From power generation to lithography machines, these 14 stocks have overcome every bottleneck in AI expansion.

Author: George Kikvadze

Translation: Deep潮 TechFlow

Deep潮 Guide: Vice Chairman of Bitfury Group George Kikvadze proposes a reverse thinking: the most profitable opportunities in the AI track are not at the model level, but in infrastructure bottlenecks such as electricity, cooling, memory, and networking. He outlines 7 critical “choke points” in AI systems and publicly shares his portfolio of 14 targets, currently yielding about 60%. This “bottleneck investment” framework is worth every AI investor’s serious review.

To understand where AI can make money, don’t look at headlines; look at where the system is under pressure.

The simplest analogy: today’s AI is like a factory with unlimited orders, but electricity, cables, and cooling can’t keep up.

This mismatch itself is an opportunity.

After detailed due diligence, we bet on the following “AI bottleneck” portfolio:

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

The real question to ask

Most investors ask: “Who will win AI?” This question is wrong.

The right question is: Where will the system break? Who is making money fixing it?

In the market, dependencies are leverage.

AI dependencies are not abstract; they are all physical:

Megawatt-level electricity

Transformer delivery cycles

Cooling capacity per cabinet

Memory bandwidth

The economic focus is shifting toward these areas.

The only analytical framework needed:

AI expansion → infrastructure under pressure → forced investment → bottleneck → pricing power → profit upgrades

When demand is rigid and supply is constrained: prices move first, profits follow, and stock prices are finally revalued.

Why now

A few numbers explain the entire issue:

Nearly 50% of data center projects in the US are currently delayed, not due to lack of demand or funding, but because of unavailable electricity. Transformer delivery times have extended from 24 months before 2020 to over 5 years now. Data center construction takes 18 months. This calculation doesn’t add up.

By 2026, large-scale vendors will spend nearly $700 billion on AI infrastructure alone, close to 6 times 2022’s figure. Amazon: $200 billion, Google: $175-185 billion, Meta: $115-135 billion. No one is slowing down.

Semiconductors now account for 42% of the total market cap of the S&P 500 IT sector, more than doubling from the bottom of the 2022 bear market, and over four times the weight in 2013. Semiconductors also contribute 47% of the forward-looking EPS of the IT sector, nearly tripling since 2023.

The market is rushing into compute power at an unprecedented density.

But compute power is no longer the bottleneck.

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

This gap is the trading opportunity.

Bottleneck Map: Where is the pressure?

Electricity: The foundation

AI cannot expand without power. Period.

The US needs to add capacity equivalent to the entire current data center power infrastructure every two years to meet AI demand forecasts before 2030. Nuclear power is the only reliable, large-scale baseload power source capable of supporting mega-vendors, but even the fastest nuclear restart takes years.

Targets: $CEG $GEV $VST $WMB

These are not utility stocks; they are AI capacity providers. The market has not yet reclassified them. This mispricing is an opportunity.

Constellation Energy ($CEG) operates the largest nuclear fleet in the US and is one of the few providers offering large-scale, reliable, zero-carbon baseload power. Mega-vendors are accelerating long-term power purchase agreements with nuclear suppliers, and Constellation is directly on this demand path.

GE Vernova ($GEV) is building the next energy cycle’s generation backbone, covering gas turbines, renewables, and grid solutions. When AI demand accelerates, the ability to deploy power quickly and at scale becomes critical, and GE Vernova’s gas turbines and electrification capabilities are at this core.

Vistra Corp ($VST) has a diversified generation portfolio, including nuclear, gas, and retail power, capable of meeting both baseload and peak demands. AI workloads cause highly volatile power needs, making this flexibility especially valuable.

Williams Companies ($WMB) operates one of the largest natural gas pipelines in the US, providing fuel to bridge the current gap between demand and future nuclear capacity. In expanding AI infrastructure, natural gas is the fastest way to bring incremental power online. Williams is essentially an energy raw material supplier for AI growth.

Power grid and electrification: Constraints behind electricity

Power generation is one thing; transmission is harder.

The US grid interconnection queue now extends beyond 2030. Over $50 billion in transmission investments are needed in the next decade just to meet existing commitments, not counting new AI data centers.

Targets: $PWR $ETN

Schedule delays here, profit margins also expand here. Companies solving the “last mile” delivery problem have long-term pricing power.

Quanta Services ($PWR) is a leading contractor building and upgrading transmission infrastructure, connecting power generation and consumption. When grid congestion becomes the main bottleneck for AI expansion, Quanta is directly on the multi-year, non-discretionary capital expenditure path. Its backlog is a forward indicator of grid pressure.

Eaton Corporation ($ETN) provides distribution systems, switchgear, and power management tech, enabling large-scale, safe, and efficient power delivery. As data centers move toward higher power densities and more complex energy flows, Eaton’s components shift from standardized hardware to critical infrastructure.

Cooling: The silent ceiling

Heat kills performance. Thermodynamics have no software patch.

The goal for next-generation AI facilities is 250 kW per cabinet, whereas a standard enterprise data center a decade ago only had 10-15 kW. Liquid cooling is no longer optional; it’s essential infrastructure. Every GPU sold requires corresponding cooling capacity, and this ratio will not change.

Target: $VRT

Vertiv is close to a monopoly in large-scale data center cooling. This is one of the most underestimated links in the entire AI stack because no one cares about cooling until a cluster crashes.

Vertiv Holdings ($VRT) designs and deploys thermal management systems to keep high-density AI clusters operational under extreme power loads. As cabinets shift from air cooling to liquid cooling, Vertiv is at the heart of this structural upgrade cycle, expanding in tandem with AI compute deployment. This is not optional spending; it’s a prerequisite for normal operation.

Memory: The next bottleneck

AI is shifting from compute-limited to memory-limited.

As models grow larger and inference volumes explode, memory bandwidth and capacity become constraints, not raw processing power. HBM (High Bandwidth Memory) supply is tight. The top three global AI memory suppliers control over 90% of worldwide HBM output. Micron is the main Western beneficiary.

Core target: $MU

This is the next wave of profit upgrades. Most portfolios are not yet positioned for this. When the market reacts, they will be.

Micron Technology ($MU) is one of the few global manufacturers capable of mass-producing advanced HBM at scale. HBM is critical for AI training and inference workloads. When memory becomes a system performance bottleneck, Micron shifts from cyclical supplier to structural beneficiary of AI demand. This shift is not yet fully reflected in valuations, leaving room for sustained profit upgrades and multiple expansion.

Networking: The throughput layer

The speed of AI clusters depends on the slowest link.

A single network bottleneck can stall an entire cluster of thousands of GPUs, wasting hundreds of millions of dollars of capital per facility. As cluster sizes expand toward 100k GPUs, interconnect issues grow exponentially. One bottleneck halts everything.

Target: $ANET $ALAB

Quiet, critical, under-allocated. No one talks about networking until it fails.

Arista Networks ($ANET) builds high-performance network infrastructure, enabling seamless data flow in large-scale AI clusters. When workloads demand ultra-low latency and high throughput, Arista’s software-defined networking becomes key to maintaining cluster efficiency. Downtime or inefficiency costs are high; Arista captures value by ensuring systems run at full speed.

Astera Labs ($ALAB) operates within data pathways, ensuring high-speed connections between GPUs, CPUs, and memory in AI systems. As cluster density increases, bottlenecks shift from network edges to chip-to-chip communication, which is exactly where Astera’s position is. In high-performance AI environments, slow communication between components slows down the entire system.

Manufacturing: Long-cycle constraints

Without chip manufacturing capacity, AI cannot scale. Without manufacturing tools, advanced chips cannot be made.

ASML’s EUV lithography machines have a cycle time exceeding a year, costing over $200 million each, with no credible substitutes. Every advanced chip, from NVIDIA’s H100 to Apple’s M-series, depends on these tools. Lam Research’s etching and deposition equipment are embedded in every major wafer fab worldwide.

Target: $ASML $LRCX

Long-cycle constraints. Structurally more resilient than any software moat. Discussion of this is far below the level it deserves.

ASML Holding ($ASML) is the sole supplier of EUV lithography systems, the most advanced chip manufacturing tools available, and a prerequisite for producing cutting-edge semiconductors. With multi-year order backlogs and no real competition, ASML controls a critical choke point in the global chip supply chain.

Lam Research ($LRCX) supplies etching and deposition equipment that form the backbone of semiconductor manufacturing. Its tools are deeply embedded in all major wafer fabs, making it an indispensable partner in capacity expansion. As AI demand drives continuous capacity growth, Lam’s long-cycle revenues are directly tied to global semiconductor manufacturing expansion.

Misclassification: The source of alpha

This is an area most investors overlook and the most asymmetric opportunity on the map.

There are companies that the market prices as A, but their operations and financials are already B.

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

The market still sees them as Bitcoin miners.

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

These companies locked in low-cost power and built infrastructure before demand materialized. Today, mega-vendors are rushing to acquire exactly these two assets.

Cipher Digital has begun a transformation, signing 15-year leases with tier-one mega tenants (the third AI/HPC park), and secured $200 million in revolving credit from top global banks. These are not speculative moves; they are long-term revenue commitments.

IREN is executing similar strategies across multiple sites, combining energy procurement with scalable data center construction. Its advantage is speed: it has already secured land, power, and infrastructure needed for AI workloads.

The market still sees them as miners. Their balance sheets look more like infrastructure companies.

This gap will close. When it does, it won’t be slow.

Portfolio overview

This is not just a bunch of stocks; it’s a system.

Each position corresponds to a specific constraint in the AI stack, and each must be addressed for the system to operate. That’s discipline.

Electricity: $CEG $GEV $VST $WMB

Power grid: $PWR $ETN

Cooling: $VRT

Memory: $MU

Networking: $ANET $ALAB

Manufacturing: $ASML $LRCX

Misclassification: $CIFR $IREN

Most investors have not yet completed this cognitive shift

We are shifting from compute scarcity to infrastructure scarcity.

This means:

GPUs are no longer the only narrative

Electricity, power grids, memory, and cooling are becoming the main profit drivers

Returns follow constraints, not hype

Most portfolios are still stuck in the old paradigm.

Risks: Discipline is equally important

This framework can fail under certain conditions. It’s worth being honest about them.

Mega-vendor capital expenditure slowdown. If Amazon, Google, and Meta slow infrastructure spending due to profit margin pressures or weaker-than-expected demand, the rigid demand assumption weakens. Monitoring quarterly capital expenditure guidance is the top priority risk indicator.

Rapid resolution of bottlenecks. Government intervention in transformer manufacturing, accelerated nuclear approvals, or restructuring grid interconnection queues could compress the premium on constrained infrastructure. These changes are slow but real.

Regulatory friction. Power and grid infrastructure intersect with utility regulation, environmental reviews, and rate-setting agencies. When regulation turns unfavorable, it can structurally and persistently limit return caps.

The key difference: this is not a product cycle bet. Product cycles can reverse in a quarter. Industrial constraints take years to build and years to resolve. This asymmetry is the point.

Finally

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

It’s created by those owning the rails, coal, and rights.

The rails of AI are measured in megawatts, transformer delivery cycles, and cooling capacity per cabinet.

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

In every system, headlines follow innovation, profits follow constraints. We focus on constraints, not narratives, and currently yield about 60%. As AI infrastructure accelerates, this is not the end of the trade but still early. We believe we are only in the third game.