The Future of AI Infrastructure: Where the Economics Accrue, 2026–2031

The next great AI investment opportunity may not be in the company building the world's most powerful GPU. It may be in the company selling the electricity, cooling, networking, and physical infrastructure required to keep that GPU running.

There is a fundamental mispricing occurring in the AI investment narrative.

For the past three years, investors have primarily viewed artificial intelligence through the lens of semiconductor demand. NVIDIA sells GPUs, hyperscalers purchase them, data centers are constructed, and investors extrapolate increasingly aggressive revenue forecasts.

That trade has worked extraordinarily well. But it is incomplete.

The AI infrastructure cycle is transitioning into something considerably larger and more complex. The industry is moving from building increasingly powerful models toward deploying autonomous systems that continuously consume computing resources. At the same time, the physical infrastructure required to support those systems is running into constraints that cannot be resolved simply by manufacturing more semiconductors.

Between 2026 and 2031, I expect the most attractive opportunities to emerge across five interconnected categories:

1. Hyperscale cloud platforms with enormous financial resources.
2. Semiconductors, custom accelerators, memory, and networking.
3. Data center electrical infrastructure and liquid cooling.
4. Electricity generation, nuclear power, and grid modernization.
5. A small number of specialized AI-native cloud providers.

My central argument is that investors should gradually shift their perspective from owning just the companies producing AI compute to owning the companies that control the entire infrastructure supply chain.

The distinction matters because the economic winners of the next five years may be very different from the winners of the previous three.

---

1. The Macro Shift: From Training AI to Running AI

The biggest change in AI infrastructure is not model size. It is the frequency with which models are being used.

The first phase of the generative AI boom was predominantly a training investment cycle.

Companies such as OpenAI, Google, Anthropic, and Meta spent enormous sums constructing computing clusters capable of training increasingly sophisticated foundation models.

Training is extraordinarily resource-intensive, but it has a relatively distinctive economic structure.

You build a large computing cluster, train a model over weeks or months, fine-tune it, and eventually deploy it.

The infrastructure must be powerful, but the training process is concentrated in identifiable projects.

The next phase operates differently.

We are moving toward agentic AI: systems that do not simply respond to questions but independently perform sequences of tasks.

Consider the difference between these two interactions:

Traditional AI:

• A user asks a question.
• The model generates a response.
• The interaction ends.

Agentic AI:

• A user requests a financial analysis.
• The agent retrieves market information.
• It searches multiple data sources.
• It evaluates the information.
• It executes calculations.
• It identifies missing information.
• It performs additional searches.
• It generates a report.
• It checks its own output.
• It potentially repeats the process.

One user request can trigger numerous inference operations.

The agent may also remain active, monitor events, maintain context, and independently initiate subsequent tasks.

This creates a fundamentally different computing demand profile.

Inference is becoming the dominant spending category

One particularly important data point comes from Gartner's August 2026 forecast of spending on AI-optimized cloud infrastructure (IaaS).

Gartner expects $23.3 billion of AI-optimized IaaS spending to support inference in 2026, exceeding the approximately $19 billion supporting training. It is the first year inference takes the larger share. These figures cover AI-optimized cloud infrastructure only, not total AI spending.

Inference is expected to account for 55% of AI-optimized IaaS spending in 2026, increasing to 59% in 2027. Gartner puts total AI-optimized IaaS spending at about $42.3 billion in 2026, up 96%.

Source: Gartner – Worldwide AI-Optimized IaaS Forecast, August 10, 2026.

This is an important transition. It does not mean training becomes irrelevant. Frontier models will continue requiring enormous training clusters. Rather, the economics of AI infrastructure are increasingly being driven by the ongoing cost of serving models to millions of users and autonomous agents.

The implications are substantial.

First, compute utilization becomes more persistent.

Training clusters can experience highly concentrated periods of utilization. Inference infrastructure must be available continuously, including outside traditional business hours.

Second, latency becomes economically important.

An autonomous agent performing a multistep task cannot afford to wait several minutes between operations. This encourages investment in regional data centers, faster networking, memory infrastructure, and optimized inference hardware.

Third, inference efficiency becomes a competitive advantage.

The cheapest provider per token will not necessarily be the provider with the cheapest GPU. It will be the provider capable of delivering the lowest total cost per useful unit of inference.

This includes electricity, hardware utilization, networking, memory, cooling, and software optimization.

Fourth, AI infrastructure becomes more geographically distributed.

Not every inference workload needs to run in a massive centralized training cluster. Enterprise applications, financial services, industrial automation, and consumer-facing agents increasingly have latency, security, and data-residency requirements.

This creates opportunities for regional cloud operators and specialized infrastructure providers.

What this means for investors

The market is beginning to transition from a semiconductor scarcity trade into a broader infrastructure-utilization trade.

NVIDIA remains enormously important, but the investment opportunity expands to:

• High-bandwidth memory and storage.
• Custom AI accelerators and ASICs.
• High-speed Ethernet and optical networking.
• Liquid cooling and thermal management.
• Power distribution and electrical equipment.
• Data center construction and grid infrastructure.
• Electricity generation and storage.

The opportunity is no longer confined to the chip sitting inside the server.

It extends to everything required to operate that server economically.

---

2. Hyperscalers vs. Neoclouds: Who Actually Wins?

This is arguably the most important capital allocation decision within AI infrastructure.

There are two fundamentally different business models competing for the same AI compute demand.

Hyperscalers: Microsoft, Amazon, Google, and other established technology platforms.

Neoclouds: Specialized AI-native cloud operators such as CoreWeave and Nebius.

Both are experiencing extraordinary demand. But their ability to finance growth and survive an infrastructure downturn is dramatically different.

A. The Hyperscalers: Financial Strength Is Their Biggest Competitive Advantage

The most underappreciated advantage of the hyperscalers is not necessarily their technology.

It is their balance sheets.

Microsoft, Amazon, and Alphabet operate enormous businesses that generate recurring cash flows independently of AI.

Microsoft has enterprise software subscriptions and productivity applications.

Amazon has retail, advertising, subscriptions, and AWS.

Google has Search, YouTube, advertising, subscriptions, and Google Cloud.

These businesses give them access to capital that specialized AI infrastructure companies simply cannot replicate.

They can tolerate periods of negative incremental infrastructure returns while waiting for demand to mature.

More importantly, they can finance enormous projects without becoming immediately dependent on external investors.

Microsoft (MSFT): Probably the strongest overall infrastructure platform

Microsoft is exceptionally well positioned because it controls several layers of the AI economy.

It owns:

• Azure cloud infrastructure.
• Microsoft 365 distribution.
• GitHub and developer ecosystems.
• Enterprise software relationships.
• AI development and deployment services.
• A massive installed base of business customers.

The distribution advantage is significant.

Microsoft does not need to convince millions of enterprises to purchase an entirely new product category. It can embed AI into existing software relationships and charge customers through subscriptions and consumption-based pricing.

Its latest results are particularly revealing.

For the quarter ended September 30, 2026:

The $677 billion commercial remaining performance obligation is especially important.

This represents contracted future revenue that has not yet been recognized. Its weighted average duration is approximately 2.3 years, with roughly 30% expected to be recognized over the next 12 months.

It is not all GPU rental revenue, and it should not be interpreted as guaranteed immediate cash flow. Nevertheless, it demonstrates the extraordinary visibility Microsoft has into future commercial demand.

Sources: Microsoft FY27 Q1 earnings and earnings call.

Now look at the capital expenditures.

Microsoft expects calendar-year 2026 CapEx of approximately $170 billion, with second-quarter spending expected to exceed $50 billion.

For the September quarter, operating cash flow was $58 billion against approximately $39.4 billion in additions to property and equipment. That implies roughly $18.6 billion in quarterly free cash flow using the standard operating cash flow less CapEx calculation.

Microsoft ended September with approximately $20.3 billion in cash and $44.2 billion in short-term investments, against approximately $35.5 billion in long-term debt including the current portion.

That is a very different financing profile from an infrastructure company that must borrow billions merely to purchase its next generation of GPUs.

Microsoft's September 2026 Form 10-Q provides the underlying balance-sheet and cash-flow data.

My view: Microsoft is a natural foundation for an AI infrastructure thesis. It has distribution, enterprise relationships, software monetization, cloud infrastructure, and the financial resources to keep investing.

Amazon (AMZN): The strongest direct infrastructure-volume opportunity

Amazon deserves more attention than it frequently receives in AI discussions.

AWS is one of the world's largest cloud infrastructure businesses, and its latest results demonstrate the acceleration in demand.

For Q3 2026:

• AWS revenue reached $46.0 billion.
• AWS grew 39.3% year over year.
• Its annualized revenue run rate reached approximately $184 billion.
• AWS operating income reached $17.8 billion.

This is not an early-stage business trying to establish product-market fit.

AWS already has a huge customer base and an established global infrastructure footprint.

Amazon also has an important strategic advantage: it can offer multiple computing architectures, including NVIDIA GPUs and its own custom silicon.

That potentially reduces dependence on any single accelerator supplier.

Its backlog is another important indicator.

Amazon reported approximately $578 billion in remaining performance obligations, primarily associated with AWS long-term customer contracts. The weighted average remaining life of its long-term contracts was 6.1 years.

Sources: Amazon Q3 2026 results and Form 10-Q.

However, there is a significant financial consideration.

Amazon generated $168.3 billion in trailing-twelve-month operating cash flow, but spent approximately $210.7 billion on property and equipment, net of proceeds and incentives.

Its trailing-twelve-month free cash flow consequently fell to negative $42.4 billion.

Management continues to expect approximately $220 billion in 2026 cash CapEx and expects spending to increase meaningfully in 2027.

Amazon also had approximately $173.1 billion in long-term debt at September 30, 2026, although it retained substantial liquidity: approximately $114.2 billion in cash, cash equivalents, and marketable securities.

This is an enormous investment program.

But I would not confuse negative free cash flow during an aggressive expansion with financial distress.

Amazon is consciously investing in infrastructure intended to produce recurring revenue over many years.

The real question is whether AWS revenue growth and infrastructure utilization eventually justify the investment.

My view: Amazon is one of the strongest long-term ways to capture the growth of AI infrastructure without betting exclusively on semiconductor manufacturers.

Alphabet (GOOGL): An increasingly underestimated AI infrastructure competitor

Alphabet is particularly interesting because the market frequently evaluates it predominantly through the advertising business.

That misses the emerging importance of Google Cloud and Google's proprietary AI infrastructure.

Google has three substantial advantages:

1. Its own AI models and research capabilities.
2. Its proprietary TPU accelerator ecosystem.
3. A massive global cloud infrastructure footprint.

Google Cloud's Q3 2026 results were extraordinary.

The roughly $561 billion Cloud backlog is particularly impressive. Management expects just over 55% of the total Google Cloud backlog to be recognized as revenue over the next 24 months.

Sources: Alphabet Q3 2026 results and earnings call.

Alphabet expects 2026 CapEx of approximately $195–205 billion, with management anticipating another significant increase in 2027.

Its trailing-twelve-month operating cash flow reached approximately $179.6 billion, while free cash flow was approximately $19.7 billion.

However, Q3 alone generated negative $9.1 billion in free cash flow.

Alphabet has also substantially increased its long-term debt, from $46.5 billion at year-end 2025 to $125.1 billion at September 2026. In June it raised approximately $20.5 billion through an underwritten common stock offering and about $19 billion through mandatory convertible preferred securities. Including at-the-market share sales and a private placement, net proceeds from issuing common stock totaled about $42.9 billion in the first nine months of 2026.

This is worth emphasizing because it challenges the idea that hyperscaler expansion is entirely self-funded.

These companies are using operating cash flow, balance-sheet capacity, debt, equity-related financing, and other capital sources.

The difference is their ability to access those markets at scale and withstand temporary cash-flow deterioration.

My view: Alphabet may offer one of the most interesting combinations of AI infrastructure growth, proprietary silicon, cloud profitability, and valuation.

Google's TPU ecosystem also makes it an important counterweight to an investment strategy that relies entirely on NVIDIA.

---

B. Neoclouds: Extraordinary Growth, But Dangerous Financial Structures

Now comes the more controversial part of the thesis.

The neocloud market is growing rapidly, but I believe investors are underestimating the financial risks involved.

Companies such as CoreWeave and Nebius are essentially building specialized infrastructure businesses designed around AI workloads.

Their business model is relatively straightforward:

1. Secure large quantities of GPUs.
2. Obtain access to data center capacity and electricity.
3. Finance the infrastructure through debt, equity, leases, and other arrangements.
4. Rent GPU computing capacity to AI laboratories, enterprises, and developers.
5. Generate returns from utilization and pricing.

The attraction is obvious.

They can generate extraordinary revenue growth without needing to develop their own foundational AI models.

But the financing structure creates significant vulnerability.

The CoreWeave example

CoreWeave is probably the clearest example of both the opportunity and the danger.

Its Q2 2026 results showed:

There is an interesting discrepancy here.

CoreWeave reports an impressive adjusted EBITDA margin of 59%, but its GAAP operating result was negative.

Depreciation and amortization alone reached approximately $1.39 billion during the quarter.

This is not a minor accounting consideration.

The company owns and operates extremely expensive infrastructure that must eventually be refreshed.

Its quarterly net interest expense of $640 million is also substantial relative to its underlying operating earnings.

At June 30, CoreWeave's balance sheet showed approximately $35.1 billion of recourse and non-recourse debt (excluding finance leases) against $5.5 billion in cash.

Sources: CoreWeave Q2 earnings and Q2 2026 Form 10-Q.

There is another important consideration.

CoreWeave reported approximately $104 billion in revenue backlog at the end of Q2, with more than $25 billion in net new customer commitments added early in Q3.

It also had 1.5 GW of active power capacity and approximately 3.7 GW of contracted power.

That is extraordinary infrastructure demand.

But backlog is not the same thing as collected revenue.

The company must actually construct facilities, secure electricity, purchase equipment, and deliver usable computing capacity.

The contracts are subject to delivery and service-availability requirements.

This introduces execution risk.

CoreWeave's Q3 report is expected around November 10, 2026, though the company had not confirmed the date at the time of writing, so its next quarterly update is still pending.

Nebius: A potentially stronger capital position, but still capital hungry

Nebius is another important company worth monitoring.

Its Q2 2026 consolidated revenue reached approximately $582 million, compared with $105 million in the prior-year quarter.

Adjusted EBITDA reached $236 million.

Adjusted net loss improved to approximately $33 million.

Those are encouraging developments.

However, examining its cash flows reveals how capital intensive this expansion remains.

During the first half of 2026:

• Operating cash flow was approximately $4.5 billion.
• Net cash used in investing activities was approximately $8.3 billion.
• Net cash provided by financing activities was approximately $9.2 billion.

The financing activities included approximately $4.34 billion from convertible-note issuance and $2.85 billion from the sale of treasury shares.

Nebius ended Q2 with approximately $9.1 billion in cash and restricted cash.

This provides substantial funding capacity, but it should not be mistaken for cash generation from a mature, self-financing operating business.

Its investment requirements are enormous, and capital-market access remains important.

Source: Nebius Q2 2026 financial results.

Nebius releases Q3 results on November 12, 2026. That report will be important for reassessing its funding requirements and infrastructure expansion.

The theoretical neocloud balance-sheet problem

Consider an illustrative AI infrastructure company with the following characteristics (a hypothetical, not any specific company):

• $5 billion in annual revenue.
• 50% adjusted EBITDA margins.
• $15–20 billion in debt.
• Large GPU lease obligations.
• Concentrated customer relationships.
• Heavy dependence on external financing.

On the surface, this can look like an extraordinarily attractive growth business.

But consider what happens when the next GPU generation arrives.

The company may need to finance new hardware while still servicing debt associated with its previous generation.

Its existing equipment is becoming less competitive.

A major customer may negotiate lower prices.

Interest rates may remain elevated.

Electricity connections may be delayed by two years.

Suddenly, a company growing revenue at 100% annually can experience a severe liquidity problem.

This is why EBITDA margins alone are insufficient for evaluating neoclouds.

I would pay considerably more attention to:

• Free cash flow after infrastructure spending.
• Interest expense relative to operating income.
• Debt maturities and borrowing costs.
• Share-count growth and dilution.
• GPU asset life and resale value.
• Customer concentration.
• Contracted versus operational power.
• Backlog conversion into actual revenue.

Hyperscaler versus neocloud: The financial resilience comparison

Figures are based on Microsoft's Q1 FY27 (September 2026 quarter) and CoreWeave's Q2 2026 disclosures. Backlog definitions differ and should not be treated as directly comparable.

Will the neoclouds survive, or will we see a shakeout?

My expectation is consolidation.

I do not believe every specialized AI cloud operator will disappear. There is a legitimate market for companies providing specialized GPU infrastructure, regional compute capacity, optimized inference environments, and cloud alternatives to hyperscalers.

However, I expect a relatively small number of operators to capture most of the sustainable economics.

The survivors will likely have:

• Long-term anchor customers.
• Access to relatively inexpensive financing.
• Reliable, secured electricity.
• High hardware utilization.
• Strong engineering capabilities.
• Diversified customer relationships.

The smaller operators that depend on expensive debt, speculative data center construction, and continuous equity issuance are vulnerable.

Some will be acquired.

Some will restructure.

Some will simply run out of money.

The strongest specialists could become valuable acquisition targets for larger cloud providers or infrastructure operators.

My stance: I treat the hyperscalers as the core of this thesis and the neoclouds as speculative rather than foundational.

High revenue growth is not particularly valuable when most of the economic benefit is consumed by interest expense, depreciation, and replacement CapEx.

---

3. The Physical Constraint: Electricity Is Becoming More Valuable Than Another Generation of Silicon

What the market is fundamentally mispricing right now is the physical limitation of AI infrastructure.

For years, investors discussed GPU shortages and semiconductor manufacturing constraints.

Those remain relevant.

But the bottleneck is increasingly shifting toward the ability to deliver electricity to data centers at the required location, density, and reliability.

A GPU cannot generate revenue sitting in a warehouse.

A data center cannot generate revenue waiting for a grid connection.

And a 130-kilowatt rack cannot operate in a facility designed around traditional server configurations.

The physical infrastructure problem is enormous.

A. The electricity requirement

The International Energy Agency's updated outlook projects global data center electricity consumption to increase from approximately:

That represents approximately a doubling over five years.

Even more importantly, electricity consumption from AI-focused data centers is expected to triple during that period.

By 2030, data centers could account for around 3% of global electricity consumption.

Source: IEA – Key Questions on Energy and AI.

This is an extraordinary amount of electricity demand.

But electricity generation itself is only part of the problem.

The real issue is delivering that electricity where it is needed.

B. The 130-kilowatt rack changes everything

Consider the NVIDIA GB200 NVL72 rack.

HPE's published technical specifications indicate:

• Approximately 132 kW nominal rack thermal design power.
• Approximately 192 kW peak electrical design power (1.5 times the nominal figure), which facilities must size their power distribution for.
• Direct liquid cooling for most of the load.
• Dedicated power distribution infrastructure.

Source: HPE NVIDIA GB200 NVL72 QuickSpecs.

This is a major departure from traditional data centers.

A facility designed around relatively low-density computing cannot simply install these racks and expect everything to work.

It needs redesigned:

• Power distribution systems.
• Electrical switchgear.
• Transformers.
• Busbars and power shelves.
• Uninterruptible power supplies.
• Cooling distribution units.
• Liquid cooling loops.
• Water and heat-rejection infrastructure.

The IEA also estimates that AI server power density increased approximately elevenfold between 2020 and 2025, with a further fourfold increase expected by 2027.

This creates an enormous opportunity for companies that manufacture the equipment required to distribute and manage electricity.

C. Grid infrastructure could be one of the biggest beneficiaries

The US electricity grid was not originally engineered to accommodate this pace of concentrated demand growth.

The International Energy Agency estimates that more than 2,500 GW of generation, storage, and large-load projects are currently stalled in electricity connection queues worldwide.

It also estimates that annual grid investment needs to increase approximately 50% by 2030 from today's roughly $400 billion.

Source: IEA – Electricity 2026.

Meanwhile, Reuters reported in July 2026 that utilities were scrambling to secure transformers, turbines, and other grid equipment.

Lead times for some high-voltage transformers have stretched to multiple years, and large substation transformers are pushing past three years.

Source: Reuters – US Power Companies Scramble to Secure Equipment, July 9, 2026.

Think about the investment implications.

Even if NVIDIA can manufacture every GPU required over the next five years, data center developers cannot necessarily install and operate them without electrical infrastructure.

That makes the manufacturers of transformers, power-management systems, switchgear, cooling equipment, and electrical distribution components strategically important.

These companies may have less technological glamour than semiconductor manufacturers, but they are positioned at a genuine physical bottleneck.

D. The multi-billion-dollar nuclear opportunity

This is where the long-term energy investment thesis becomes particularly interesting.

Solar and wind will contribute meaningfully to incremental electricity supply.

Battery storage can smooth demand fluctuations and support power availability.

Natural gas will likely remain an important source of dispatchable electricity.

But large AI data centers require reliable electricity continuously.

Nuclear power offers an attractive combination of firm generation, high availability, and relatively low operational emissions.

The hyperscalers recognize this.

Amazon and X-energy

Amazon has invested in X-energy and is collaborating on projects targeting more than 5 GW of new US nuclear capacity by 2039.

Its initial Energy Northwest project involves four small modular reactors providing 320 MW, with an option to expand to 960 MW.

In October 2026, Amazon, AEP, and X-energy also agreed to evaluate a 960 MW project at a former coal site in Indiana. It is an exploratory agreement: the companies have not given a timeline or cost estimate.

There is no firm construction start date for these reactors, and power is expected in the 2030s.

This is a long-term infrastructure commitment, not an immediate electricity solution.

Sources: Amazon's Indiana nuclear initiative, Utility Dive on the Indiana project and Power Magazine on the 5 GW plan.

Google and Kairos Power

Google has signed an agreement with Kairos Power to deploy an advanced nuclear reactor fleet delivering up to 500 MW by 2035.

The Hermes 2 plant in Tennessee, which broke ground in April 2026, is targeted to begin supplying up to 50 MW in 2030.

Source: Kairos Power – Google Partnership.

The broader SMR opportunity

The regulatory environment is also progressing.

On September 29, 2026, the Nuclear Regulatory Commission issued a construction permit for TVA's Clinch River small modular reactor project, an important licensing milestone for commercial SMR deployment. The permit covers construction only, not operation.

However, investors should understand the timeline.

SMRs are not going to solve the electricity shortage facing data centers being constructed today.

Most significant commercial deployments are targeting the early 2030s and beyond.

Their economics, construction costs, fuel supply, licensing processes, and first-of-a-kind execution still need to be proven at scale.

That makes existing nuclear generators and electrical infrastructure manufacturers much more immediately investable than speculative SMR development companies.

E. Long-duration energy storage

One additional area deserves attention.

AI infrastructure creates substantial variations in electricity demand.

The IEA highlights the importance of storage in managing the rapid power fluctuations associated with training and model use.

It estimates that approximately 20–25 GW of battery storage could be installed at data centers globally by 2030, up from around 5 GW today.

The opportunity extends beyond traditional lithium-ion batteries.

Longer-duration storage, grid-scale batteries, advanced power electronics, and systems that coordinate data center consumption with electricity availability could all become important.

This creates opportunities in companies involved in grid-scale electricity storage, power conversion, and energy management.

The energy transition and the AI infrastructure boom are increasingly becoming the same investment story.

---

4. The Five-Year Framework: Where the Economics Accrue

The objective is not to find twelve companies that all benefit from AI.

It is to understand how to capture multiple layers of infrastructure spending while avoiding excessive exposure to any single technology or financing model.

I would divide the opportunity into four primary categories.

Category 1: Cash-generative hyperscalers

Names I watch most closely: Microsoft, Alphabet, Amazon.

In my framework, these are the foundation.

They have established commercial relationships, diversified revenue streams, enormous cash-generating businesses, and the ability to finance infrastructure expansion.

What each brings:

• Microsoft: enterprise distribution and Azure.
• Alphabet: proprietary TPUs, Google Cloud, and relatively attractive valuation.
• Amazon: AWS infrastructure scale and custom silicon.

I particularly like Alphabet's strategic positioning because it is competing across multiple layers of AI infrastructure rather than relying on a single accelerator architecture.

Category 2: Compute and networking

Names I watch most closely: NVIDIA, Broadcom, Arista Networks.

NVIDIA remains the dominant accelerated-computing ecosystem.

However, the next five years will likely involve increasing architectural diversification.

Hyperscalers are investing in custom accelerators to reduce dependence on general-purpose GPUs and potentially improve cost per inference operation.

Broadcom is particularly interesting in this environment because of its custom ASIC and networking exposure.

Arista provides exposure to the networking infrastructure required to connect increasingly large computing clusters.

I would rather think of this layer as NVIDIA plus custom silicon than bet exclusively on one hardware architecture.

Category 3: Electrical infrastructure, thermal management, and grid construction

Names I watch most closely: Vertiv, Eaton, Quanta Services.

This is the category I believe many traditional AI investors are underweight.

Consider the economics.

A hyperscaler can purchase GPUs, but it still needs electrical distribution, cooling, grid connections, and data center construction.

These requirements are not optional.

The key companies:

Vertiv (VRT)

Vertiv provides critical power and thermal-management infrastructure for data centers.

Its exposure includes power management, liquid cooling, and integrated infrastructure solutions.

Its forward valuation is particularly interesting relative to the growth opportunity.

Eaton (ETN)

Eaton operates across electrical power management, distribution, and industrial infrastructure.

The data center buildout creates demand for its electrical equipment and power management products.

Its competitive advantage is that it participates in broader electrification demand, not exclusively AI.

Quanta Services (PWR)

Quanta provides the physical infrastructure required to expand electricity transmission and distribution networks.

If utility companies and data center developers must build new substations, transmission lines, and electrical connections, Quanta participates in that investment cycle.

I particularly like this category because the grid expansion cycle could continue even if the market experiences periods of slower GPU demand.

Category 4: Electricity generation

Name I watch most closely: Constellation Energy (CEG).

Constellation is an interesting way to gain exposure to existing nuclear generation capacity.

Its strategic advantage is that it operates a large fleet of established power-generating assets.

Unlike an SMR startup, it does not need to prove that its underlying reactor technology can work commercially.

Its challenge is valuation, power-market regulation, and the risk that future electricity price expectations become excessively optimistic.

I would also monitor GE Vernova for its exposure to generation equipment, gas turbines, and electrification infrastructure.

However, valuation discipline matters here.

---

5. Valuation: What Are Investors Actually Paying?

The following forward P/E ratios are approximate market snapshots using Stock Analysis data at the November 6, 2026 market close.

These are consensus-estimate multiples, not guaranteed earnings outcomes. Different financial data providers may calculate forward P/E differently.

Sources: Microsoft, Amazon, Alphabet, NVIDIA, Broadcom, Arista, Vertiv, Eaton, Quanta, Constellation.

There are several interesting observations.

First, NVIDIA's forward multiple is considerably lower than that of several infrastructure-related companies.

This does not automatically make NVIDIA cheap. Earnings forecasts can change rapidly, and semiconductor cyclicality remains a meaningful risk.

Second, Broadcom offers exposure to custom silicon at a relatively low earnings multiple.

Third, Vertiv and Eaton are trading at multiples that are not obviously excessive relative to their strategic exposure to data center construction and electrification.

Fourth, Arista and Quanta command higher multiples. Their growth expectations are already reflected more aggressively in their valuations.

Finally, CoreWeave and Nebius do not have meaningful forward P/E ratios because they are not yet generating consistent GAAP profitability.

For CoreWeave, enterprise value is also substantially greater than market capitalization, reflecting the importance of financing obligations in evaluating the equity.

I would not use forward P/E to compare profitable hyperscalers with unprofitable neoclouds.

For the latter, debt-adjusted enterprise value, future operating cash flow, dilution, and infrastructure utilization are more important.

---

6. Positioning: How I Would Think About the 2026–2031 Cycle

This is a framework for a dedicated thematic view, not a replacement for an otherwise diversified long-term portfolio, and not a recommendation for anyone's specific situation. How much exposure, if any, makes sense depends on your own goals, time horizon, and risk tolerance.

Rather than precise weights, I would group the names by conviction and risk.

Why this structure?

The hyperscalers sit at the center because they are the financial anchors. They have the broadest revenue bases and can finance enormous infrastructure projects.

Compute and networking capture demand for GPUs, specialized accelerators, and the infrastructure that connects AI clusters.

Electrical infrastructure, grid construction, and electricity generation deserve far more weight in an AI infrastructure thesis than most technology investors give them, because I expect physical constraints to remain relevant throughout much of the period.

The neoclouds sit in the speculative tier deliberately. There is enormous upside potential, but the capital structure and refinancing risks are too significant to treat them as core.

Keeping some cash or short-term Treasuries provides flexibility during market corrections.

Building exposure over time

In a cycle like this, I would not commit to a theme in a single transaction.

Building exposure gradually, in several steps over months rather than all at once, reduces the risk of buying at a point of peak optimism about infrastructure spending.

The more established companies come first in that sequence, with power infrastructure and the speculative tier built more slowly.

I would also reassess valuations after every earnings cycle.

What would raise or lower conviction

The most important numbers to monitor each quarter:

1. Hyperscaler CapEx versus operating cash flow: Are infrastructure investments producing corresponding cloud revenue growth?
2. Backlog conversion: Are contractual commitments becoming actual revenue?
3. Neocloud debt and dilution: Is revenue growth improving financial health, or simply creating a larger financing requirement?
4. GPU utilization and pricing: Are specialized clouds maintaining attractive economics as hardware supply expands?
5. Power availability: Are contracted electricity and data center projects becoming operational?
6. Hardware economics: Are customers migrating toward custom ASICs and more efficient inference architectures?

These metrics matter considerably more than another announcement about a record-breaking GPU cluster.

---

7. The Biggest Risks to This Thesis

A credible investment thesis needs to acknowledge the possibility that the underlying assumptions prove wrong.

There are five material risks.

1. AI monetization disappoints.

Companies can spend hundreds of billions building infrastructure, but if enterprise customers do not generate meaningful economic returns from AI, the willingness to pay for compute could decline.

This is the most important macroeconomic risk.

2. Inference efficiency improves faster than demand.

Model compression, quantization, specialized accelerators, and improved inference software could dramatically reduce the amount of computing required per task.

Total AI demand could still increase, but infrastructure revenue per unit of computing activity could fall.

3. Hyperscaler capital expenditure overshoots demand.

A significant increase in data center capacity could eventually produce excess computing supply.

If this happens, GPU rental prices and cloud margins could decline.

4. Electricity projects encounter delays.

Grid interconnections, transformers, nuclear projects, and transmission infrastructure all involve lengthy development processes.

A data center construction project without reliable electricity is a stranded investment.

5. Interest rates and financing conditions deteriorate.

This is especially important for neoclouds.

Companies with large debt balances, collateralized equipment, and substantial lease commitments are much more sensitive to changes in financing conditions.

A 2027 or 2028 credit-market downturn could accelerate consolidation across the industry even if underlying AI demand remains healthy.

---

The Investment View

My base case for 2026–2031 is that the AI infrastructure industry becomes substantially larger, more geographically distributed, and more capital intensive.

The spending commitments already demonstrate the scale.

Microsoft expects approximately $170 billion in calendar-year 2026 CapEx. Amazon expects approximately $220 billion. Alphabet expects $195–205 billion, while Meta expects $135–145 billion (including principal payments on finance leases).

Together, these four companies are signaling approximately $720–740 billion in 2026 capital expenditure.

These budgets include infrastructure beyond AI alone, and each company defines CapEx slightly differently, so they should not be interpreted as pure AI spending. Nevertheless, they demonstrate the extraordinary scale of the investment cycle.

Microsoft, Amazon, Alphabet, and Meta have all disclosed these investment plans or expectations.

The cumulative infrastructure investment opportunity over five years is therefore comfortably a trillion-dollar-scale theme.

But I would make one important distinction.

I am not betting that every company associated with AI will become a great investment.

I am betting that the infrastructure required to support commercially useful AI will become an increasingly essential part of the global economy.

My highest conviction is in Microsoft, Alphabet, Amazon, Broadcom, Vertiv, and Eaton.

NVIDIA remains essential, but I would not assume GPUs will capture all future AI economics.

I would also pay much closer attention to electricity generation and grid infrastructure than the average technology investor.

As for the neoclouds, I expect consolidation. There will be winners, but I would rather own the companies supplying the indispensable infrastructure than become dependent on the financial survival of every operator renting GPUs.

The most important question to ask over the next five years is not simply who will build the smartest AI model, or who will sell the most GPUs.

It is this:

When every major technology company wants to run millions of autonomous AI agents simultaneously, who owns the scarce resources required to make that possible?

The companies controlling electricity, infrastructure, networking, and economically viable compute will have considerably more negotiating power than companies simply competing to sell another generation of servers.

By 2031, the most valuable AI infrastructure company may not be the one with the largest GPU cluster.

It may be the one that can deliver the cheapest reliable kilowatt-hour, convert it into the cheapest useful token, and finance the next generation of infrastructure without having to dilute its shareholders to survive.

Data: company filings, earnings releases and calls as linked above; Gartner and IEA forecasts; valuation snapshot from Stock Analysis at the November 6, 2026 close.