TSMC Pays for Redundancy

If You Only Read One Thing

The chip boom’s most important product is optionality. TSMC’s record quarter gives it the cash to duplicate capacity in Arizona; India’s Semicon 2.0 uses public money to assemble the suppliers, skills and design firms a fab cannot create alone. TSMC Funds the Hedge and India Buys the Missing Layers are two versions of the same industrial bargain.

TSMC Funds the Hedge

TSMC is using the world’s best semiconductor economics to pay for a worse version of itself abroad. That is not a contradiction. It is the price of making its indispensability politically durable.

The foundry reported second-quarter revenue of $40.2 billion, up 33.7% in dollars, while net income rose 77.4% to NT$706.6 billion. Gross margin reached 67.7%; advanced nodes accounted for 77% of wafer revenue, high-performance computing for 66%, and North American customers for 78%. TSMC raised 2026 capital spending to $60 billion-$64 billion and added another $100 billion to its Arizona plans, taking its total U.S. commitment to $265 billion, according to the Associated Press.

Why it matters: The numbers expose the financing mechanism behind semiconductor reshoring. TSMC’s Taiwan network produces scarcity rents through supplier density, accumulated yield knowledge and leading-node scale. Management can spend part of those rents reproducing capacity in a more expensive location, accepting near-term overseas-fab margin dilution in exchange for customer and political insurance.

This is less a retreat from Taiwan than a hedge against the foundry becoming too concentrated to remain commercially usable. The customers demanding geographic redundancy are already the source of nearly four-fifths of TSMC’s revenue. Arizona places production closer to their political jurisdiction while Taiwan retains the deeper manufacturing system that makes leading-edge economics possible. Production follows demand, but the operating advantage stays anchored in the original cluster.

Room for disagreement: Arizona may become ordinary capacity rather than an insurance premium. AI demand is real, customers want domestic supply and a larger U.S. cluster could eventually close the cost gap. The risk is that TSMC is extrapolating today’s extraordinary margins into a construction program that arrives after the shortage has eased.

What to watch: Follow the gross-margin drag from overseas fabs, not the pledge total. If dilution rises even while utilization remains high, geographic resilience is carrying a measurable and persistent tariff.

India Buys the Missing Layers

India’s first semiconductor push financed factories. Its second is an attempt to finance the industrial system that makes factories productive.

The cabinet approved Semicon 2.0 with an outlay of ₹1.275 trillion, about $13.3 billion. The program covers design, wafer fabrication, advanced packaging, materials, equipment, research and talent. That breadth reflects the first phase’s results: 12 projects with roughly ₹1.64 trillion in planned investment, but just three facilities in commercial production so far. Twenty-four chip-design projects have also been approved.

Why it matters: A semiconductor fab is not a standalone asset. It is a demanding customer for tools, gases, specialty chemicals, construction expertise, reliable utilities and engineers who learn by debugging production. Subsidizing only the building can produce an expensive island whose critical inputs and process knowledge remain imported. Semicon 2.0 is an effort to create network effects on the supply side, where each additional supplier, designer and packaging plant makes every other investment more viable.

The dependency is visible in India’s flagship project. Tata’s $11 billion Dholera fab relies on PSMC’s process technology and an ASML partnership for lithography systems, training and manufacturing support across mature nodes. That is sensible sequencing: India must rent the operating system before it can localize more of it. The policy’s real return will come from domestic firms learning around those imported anchors, not from declaring self-sufficiency.

Shared suppliers also improve the subsidy multiplier. A metrology lab, packaging line or trained process engineer can support several manufacturers, while a factory-specific grant creates value only if that single project reaches competitive yields.

Room for disagreement: Governments are better at announcing capital than manufacturing yields. Semiconductor founders told The Economic Times [paywalled] that rigid financing terms and slow administration can starve design startups while politically legible factories receive large checks. A broader mandate could disperse accountability instead of deepening the cluster.

What to watch: The first implementation rules will show whether “ecosystem” is operative or decorative. Track approval speed and the share of support that reaches design tools, materials, equipment and workforce programs rather than another round of fab announcements.

The Contrarian Take

Everyone says: National chip subsidies are steadily eroding TSMC’s moat by copying its capacity across the United States, Europe, Japan and India.

Here's why that's wrong (or at least incomplete): Subsidies can copy cleanrooms faster than they copy the learning system around them. TSMC’s 67.7% gross margin and 77% advanced-node mix are evidence of accumulated process advantage, not merely scarce floor space. Even India’s localization strategy begins with ASML tools and PSMC know-how. Geographic duplication can make supply more resilient while increasing dependence on the few companies able to transfer technology and operate new plants. The likely result is a more distributed footprint with TSMC still capturing the coordination premium.

Under the Radar

Britain now treats cloud providers like financial infrastructure — The Bank of England, PRA and FCA have begun direct oversight of AWS, Google Cloud, Microsoft and Oracle as the country’s first Critical Third Parties. Regulators can demand resilience testing, incident reporting and self-assessments. The designation recognizes systemic dependence; it does not solve concentration, because the same four firms remain the recovery options when one fails.
Preventive scanning just raised a deployment round — Neko Health raised a $700 million Series C ahead of its U.S. launch after scanning more than 100,000 people in Sweden and Britain. The company is turning full-body imaging and bloodwork into a repeat service. Its 350,000-person waitlist and appointment book suggest demand; U.S. expansion will test whether utilization economics survive a more fragmented clinical and reimbursement system.

Quick Takes

Thinking Machines finally has a product loop — Inkling is a 975-billion-parameter open-weight model designed to be customized through the startup’s Tinker service, which is offering a 50% launch discount on fine-tuning. Open weights make the model easy to inspect and distribute; the business sits in repeated adaptation. Mira Murati’s company is betting enterprises will pay for a model workshop, not another closed-model tollbooth. (Source)
Chip controls are becoming legislation, not discretion — The Senate NDAA now includes the AI OVERWATCH Act: an 18-month ban on exports of the most advanced chips to countries of concern plus Commerce certifications before licensing lesser chips. It still must survive conference and enactment. If it does, Congress will have converted a reversible executive policy into a more durable constraint on both exporters and future administrations. (Source)
The phone shortage is selecting brands — OnePlus reportedly plans to wind down U.S. and European operations as parent Oppo reorganizes, while keeping the brand in China and using Realme abroad. This materially advances yesterday’s memory-shortage story: component inflation does not merely raise handset prices; it gives large vendors first claim on supply and forces marginal brands to abandon expensive distribution. Consolidation is the shortage’s second-order effect. (Source)

The Thread

Today’s stories are about who pays for redundancy before a crisis makes it mandatory. TSMC can duplicate its Taiwan capacity because the original cluster produces enough cash and know-how to finance the copy. India has no equivalent profit pool, so the state is spreading the cost across suppliers, designers and training. Britain’s cloud designation applies the same logic to services: concentrated vendors created public tail risk, and regulators are making them carry more of the resilience bill. Efficiency built these systems; assigning the cost of a backup will shape the next phase.

Predictions

New predictions:

I predict: The AI OVERWATCH Act's 18-month advanced-chip restriction, or substantially equivalent language, will survive NDAA conference and be enacted by December 31, 2026. (Confidence: medium; Check by: 2026-12-31)

Issue date: July 16, 2026 · Generated: 3:27 AM ET