Great earnings, worse share prices: the semiconductor market has raised its proof standard

TL;DR

• Taiwan Semiconductor Manufacturing Co. (TSMC) reported record second-quarter earnings and announced an additional US$100 billion for US capacity, yet the news did not stop a global semiconductor retreat.¹²
• South Korea’s KOSPI fell 6.4% on 16 July; SK hynix fell 11.5%, Samsung Electronics 8.8%, Japan’s Kioxia 15%, and Tokyo Electron and Advantest also declined.²
• Europe joined the move: ASML’s strong outlook did not prevent pressure on chip-equipment names, while STMicroelectronics fell sharply.³
• This is a valuation and capital-intensity reset, not evidence of a sudden collapse in chip demand. But the two are connected: exceptionally high expected returns are what justified the valuations.

The number that changed the conversation

US$100 billion. That is the additional US investment TSMC announced alongside a second-quarter profit increase of more than 77% year on year and stronger revenue-growth expectations.² In an earlier phase of the AI trade, that combination would have been treated as uncomplicatedly bullish.

It was not. The result landed during a broad semiconductor sell-off stretching from Seoul and Tokyo to Amsterdam and European markets. That is the story.

What happened

TSMC held its second-quarter 2026 earnings call on 16 July Taiwan time.¹ AP reported that it delivered record quarterly results, lifted its revenue-growth outlook and added US$100 billion of planned US chipmaking investment.²

Instead of lifting the broader complex, the numbers arrived into an already brittle market. The KOSPI fell 6.4%; SK hynix and Samsung dropped 11.5% and 8.8% respectively. Japan’s Kioxia fell 15%, while Tokyo Electron and Advantest also lost ground.² AFP reported the same pattern across major markets: TSMC’s own US-listed shares fell despite the results, and ASML also struggled to sustain an initial gain.³

That constitutes clear international virality in market terms: one earnings event became the shared subject of a repricing across Asian manufacturing, European capital equipment and US-listed AI infrastructure. It was rapidly carried by AP, AFP, Reuters syndication and regional business press. Direct native-platform engagement totals were not independently available.

What it actually means: the burden of proof has moved

Here is the institutional dialogue now being enacted:

Company: Demand is strong, capacity is tight, and we are investing more.

Investor: Precisely. Show me why the return on that extra capacity will still be exceptional after everyone else also spends more.

That is not bearishness about semiconductors. It is a change in the question.

The first question of the AI-infrastructure boom was whether demand existed. TSMC’s figures strongly support the answer: yes. The second is whether the earnings pool will remain concentrated enough, and long enough, to justify today’s prices after a spending wave that includes chipmakers, equipment suppliers, cloud providers and customers. TSMC’s US$100 billion plan is economically rational if demand persists; it also makes future utilisation, depreciation and supply-discipline questions more important.

The immediate cross-layer connection is non-obvious but consequential: capex is both proof of demand and the mechanism by which a scarcity premium eventually disappears. Investors can cheer the first and fear the second on the same day.

What this is not

This is not proof that AI infrastructure has become a failed investment theme. TSMC’s reported results and upgraded outlook point the other way.²

Nor is it a simple “AI bubble burst” story. Different parts of the chain carry different risks:

• Foundries face utilisation and geographic-capex execution.
• Memory makers face the most familiar cycle risk: tight supply can turn into excess supply.
• Equipment suppliers depend on customers converting intentions into multi-year orders.
• Cloud buyers must turn compute expenditure into revenue or durable productivity gains.

A broad sell-off treats these as one trade. They are not one business.

Stakeholder landscape

Durability: one week, one month, one year

• One week: earnings reports from semiconductor and hyperscale peers will determine whether the sell-off becomes a sector-wide de-rating or a sharp rotation.
• One month: the key evidence will be capex guidance and memory-pricing data, not daily share moves.
• One year: the decisive issue is whether supply expansion arrives faster than AI workloads and monetisation. That is the point at which today’s capital plans become either strategic bottlenecks or excess fixed cost.

Recommendations

For investors and market observers

1. Do not use a single benchmark as the thesis. Separate foundry, memory, equipment and cloud exposure. A falling semiconductor index does not describe each sub-sector’s cash-flow risk.
2. Track four hard indicators: TSMC advanced-node utilisation and gross-margin guidance; HBM/DRAM contract prices; ASML net bookings; and aggregate hyperscaler capex versus incremental cloud/AI revenue.
3. Treat ‘record earnings’ and ‘good entry point’ as separate statements. They can both be false, or only one can be true.

For business leaders buying compute

1. Do not pause valid workloads because public chip prices fell. Re-price the workload economics instead: GPU-hours, model-inference volume, latency requirements and committed-spend terms.
2. Avoid signing capacity commitments solely to beat a perceived shortage. Tie any multi-year reservation to utilisation thresholds, portability rights and exit provisions.
3. Model a 12–24 month cost decline. Chip and cloud capacity often becomes cheaper after capital cycles; do not make permanent pricing promises based on temporary scarcity.

Uncertainty ledger

• The sources establish price moves and reported investment plans, but they do not establish a single causal trigger for every decline.
• The exact division between valuation concerns, portfolio rebalancing and fears of future oversupply remains unresolved.
• Native social engagement data was unavailable; this score relies on multi-market pickup and observable market participation rather than claimed view counts.
• A reversal in subsequent earnings guidance would weaken the ‘valuation reset’ interpretation; sustained cuts to capex or demand forecasts would strengthen a more negative reading.

Bottom Line

The market did not reject TSMC’s numbers. It rejected the idea that extraordinary numbers alone can justify extraordinary valuations forever. AI-chip demand remains real; the investment case has become less about proving demand and more about proving that a US$100 billion expansion will earn its cost of capital through the next supply cycle.

Sources 

1. Tier 1 — Primary source: TSMC, Second Quarter 2026 Earnings Conference, 16 July 2026.
   Record quarterly results, forward outlook and the additional US$100 billion US-capacity plan.

2. Tier 1 — Associated Press: Asian shares mostly decline with South Korea's Kospi down 6.6%, 16 July 2026.
   Market moves across South Korea and Japan, including declines in SK hynix, Samsung Electronics, Kioxia, Tokyo Electron and Advantest.

3. Tier 2 — Agence France-Presse: Stocks drop on tech sell-off, oil yo-yos on Mideast, 16 July 2026.
   European spillover, including pressure on semiconductor-equipment companies and STMicroelectronics.