The Trade That Got Away

Why a Legend Admits Defeat — and Moves On

Stanley Druckenmiller doesn't miss often. The man who broke the Bank of England alongside George Soros, who ran one of the most successful hedge funds in history without a single down year, made what he himself calls a "big mistake."

In late 2022 and early 2023, his Duquesne Family Office built a massive Nvidia stake at a split-adjusted $22-24 per share. A $210-220 million bet on the AI wave. By mid-to-late 2024, he was out — selling the entire position at around $73.50. A triple-plus return. Clean, respectable, profitable.

Except Nvidia kept climbing. By June 2026, it's trading at $190. That position would be worth $1.7 billion today. He walked away from over a billion dollars in unrealized gains.

Most investors would bury that story. Druckenmiller acknowledged it publicly and moved on. Because in Q1 2026, he made a new bet: 196,000 shares of Broadcom, purchased at an average price of $330. As of this writing, it's trading at $365.

The Rotation Thesis: From Training to Inference

Druckenmiller hasn't commented publicly on the Broadcom buy, but the logic is clear if you understand the architecture of AI infrastructure.

Nvidia dominates the training market — the phase where massive GPU clusters churn through petabytes of data to build large language models. General-purpose GPUs are ideal here: flexible, powerful, parallel. Nvidia owns this layer.

But once a model is trained, the economics flip. Inference — the act of running that trained model to answer queries, generate text, analyze images — is where the real volume lives. Billions of queries, every second, at scale. And for inference, application-specific integrated circuits (ASICs) can outperform Nvidia's GPUs on both speed and cost per operation.

That's Broadcom's lane. It designs custom ASICs tailored to specific AI workloads. Meta, Alphabet's Google, OpenAI, and Anthropic are all deploying Broadcom's chips. These aren't off-the-shelf components — they're hyper-optimized silicon built to do one thing faster and cheaper than anything else.

In fiscal 2025 (ended November 2025), Broadcom's AI chip revenue hit $20 billion, up 65%, representing 31% of total revenue. By fiscal 2027, the company projects that figure will exceed $100 billion — more than 58% of an expected $171 billion top line. From fiscal 2025 to 2028, analysts expect revenue to more than triple and EPS to more than quadruple.

Nvidia vs. Broadcom: Two Plays, Different Layers

Nvidia (Training)

General-purpose GPUs power the compute-intensive model training phase. Flexible, scalable, dominant market position. High valuation reflects its central role in the AI buildout.

Broadcom (Inference)

Custom ASICs optimized for inference deliver faster, cheaper operation at scale. Lower profile, but essential for deploying AI in production. Trading at 22x forward earnings despite explosive growth ahead.

Is This a Trade or a Hold?

Druckenmiller has a history with Broadcom — he traded in and out of the stock through 2023, 2024, and into 2025. He held no shares at the end of 2025. The Q1 2026 purchase marks a fresh entry.

It's impossible to know whether this is a short-term trade or a conviction hold. But the timing is notable. The AI market is maturing past the early training frenzy. The focus is shifting to inference, to deployment, to economics at scale. Broadcom sits at the center of that transition.

At 22 times next year's earnings — modest for a company projecting that kind of growth — there's room for multiple expansion if the inference thesis plays out. Druckenmiller may have missed the Nvidia moonshot, but Broadcom offers a different angle on the same secular wave.

FAQ

Why did Druckenmiller sell Nvidia so early?

He hasn't detailed his exact reasoning publicly, but timing exits is notoriously difficult even for elite investors. A triple-plus return looked strong in mid-2024 — the explosive move to $190 came later. He's acknowledged the early exit as a mistake.

What's the difference between training and inference in AI?

Training is building the model — crunching vast datasets to teach the AI. Inference is running that trained model to answer real-world queries. Training needs flexible, powerful GPUs. Inference at scale benefits from custom ASICs optimized for speed and cost efficiency.

Is Broadcom a better buy than Nvidia now?

They're different plays. Nvidia still leads training infrastructure; Broadcom is betting on inference dominance. Broadcom trades at a lower multiple and is earlier in its revenue ramp. It's not better or worse — it's a different layer of the same stack.

This content is for informational purposes only and does not constitute investment advice. All investment decisions carry risk. Consult a qualified financial advisor before making any investment decisions.