TSMC and the AI Chip Shortage: A Better Capacity Planning Model

Suppose a buyer secures 20,000 accelerators but receives only half the required high-bandwidth memory. Those chips are inventory. Suppose the complete servers arrive at a site waiting 18 months for a substation. They are expensive furniture. The AI chip shortage becomes intelligible only when the unit of supply changes from silicon dies to usable accelerator hours.

Asianometry's response to claims that TSMC is braking the AI boom uses the MIT Beer Game to explain why demand swings grow as they travel through a supply chain. That is the right starting point. The sharper operating conclusion is that no participant should optimize its own capacity forecast in isolation. Hyperscalers, chip designers, foundries, memory vendors, system builders, and utilities must synchronize commitments with long and unequal lead times.

TSMC deserves scrutiny because it manufactures most leading AI accelerator logic and supplies the CoWoS packaging used to join logic with HBM. Calling its caution a failure skips the 2022 lesson: capacity built for apparently durable demand can become a multibillion-dollar depreciation problem before it reaches volume. The remedy is better contracting and visibility across the chain, plus credible foundry alternatives. Exhorting one supplier to believe harder does not manufacture an EUV scanner, qualify an HBM stack, or energize a data hall.

One AI Server Contains Several Capacity Markets

An AI accelerator shipment sits at the end of several specialized production systems. Start with a chip design and electronic design automation tools. Add leading-edge logic wafers, reticles, photoresist, deposition and etch equipment, metrology, and hundreds of process steps. Known-good logic dies then meet HBM stacks on an advanced package with an interposer or redistribution structure. The package needs substrates, assembly, and test. Systems add network adapters, switches, optical modules, power shelves, cooling manifolds, firmware, and mechanical integration.

The data center adds another chain: land, permits, transformers, switchgear, generation or grid capacity, cooling plant, water where applicable, and staff. A delay anywhere can strand everything upstream.

This chain has three properties that software planning often misses:

1. Capacity is specific. A deposition tool configured and qualified for one process step is not a general compute pool. A packaging line qualified for one package size cannot instantly build another.
2. Yield is capacity. Installed tools do not equal good output. New processes and larger packages require learning before the line reaches planned yield and cycle time.
3. Timing determines value. A component delivered six months before the rack or six months after the powered site can both destroy return on capital.

TSMC's own disclosures show the coordinated expansion. It said in April 2025 that it was working to double CoWoS capacity during 2025. Its 2025 annual report records $122.42 billion in revenue and $55.21 billion in net income while describing continued work on CoWoS, SoIC, and silicon photonics. By the first quarter of 2026, management was guiding capital spending toward the top of a $52 billion to $56 billion range.

Those numbers document aggressive current investment without establishing that every prior decision was optimal. Capacity can be too late even when the eventual spending is enormous. The decision question is what evidence was available when an earlier commitment had to be made, and which capacity would have been useful once built.

Use a constraint register for every large deployment:

Report the lowest dated number in the chain. Aggregate accelerator orders hide the real constraint.

The Bullwhip Punishes Fast Followers and Cautious Suppliers

The Beer Game developed at MIT demonstrates how order delays and local decisions amplify a modest change in customer demand. Retailers see empty shelves and raise orders. Wholesalers see larger orders and raise theirs further. By the time new supply arrives, the original shortage may have passed. The MIT System Dynamics Society's Beer Distribution Game is simple because the behavior appears without malicious actors or bad technology.

Semiconductors add long construction and qualification cycles. A foundry must decide on buildings and tools before customers can give reliable volume orders. Equipment vendors must decide whether to expand factories before the foundry places firm orders. Material vendors sit another step back. Each firm sees a noisy order signal mixed with real demand, safety stock, double booking, and bargaining tactics.

The COVID cycle is a useful test. Automakers cut chip orders during shutdowns, then tried to restore them as vehicle demand recovered. Suppliers had reassigned capacity. Customers placed larger orders to protect themselves, which made final demand harder to read. In parallel, PC and consumer demand surged and then fell. TSMC spent heavily into the shortage, only to face weaker utilization at some mature and advanced nodes when the cycle turned.

ChatGPT appeared in November 2022 during that correction. Five months later, TSMC's April 2023 call recorded a customer's sudden request for much more CoWoS capacity. In July 2023, management described AI accelerators as about 6 percent of revenue and forecast rapid growth, but the company still expected full-year revenue to decline. A board approving years of capacity on that evidence had to price two outcomes: an AI surge and another short-lived customer forecast.

By 2024, system readiness added uncertainty. NVIDIA disclosed a Blackwell production issue that required a mask change, and rack-scale systems had to qualify liquid cooling, networking, firmware, and power delivery. A good package built before a system fix can sit in inventory. Scaling the package line before design stability would not guarantee earlier revenue.

Forecasting still deserves scrutiny. A useful forecast must include:

• Firm system orders separated from reservations and duplicate requests
• Customer financing and data center power dates
• Package design maturity and system acceptance status
• End-user workload demand, utilization, and revenue per accelerator hour
• Cancellation rights and who absorbs idle capacity
• Signals from equipment, materials, memory, and utility suppliers

The supplier with the longest lead time should not carry all forecast risk. Long-term agreements can fund expansion, but customers must pay for meaningful reservations or minimum volumes. Otherwise optimistic orders are free options on a foundry's balance sheet.

More Foundries Help, but Shared Suppliers Still Set the Pace

TSMC's concentration creates real economic and geopolitical risk. A qualified Samsung or Intel process gives chip designers bargaining power, regional diversity, and additional output. Public investment through the US CHIPS program is intended to expand that industrial base, and the National Institute of Standards and Technology tracks awards and program goals.

Yet a second foundry is not interchangeable capacity. A leading accelerator designed around one process uses that foundry's transistor models, design rules, libraries, packaging, and yield behavior. Porting can require physical redesign, new masks, validation, firmware work, and a separate ramp. Even when a chip designer maintains two designs, the resulting products may differ in cost and performance.

Foundries also buy from a concentrated equipment chain. ASML supplies EUV lithography. Applied Materials, Lam Research, Tokyo Electron, KLA, and a small group of peers dominate many process categories. Leading HBM comes primarily from SK hynix, Samsung, and Micron. More foundries place more orders into those shared suppliers, so competition changes resilience faster than it changes short-term total output.

Multiple foundries provide resilience, learning, and long-run capacity. Instant fungibility remains unrealistic.

A buyer can act on that distinction:

At product design: fund a true alternate path for selected workloads. This may be a second foundry, a different accelerator vendor, or a model architecture that can run on several backends. Paper compatibility is not enough. Benchmark and validate it.

At procurement: reserve packages, memory, networking, and rack integration together. A logic-wafer agreement without matching HBM and packaging is incomplete.

At operations: keep models and serving layers portable enough to redirect some demand. Portability has a cost, so apply it to valuable fallback capacity rather than forcing every optimization onto the lowest common denominator.

At finance: value diversity as insurance. A backup with a higher unit price can be rational if it prevents months of lost service. Compare the premium with expected interruption cost.

Foundry share should fall as qualified alternatives earn production. Customers still have to fund the qualification work.

Power and Silicon Are Bottlenecks on Different Calendars

Arguments that choose either power or chips compress a project schedule into one moment. A data center can be silicon-constrained in 2026 and power-constrained in 2028. Another site may have an energized shell and no accelerators. The binding constraint changes as deliveries and approvals move.

The International Energy Agency's 2026 update reports that data center electricity demand grew 17 percent in 2025, while demand from AI-focused facilities rose 50 percent. Its central case has total data center use rising from 485 TWh in 2025 to about 950 TWh in 2030. The IEA also identifies tighter supply for transformers, gas turbines, chips, and other components. Its earlier work estimated that grid risks could delay around 20 percent of planned data center projects.

Power equipment and semiconductor capacity respond differently. A packaging line may add tools inside an existing plant. A new leading-edge fab takes longer. A major transmission connection can require studies, rights of way, substation work, and generation changes across several organizations. Gas turbines and large transformers have their own factory queues.

Track each constraint as a dated capacity curve:

• Good accelerator packages per quarter
• Accepted racks per quarter
• Network ports and optical links per quarter
• Energized IT megawatts by site and date
• Cooling capacity at required rack density
• Staffed operating capacity

Then calculate deployable accelerators as the minimum supported by all curves. Do this monthly. A permanent label such as "chip constrained" encourages teams to keep buying the last scarce item after the bottleneck has moved.

Efficiency belongs in the supply plan too. Better quantization, batching, scheduling, model routing, and network utilization can create usable capacity faster than a new fab or substation. These measures leave physical demand intact while providing a buffer until long-lead assets arrive. They also lower exposure if demand changes.

Contract for Learning, Not Perfect Forecasts

Nobody can forecast AI demand, process yield, model efficiency, and power delivery five years out with useful precision. Capacity planning should make forecast errors survivable.

Use three commitment horizons.

The locked horizon covers components already inside manufacturing lead time and facilities under construction. Require firm quantities, named configurations, delivery windows, and financial responsibility for changes.

The adjustable horizon covers capacity that can still move among products or dates. Set bands rather than one number, with option prices that reflect the supplier's cost of holding flexibility.

The learning horizon records scenarios without pretending they are orders. Tie the next commitment to evidence such as workload revenue, accelerator utilization, package yield, rack acceptance, and utility milestones.

Each capacity review should answer five questions:

1. Which component limits deployable accelerator hours this quarter?
2. Which component will limit them when today's purchase arrives?
3. What demand evidence changed since the last commitment?
4. Who pays if an order is canceled or a dependent component is late?
5. Which substitution has been tested in production conditions?

Add a pre-mortem for surplus. If model efficiency cuts required compute, a competitor ships a better architecture, or application revenue disappoints, decide where capacity can be redirected. A wafer line, packaging tool, data hall, and power contract have different reuse values.

For the next board-level AI infrastructure proposal, replace the accelerator count with a quarterly table of usable hours, synchronized inputs, confidence ranges, and stranded-capital exposure. Approve the next reservation only when its dependent memory, package, rack, cooling, and power dates fit the same deployment window.