AI Infrastructure Bottleneck Tracker: Power, HBM & Networks

The biggest cloud platforms have a capex envelope of $720 billion to $740 billion for 2026 (tracker, with Meta's range from its third-quarter results). None of that money runs an AI workload until the slowest part of the system shows up.

A usable AI cluster needs chips, high-bandwidth memory, packaging, networking, power equipment, cooling, a grid connection and a finished building, all at once. The slowest qualified piece sets the pace for everything else, and power is becoming that piece. A simple hypothetical further down shows what one late piece costs everyone else.

Watch it play outThe slowest qualified piece sets the pace for everything else

A usable AI cluster needs chips, HBM, packaging, networking, cooling, a building, power equipment and a grid connection, all at once.

Six pieces arrive. The cluster still can't run a single workload.

Power is becoming the slowest piece, and it holds up everything else.

Only when the last piece lands does any of the capex envelope run an AI workload.

Replay

The short version

• The International Energy Agency (IEA) sees data-center electricity use rising from about 485 TWh in 2025 to 950 TWh in 2030. Grid connections and equipment are the near-term limit.
• HBM supply is a chain of yields, not wafer starts. Samples, qualification, volume production and revenue shipments are four different milestones.

Watch it play outData-center power demand could nearly double, through a narrow grid

The IEA puts data-center electricity use at about 485 TWh in 2025.

By 2030 it could reach 950 TWh, nearly double. Circle areas show the size.

All of it has to come through grid connections and equipment, the near-term limit.

Replay

• Broadcom's $10.8 billion of fiscal Q2 2026 AI semiconductor revenue shows the spending reaches well beyond GPUs.
• Bottlenecks move. When one eases, the next one (packaging, power or customer deployment) takes over.

Capex is the money companies spend on long-lived assets like data centers. The hyperscaler AI capex tracker follows the spending itself; this page tracks what stands between that spending and capacity that can actually run.

The bottleneck dashboard

This is a research sequence, not a ranking of stocks. A supplier can profit from scarcity while its customers earn poor returns. A bottleneck can also attract so much investment that it destroys its own pricing power.

Power is becoming the system constraint

The IEA estimates data-center electricity use could rise from about 485 TWh in 2025 to 950 TWh in 2030, while electricity use from AI-focused data centers triples. Its 2026 energy and AI update also flags near-term limits in grid connections, energy equipment, approvals and technology supply chains.

Data-center electricity use could nearly double by 2030

2004006008001,000 TWh0

485 TWh

950 TWh

20252030

The U.S. Energy Information Administration's 2026 data-center analysis models a wide long-term range for server electricity use, because installed servers, power draw and efficiency are all uncertain. That range is a useful guard against treating any single forecast as guaranteed revenue.

For a project, the number that matters is energized capacity with a reliable delivery date and a contract that assigns construction and stranded-asset risk. Track five milestones separately:

Read closely

A campus update lists five milestones. Tap the ones that can support a live workload.

A hypothetical AI campus update

The campus has requested interconnection capacity, contracted its power, permitted generation and grid equipment, energized its first megawatts and has compute running at productive utilization.

Tap the phrase you think is the tell.

Spotted it.Not that phrase. Only the last two milestones can support a live workload.

• “The campus has requested interconnection capacity” A place in the queue, not power.
• “contracted its power” Contracted power can be years from delivery.
• “permitted generation and grid equipment” Permission to build, not power at the site.
• “energized its first megawatts” Power is flowing to the site.
• “and has compute running at productive utilization” The workload is live.

Explain every phraseHide the other explanations

1. Requested interconnection capacity
2. Contracted power
3. Permitted generation and grid equipment
4. Energized megawatts
5. Compute running at productive utilization

Only the last two can support a live workload. The AI data-center power guide explains how utilities, generators, equipment makers and developers split the economics.

HBM is a chain of yields

High-bandwidth memory (HBM) is stacked memory that sits beside an AI chip. It has to survive DRAM fabrication, die thinning, through-silicon-via processing, stacking, base-die integration, packaging, testing and customer qualification. More wafer capacity doesn't produce the same percentage increase in saleable HBM attached to a chip.

Every HBM stack must survive a chain of yields

1. DRAM fab to Thin and TSV (Through-silicon vias)
2. Thin and TSV (Through-silicon vias) to Stack (Plus base die)
3. Stack (Plus base die) to Package, test, qualify (Saleable HBM)

The most common mistake is collapsing four milestones into one:

• Samples reached a customer.
• A product finished qualification.
• High-volume production began.
• Finished stacks shipped in revenue volume.

Micron's fiscal second-quarter 2026 materials discuss volume shipments for one configuration while describing sampling for another. So "HBM supply" needs a generation, a configuration, a customer status and a date.

Read the full HBM yield and supply-chain guide. The HBM demand calculator turns an accelerator fleet into memory capacity, though its output is not a supplier revenue forecast.

Packaging decides usable output

Chips and HBM aren't a deployable system until they are joined, tested and connected. Advanced packaging needs specialized equipment, substrates or interposers (the base layers that link chip and memory), process control and known-good parts.

That creates lopsided risk: a shortage of one high-value part can leave other expensive parts sitting idle. Expanding packaging takes time, and the mix shifts as designs use bigger packages, more HBM stacks and faster interconnects.

One short part leaves expensive parts idle

1. Accelerators (Ready, waiting) to Packaging capacity (The short part)
2. HBM stacks (Ready, waiting) to Packaging capacity (The short part)
3. Packaging capacity (The short part) to Deployable system (Delayed)

Announced capacity isn't enough. Watch finished package shipments, customer qualification, cycle time, yield, and whether chip vendors turn component supply into complete systems.

Networking turns chips into a cluster

A rack of accelerators can't train or serve large models well if the network can't move data, keep work in sync and recover from failures. Bigger clusters mean more switches, network cards, optical parts, cables and software.

Broadcom reported $10.8 billion of fiscal second-quarter 2026 AI semiconductor revenue, driven by custom accelerators and AI networking, in its official results.

Networking revenue alone doesn't prove the customer earns a return. The test is whether bigger clusters improve training time, inference throughput, reliability and cost per useful result. A faster link feeding underused compute is still part of a weak system.

Networking revenue is not the customer's return

Proves the supplier sells

• Broadcom's $10.8B fiscal Q2 2026 AI semiconductor revenue
• More switches, optics and cables per cluster

vs

Proves the customer earns

• Faster training time
• Higher inference throughput
• Better reliability
• Lower cost per useful result

Cooling and construction decide time to service

Denser racks push more heat through a smaller footprint. Liquid cooling, power distribution, backup, water supply and local climate become design limits, not accessories.

Construction risk shows up in several ways:

• Powered land is unavailable where demand is strongest.
• Grid studies and permits take longer than the servers' product cycle.
• Transformers, switchgear, turbines or cooling gear arrive late.
• A building is finished before enough power is available.
• A customer changes the cluster design after site work starts.

Track construction in progress and management's timing for placing assets in service. A rising asset balance can mean productive expansion or a growing queue of unpowered projects. Revenue timing tells them apart.

Revenue timing tells growth from a queue

1. Construction in progress rises to Energized (Placed in service)
2. Construction in progress rises to Waiting for power (Unpowered queue)
3. Energized (Placed in service) to Revenue starts
4. Waiting for power (Unpowered queue) to Cost pushed later

Member research: Equipment usually starts depreciating only once it is placed in service, so a queue of unpowered projects pushes the cost into later earnings. Issue 03 finds four hyperscalers put almost $4 into Q1 infrastructure for every $1 of depreciation, and tracks where the margin test lands next.

Worked example: the slowest link sets the output

Hypothetical example: a cloud platform buys enough chips for 100 MW of computing. Every other layer has to keep up.

Power caps a 100 MW chip order at 60 MW

Chips purchased100 MW

HBM supplied90 MW

Packaging completed80 MW

Power energized at the site60 MW

Usable cluster = the smallest number: 60 MW. Hypothetical example.

Now price the wait. Say the equipment cost $10 billion and the company's cost of capital, the yearly return its investors expect for tying up money, is 10%.

Make a guess

$10 billion of equipment, 40% of it idle while the site waits for power, at a 10% cost of capital. What does each month of waiting cost?

Tap your guess

$0–10mYour guess$10–25mYour guess$25–50mYour guessOver $50mYour guess

$01234567890123456789mActual

Spot on.Close, a little low.Close, a little high.Way off, too low.Way off, too high. It was $33m.

$10bn × 40% = $4bn idle, × 10% = $400m a year, ÷ 12. That's before a dollar of AI revenue.

Two lessons:

• Fixing the wrong layer adds nothing. Lifting packaging from 80 MW to 100 MW leaves the cluster at 60 MW while power is the limit.
• Delay is expensive even when nothing breaks. Every month the site waits for power costs about $33 million in this example, before a dollar of AI revenue. The ROIC vs. WACC calculator shows how that kind of drag erodes value creation.

Fixing the wrong layer adds nothing

Expand packaging or power and watch the usable cluster

Packaging completed

80 MW100 MW

Power energized

60 MW80 MW100 MW

Usable cluster60 MW80 MW80 MW60 MW80 MW90 MW

0 MW+20 MW+20 MW0 MW+20 MW+30 MWin line with today's 60 MW clusterabove today's 60 MW clusterabove today's 60 MW clusterin line with today's 60 MW clusterabove today's 60 MW clusterabove today's 60 MW cluster

today's 60 MW cluster 60 MW

60 MW90 MW

How bottlenecks move

One constraint can hide another. If HBM is scarce, networking demand can look weak simply because fewer complete systems ship. Once memory improves, the next limit may be packaging, power or customer deployment.

When one bottleneck eases, the next takes over

1. HBM scarce to Networking looks weak (Fewer systems ship)
2. Networking looks weak (Fewer systems ship) to Memory improves
3. Memory improves to Next limit (Packaging, power or deployment)

That's why single-point forecasts fail. Track the system as a sequence:

1. Hyperscaler demand and financing
2. Chip and memory production
3. Package integration and networking
4. Site, power and cooling readiness
5. Software deployment and utilization
6. Revenue, margin and cash return

The AI infrastructure investment map ties each stage to its business model and failure mode.

Try it yourself

A 10-minute exercise with the free tools:

1. Open the AI capex dataset and pick one hyperscaler, such as Microsoft. Note how its spending has moved.
2. Open the AI power exposure model. Enter an announced campus size, then lower the completion probability and add a delay to see how much load survives.
3. Use the HBM demand calculator to turn an accelerator fleet into HBM stacks.
4. Skim the EIA's data-center analysis and note how wide its range of server electricity scenarios is.
5. Copy the dashboard table and fill in the "good sign" and "warning sign" columns with what you found.

You should now be able to answer: which layer is the binding constraint right now, and what signal would tell you it has moved?

Our read

In an infrastructure boom, the winners are often the companies that control the binding constraint. That edge is temporary unless the supplier also has durable technology, strong customer relationships and capital discipline. Today the binding constraint is shifting toward power and grid connections.

The current bottlenecks would ease if several of these showed up together:

• Efficiency cuts total compute demand faster than usage grows.
• HBM and packaging output catches demand without new qualification delays.
• Grid connection and equipment delivery times normalize.
• Hyperscalers slow new commitments because existing capacity is enough.
• Cluster utilization weakens after new sites come online.

The first four would reduce scarcity. The fifth is more serious: it would mean the physical build outran economic demand.

Questions readers ask

What is the biggest AI bottleneck right now?

Power and grid connections are becoming the system constraint. But bottlenecks move: when one eases, the next limit may be packaging, power or customer deployment, which is why this tracker watches every layer.

Why should stock investors care about bottlenecks?

The companies that control the binding constraint often capture the best economics. That edge fades if scarcity attracts enough new capacity, so it lasts only with durable technology, strong customer relationships and capital discipline.

How can I tell if a new data center is actually running?

Look for energized megawatts, management's timing for placing assets in service, and revenue. A rising construction-in-progress balance without energization can mean a growing queue of unpowered projects.

Could the AI build-out end up too big?

That is the fifth warning signal above. If cluster utilization weakens after new sites come online, the physical build has outrun economic demand, which is more serious than any single shortage.

Go deeper with member research

A bottleneck only matters to investors if someone can charge for it. Member reports test who can:

• 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.
• The HBM Capacity Trade: Micron’s HBM capacity thesis as the evidence stood on December 18, 2025, with timing, valuation and later results kept separate.
• GE Vernova and the value of delivering dependable power: Gas-power and grid economics as of January 29, 2026, and why it was already a recognized trade by then.

Limits of this data

• Forecasts, capacity plans, qualifications and company guidance can change after the research cutoff.
• Electricity forecasts span wide ranges and should be read as scenarios.