For every $100 spent on new artificial intelligence infrastructure, roughly $50 ends up in semiconductors, according to an estimate from BNP Paribas Equity Research and SankeyMATIC. The chart puts numbers to a reality that is shaping data center expansion: buying accelerators is only part of the bill. Networking, power, cooling, buildings, and grid connections make up a much larger investment.
The AI infrastructure spending breakdown in 30 seconds
- Of every $100 in AI CAPEX, the chart allocates $50 to semiconductors, $20 to power, $15 to networking, $7.5 to cooling, and another $7.5 to construction and facilities.
- Accelerators alone account for $25, and memory another $15.
- Electrical infrastructure gets more money than cooling itself.
- The chart is a high-level estimate, not a universal accounting for every data center.
- The breakdown helps identify which parts of the chain actually absorb AI investment.
The picture is especially useful because it separates two ideas that often get mixed together. Big tech operators’ capital spending isn’t just about buying GPUs. Putting those GPUs to work requires servers, high-speed networks, optical transceivers, electrical systems, cooling, and buildings built to handle densities that were unusual just a few years ago.
A word of caution about the figures is also worth adding. Large hyperscalers don’t publish a uniform line item called “AI CAPEX,” so it can’t be assumed that all of their capital spending goes to artificial intelligence. Combined 2026 investment estimates for Microsoft, Amazon, Alphabet, and Meta already exceed $700 billion, but that includes infrastructure and other assets beyond AI.
Semiconductors absorb 50 of every 100 dollars
The Sankey diagram’s first big takeaway is spending concentration.
Of the initial $100, $50 goes to semiconductors. Within that amount, BNP allocates roughly $25 to accelerators, $15 to memory chips, and $10 to CPUs, other chips, and server manufacturing.
That means accelerators account for roughly a quarter of all the investment considered in the model.
The flow continues because manufacturing an accelerator or a memory chip requires its own set of industrial processes. The chart sets aside about $8 for wafer fab equipment (WFE).
From there, several categories appear: deposition, lithography, etching, inspection, packaging, and other processes.
The Sankey diagram helps visualize something an aggregate figure can hide. When a hyperscaler buys thousands of AI systems, the money starts moving through an extensive industrial chain that runs from chip designers to semiconductor equipment makers.
Memory deserves particular attention.
AI systems need enormous amounts of bandwidth to continuously feed the accelerators. Hence the growing importance of HBM (High Bandwidth Memory), which sits alongside GPUs and other advanced accelerators — a bottleneck that has turned suppliers like Micron into a critical piece of the AI supply chain.
Then there’s advanced packaging. Integrating accelerators, memory, and other components into systems capable of exchanging massive amounts of data requires increasingly complex manufacturing technologies.
That’s why AI spending ends up reaching companies that are seemingly several tiers below the vendor selling the final accelerator.
Networking, power, and cooling take the other half
The second part of the chart is almost as interesting as the chips.
BNP estimates that 15 of every 100 dollars go to networking equipment. Within that category, roughly $3 goes to network processors, $2 to cabling, $4.5 to switches, and $5.5 to optical transceivers.
The growth of this category can also be seen in NVIDIA’s results. In its fiscal 2027 first quarter, the company reported $14.8 billion in quarterly data center networking revenue, up 199% year over year — part of a broader shift in which NVIDIA’s data center business now accounts for 92.5% of its total revenue.
The figure helps explain why networking has become another major area of competition around AI.
Thousands of accelerators working together need to exchange data with very low latency and enormous bandwidth. The larger the cluster, the more important the network connecting its nodes becomes.
But the Sankey diagram sets aside even more money for another category: power.
Of every $100, about $20 is tied to electrical infrastructure.
BNP allocates roughly $6.5 to power distribution, $10 to grid connection and independent generation, and another $3.5 to on-site electrical facilities within the data center itself.
It’s a particularly relevant figure because power availability is increasingly determining where and when new facilities can be built.
A server can be bought in months. Securing tens or hundreds of additional megawatts of power capacity can take much longer, a constraint that is already colliding with the grid as transformers grow scarce for new data centers.
That’s why the AI infrastructure debate is progressively shifting from GPU availability toward the combined availability of compute, networking, and power.
Cooling adds another $7.5
The next block corresponds to cooling, at $7.5 of every $100.
The chart splits this amount among cold plates, coolant distribution units (CDUs), chillers, and cooling towers.
New generations of accelerators have significantly raised the thermal density of racks. That’s driving the adoption of liquid cooling in facilities that traditionally relied mainly on air-based systems.
The remaining $7.5 goes to facilities and construction.
BNP splits it roughly between land acquisition and site preparation, building structure and envelope, and interior fit-out and auxiliary infrastructure.
The full breakdown looks like this:
| Destination | Dollars per 100 |
|---|---|
| Semiconductors | 50 |
| Power | 20 |
| Networking equipment | 15 |
| Cooling | 7.5 |
| Facilities and construction | 7.5 |
The table shows an interesting feature: the building represents a relatively small fraction compared with the technology installed inside it.
An AI-ready data center can be a considerable engineering project, but servers, accelerators, networking, and electrical systems end up absorbing most of the capital.
The chart also shows where the bottlenecks are
BNP’s Sankey diagram is useful precisely because it shouldn’t be read as an exact formula applicable to any project.
The chart itself describes its figures as high-level estimates subject to change due to the rapid evolution of costs, demand, and supply.
An existing data center installing new servers will have a different cost structure than a facility built from scratch. The breakdown will also change depending on the type of accelerator, available power, network architecture, rack density, or cooling system.
But the proportions help convey the economic scale of the AI race.
Aggregate CAPEX from major U.S. operators keeps growing — a trend that has already made hyperscale companies commit close to $2 trillion and reshaped who buys chips. A recent estimate puts the four largest hyperscalers’ 2026 investment between $720 billion and $745 billion. Including Oracle, the figure would be around $835 billion.
Other estimates use different groups of companies and arrive at different figures. Moody’s, for instance, had raised its forecast in May to $785 billion for six hyperscalers during 2026.
Not all of that money can be directly labeled AI investment. It does, however, reflect the scale of the infrastructure build-out cycle that artificial intelligence is accelerating.
And BNP’s chart lets us see where a considerable share of that capital can end up.
The GPU sits at the center of the conversation, but behind it, it needs memory, packaging, servers, and manufacturing equipment. Then it needs switches and optics to communicate, power to run, and cooling systems to remove heat. Finally, it needs a building and an electrical connection capable of supporting it all.
The race for AI infrastructure makes a lot more sense once you follow those $100 all the way to the end.
Frequently asked questions
How much of AI infrastructure spending ends up in chips?
According to the BNP Paribas Equity Research model shown in the chart, roughly 50 of every 100 dollars goes to semiconductors. Accelerators account for about $25 and memory chips for another $15.
How much does the electrical infrastructure of an AI data center cost?
The Sankey diagram allocates roughly $20 of every $100 to power, including power distribution, grid connection, independent generation, and on-site electrical facilities.
How much weight does networking carry in AI investment?
The model estimates around $15 of every $100, split among network processors, cabling, switches, and optical transceivers.
Do these figures apply to every data center?
No. BNP presents them as high-level estimates and warns that they can vary due to changing costs and supply-and-demand conditions. The actual breakdown will depend on each project and on whether it’s a new facility or an expansion of an existing one.
Sources:
- BNP Paribas Equity Research / SankeyMATIC, Where does $100 of AI Capex flow?
