Chips Down, Fear Up: Making Sense of the AI Selloff Before Big Tech Earnings

Semiconductors just had one of those weeks where the price action and the fundamentals seemed to be reading from different scripts. TSMC delivered strong results and got sold. A Chinese AI lab dropped a model that briefly spooked the entire compute trade. Korean volatility bled directly into Micron and SanDisk in real time. And underneath all of it, Wall Street quietly raised its capex forecasts for the hyperscalers.

Let's untangle it, because the gap between the headlines and the underlying data is where the opportunity usually lives.

TSMC's "Bad" Quarter That Wasn't

Start with the strangest trade of the week: TSMC reported excellent numbers and fell hard anyway. The trigger wasn't revenue or margins — it was guidance. Management sharply raised its capital expenditure outlook through 2026, and the market's reflex was to punish it: if they're spending that much, margins must be at risk.

Think about that logic for a second. The most important foundry on Earth, with the deepest visibility into real AI chip demand, says it needs to spend dramatically more to keep up — and the market treats the spending itself as the problem. The same reflex dragged down Micron, Samsung, and ASML in sympathy. This has the texture of pre-earnings-season nervousness hunting for a bear case, not a genuine deterioration in the business. When companies with order books get sold for investing in capacity, that's sentiment, not analysis.

Kimi K3: DeepSeek Déjà Vu, With the Same Fine Print

The week's psychological shock came from China. Moonshot AI released Kimi K3, and early benchmark claims had it edging out frontier US models — reports cited wins over Anthropic's Claude Opus 4.8 and GPT-5.5 on select tests. Cue the familiar spiral: if China can match frontier performance cheaply, does the AI capex supercycle still need all those chips?

We ran this exact drill with DeepSeek. And just like then, the fine print matters more than the headline. Kimi K3 is currently so overwhelmed by traffic that Moonshot has halted new signups for lack of computing power. Read that again — the supposed proof that the world needs less compute can't serve its own users because it doesn't have enough compute. Meanwhile, other Chinese AI players like Zhipu are reporting growing revenue alongside widening losses. Impressive engineering, yes. A threat to the demand side of the AI hardware ecosystem, no — if anything, it's more evidence that global compute demand exceeds supply everywhere.

The Korea Transmission Channel Is Now Live

Here's a structural change US chip investors need to internalize: with SK Hynix ADRs and leveraged products now trading on US exchanges, Korean market volatility no longer arrives with a time-zone delay. It transmits instantly. This week, turbulence in Seoul propagated straight into US memory names — Micron and SanDisk got whipsawed by flows that had little to do with their own fundamentals.

I covered the SK Hynix ADR premium story in my last piece; this is the flip side. A tighter link between the two markets means better price discovery over time, but in the short run it means US memory investors have effectively imported Korea's volatility profile. Size positions accordingly.

Big Tech Scorecard Heading Into the Main Event

• Alphabet is under pressure on reports that Gemini 3.5 has slipped by months. Its earnings and capex disclosure this week is arguably the single most important data point for the entire AI trade.
• Anthropic is reportedly so compute-constrained that access to its top model is limited to premium tiers, and it's renting cloud capacity from Meta — with an IPO push taking shape for October. OpenAI's Sam Altman, for his part, publicly acknowledged a year of missteps while signaling a more aggressive posture ahead. Again: every scarcity story in AI right now is a compute scarcity story.
• Apple reclaimed the #1 market cap spot doing the opposite of everyone else — minimal direct AI capex, price increases on devices and services, and partnerships for AI capability. HSBC upgraded it to buy. It's the low-capex way to own the AI cycle, and the market is rewarding it this week.
• Netflix beat expectations but sold off on free cash flow concerns. Michael Burry's jab — comparing Disney's durable catalog to wine and Netflix's fast-expiring content to milk — is quotable, but the FCF pressure is the real story for income-oriented investors.

Macro: The Economy Is Too Strong for Its Own Good

The geopolitical backdrop darkened — two US soldiers killed in an attack in Jordan, additional refueling aircraft deployed, and Iran warning it could close the Bab el-Mandeb and Hormuz chokepoints. Container shipping rates have already snapped higher as traffic thins out. That's an inflation input, not just a headline.

Meanwhile the domestic data — Michigan sentiment, retail sales, Philly Fed, jobless claims — all pointed the same direction: the US economy is running hot. Hot enough that Dallas Fed's Logan and others have put rate hikes back into the conversation, sending Treasury yields lurching in both directions. Strong economy, sticky rates, rising shipping costs: that mix argues for quality balance sheets and pricing power, which conveniently describes most of the mega-cap AI complex.

For dividend investors watching rate sensitivity, this environment is worth revisiting how interest rates affect REIT valuations and the broader income landscape.

What Wall Street Actually Did While Retail Panicked

Two details cut through the noise. Citigroup, days before earnings, raised its combined capex forecast for Google, Meta, and Amazon to over $800 billion. And Bank of America shrugged off the semiconductor drawdown as seasonal summer chop.

So while the tape screamed "AI peak," the institutions closest to the spending data revised their numbers up.

The Bottom Line

The bearish case this week was built on vibes: TSMC spends too much, China builds too cheap, Korea is too volatile. The bullish case is built on receipts: record foundry capex, compute shortages at every AI lab on the planet, and $800B+ in hyperscaler spending forecasts. The framework I'd carry into earnings week is the one that matters long term — stop viewing this as a "memory peak-out" story and start viewing it as the buildout of an entire physical-AI hardware ecosystem.

Alphabet reports this week. If the capex number is big and the guidance is confident, this selloff will look, in hindsight, like the market flinching at its own shadow.