Concentration and risk: single points of failure

After five lessons of category-by-category concentration, an aggregate map emerges.

• Taiwan. TSMC's leading-edge fabs, the densest cluster of fabless customers, and a large fraction of advanced packaging.
• South Korea. Samsung and SK Hynix: leading-edge memory and the second-largest logic foundry.
• Japan. Photoresists, silicon wafers, specialty gases, etch and deposition equipment, advanced packaging materials.
• The Netherlands. ASML (sole-vendor EUV) and its lithography supply chain.
• United States. EDA tools, much of the equipment market (Applied Materials, Lam Research, KLA), advanced design IP, and a recovering leading-edge fab presence.
• Germany. ASML's principal optics partner (Carl Zeiss SMT, Oberkochen), specialty chemicals, silicon wafers (Siltronic).
• China. Mature-node manufacturing, packaging, materials, and a substantial assembly-and-test base.

The pattern: each category's leading suppliers concentrate in two or three countries, and the overlap across categories produces a small set of countries on which most of the global supply chain depends. The next step is to express this as engineering risk.

A single point of failure (SPOF) is, in reliability engineering, a component whose failure causes the entire system to fail. The standard remedies are:

• Redundancy — operate two or more instances in parallel, so that one failure does not cascade.
• Diversity — make the redundant instances independent in their failure modes (different vendors, different sites, different designs).
• Buffering — hold inventory that absorbs short-duration failures while the upstream supply restores.
• Graceful degradation — design so that partial loss reduces capacity rather than eliminating it.

Applied to the chip supply chain, the analysis runs:

• The leading-edge logic fab cluster in northern Taiwan is a SPOF for advanced AI accelerators and high-end mobile and PC silicon.
• The ASML EUV scanner production at Veldhoven is a SPOF for the equipment that produces those chips.
• The Japanese photoresist suppliers are SPOFs for the materials that go into those scanners' output.
• The Carl Zeiss SMT optics plant is a SPOF for the mirrors inside the scanners.

These are statements of system topology, not predictions about probabilities. Whether and how often a SPOF actually fails is a separate empirical question.

Some of the largest concentrations sit in regions with measurable natural-disaster exposure.

• Taiwan. Located on the boundary between the Eurasian and Philippine Sea plates. Seismic events of magnitude 6 or greater occur multiple times per decade in the broader region. The April 2024 magnitude-7.4 Hualien earthquake disrupted TSMC operations for hours; reported industry losses ran into the hundreds of millions of dollars before recovery.
• Japan. Seismic and tsunami exposure across the Honshu coastline where many materials and equipment suppliers operate. The March 2011 Tōhoku earthquake disrupted silicon wafer, chemical, and specialty material supply for months.
• South Korea. Lower seismic exposure than Taiwan or Japan but with concentrated industrial sites whose failure modes are correlated.
• The Netherlands. Low seismic exposure but concentrated to a single industrial campus whose disruption would propagate to global lithography supply.

Fabs have engineered for seismic and disaster events to a high standard — base isolation, vibration damping, redundant utilities — and the historical record shows recovery from major events. The structural point is that the concentration of multiple critical capabilities in seismically active regions is itself a topology feature, independent of any particular event probability.

Major fab capacity is currently being built outside the traditional concentration regions. Examples include:

• TSMC Arizona (Phoenix). Initial fab at 4 nm, with later phases at more advanced nodes. Construction began in 2021.
• TSMC Kumamoto (Japan). 28/22 nm and 12/16 nm production with majority TSMC equity and Japanese government cost participation.
• TSMC Dresden (Germany). 28/22 nm and 16/12 nm specialty-process production, partnership with Bosch, Infineon, NXP.
• Samsung Taylor (Texas). Leading-edge logic capacity announced for the mid-2020s.
• Intel Ohio. Leading-edge logic fab construction announced, with completion timelines repeatedly adjusted.
• GlobalFoundries Malta (New York) and Singapore capacity expansions for trailing-edge specialty nodes.

These projects are structural hedges: they create capacity at the relevant node in regions geographically uncorrelated with the original concentration. The hedge value depends on yield ramp, sustained operating cost, and customer qualification — none of which is automatic. Public subsidies (US CHIPS Act, EU Chips Act, Japanese government participation) cover a portion of the capex differential between the new geographies and the established ones.

Redundancy is not free. A duplicate fab built outside the concentration zone produces the same wafers as the original; combined output is not double what is needed. The economic problem is straightforward:

$\text{capacity}_\text{combined} = \text{capacity}_\text{original} + \text{capacity}_\text{redundant}$

If the market demands less than the combined capacity, both fabs operate below break-even. If only the original is needed, the redundant fab is a fixed cost without revenue.

Three mechanisms address this:

• Subsidies. Public funding covers a portion of the capex and operating losses, treating the redundancy as a public good (security, supply-chain resilience).
• Long-term contracts. Customers commit to take volume from the redundant fab at premium prices, accepting higher cost in exchange for assured supply.
• Differentiated products. The redundant fab specializes in something the original does not — automotive-grade, lower-volume specialty, particular packaging — that has its own demand.

The arithmetic implies that geographically diversified manufacturing capacity is structurally more expensive per wafer than fully concentrated capacity. The question is not whether redundancy raises cost but how the additional cost is allocated between producers, customers, and taxpayers.

Holding inventory of finished chips and critical materials absorbs short-duration supply disruptions.

• Finished-chip inventory. Hyperscalers and major OEMs hold weeks to months of inventory of their highest-volume parts. Strategic buyers held more after the 2020–2022 shortage cycle than before.
• Wafer-bank arrangements. Foundries hold partially processed wafers (bank wafers) at intermediate steps to absorb late-stage demand shocks.
• Material reserves. Some governments and firms maintain reserves of critical materials (rare gases after the 2022 neon disruption, specific specialty chemicals).

The limits of inventory:

• Time. A multi-week disruption is absorbable; a multi-month or multi-year disruption is not, because inventory exhausts.
• Specificity. Inventory protects against shocks to known parts. New designs need new fab runs that inventory cannot substitute for.
• Working capital. Holding inventory ties up capital that could fund R&D, capex, or returns. The opportunity cost limits how much inventory a firm holds.
• Obsolescence. Chip generations turn over rapidly; inventory of two-generation-old parts can be worthless before it is needed.

Inventory is a buffer, not a substitute for capacity diversification. The two strategies complement each other, with inventory covering short-term shocks and diversified capacity covering long-term ones.

The supply side is concentrated. So is the demand side.

• Hyperscale cloud and AI buyers. A small set of firms (AWS, Google, Microsoft, Meta, ByteDance, Alibaba, Tencent, Oracle) accounts for a large share of the demand for leading-edge AI accelerators and the supporting silicon. Their capex announcements move multi-quarter foundry allocations.
• Major mobile OEMs. Apple and Samsung's mobile divisions, plus a handful of Chinese OEMs (Xiaomi, Oppo, Vivo), account for most leading-edge mobile SoC volume.
• Major PC and gaming. A small set of GPU and CPU buyers anchors the second-largest leading-edge segment.

The demand side's concentration produces its own structural effects. A foundry that loses one major fabless customer can struggle to fill that capacity, because the next customer is also large and has its own preferred supplier. The bilateral concentration — small number of foundries serving a small number of major buyers — produces multi-year volume commitments that resemble offtake agreements more than spot purchases. The negotiation dynamics between the two sides shape both pricing and capacity-allocation decisions in ways that aggregate market statistics do not show.

Demand-side concentration also produces correlated demand shocks. When the dominant buyers all rotate capex up or down simultaneously, the foundries see a more volatile order book than the underlying end-market demand would imply.

The six lessons of this cursus describe a system: three business models (fabless, foundry, IDM); one wavelength-limited printing process; thousands of supporting equipment, material, and chemical inputs; a small set of jurisdictions whose policies can act on the chokepoints; multi-decade catchup campaigns whose ingredients are knowable in advance; and a customer base concentrated enough to mirror the supply side.

The deliberate omissions:

• No prediction of how policy or industry will evolve over any specific horizon.
• No claim about which firm or country will lead a future node.
• No advocacy of a particular policy mix as correct.
• No characterization of geopolitical actors as protagonists or antagonists.

These were left out because they are contested, dated, or both. The structural framework presented here — concentration as compounded R&D, R&D as compounded learning curves, learning curves as compounded yield — outlasts the particular numbers in any given year.

What the framework supports: reading news about the industry with an understanding of which claims describe structural features and which describe contingent events; estimating the probable response of one chokepoint to a change in another; and recognizing the difference between a binding constraint and an apparent one. Whether and how that understanding informs a particular policy or business decision is left to the reader.