From Semiconductors to Nuclear Power — The AGI Infrastructure Value Chain

>

TL;DR — The AGI infrastructure value chain comprises five layers: Energy-Chips-Cloud-Models-Apps

>

– $600B pouring into AI infrastructure in 2026 alone — nuclear restarts, data center construction rush

– Korea dominates 87% of the HBM market, anchoring Layer 2 (chips) — Layer 1 (energy) remains the weak link

– Even after DeepSeek’s efficiency breakthrough, physical infrastructure investment has reached an irreversible scale

>

Read time: ~10 min

Intro

The contours of the AGI infrastructure value chain are emerging. While Davos debated whether AGI arrives in one year or ten, something else was already in motion.

Ground is being broken, power lines are being laid, and nuclear contracts are being signed. Anthropic CEO Dario Amodei captured the dynamic at Davos: “The bottleneck is no longer intelligence — it’s chips and factories.”

NVIDIA CEO Jensen Huang defined AI on the same stage as “the largest infrastructure build in human history” (WEF). Both statements point to the same reality.

The AGI timeline debate will continue. But regardless of its outcome, physical infrastructure investment has already reached an irreversible scale. In 2026 alone, $600B is projected to flow into AI infrastructure (CarbonCredits).

Today, we deconstruct the AGI infrastructure value chain layer by layer and map where each nation stands within it. Jensen Huang’s five-layer cake is our starting point.

Background: Jensen Huang’s Five-Layer Cake

JENSEN HUANG’S 5-LAYER AI INFRASTRUCTURE

L5 Applications — AI agents, SaaS

L4 AI Models — LLMs, foundation models

L3 Cloud/Data Centers — Hyperscalers

L2 AI Semiconductors — GPU, HBM, CoWoS

L1 Energy/Power — Nuclear, SMR, cooling

At Davos in January 2026, Jensen Huang compared AI infrastructure to a five-layer cake (NVIDIA Blog). Each layer builds on the one below.

Layer 1 is energy. The heaviest, slowest to build, and first to be needed. AI “generates intelligence in real time,” and sustaining that intelligence requires continuous power supply.

Layer 2 is chips and computing infrastructure. GPUs, HBM, AI-specific semiconductors. Huang assessed that the US is “generationally several steps ahead” at this layer.

Layer 3 is cloud data centers — the vessels that house the chips. This is also the bottleneck Huang emphasized most — building a data center in the US takes approximately three years. Permits, power contracts, construction crews, cooling systems — hundreds of procedures chained sequentially.

Layer 4 is AI models. LLMs, multimodal models, inference engines. Huang estimated the US leads by about six months in frontier models — but this is the layer where the gap is narrowing fastest.

Layer 5 is applications. The products consumers and enterprises actually use. The top of the cake, but one that can only stand if the four layers below are solid.

Huang’s core argument was straightforward: building all five layers simultaneously demands physical infrastructure at a scale fundamentally different from previous IT revolutions. Plumbers, electricians, steelworkers, cooling engineers — many of the jobs this AI revolution creates are not white-collar. That is the point.

Five-Layer Cake Status at a Glance

AGI Infrastructure Value Chain: The Investment Surge

Numbers tell the story fastest. In 2024, Amazon, Microsoft, Google, and Meta spent a combined $200B in capex — a 62% year-over-year increase (Deloitte).

Across the full AGI infrastructure value chain, the 2026 projection is 3x that: $600B, of which $450B goes directly to AI infrastructure. Capex tripled in two years.

Through 2030, cumulative data center investment is forecast to reach $3T (ZDNet Korea). An amount equivalent to roughly 15x South Korea’s annual GDP, flowing into global data centers within a decade.

The Power Crisis: Nuclear Returns

One of the biggest beneficiaries of this investment wave is nuclear power. The reason is simple: AI data centers require stable, 24/7 power. Wind and solar cannot guarantee that stability. Nuclear can.

According to the IEA, US data center power demand is projected to rise from 200 TWh today to 640 TWh by 2035 — a 3.2x increase (IEA). Globally, demand grows from 176 TWh in 2023 to 325–580 TWh by 2028.

To close this gap, Big Tech is signing nuclear contracts directly.

Microsoft invested $1.6B to restart the Three Mile Island nuclear plant (835 MW) — reopening a decommissioned nuclear facility specifically to power AI data centers. The contract spans 20 years.

Meta went further: a 1.1 GW nuclear deal with Constellation Energy (20 years), plus additional contracts with Vistra, TerraPower, and Oklo to secure a total of 6.6 GW of nuclear capacity by 2035 (Tortoise Capital).

The US Department of Energy announced $2.7B in support for domestic uranium enrichment services in January 2026. AI infrastructure is pulling even the nuclear fuel industry into its orbit.

The Situational Awareness report (IIIa. Racing to the Trillion-Dollar Cluster) mapped this power demand escalation by tier. The largest AI cluster in 2026 requires approximately 1 GW — equivalent to Hoover Dam or one large nuclear reactor. By 2028: 10 GW (an entire small US state). By 2030: 100 GW — 20% of total US electricity generation.

If Layer 1 collapses, the other four layers collapse with it. Energy has transformed from a basic utility into a strategic asset in the AI competition.

The Construction Bottleneck: Three Years

Even with capital available, speed is the constraint. As Huang noted, building a data center in the US takes roughly three years. Power contracts, site permits, transformer procurement, construction labor — all chained sequentially.

According to The Meridiem, private data center construction spending in the US runs at approximately $41B annually — rivaling total state and local government spending on transportation infrastructure (The Meridiem). AI data centers are crowding out roads.

Japan’s Play: AI as Diplomatic Currency

Japan has made a notable move. In early 2026, Japan pledged $550B in AI infrastructure investment in the US (Construction Dive). Of this, $332B is earmarked for energy infrastructure.

As TheByteDive previously analyzed (AI Agent Update), the first tranche of Japan’s US investment ($36B) comprises gas power plants ($33B), crude oil facilities ($2B), and synthetic diamonds ($600M). The last item is noteworthy.

Synthetic diamond has a thermal conductivity of 2,000 W/m-K — five times that of copper. It is the ideal cooling material for dissipating the massive heat generated by AI chips. Japan’s AI diplomacy with the US spans the full stack, from energy to chip cooling materials.

Domestically, Japan is committing $330B over 10 years (public-private combined) to AI infrastructure and quadrupled its budget for Rapidus, the domestic foundry startup (Digitimes).

Where Key Nations Stand in the AGI Infrastructure Value Chain

Where does each nation sit in this five-layer cake? The short answer: some lead in Layer 2 (chips) as global powerhouses, face constraints in Layer 1 (energy), and are early-stage in Layer 3 (data centers).

HBM: Samsung and SK Hynix

The core of Layer 2 (chips) is HBM (High Bandwidth Memory). HBM feeds data to GPUs at ultra-high speed during AI computation. If the GPU is the brain, HBM is the ultra-fast short-term memory attached directly to it — reducing the idle time that consumes 70–80% of GPU operating cycles.

Korea effectively monopolizes this market. As of 2026: SK Hynix 63% + Samsung 24% = Korean companies hold 87% of the global HBM market (SK Hynix Newsroom).

SK Hynix is set to supply approximately 70% of HBM4 volume for NVIDIA’s next-generation Rubin GPU platform (TrendForce). NVIDIA defines AI chip standards; SK Hynix supplies the memory essential to those chips.

Market size is exploding. The total HBM market in 2026 is projected at $54.6B — a 58% year-over-year increase (BofA). SK Hynix’s HBM revenue alone is expected to surge from $9.4B in 2025 to $21.1B in 2026.

Next-Gen Memory: The Rise of HBF

The next generation is already in development. HBF (High Bandwidth Flash) — HBM’s successor. While HBM is DRAM-based ultra-fast memory, HBF stacks NAND flash in an HBM-like configuration to achieve both high capacity and high speed. SK Hynix is co-developing with SanDisk, targeting mass production in early 2027. Samsung is pursuing an independent development path.

As TheByteDive’s HBM analysis detailed, Korea’s hold on the HBM-to-HBF memory semiconductor innovation chain is a defining variable in the AGI infrastructure value chain.

Nuclear/Energy

Layer 1 (energy) remains a vulnerability. Korea leads the world in semiconductors, but the power infrastructure needed to manufacture those semiconductors and run AI data centers has not kept pace (econmingle).

The government’s 11th Basic Plan for Long-Term Electricity Supply and Demand explicitly includes AI data center power demand as a driver for building two additional large-scale nuclear reactors. This is not simply energy policy — it is a survival strategy for remaining in the AGI infrastructure value chain.

Korea’s nuclear technology is proven: UAE Barakah plant exports, and a global top 3–4 ranking in SMR (Small Modular Reactor) development. Export potential exists.

Supply chain risks loom, however. Korea holds only 56 days of rare earth reserves. Compare that to Japan’s unofficial 20-year stockpile. If the US-China conflict intensifies and China blocks gallium and germanium exports, HBM production faces a direct hit.

Data Centers

At Layer 3 (cloud data centers), Korea is still in its infancy. Insufficient power infrastructure, limited cooling-capable sites, and complex permitting create compounding constraints.

Opportunity exists, though. Korea’s geographic position as a Northeast Asian hub could make it a viable data center alternative to Japan, Taiwan, and Singapore. The government is moving to reform data center power pricing and permitting processes.

The headline domestic investment is SK Hynix’s Yongin cluster — originally budgeted at roughly $86B, now expanded fivefold to approximately $430B. While this is a semiconductor production cluster, its infrastructure could serve as a domestic AGI infrastructure anchor.

in the AGI infrastructure value chain. Investors should watch companies resolving bottlenecks in the L1 (energy) and L2 (chips) layers.

should watch companies resolving bottlenecks in the L1 (energy) and L2 (chips) layers.

Conclusion

Whether the AGI timeline follows Amodei’s 1–2 years or Hassabis’s 5–10 years, Jensen Huang’s five-layer cake is being stacked right now. $600B is moving in a single year, nuclear plants are restarting, and Japan is deploying $550B in AI diplomacy.

Korea occupies a decisive position in Layer 2 (chips/HBM) of this AGI infrastructure value chain. The fact that Korean companies manufacture 87% of the memory semiconductors that power the world’s AI infrastructure means Korea is an indispensable node in the AI infrastructure war.

Bottom line. In Jensen Huang’s five-layer cake, Korea is already entrenched as the core Layer 2 (chip) supplier — the question is how quickly it catches up on the other four layers, especially Layer 1 (energy).

Takeaway for professionals. Understanding the AGI infrastructure value chain as five layers — energy-chips-cloud-models-apps — sharpens career direction. Semiconductor materials and process engineers, nuclear and energy specialists, data center design and operations teams — these roles are positioned for direct benefit from the AI revolution. The moment you shift from seeing AI as a “software problem” to a “physical infrastructure problem,” opportunities that were previously invisible come into view.

Frequently Asked Questions (FAQ)

What is the AGI infrastructure value chain?

The AGI infrastructure value chain refers to the physical infrastructure needed to build Artificial General Intelligence. Jensen Huang’s five-layer model explains it: Layer 1 Energy (nuclear/power), Layer 2 AI Semiconductors (GPU/HBM), Layer 3 Cloud Data Centers, Layer 4 AI Models (LLMs), Layer 5 Applications. All five layers must be built simultaneously for AGI infrastructure to be complete.

Why does AI need nuclear power?

AI data centers require massive, stable power 24/7. US data center power demand is projected to grow 3.2x to 640 TWh by 2035, and wind/solar cannot guarantee that stability. This is why Microsoft (Three Mile Island restart), Meta (6.6 GW nuclear capacity), and others are signing nuclear contracts directly.

Why is HBM critical to AGI infrastructure?

HBM (High Bandwidth Memory) feeds data to GPUs at ultra-high speed during AI computation. GPUs spend 70–80% of their operating time waiting for data — HBM reduces this bottleneck. As of 2026, SK Hynix (63%) and Samsung (24%) control 87% of the global HBM market, making Korea the dominant force at this critical node.

Where does Korea stand in the AGI infrastructure value chain?

Korea leads the world in Layer 2 (AI semiconductors) with 87% HBM market share. In Layer 1 (energy), nuclear technology is proven but power infrastructure expansion is needed. Layer 3 (data centers) is nascent, though major investments like the Yongin SK Hynix cluster (~$430B) are underway.

Recommended Reading

The Day After AGI — Two AI Titans at Davos Reveal Diverging Paths

HBM Deep-Dive — The $54.6B Market That Solved GPU Idle Time

AI Agent Update — $3T Data Centers, MCP Security, De-aging

Sources

Largest Infrastructure Buildout in Human History: Jensen Huang on AI’s Five-Layer Cake (NVIDIA Blog, 2026-01)

Jensen Huang brings a 5-layer AI pitch to Davos (Quartz, 2026-01)

Davos 2026: Nvidia CEO Jensen Huang on the future of AI (WEF, 2026-01)

AI Demand to Drive $600B From the Big Five for GPU and Data Center Boom by 2026 (CarbonCredits, 2025)

Energy supply for AI (IEA, 2025)

Nuclear Infrastructure Investing: Why AI’s Power Demand Changes Everything (Tortoise Capital, 2025)

Can US infrastructure keep up with the AI economy? (Deloitte, 2025)

Data Center Boom Crowds Out Roads (The Meridiem, 2025-12)

Bechtel, Kiewit snare piece of $550B US-Japan AI infrastructure deal (Construction Dive, 2026)

Japan quadruples chip and physical AI spending (Digitimes, 2025-12)

Outlook 2026 Series IV: The AI Power Endgame (MacroMicro, 2026)

2026 Market Outlook — HBM-led Memory Supercycle (SK Hynix Newsroom, 2026)

Samsung-SK Hynix $50B+ Investment Race 2026 (Global Economic, 2026-01)

Korea’s Semiconductor Power: A Reckoning (econmingle, 2026)

WEF 2026: What the Most Powerful AI Leaders Say About AGI (Context Studios, 2026-01)

IIIa. Racing to the Trillion-Dollar Cluster (Situational Awareness, 2024)

AGI’s Last Bottlenecks (AI Frontiers, 2025)

3 Nuclear Power Stocks Set to Flourish in 2026 (Nasdaq, 2026)

Disclaimer: This article is for informational purposes only and does not constitute investment advice. All data cited is sourced from publicly available reports and filings.