With chipmakers posting earnings that are lapping the rest of the market by multiples, the infrastructure bet on AI is paying off faster and more durably than almost anyone predicted.
Micron Technology just reported $41.5 billion in fiscal Q3 2026 revenue , quadruple the year-ago figure. GAAP net income reached $28.24 billion for a single quarter. SK Hynix posted 52.58 trillion won ($35.53 billion) in Q1 revenue, up 198% year-over-year, with operating margins around 72%. These are not forward multiples or speculative valuations. They are reported earnings, and they are reshaping where the intelligent money on the AI trade actually sits.
The benchmark SOXQ semiconductor ETF is up roughly 90% year-to-date through late June. That number alone is striking, but the more important story is underneath it. This isn't a crowded momentum trade running ahead of fundamentals. Earnings are keeping pace because datacenter demand is outstripping chip supply, forcing higher selling prices across the board. TSMC CEO C.C. Wei told shareholders this year that global chip supply will continue to lag AI demand for years, not quarters. He posted Q1 2026 revenue of $35.7 billion on 35% year-over-year growth and refused to sandbag the full-year outlook, projecting more than 30% revenue growth for all of 2026.
The structural constraint isn't just fabrication capacity. TSMC's CoWoS advanced packaging is sold out through the end of 2026, with Nvidia locking in over 70% of the CoWoS-L allocation and the remaining slice split among AMD, Broadcom, and Marvell. You can't print more chips fast enough to satisfy this cycle, and that supply ceiling is what's holding prices up.
South Korea's KOSPI has nearly doubled since January, making it one of the world's best-performing major indexes in 2026. Goldman Sachs flagged it as its highest-conviction equity market in the region, forecasting 300% earnings growth for the index this year , the strongest annual profit expansion in any Asian market since the post-Asian financial crisis recovery of 1999. Samsung and SK Hynix together represent roughly 40% of the index, and SK Hynix alone holds approximately 57% market share in high-bandwidth memory, acting as a primary supplier to Nvidia's GPU stack.
The point here isn't that buying Korean equities is the obvious move. It's that most retail and institutional exposure to the AI trade is concentrated in US-listed model companies and cloud hyperscalers, while the manufacturers actually translating demand into profit are disproportionately Asian. If you're long the AI theme through software multiples and US megacap tech, you may be capturing less of the durable earnings power than the positioning implies.
Micron's numbers make the argument in concrete terms. Its Core Data Center Business Unit, which includes high-bandwidth memory for AI accelerators, posted $11.52 billion in revenue at an 87% gross margin this quarter, up from $1.53 billion at 38% margins a year earlier. The company has also locked in roughly $100 billion in minimum contracted revenue through 16 take-or-pay Strategic Customer Agreements, with $22 billion in upfront customer cash. Micron's HBM production through calendar 2026 is sold out, and a large share of forward sales is already under multiyear contracts. That's not a momentum stock. That's a business with a locked order book.
As CNBC's analysis noted after Micron's Q3 report, the company's market valuation has now passed Meta, a shift that would have seemed implausible as recently as 2024, when the memory cycle was widely expected to remain volatile and margin-pressured. The cycle has not merely turned , it has been structurally altered by the AI infrastructure buildout in a way that veteran semiconductor analysts are still revising their models to reflect.
Frankly, the conventional framing around AI investing has been too narrow. The debate has centered on which model provider wins, which cloud hyperscaler captures the most inference revenue, which software application layer monetizes fastest. These are real questions, but they've pulled attention away from the layer underneath all of it. You can't run inference or training without memory. You can't scale a datacenter without advanced packaging. You can't build the next Nvidia GPU cluster without HBM3E chips that are already backordered. The semiconductor infrastructure layer doesn't need AI to
Also read: Ardian is betting over a billion dollars that the Nordics will power Europe's AI infrastructure buildout • Proception settles Tesla's trade secret lawsuit and closes an $11 million seed round on the same day • OpenAI scrambles to fix Codex as coding agent usage blows past its own capacity models
