios are still holding positions from the old world.

This framework will fail under certain conditions. They deserve honest acknowledgment.

Large-scale vendors are slowing capital expenditures. If Amazon, Google, and Meta reduce infrastructure spending due to margin pressures or weaker-than-expected demand, the assumption of rigid demand weakens. This is the primary risk to monitor, with quarterly capital expenditure guidance serving as a leading indicator.

The resolution of bottlenecks is occurring faster than expected. Government intervention in transformer manufacturing, accelerated nuclear power approvals, or restructuring the grid interconnection queue could all reduce the premium on constrained infrastructure. These changes are gradual but real.

Regulatory friction. Power and grid infrastructure intersect with utility regulation, environmental reviews, and rate-setting authorities. When regulatory shifts turn unfavorable in this area, they structurally and persistently limit upside returns.

The key difference is that this is not a bet on a product cycle—product cycles can reverse within a quarter. Industrial constraints take years to build and years to resolve. This asymmetry is the point.

In every industrial era, wealth is not created by the companies that build trains.

But were created by companies that owned the railways, coal, and rights-of-way.

AI's rail is measured in megawatts, transformer lead times, and cooling capacity per rack.

Most investors are chasing AI. The real opportunity lies in owning what AI cannot do without.

In every system, headlines follow innovation, but profits follow constraints. We focus on constraints, not narratives—current returns are around 60%. With AI infrastructure accelerating, this is not the end of trading; we’re still in the early stages. We believe we’re only in the third inning.