The Chip Buyers Are Becoming the Chipmakers

In May, Michael Burry warned that semiconductor stocks were mirroring the dot-com bubble, propelled by momentum rather than fundamentals ^[1]. By July he had put money behind a short of the iShares Semiconductor ETF, riding a slide that had already wiped hundreds of billions from chip stocks ^[2]. The bubble frame was familiar: demand is hype, valuations are detached, a reckoning is coming. What was actually getting priced, though, was something more specific and more durable than a sentiment correction. Across the supply chain, the same pattern kept appearing. The largest buyers of AI chips were designing their own. Amazon signed a deal to buy one million Nvidia GPUs, tens of billions of dollars in revenue for Nvidia, while simultaneously deploying more of its own Trainium chips than Nvidia's inside its data centers ^[3]. Andy Jassy framed the two moves as perfectly compatible.

While the largest number of AI chips we are bringing in are Trainium, we continue to have a deep partnership with Nvidia. — Andy Jassy

He also put a number on what Trainium means for Amazon's economics: a 30% price-performance improvement over comparable GPUs, and a roughly $50 billion run-rate chip business if sold to third parties ^[4]. Alphabet went further. It began selling its custom TPU chips directly to external customers for installation in their own data centers, the first time a hyperscaler moved from building silicon for internal use to acting as a merchant chip vendor ^[5][6]. Meta started selling surplus GPU compute under a product called Meta Compute, placing it in direct competition with AWS, Azure, Google Cloud, and the neoclouds it still buys capacity from; CoreWeave and Nebius dropped 14% and 17% on the news ^[7]. And DeepSeek, the Chinese AI lab, is developing its own inference chip to escape dependence on both Nvidia and Huawei, citing U.S. export controls as the catalyst ^[8]. Four of the most important chip buyers on the planet, moving in the same direction at the same time, for different reasons. Cost for Amazon and Alphabet. Compute monetization for Meta. Geopolitics for DeepSeek. The convergence is the pattern. The mechanism connecting all of this to the selloff is a shift in what AI chips actually do. Nvidia's high-margin GPUs dominate training, the phase where a model learns from vast datasets. But once a model is trained and deployed, it enters inference: answering queries, generating text, running agents. Inference at scale favors ASIC, and the hyperscalers designing their own ASICs can optimize the cost-per-inference in ways a merchant GPU cannot match ^[9][10]. The market is rotating from training toward inference, which shifts the chip mix from high-margin merchant silicon to lower-margin custom silicon designed by the customers themselves. Margin compression is already visible at every layer. Broadcom, which designs custom ASICs for Alphabet, Meta, OpenAI, and Anthropic, saw its gross margin slip from 77.1% to 74% on the custom-chip mix ^[11]. Its Q2 results were strong enough on AI revenue, growing 143% to $10.8 billion, yet triggered a $1.3 trillion sector sell-off because total sales slightly missed estimates and investors focused on the margin trajectory ^[12][11].

We expect this momentum to continue into fiscal year 2027 and reiterate our AI semiconductor revenue guidance to be in excess of $100 billion. — Hock E. Tan

Dell's gross margin fell from 21% to 18% as it shifted toward lower-margin AI servers ^[13]. Nvidia's forward price-to-earnings ratio hit a multiyear low of 22.22x despite record fiscal 2026 revenue of $215.9 billion, up 65% year over year ^[9]. Samsung plunged 7% in Asia despite a nineteenfold profit jump, as capital rotated out of semiconductor stocks even as profits hit records ^[14]. The sell-off is not random. It traces a line from the customer-turned-rival dynamic through each layer of the supply chain, compressing the margin share that merchant chipmakers capture. Leopold Aschenbrenner's short of Broadcom, shared by Ken Griffin and D.E. Shaw, was explicitly predicated on the concern that cloud customers may eventually replace Broadcom's custom silicon with fully proprietary internal designs ^[15]. Here is the catch, and it matters. Nvidia and TSMC both project parabolic decadal demand. Jensen Huang said demand has gone parabolic. TSMC raised its gross margin floor from 53% to 56% and projects a $1.5 trillion processor market by 2030 ^[16].

Demand has gone parabolic. — Jensen Huang

Nvidia is also counter-attacking: diversifying beyond GPUs into the $200 billion CPU market with Vera Rubin, projecting nearly $20 billion in CPU revenue this year ^[17]. The analyst community splits precisely along this fault line. Some warn that custom chips could disrupt Nvidia's market dominance. Others argue the rise of AI agents has pushed computing demand so high that capacity, not competition, remains the binding constraint ^[10]. Both camps can be right. Total AI chip demand can grow parabolically while the margin share captured by merchant chipmakers shrinks, because their own customers are vertically integrating the value away. The volume-bull case and the margin-bear case operate on different axes: one is about how many chips the world needs, the other about who captures the profit inside each chip. The selloff of June and July 2026 is the first time the market began distinguishing between those two questions. What comes next is observable. If Alphabet's external TPU sales gain traction and Amazon follows through on selling Trainium to third parties, hyperscalers will have crossed from buyer to vendor in the merchant silicon market, not just inside their own walls. The next earnings cycle will show whether Broadcom's custom-ASIC margins stabilize or continue their slide, and whether Nvidia's CPU diversification can replace the margin it is losing as inference shifts to chips its customers design themselves.