AI Infrastructure Investing Map: Where the Money Flows in 2026

Microsoft, Amazon, Alphabet and Meta plan $720 billion to $740 billion of combined 2026 capex, the money spent on long-lived assets like data centers (tracker, with Meta's range from its third-quarter results). That is the starting pool of demand, and it is unusually large.

Watch it play outFour platforms pour their 2026 capex into one pool of demand

Four platforms sit at the top of the chain: Microsoft, Amazon, Alphabet and Meta.

Their 2026 capex, the money spent on long-lived assets like data centers, streams into one pool.

Combined: $720 billion to $740 billion. That is the starting pool of demand, and it is unusually large.

Replay

But AI infrastructure investing isn't one bet on GPUs. The money flows through chips, memory, networking, data-center gear, electricity, construction and financing. The best-positioned business is the one that captures scarce economics without taking on more capital, customer-concentration or obsolescence risk than its share price already reflects. Later, a two-line example shows how a company can grow revenue 30% and still disappoint shareholders.

Watch it play outThe money cascades through seven layers, not one bet on GPUs

At the center, the cloud platforms sign the contracts and pay the bill.

Their money flows out through chips, memory, networking, data-center gear, electricity, construction and financing.

Seven layers, and each makes money differently and fails differently. AI infrastructure investing isn't one bet on GPUs.

Replay

The short version

• The chain has seven layers, from cloud platforms down to construction and finance. Each makes money differently and fails differently.
• Demand is real: Microsoft reported continued capacity constraints, and Amazon reported faster AWS growth. Amazon also showed a large free-cash-flow cost from building it.

Watch it play outData-center electricity use is set to roughly double by 2030

Picture data-center electricity use in 2025 as one block.

The IEA projects it roughly doubling between 2025 and 2030: a second block as large again.

That power has to arrive through the grid, so power and grid companies sit on the map too.

Replay

• The IEA projects data-center electricity use roughly doubling between 2025 and 2030, so power and grid companies sit on the map too.
• Scarcity creates pricing power only while the constraint lasts. A great business can still be a poor investment if the price assumes scarcity forever.

The AI capex tracker keeps the company guidance and accounting caveats behind the headline number. This page maps where the money goes and what has to happen before it becomes a shareholder return.

Check the claimThree claims to test as you read

ClaimAI infrastructure investing is a bet on GPUs.Tap to checkMythThe money flows through seven layers, from chips and memory to power, construction and finance. Each makes money and fails differently.Flip backClaimIf a supplier's revenue grows 30%, shareholders win.Tap to checkMythGrowth can eat cash. In the worked example below, revenue rises 30% while free cash flow turns negative.Flip backClaim"Sold out" HBM means memory is shipping today.Tap to checkMythA sold-out claim may describe future contracted capacity, not finished shipments today.Flip back

The AI infrastructure value chain

This is a causal map, not a list of recommendations. Each layer needs different evidence.

Platforms: the buyer has to earn a return too

Hyperscalers, the giant cloud platforms, sit at the center. They sign the contracts, build the capacity and run the products that ultimately pay the bill.

Microsoft's fiscal third-quarter 2026 earnings call reported continued capacity constraints, rapid Azure growth and concrete progress bringing GPUs online. Amazon's second-quarter results showed faster AWS growth alongside a large free-cash-flow cost from property investment.

Read those facts together. Strong demand proves the need for capacity. Cash and margin decide whether the company keeps enough of the value.

Demand proves the need; cash decides the return

1. Strong demand (Capacity constraints, faster AWS) to Capacity build (Large free-cash-flow cost)
2. Capacity build (Large free-cash-flow cost) to Cash and margin (How much value is kept)

Compute: flexibility vs. specialization

General-purpose GPUs benefit from broad software support and can handle changing workloads. Custom accelerators can lower costs on huge, stable workloads and give a platform more control over supply.

Expect a mixed fleet, not a single winner. Frontier training tends to reward flexibility, while repetitive inference (running a trained model to answer requests) can justify specialization. The unit that matters is the whole system: chip, memory, network, power, software, reliability and time to deploy.

Expect a mixed fleet, not a single winner

General-purpose GPUs

• Broad software support
• Handle changing workloads
• Frontier training rewards flexibility

Flexibility

vs

Custom accelerators

• Lower costs on huge, stable workloads
• More control over supply
• Repetitive inference can justify them

Specialization

Track performance per dollar on real workloads, outside adoption, software migration cost, and how quickly each new generation replaces the last. Peak specs are not a return metric. The custom chips versus GPUs guide covers the workload-level math.

Memory and packaging: usable supply matters

HBM, high-bandwidth memory stacked beside the chip, matters because accelerators must move enormous amounts of model data quickly. It's hard to make because several dies must be fabricated, stacked, connected, packaged with the chip, tested and qualified.

HBM only counts as supply once it clears every step

1. Fabricate dies (Several per stack) to Stack and connect (Dies bonded together)
2. Stack and connect (Dies bonded together) to Package (Beside the accelerator)
3. Package (Beside the accelerator) to Test and qualify (Usable supply)

Scarcity can lift supplier prices and margins, but it also invites new capacity and pushes customers to diversify their suppliers. A "sold out" claim may describe future contracted capacity, not finished shipments today.

Track the HBM generation, stack height, qualification stage, packaging availability, customer platform and shipment timing. The HBM bottleneck guide separates wafer capacity from qualified output, and the HBM demand calculator converts chip assumptions into physical memory demand.

Networking: the cluster is the product

Large AI systems need chips to talk at high speed and recover when parts fail. Demand spreads across switches, optical links, cables, network cards, signal-processing chips and software.

Broadcom's fiscal second-quarter 2026 results show custom accelerators and AI networking becoming a material revenue pool outside the merchant GPU market.

The core risk is concentration. A few platform buyers can drive extraordinary growth, then change architecture, delay a cluster or bring design work in-house. Backlog is strongest when it turns into diversified shipments without unsustainable concessions to customers.

Data-center systems: density creates new spending

AI chips raise rack power and heat, which increases the value of power distribution, liquid cooling, backup, monitoring and design expertise.

Denser racks create new spending

1. AI chips (More power per rack) to Higher rack power and heat
2. Higher rack power and heat to New spending (Power gear, liquid cooling, backup)

The best evidence is order growth tied to funded sites, delivery schedules that match when customers get power, and margins that survive new competitors adding capacity. Watch concentration here too: several equipment vendors can look diversified by product while selling into the same four or five hyperscaler budgets.

Energy and grid: contracted is not energized

The International Energy Agency's 2026 update projects data-center electricity use roughly doubling between 2025 and 2030, with faster growth in AI-focused facilities. It names grid connections, energy equipment, planning and approvals as near-term limits.

Many kinds of companies can benefit, with very different economics:

• A regulated utility may grow its rate base (the investment it's allowed to earn a return on) but face limits on passing costs to households.
• A generator may win a long contract but carry construction, fuel and counterparty risk.
• An equipment supplier may earn scarcity margins that normalize when new capacity arrives.
• A developer may own valuable powered land but carry debt while waiting for a tenant and a connection.

Separate requested, contracted and energized megawatts. The AI data-center power guide explains the difference.

Read closely

A developer's update mixes three kinds of megawatts. Tap the one that can run servers today.

A hypothetical campus update

Our campus has requested a multi-gigawatt grid connection, signed long-term power contracts and energized its first phase.

Tap the phrase you think is the tell.

Spotted it.Not that phrase. Separate requested, contracted and energized megawatts. Only energized capacity can run a workload now.

• “Our campus has requested a multi-gigawatt grid connection” Requested capacity is a place in the queue, not power.
• “signed long-term power contracts” Contracted power can be years away.
• “and energized its first phase” Energized megawatts are the only ones that can run a workload today.

Explain every phraseHide the other explanations

Member research: Each of those four business models turns grid demand into cash differently. The GE Vernova case study tests gas-power and grid economics as the evidence stood on January 29, 2026.

Construction and finance: the hidden balance sheet

The cycle extends beyond reported capex. Campuses can involve leases, joint ventures, project debt, private credit, supplier financing and long purchase commitments.

Financing can speed deployment, but it moves risk rather than removing it. A long lease can lower near-term cash capex while creating a fixed obligation. Project debt can match an asset's life yet turn fragile if power or tenants arrive late.

Read cash capex, finance leases, purchase commitments, construction in progress and debt together. The AI capex accounting comparison shows why headline totals can't be compared without those reconciliations.

Reported capex is only one layer of the commitment

1. Cash capexThe headline number
2. Finance leasesLower cash capex now, fixed obligation later
3. Purchase commitmentsLong supplier contracts
4. Construction in progressBuilt but not yet in service
5. Project debt and private creditFragile if power or tenants arrive late

Read together

Worked example: when growth disappoints shareholders

Part 1: growth that eats cash. Hypothetical supplier with round numbers. Free cash flow is the cash left after paying for new equipment: operating cash flow minus capex.

Make a guess

Revenue grows 30% to $13.0bn. Operating cash flow rises to $2.4bn and capex jumps to $2.5bn. Where does free cash flow land?

Tap your guess

Below zeroYour guessDown, but positiveYour guessUp from $1.0bnYour guess

−$0123456789.0123456789bnActual

Spot on.Close, a little low.Close, a little high.Way off, too low.Way off, too high. It was −$0.1bn.

$2.4bn of operating cash flow minus $2.5bn of capex. Revenue grew 30%, yet the business produced less cash for shareholders.

Actual−$0.1bn

Year 1 free cash flow$1.0bn

Revenue grew 30%, yet the business produced less cash for shareholders because serving the growth took so much new capital.

Revenue rose 30% while free cash flow turned negative

Operating cash flow, Year 1$2.0bn

Operating cash flow, Year 2$2.4bn

Capex, Year 1$1.0bn

Capex, Year 2$2.5bn

Free cash flow, Year 1$1.0bn

Free cash flow, Year 2−$0.1bn

Hypothetical supplier with round numbers. Revenue rose from $10.0bn to $13.0bn.

Part 2: scarcity priced as permanent. The price-to-earnings ratio (P/E) is the share price divided by a year of earnings per share. Hypothetical stock: $5 of earnings per share, valued at 40 times earnings.

Earnings fell 20%, but the share price halved, because the multiple fell too. That is what "a great business can be a poor investment" looks like in numbers.

A great business can still be a poor investment

Pick earnings per share and the multiple investors pay

Earnings per share

$5$4

P/E multiple

40×25×

Share price$200$125$160$100

0.0%−37.5%−20.0%−50.0%in line with Todaybelow Todaybelow Todaybelow Today

Today $200

$100$200

Try it yourself

A 10-minute exercise with the free tools:

1. Open the AI capex dataset and one buyer's page, such as Amazon. Note how capex compares with the cash the business generates.
2. Enter the Part 1 numbers into the capex and free-cash-flow coverage calculator, then try figures from Amazon's second-quarter results.
3. Pick one supplier from the value chain, such as Broadcom or Micron, and decide which layer it sits in.
4. Use the P/E multiple compression calculator to see what a drop from 40 to 25 times earnings would do to its price.
5. Run the company through the five questions below.

You should now be able to answer: which layer does this company sit in, what is its binding constraint, and how long does its share price assume that scarcity lasts?

Our read

AI infrastructure spending creates a long chain of revenue before it creates a satisfying shareholder return. The spending is real; whether each layer keeps its share of the profit is not settled. The research that matters ties every supplier claim back to a funded customer, a shipped system, an energized site, a productive workload and cash.

Five questions to ask before treating exposure to the theme as a thesis:

1. Is the product tied to a binding bottleneck? Scarcity gives pricing power only while it lasts.
2. Is demand turning into shipments or energized capacity? Announcements and backlog come earlier.
3. How concentrated are the customers? Several end markets may still rest on the same hyperscaler budgets.
4. How much capital does growth need? Revenue can rise while free cash flow falls, as Part 1 of the worked example shows.
5. What breaks the thesis? Efficiency, new supply, architecture changes, customer insourcing, regulation and financing can all move the profit pool.

Then check the valuation. Use the bottleneck tracker to see which constraint sets the pace today, and the AI capex ROI scorecard to test whether the buyers earn enough to keep the cycle going.

Questions readers ask

Which layer of AI infrastructure is safest?

None is risk-free; each layer fails in its own way, as the value-chain table shows. The better question is which business sits on a binding bottleneck without taking more capital, customer-concentration or obsolescence risk than its price reflects.

How can a company's revenue rise while its stock falls?

Two ways, both in the worked example. Growth can need so much new capital that free cash flow shrinks, and investors can cut the earnings multiple once they stop believing scarcity will last.

Are the hyperscalers themselves AI infrastructure investments?

They are the buyers at the center of the chain, and they have to earn a return on what they spend. Microsoft reported capacity constraints and rapid Azure growth, while Amazon showed faster AWS growth alongside a large free-cash-flow cost. Cash and margin decide whether they keep enough of the value.

Why does so much depend on four companies?

Microsoft, Amazon, Alphabet and Meta supply the $720 billion to $740 billion demand pool. Even suppliers that look diversified by product can be selling into the same four or five hyperscaler budgets, so one budget cut can ripple through many layers.

Go deeper with member research

The map shows where the money flows. Member reports go deep on the links where profits may concentrate:

• Issue 06: Who keeps the profits from AI's photonics buildout?: A five-company study of optical architectures, manufacturing yields, reliability, cash conversion and valuation, with worked models and the evidence that would break the thesis.
• Issue 02: The AI Capex Shock: Where the next AI infrastructure dollar goes, which bottlenecks keep pricing power, how the boom is financed and what could break first.
• Comfort Systems and the scarce capacity behind AI construction: Whether the backlog and cash conversion looked investable on February 20, 2026, and what the later gains really rewarded.

Limits of this data

• Company guidance, forecasts, capacity plans and market prices can change after the research cutoff.