Druckenmiller Dumped Broadcom, Intel, and Micron. The Chip He Bought Instead Says the AI Trade Has Two More Years Left.

The AMD buy is small relative to the $5.2 billion portfolio — 0.8% of NAV — but that's the point of a 13F: it shows the entry point, not the conviction ceiling. Druckenmiller bought the one AI chip designer in the cohort that had the cleanest 2027 ramp, backed by a customer pipeline AMD itself has been printing all year.

Taiwan Semiconductor Manufacturing Company was the chip name Druckenmiller added to rather than trimmed. The 19.1% bump lifted the stake to 589,680 shares — about $281.6 million, or 5.4% of the 13F book. That's his largest non-healthcare position. The thesis logic: even if AMD loses share to NVIDIA, every AI accelerator that ships between 2026 and 2028 is still fabricated in Hsinchu, Tainan, or Phoenix.

Taiwan Semiconductor Manufacturing Company delivered Q2 2026 revenue of NT$1.27 trillion (about $40.2 billion), up 36% year-over-year, with a 67.7% gross margin and operating margin of 60.3%. Management raised full-year 2026 capex to $60–64 billion (from $52–56 billion), plus an additional $100 billion U.S. investment in Arizona. HPC revenue grew 20% quarter-on-quarter — the fastest expansion of any platform segment.

Four new positions all share a profile: a legacy bitcoin-mining business that has been converting its cheap power contracts, substation connections, and land holdings into AI-cloud capacity. Bitdeer Technologies, Riot Platforms, Hut 8, and IREN all traded the same way — they had sites before hyperscalers needed them, and now they're monetizing them as AI colocation. Combined new exposure: roughly $125 million across the cluster.

The pattern: the chips that already printed the 2025–26 AI capex wave — Broadcom up triple-digits off a $58 billion FY2026 AI guide, Micron up 309% YTD, Intel] a structural write-down story — got trimmed. The names still tied to incremental build-out — AMD's MI400 ramp, TSMC's 2nm node, miner-converted AI sites — got fresh or enlarged positions. This is not a call that hyperscalers stop spending. It is a call that hyperscalers' money now flows to different chip vendors and to the physical layer that delivers compute.

The macro math behind it: combined hyperscaler capex for 2026 sits in a $680–720 billion band, with Amazon guiding ~$220 billion, Alphabet ~$175–185 billion, and Microsoft targeting ~$175 billion in AI infrastructure. AMD's locked-in 14+ gigawatts of customer commitments maps to roughly $90 billion in revenue over the deal lives, per analyst estimates — a number that lines up with the doubling-in-2027 guidance management gave on the Q2 FY2026 call. Druckenmiller isn't betting the cycle ends. He's betting which suppliers still have to ramp.