AI Capital Cycle ④ — Empire's Periphery

There is one number that defines the AI cycle. It is NVIDIA's quarterly data center revenue.

FY23 Q4 (ended January 2023): $4.3B. FY26 Q2 (ended July 2025): $46.7B. An 11x increase. The same business segment of the same company went from low single-digit $B to nearly $50B in ten quarters. 90% of the company's total revenue now comes from that one segment.

This curve says two things. First, *how big* the AI capital cycle is quantitatively — one NVIDIA segment in one quarter makes roughly twice what Lucent made in all of 1999. Second, the curve's quarterly trajectory is a function of *supply constraint* — if there were quarters where NVIDIA shipped less, it wasn't because demand was missing. It was because there weren't enough GPUs to ship.

For GPU supply to flow, two things must unlock simultaneously. HBM and CoWoS. NVIDIA makes neither. Neither is made in the United States. One comes from Korea. The other comes from Taiwan.

So this curve — on the surface NVIDIA's revenue — is actually *the curve of how much Korean and Taiwanese supply unlocked*.

Q1 2025. SK Hynix's memory revenue passed Samsung's for the first time in history. Viewed from KOSPI, this isn't just a quarterly data point. It is the first proof of a 40-year power shift.

That power has a name: HBM (High Bandwidth Memory). The memory stacked 12 or 16 layers high on top of AI training GPUs. Priced at 3-5x normal DRAM. NVIDIA H100 carries 80GB; B100/B200 carries 192GB. More than half of a GPU's essential value comes from the HBM stacked on top.

The HBM market is effectively a duopoly + 1. SK Hynix, Samsung, Micron. But Q2 2025 share was — SK Hynix 62%, Samsung 17%, Micron 21% (Counterpoint Research). A year earlier (Q2 2024): SK Hynix ~50%, Samsung 41%, Micron 9%. In four quarters, Samsung lost 24 percentage points.

What happened? *Samsung failed NVIDIA's HBM3E 12-Hi qualification.* NVIDIA stacks 8-12 HBM modules on each GPU; their reliability determines the entire GPU's reliability. Thermal control, signal integrity, TSV (through-silicon via) yield — Samsung was a generation behind SK Hynix on each. Meanwhile, Micron quietly built share.

The market sees this as "Korea vs Korea." More precisely: *NVIDIA's single qualification decision moved over ₩6 trillion in memory revenue from one Korean company to another*.

If HBM is Korea, CoWoS is Taiwan.

CoWoS (Chip-on-Wafer-on-Substrate) is TSMC's advanced packaging process that mounts a GPU die and HBM stacks on a single interposer. Without it, a GPU does not exist as a single chip. NVIDIA H100, B100, B200, Rubin — every data center GPU passes through CoWoS.

CoWoS is, in practice, sole-sourced from TSMC. Samsung and Intel attempt similar packaging, but neither has meaningful capacity or yield. So how much NVIDIA can *ship* in 2025-26 is defined by TSMC's CoWoS capacity.

The capacity curve:
- End 2024: ~35,000 wafers/month
- End 2025: ~70,000 wafers/month (+100%)
- End 2026: ~130,000 wafers/month (+86%)
- CAGR: ~80%

NVIDIA pre-allocated more than 60% of that capacity (2025-26 basis). Meaning AMD MI300X, Broadcom TPUs, AWS Trainium and other ASICs compete for the remaining 40%. TSMC capex: $40-42B in 2025, $50B in 2026-27 — more than half going to CoWoS expansion.

The implication: the AI cycle's quarterly trajectory is defined not by US capital's will but by Taiwanese factory capacity ramp curves. Big Tech can promise a trillion dollars, but if TSMC can't add capacity, that capital does not convert into quarterly revenue.

To understand this dependence, we have to see one historical pattern. No hegemon in history has ever been self-sufficient within its own territory.

Look at 19th-century Britain. The two core resources of the Industrial Revolution were cotton and opium. Cotton fed the Lancashire mills. Opium was a weapon to close Britain's trade deficit with China (the deficit caused by tea imports). But Britain didn't grow cotton. It imported from American slave plantations. Britain didn't grow opium either. It produced opium in Bengal, India.

The industrial engine of the British Empire depended on American cotton. The trade surplus of the British Empire depended on Indian opium. The moment both dependencies vanished — the American Civil War (1861-65) cut off cotton supply, and post-Sino-Japanese War China regained control of opium — two core legs of British imperial economics collapsed.

The same pattern held for 20th-century American oil hegemony. The US imported 50%+ of its oil through the 1970s. The core supplier was Saudi Arabia. When Saudi Arabia imposed the 1973 oil embargo, American hegemony shook within one quarter.

Now look at American AI hegemony in 2026. Capital is American (Big Tech). Models are American (OpenAI, Anthropic). Chip design is American (NVIDIA, AMD, Broadcom). But HBM is Korean. CoWoS is Taiwanese. Same position as Britain depending on cotton and opium during the Industrial Revolution.

This is not a flaw of hegemony. It is the *structure* of hegemony. A truly strong hegemon does not self-supply within its own territory. When it tries to, the attempt itself becomes the cost of hegemony (CHIPS Act is that attempt). Accepting the dependence and stabilizing it — rather than trying to break it — usually preserves hegemony longer.

Two attempts try to unwind the dependence. One says *let's break NVIDIA's GPU monopoly with other chips* — ASICs. The other says *let's break the Korea-Taiwan supply chain monopoly by building in the US* — CHIPS Act.

The ASIC counter-attack. Broadcom makes Google TPU, Meta MTIA, OpenAI Titan. Marvell makes AWS Trainium2/3, Microsoft Maia. Both are hyperscaler decisions to replace NVIDIA dependence with in-house silicon.

Broadcom AI revenue: FY25 Q4 $6.5B → FY26 Q1 $8.2B (+74% YoY), backlog $73B. CEO Hock Tan stated "line of sight to $100B AI revenue by 2027." If real, Broadcom's AI revenue would be roughly half of NVDA's DC revenue. The ASIC vs GPU split is shifting fast across the industry — ASIC-based AI servers projected to be 28% of the market by 2026 (Trendforce).

But — *ASICs also depend on TSMC packaging in the end*. Broadcom's Google TPU, Marvell's AWS Trainium — all pass through TSMC's CoWoS line. Even when ASICs take share from NVDA, *the Taiwan dependence does not break*. The fish change in the same lake.

The knot CHIPS Act tries to untie is more fundamental. TSMC Arizona Fab 21 (3nm, 2024 production), Samsung Taylor TX (4nm, 2025), Micron Boise/Clay NY (DRAM, 2026-27). Combined $200B+ capex. The intent — produce leading-edge nodes on US soil and reduce Korea/Taiwan dependence.

But the math doesn't work. US new capacity build pace: ~25% per year. AI demand growth pace: ~60% per year. *Demand is growing more than 2x as fast as capacity*. CHIPS Act doesn't unwind dependence — it merely slows the *ratio* slightly. The quantitative conclusion: Korea-Taiwan dependence cannot break within the next 5 years.

From the Korean capital markets perspective, this is good news. The more the US tries to break the dependence, and the more those attempts are delayed, the longer SK Hynix, Samsung, and TSMC keep their pricing power.

The single proposition of this memo is simple — a truly strong hegemon does not self-supply within its own territory.

The British Empire was like that. Its industrial engine depended on American cotton. Its trade surplus depended on Indian opium. The dependence did not weaken the empire — the *capacity to stabilize* that dependence is what extended the empire's life.

American AI hegemony is the same. Every quarter of NVDA's data center revenue is a direct function of Korean SK Hynix HBM shipments and Taiwanese TSMC CoWoS capacity. That fact is not a weakness of hegemony. It is its *structure*.

From an investor's view, two implications.

First, the Korean memory duopoly maintains pricing power while the cycle runs. No matter how hard CHIPS Act tries, as long as demand grows 2x faster than capacity, SK Hynix's 62% HBM share does not shake in the short term. KOSPI's record highs are not coincidence — they are a function of Big Tech capex.

Second, a single catalyst — Samsung HBM4 12-Hi NVIDIA qualification — is the largest single event for the Korean market in 2026. Pass = duopoly restored, Samsung +20-30% catalyst, SK Hynix gives up some share. Fail = SK Hynix locks in monopoly, Samsung memory business in structural crisis. Either way, *NVIDIA decides*. The fact that the largest single decision variable for the Korean market is *an American company's qualification verdict* — that itself shows the structure of dependence more clearly than anything.

The self-revenue circuit Lucent ran in 1999 (Memo 1) was *a capital circuit*. What SK Hynix and TSMC are running in 2026 is *a supply circuit*. American capital must pass through Korean and Taiwanese supply to become American revenue.

*The circuit does not close within America.* Just as it did not for the British Empire.