The Philadelphia Semiconductor Index rose 1.6 percent at Wednesday's open, as a more than 6 percent rally in Advanced Micro Devices Inc. to a new high underscored relentless investor appetite for artificial intelligence exposure.
The move extends a blistering 2026 rally for the semiconductor space, which has seen the SOX index gain approximately 65 percent this year alone. "NVIDIA is benefitting from strong demand, but is selling into a concentrated set of buyers whose own demand is being distorted by a training and benchmarking phase that will not last," Michael Burry wrote in a recent Substack post, arguing the demand is a temporary "bezzle."
The gains were broad-based beyond AMD. Astera Labs and Qualcomm both rose over 5 percent, while Texas Instruments and NXP Semiconductors gained nearly 3 percent. The rally occurred even as the 10-year Treasury yield held firm and the S&P 500's Shiller PE ratio sits above 40, its highest level since the 2000 tech bubble, according to 24/7 Wall St.
The core tension for investors is timing. While bulls point to real AI-driven revenue at Nvidia, Microsoft, and Google, bears like Burry argue the valuation assumes the initial, frantic "training phase" of AI is permanent. Burry, who is short the SOXX index via put options, believes the demand profile will fundamentally change as the market shifts to a less GPU-intensive "inference" phase, creating a "Cisco moment" that led to an 80 percent stock collapse after the dot-com bubble.
The Bear Thesis
Burry's argument, detailed in a May 2026 Substack post, centers on three mechanisms: extreme customer concentration with hyperscalers, a "bullwhip effect" causing supply chain over-ordering, and a "bezzle," or a gap between perceived and actual value. He points to Nvidia's $119 billion in non-cancellable supply commitments to TSMC as a key risk if demand slows even modestly.
The Counterargument
The counterargument is that Burry has been early and wrong before, calling a market top in 2023 before a massive rally. Furthermore, the transition from AI training to inference may not be less compute-intensive. Many researchers argue inference at scale will require just as many GPUs, simply distributed differently, sustaining demand for chips from Nvidia, and increasingly, from competitors like AMD and Broadcom who are targeting the inference and custom chip markets.
This article is for informational purposes only and does not constitute investment advice.
