Chips Are in a Bear Market. The Real Question Starts Next Week.

Hey there, bargain hunter.

Let’s start with the number nobody expected to see this quickly.

The Philadelphia Semiconductor Index fell more than 20% from its June 22 record close in under four weeks, officially entering bear market territory. The index was down about 11% on the week as of Friday (July 17) — its largest weekly decline since March 2025. Global semiconductor stocks shed roughly $3.3 trillion in market value since June 22.

For context: the SOX had surged 105% between its March low and that June peak. A 105% gain in three months is not normal price discovery. When that kind of vertical move reverses, it hurts.

What triggered it? A few things collided at once.

The Kimi K3 Problem

Chinese AI startup Moonshot unveiled Kimi K3 around the World Artificial Intelligence Conference in Shanghai on Friday — a 2.8 trillion-parameter open-weight model that the company claims rivals top offerings from OpenAI and Anthropic. It jumped to first place on Arena’s front-end coding leaderboard.

Because it is open-weight, developers can download and run it themselves. That directly challenges the premise that frontier AI requires ever-more-expensive proprietary hardware. The market read it exactly the way it read DeepSeek in early 2025 — and sold chips hard.

Worth noting: reports on Kimi K3’s relative standing versus top U.S. models and its operating cost vary by benchmark and methodology. Moonshot has said full open weights are expected on July 27. Nobody has independently stress-tested those full weights yet.

That nuance did not matter on Friday morning. Sentiment moved faster than analysis.

What Else Was Already Under the Surface

The Kimi K3 headline was the match, but the kindling had been building. TSMC reported a strong quarter earlier in the week, then guided capital expenditures to $60 billion to $64 billion for 2026, above the prior range of $52 billion to $56 billion. Higher capex from the world’s most critical chipmaker raised the question of whether spending is accelerating because of demand strength — or because costs are simply rising. The market chose the bearish read.

Separately, reports that Alphabet’s Gemini 3.5 Pro launch had fallen months behind schedule added to the unease. If the hyperscalers are not delivering AI products on schedule, the urgency of the infrastructure buildout looks less certain.

And underneath all of it: hedge funds had been trimming tech hardware exposure for a fourth consecutive week heading into the selloff. The selling was not random. It was organized.

The Earnings Gauntlet Ahead

Here is where the real debate starts. Semiconductor industry earnings are expected to grow 131% year over year in Q2, according to FactSet data. That is not a sector in structural decline. That is a sector that got priced for perfection and is now being asked to prove it.

Alphabet and Tesla report July 22. Their capital expenditure numbers will be the first real data point on whether hyperscaler spending is actually pulling back or just getting questioned. Microsoft reports July 29. Meta the same day. Those four capex lines, combined, will either validate or undercut the bear case for chip demand.

Marvell Technology is the name worth watching most closely in this environment. The company builds custom silicon for cloud AI accelerators — often discussed in the industry as “XPUs.” Its Q1 FY2027 revenue was $2.418 billion, up 28% year over year, with Q2 guidance pointing to $2.7 billion. CEO Matt Murphy said the company is seeing exceptional AI-related bookings and expects revenue growth to accelerate every quarter of fiscal 2027. The stock is down more than 40% from its 52-week high and still up approximately 110% year to date. That disconnect is unusual for a business with those numbers.

Micron is the other piece. Revenue grew 346% year over year in its most recent quarter. On high-bandwidth memory, supply tightness is expected to persist into 2027, with industry commentary pointing to demand continuing to outstrip supply. The stock is down about 30% from its peak but is still up nearly 200% this year.

The Real Split in This Trade

The commodity chip story and the custom silicon story are not the same bet right now.

Memory and GPU plays are exposed to the commoditization argument. If open-weight models keep improving, if Meta has excess GPU capacity to rent out rather than buy more, if inference efficiency keeps climbing — those dynamics compress the urgency of the hardware buildout. That is a legitimate structural risk worth pricing in.

Custom silicon is a different animal. When Amazon designs its Trainium chip with Marvell’s engineering team baked in, switching costs are real. Those are multi-year commitments, not spot purchases. Cheaper open-weight models do not immediately change that relationship. They may even extend it, as hyperscalers look for more cost-efficient custom solutions rather than buying off-the-shelf at Nvidia prices.

Cheap Investor Scorecard on the Sector

• SOX index from March low to June peak: up 105%
• SOX from June 22 high to now: down more than 20%, bear market confirmed
• Semiconductor Q2 earnings growth estimate: 131% year over year (FactSet)
• TSMC 2026 capex guide: $60B to $64B, raised from $52B to $56B
• Marvell Q1 FY2027 revenue: $2.418B, up 28% year over year
• Marvell Q2 guidance: $2.7B at midpoint
• HBM supply: expected to remain tight into 2027
• Kimi K3 full weights available: July 27 — that is when independent testing begins

The bear case is not that AI is fake. The bear case is that the market priced a near-perfect demand trajectory into names that were already up 100% in a single quarter. That is a valuation reset, not a thesis collapse.

Whether this is the entry point of 2026 or the first act of something longer depends almost entirely on what hyperscalers say about their spending plans over the next two weeks. The numbers are still strong. The question is whether the market believes them again.

That answer comes July 22.