TSMC CEO: AI chip shortage will last years — and “stable” pricing still means hikes

What Wei actually said

The shareholder meeting is where TSMC's management sets expectations for the year ahead. In 2026, the dominant theme was capacity, not demand. Wei acknowledged that AI models are proliferating across consumer devices, enterprise software, and sovereign compute programs, all of which require more advanced logic and packaging than the prior smartphone cycle. His framing was candid: “Customer demand is so high, and we can only support so much. We are doing our best to ensure TSMC does not become a bottleneck.”

On pricing, a shareholder asked whether TSMC would raise prices. Wei's answer was equally direct: “I'd like to do that … we still need to make money.” He immediately qualified that TSMC is “not that kind of company” — meaning it will not shock the market with abrupt hikes like the memory sector has seen in 2026. The distinction matters. Negotiated, node-by-node increases of 3% to 10% on advanced processes for 2026 are already in client notifications, according to industry reports cited by TrendForce and Taiwanese trade press. A further 15% increase on 3nm wafers in the second half of 2026 has been widely reported, though TSMC has not published a public rate card. “Stable” in Wei's vocabulary means predictable escalators, not frozen quotes.

Arizona received its own reality check. TSMC is spending roughly $165 billion expanding U.S. fabrication, but Wei warned that even with those fabs ramping, the company will not be able to fully meet American customer demand domestically for years. Offshoring advanced packaging and leading-edge logic is a decade-long project; the AI capex cycle does not wait for ribbon-cuttings.

Why this landed after Broadcom's selloff

Context is everything. TSMC's meeting came two days after Broadcom's quarterly AI-chip outlook disappointed elevated expectations, triggering a third straight down day for Nasdaq futures and a 4.7% plunge in South Korea's KOSPI that we analyzed in our Broadcom shock and Asia spillover piece. Investors had been pricing AI semiconductors as if supply would expand as fast as model demand. Wei's comments push the opposite narrative: demand is winning the race, and the foundry that fabs Nvidia's, Apple's, and Broadcom's most advanced chips is rationing access.

That rationing shows up in wafer economics. Industry analysts estimate 3nm wafers currently run $18,000–$20,000 each; a 15% second-half increase pushes that toward $23,000. Two-nanometer wafers are already discussed at $30,000+ — a 50% step above today's 3nm pricing before yield learning curves mature. Mask sets at 3nm can cost $50–$70 million before a single die ships. New customers face lead times of 40–50 weeks unless they arrive with multi-year capacity deposits. The queue is not first-come, first-served; it is first-committed, first-served. Apple reportedly holds more than half of early 2nm allocation for M-series and A-series roadmaps heading into Monday's WWDC keynote.

TSMC shares fell roughly 2% in Thursday pre-market trading after Wei's remarks, per Yahoo Finance. The stock reaction is almost beside the point. Foundry pricing power in a shortage is not a bearish signal for TSMC; it is a tax on everyone downstream who cannot lock capacity.

Winners, squeezes, and spillovers

Tier-one AI spenders win by paying. Nvidia, Apple, AMD, and Broadcom sit at the front of TSMC's priority queue because they pre-commit billions in annual capex and accept multi-year price escalators. For them, a 10% wafer hike on a $30,000 2nm lot is annoying but absorbable inside $56 billion AI revenue guidance (Broadcom's fiscal 2026 AI semiconductor target) or a $1.75 trillion SpaceX-adjacent ecosystem. Capacity access matters more than unit cost when every month of delay costs market share.

Mid-tier chipmakers get squeezed. Automotive suppliers, networking SoC vendors, and IoT designers that relied on mature 6nm and 7nm nodes are collateral damage. TSMC has been shifting engineers and equipment toward 3nm and 2nm, tightening supply on nodes that still power industrial and automotive silicon. Paying higher prices does not guarantee allocation when Apple and Nvidia have already booked the ramp. Some firms will migrate to Samsung Foundry or Intel's 18A process — the bet Intel made with its Xeon 6+ Clearwater Forest launch at COMPUTEX — but alternative foundries carry their own yield and ecosystem risks.

Consumers and gamers pay indirectly. Memory shortages already pushed NAND and DRAM costs higher, which our gaming storage crunch analysis tied to AI datacenter demand for high-bandwidth memory. Logic price hikes compound the bill: pricier SoCs flow into console BOMs, flagship phone ASPs, and cloud instance rates. Microsoft's Xbox division cited 2.75x memory cost inflation in its recent reset; TSMC's pricing trajectory suggests logic will not offset that pressure. If component costs rise faster than willingness to pay, product cycles stretch — fewer generational leaps, more subscription services, more “Siri mode” software features over new silicon.

Hyperscalers face a margin math problem. Oracle, Microsoft, Google, and Amazon are simultaneously building AI datacenters and buying every GPU TSMC fabs. Our Oracle Q4 earnings preview flagged how remaining performance obligations and capex guidance will test whether cloud revenue scales with input costs. If foundry prices rise 3–10% annually through 2029 while cloud list prices face enterprise procurement pushback, the AI infrastructure trade stops being a pure volume story and becomes a gross-margin story. That is the subtext behind Friday's defensive rotation out of XLK and into staples: investors are repricing who captures AI rents when the bottleneck owner raises tolls.

Reading “no sudden hikes” precisely

Wei's pledge to avoid memory-style shocks is strategically important. DRAM and NAND spot prices in 2026 have whipsawed on inventory scares and AI pull-through, creating planning chaos for OEMs. TSMC is signaling a different contract structure: annual negotiated increases by node, visible years in advance, tied to capacity reservations. That is how you run a foundry monopoly without triggering regulatory panic or customer revolt.

The practical implication for buyers is that budgeting on last year's wafer quote understates 2026–2027 COGS. Procurement teams that treated leading-edge silicon like a commoditized input — negotiate quarterly, switch vendors freely — will discover the market is allocation-constrained. The winners pre-pay, accept escalators, and design products that justify higher ASPs. Everyone else waits in a 40-week line or ships on older nodes and eats competitive disadvantage.

Lead times for AI-related logic and advanced packaging are expected to stretch into 2027 before new capacity materially eases the imbalance, according to meeting summaries. TSMC's own capex is enormous, but fabs take years to yield at mature defect rates. Until then, Wei's bottleneck warning is not bearish AI demand. It is bullish TSMC pricing power and bearish anyone who assumed silicon scarcity was a 2024–2025 problem already solved.

Three scenarios through year-end

Shortage premium persists, AI stocks bifurcate (50%)

TSMC implements reported H2 2026 3nm hikes without customer churn at the top of the queue. Nvidia, Apple, and Broadcom guide revenue up, gross margins flat-to-down on input costs. Mid-cap semi and consumer hardware names miss on component availability. KOSPI and Taiwan equities outperform U.S. hyperscalers on a relative basis. The AI trade does not break, but leadership narrows to foundry-adjacent winners and capacity-locked platforms.

Demand air pocket exposes over-ordering (30%)

A hot May CPI print (10 June) and Kevin Warsh's June FOMC hawkish tilt slow enterprise AI orders. Hyperscalers delay second-wave GPU deployments, creating a temporary utilization dip that Wall Street misreads as supply relief. TSMC holds pricing but utilization falls one quarter; memory spot prices soften while contract logic prices stay sticky. Semiconductor equities chop; defensive rotation extends.

Geopolitical shock reroutes supply chains (20%)

Taiwan Strait tension or new export controls accelerate customer diversification to Intel 18A, Samsung, and U.S. onshore fabs. TSMC's Arizona ramp gets politicized subsidies and faster tooling imports; near-term global supply becomes more fragmented and expensive, not cheaper. Lead times lengthen further on logistics and qualification, not wafer starts alone.

What to watch

• WWDC keynote (8 June, 10:00 a.m. PT). Apple silicon announcements reveal how much 2nm/3nm capacity Tim Cook has secured for fall iPhone and Mac launches.
• Oracle earnings (10 June). OCI capex guidance and RPO growth test whether cloud margins can absorb rising foundry input costs.
• TSMC monthly revenue (early July). Sequential growth despite pricing hikes confirms volume rationing, not demand collapse.
• 6nm/7nm lead-time reports. Automotive and industrial chip shortages are the early warning for consumer electronics BOM stress in H2.
• Intel 18A customer wins. Mid-tier defections from TSMC's queue would signal pricing power limits; silence confirms rationing dominance.

Sources: Reuters — TSMC CEO on AI supply (Jun 4, 2026); TechTimes — capacity and pricing analysis (Jun 5, 2026); Yahoo Finance — Wei shareholder remarks (Jun 2026); Android Headlines — shortage timeline (Jun 2026).