VALUE CHAIN · AI INFRASTRUCTURE

The AI Infrastructure Value Chain

Everyone's watching the GPU queue. The people actually building AI data centers are watching the transformer queue — and as of 2026, that one's running three to five years long.

7 stagesSilicon to socketUS + Korea coverageChip → foundry → memory → power → cooling → cloud

Ask someone what's holding AI back and they'll say GPUs. That was true in 2023. It stopped being true somewhere around 2025, when Nvidia's shipments caught up with demand and a much less glamorous set of problems took over: where do you plug in 300 megawatts, and how do you stop it from melting the moment you turn it on. This is the AI infrastructure value chain — the physical path from a chip design file to a hyperscaler's live AI data center, in seven stages. Some of these stages are sold out for years. Some are just getting started. We've tagged each one so you can tell which is which at a glance.

How to read this map

Each stage below has a phase badge. Red means the stage is the acute bottleneck right now — multi-year backlogs, sold-out capacity, pricing power shifting to whoever has supply. Amber means tight but moving. Blue means scaling hard and mostly keeping up. Gray means early-stage — real, but still small relative to the rest of the chain. Company profile pages will populate under each stage as SignalsDeck builds them out; check back as the roster grows.

FIG. 01

The chain at a glance

Seven stages, one flow — node color shows how tight each link is right now, not how important it is.

The seven-stage AI infrastructure value chain from chip design to cloud, color-coded by bottleneck severity01CHIP02FOUNDRY03MEMORY04POWER05COOLING06ONSITE07CLOUDChip designFoundryMemoryPower gridCoolingOn-sitepowerCloud↑ the two tightest linksHBM sold out / 3–5yr transformer queue
BottleneckTightScalingEarly-stageDemand side
01
AI chip designFull throttle

The layer everyone actually talks about. Nvidia still designs the GPUs the rest of this entire value chain exists to house, feed, and cool, with AMD pushing a credible second option and Broadcom and Marvell quietly building the custom AI accelerators that Google, Amazon, and Meta design in-house and don't want anyone to notice. This is the stage with the most attention and, at this point, the least mystery — the real story in 2026 has moved three stages downstream, to the actual building that houses the chip.

AMD AMD · NasdaqUS
AMD (NASDAQ: AMD) holds the industry's first confirmed glass substrate roadmap — pilot production in 2026, real products by 2028.
Read more →
Broadcom AVGO · Nasdaq
The company most people have never heard of, quietly designing the custom AI chips running inside Google, Meta, OpenAI, and Anthropic's data centers.
Read more →
Marvell Technology MRVL · Nasdaq
The 'flexible' alternative to Broadcom — the custom chip designer AWS and Microsoft turn to when they want a second source.
Read more →
Nvidia NVDA · NASDAQUS
The GPU maker that became AI's operating system — and the reason every other stage in this chain exists at all.
Read more →
02
Foundry & advanced packagingCapacity-constrained

Someone has to actually manufacture the chip Nvidia only draws, and that's TSMC, running at effectively full allocation on its most advanced nodes and its CoWoS advanced-packaging lines, with Samsung Foundry chasing the same customers and ASE Technology handling a large share of the back-end packaging and testing that turns a finished die into a shippable chip. Advanced packaging — not raw wafer capacity — has been the tighter of the two constraints through 2026, which is the less obvious half of this stage's story.

ASE Technology ASX · NYSE (ADR) / TWSE: 3711
The world's largest chip packaging and testing company — TSMC's overflow valve when its own CoWoS lines can't keep up.
Read more →
TSMC TSM · NYSE (ADR) / TWSE: 2330
The one company every AI chip has to pass through — and the reason 2026's real bottleneck is packaging, not silicon.
Read more →
03
Memory (HBM)Sold out through 2026

High-bandwidth memory is the stack of DRAM sitting right next to the GPU die, and it's arguably tighter than the GPU itself — every major HBM producer's 2026 output was reportedly committed before the year even started. SK hynix holds the largest share of this market and supplies Nvidia directly, Samsung Electronics is racing to close the qualification gap with its own HBM line, and Micron is the US-based third name scaling output as fast as its fabs allow. If GPU supply looks fine on paper but AI servers still aren't shipping, HBM is usually the reason.

Micron MU · NasdaqUS
The only US-headquartered maker of the memory chips sitting inside every major AI accelerator — with its entire 2026 output already sold.
Read more →
Samsung Electronics 005930 · KRXKR
How glass substrate packaging fits into Samsung Electronics' race to win AI chip customers away from TSMC's foundry business.
Read more →
SK hynix 000660 · KRX / Nasdaq (ADR)KR
The company that turned high-bandwidth memory into the tightest, most profitable link in the entire AI hardware stack — and just pulled off the largest US listing by a foreign company in history.
Read more →
04
Power grid equipmentThe real bottleneck

This is the stage that surprised almost everyone. High-voltage transformer lead times stretched from roughly 143 weeks in 2024 to around 160 weeks by early 2026, and switchgear isn't much better — nearly half of the AI data centers planned for 2026 in the US are running late, and the cause is transformers and breakers, not chips. Korean heavy-electrical makers have become unlikely stars of this stage: KR HD Hyundai Electric holds the top share of the North American high-voltage transformer market and is expanding its Alabama plant to keep up, KR Hyosung Heavy Industries is investing over $150M into its Memphis 765kV transformer line, and KR LS Electric just landed its largest single high-voltage transformer order ever from a US utility building out data-center-linked renewable generation, alongside smaller transformer maker KR Sangil Electric. Globally, GE Vernova, Hitachi Energy, and Siemens Energy round out the field — and all of them currently have more orders than they can fill on any reasonable timeline.

GE Vernova GEV · NYSEUS
The 2024 GE spinoff that turned out to own one of the tightest chokepoints in the entire AI buildout: the grid equipment connecting data centers to power.
Read more →
HD Hyundai Electric 267260 · KRXKR
North America's #1 high-voltage transformer supplier, and the clearest single proof point that Korean industrials are now core AI infrastructure suppliers.
Read more →
Hitachi Energy Wholly owned by Hitachi, Ltd. (TSE: 6501)
Not publicly traded, but one of the three companies that effectively decide how fast US AI data centers can get connected to the grid.
Read more →
Hyosung Heavy Industries 298040 · KRXKR
Landed Korea's largest-ever single power equipment export contract in early 2026 — and is investing $150 million in Memphis to defend that lead.
Read more →
LS ELECTRIC 010120 · KRXKR
A 50-year-old Korean electrical company that just crossed 1 trillion won in data center orders, with a Honeywell partnership aimed squarely at US hyperscalers.
Read more →
Sangil Electric 062040 · KRXKR
A small Korean specialty transformer maker where the operating margin (36%) beats every one of its bigger domestic rivals — almost entirely on the back of US exports.
Read more →
Siemens Energy ENR · Frankfurt Stock Exchange (DAX)
Spun off from Siemens in 2020 nearly broke, now sitting on one of the largest order backlogs in industrial history — largely because of AI data centers.
Read more →
05
Integrated power & coolingScaling fast

Once power reaches the building, someone still has to distribute it through UPS systems and pull the heat back out — and GPU racks now running past 100 kilowatts have made air cooling obsolete almost overnight, turning liquid cooling from a niche upgrade into the default. Vertiv is the clearest structural winner here, named Nvidia's official cooling and power partner for its next-generation Vera Rubin platform, with Eaton close behind after its $9.5B Boyd Thermal acquisition and Schneider Electric and nVent rounding out the field. The Korean entrant worth watching is KR LG Electronics, which signed a multi-year, multi-billion-dollar deal to supply cold plates, coolant distribution units, and liquid chillers across Microsoft's global AI data center fleet — a direct hyperscaler pick, not a subcontractor role.

Eaton ETN · NYSE
The 115-year-old power company that just bought its way into liquid cooling, betting the AI buildout gets won grid-to-chip, not stage by stage.
Read more →
LG Electronics 066570 · KRXKR
Best known for refrigerators and TVs, LG's industrial division is quietly building AI data center chillers, liquid cooling, and immersion cooling systems.
Read more →
nVent NVT · NYSE
The decade-long liquid-cooling specialist now defending its head start against much bigger rivals racing to catch up through acquisition.
Read more →
Schneider Electric SU · Euronext Paris
The French industrial giant selling the blueprint, the hardware, and the software for AI data centers all at once — plus, lately, new partners to build them with.
Read more →
Vertiv VRT · NYSEUS
The precision-cooling company that traces back to a 1965 Columbus, Ohio garage business, now Nvidia's official cooling partner for the AI data center era.
Read more →
06
On-site power generationEmerging

The workaround for stage 4's bottleneck: skip the grid queue entirely and make your own electricity on-site. Bloom Energy's solid-oxide fuel cells have become the poster child of this approach, letting data centers generate power directly from natural gas or hydrogen without waiting years for a utility interconnection, while Caterpillar and Cummins supply the backup generators every data center needs regardless of how it gets its primary power. This stage is small relative to the grid today, but it's the fastest-growing release valve for stage 4's multi-year queue.

Bloom Energy BE · NYSEUS
Fuel cell technology built for NASA's Mars program, now selling hyperscalers a way to skip the grid queue entirely.
Read more →
Caterpillar CAT · NYSEUS
For a century Caterpillar sold the machines that build things. Now it sells the generators that keep AI's biggest buildings running when the grid can't.
Read more →
Cummins CMI · NYSEUS
Everyone talks about GPUs. Almost nobody talks about the diesel generators that have to work, or the data center doesn't open at all — and Cummins builds them.
Read more →
07
Cloud & hyperscalersWriting the checks

The demand side, and the reason every stage above this one exists at all. Microsoft, Google, Amazon, and Meta are collectively pushing combined 2026 AI infrastructure spending past $650 billion, with Oracle riding the same wave through its OpenAI-linked cloud buildout. The one to watch from Korea is KR NAVER, which has shifted from an asset-light cloud strategy to building its own AI factory with Nvidia — targeting 200 megawatts by 2028 and gigawatt scale within five to six years, putting it on a genuinely comparable track to the smaller Western neoclouds rather than staying a regional also-ran.

Amazon AMZN · NasdaqUS
How Amazon's AWS-built AI chips connect to the emerging glass substrate industry — and why Trainium is outgrowing today's packaging.
Read more →
Google GOOGL · NASDAQUS
Google is the only major hyperscaler that doesn't pay Nvidia's margins for most of its AI compute — because it designs its own chips, in-house, all the way down.
Read more →
Meta META · NASDAQUS
Meta is trying to do something no other hyperscaler has managed yet: build a chip good enough to train its own frontier AI models, not just run them afterward.
Read more →
Microsoft MSFT · NasdaqUS
The hyperscaler spending $80 billion-plus a year on AI infrastructure — and restarting a dormant nuclear plant just to keep it powered.
Read more →
NAVER 035420 · KRXKR
NAVER runs Korea's most-used search engine — but the thing Jensen Huang flew to Korea to see wasn't the search box, it was the data center.
Read more →
Oracle ORCL · NYSEUS
Two years ago Oracle was a database company nobody associated with AI. Today, more than half its entire contract backlog is tied to a single customer: OpenAI.
Read more →

Frequently asked questions

What is the AI infrastructure value chain?
It's the full physical path from an AI chip design to a running AI data center: chip design, foundry manufacturing and advanced packaging, high-bandwidth memory, the power-grid equipment that feeds the building, the cooling and power-distribution systems inside it, on-site backup generation, and finally the cloud companies operating the finished data center.
Why is power, not GPU supply, the bottleneck in 2026?
GPU shipments largely caught up with demand through 2025. What didn't scale at the same pace was the physical grid: high-voltage transformer lead times run three to five years, and nearly half of the AI data centers planned for 2026 in the US have been delayed for exactly that reason — not chip shortages.
Which Korean companies are in the AI data center supply chain?
On the power-equipment side: HD Hyundai Electric, Hyosung Heavy Industries, LS Electric, and Sangil Electric all supply high-voltage transformers and switchgear to US utilities and data-center developers. On the cooling side, LG Electronics supplies liquid-cooling hardware directly to Microsoft. On the cloud side, NAVER operates its own data centers and is building AI-factory-scale infrastructure with Nvidia.
Where does memory (HBM) fit in the AI hardware stack?
High-bandwidth memory sits physically next to the GPU die inside the same package, feeding it data fast enough to keep the chip busy. It's produced mainly by SK hynix, Samsung Electronics, and Micron, and industry-wide HBM capacity has reportedly been sold out well in advance for 2026 — making it one of the tightest single links in the entire chain.

This page describes public value-chain structure and company positioning for informational purposes only. It is not investment advice, and inclusion in this map is not a recommendation to buy or sell any security. Company facts and figures reflect public reporting as of mid-2026 and may have changed since.