Nvidia shares climbed Thursday after Amazon raised its capital expenditure forecast and downplayed the competitive threat from its in-house Trainium accelerators, easing investor concerns that hyperscaler custom silicon would erode the GPU maker's data center dominance.
Nvidia shares climbed Thursday after Amazon raised its capital expenditure forecast and downplayed the competitive threat from its in-house Trainium accelerators, easing investor concerns that hyperscaler custom silicon would erode the GPU maker's data center dominance.
Nvidia shares rose Thursday after Amazon raised its capital expenditure forecast and downplayed competition from its custom Trainium chips, reinforcing expectations that hyperscaler AI infrastructure spending will continue to flow through Nvidia's data center business.
The reassurances come as hyperscalers collectively plan roughly $700 billion in AI infrastructure capital expenditure for 2026, according to industry estimates. Amazon's raised forecast signals that its custom silicon push expands the total AI compute market rather than shrinking Nvidia's addressable share.
Amazon's Trainium 3 accelerator, reportedly targeting Intel's EMIB-T packaging platform, represents the company's most aggressive push into custom silicon. Yet Amazon executives downplayed the competitive dynamic, framing Trainium as serving workloads where cost efficiency matters more than raw performance, while Nvidia's GPUs remain the default for training-class models.
The stock move reflects a broader reassessment of the AI semiconductor complex. With TSMC's CoWoS packaging capacity sold out through 2026 and well into 2027, and lead times running 52 to 78 weeks, the constraint on AI chip supply is manufacturing capacity, not demand. Every additional dollar of hyperscaler capex translates into more packaging demand, regardless of whether the chips come from Nvidia or from custom silicon.
The competitive dynamics between Nvidia and hyperscaler custom silicon have been a persistent overhang on the stock. Amazon's Trainium line, Google's TPUs, and Microsoft's Maia chips all represent attempts by cloud providers to reduce their dependence on Nvidia's GPUs, which command premium pricing due to their performance lead and CUDA software stack lock-in.
Amazon's decision to raise its capex forecast while simultaneously downplaying the competitive threat addresses both sides of the investor equation: it confirms that AI infrastructure spending is accelerating, and it suggests that custom chips are additive rather than substitutive. This is a meaningful signal because Amazon is one of the largest buyers of Nvidia GPUs, and any shift in its procurement strategy would have outsized impact on Nvidia's data center revenue.
The broader context supports this interpretation. The advanced packaging market is projected to reach approximately $69.5 billion by 2029, according to Yole Group. TSMC's CoWoS capacity is fully sold out through 2026 and well into 2027, with lead times running 52 to 78 weeks depending on configuration. Intel's EMIB-T platform, which has reached a 98 percent yield rate, is gaining traction with Google, Amazon, and potentially Nvidia itself.
Amazon's Trainium 3, reportedly targeting Intel's EMIB-T packaging platform, is designed for inference workloads where cost per token matters more than peak performance. Nvidia's GPUs, by contrast, remain the standard for training large models, where raw compute density and interconnect bandwidth determine throughput. This division of labor — Nvidia for training, custom silicon for inference — is the framework Amazon executives used to downplay the competitive threat.
The framing has support from industry data. Nvidia's CUDA software stack, which has been in development for nearly two decades, creates switching costs that custom silicon cannot easily overcome. Hyperscaler custom chips require their own software stacks, which take years to mature. In the interim, Nvidia's GPUs remain the default choice for frontier model training.
Nvidia shares, which have been volatile as investors weigh the custom silicon threat, now have a clearer picture: hyperscaler capex is accelerating, and the largest cloud providers are explicitly framing their custom chips as complementary to Nvidia's products. This supports Nvidia's revenue outlook and could lift the broader AI semiconductor complex.
The key risk remains execution. If Amazon's Trainium line achieves performance parity with Nvidia's GPUs in training workloads — not just inference — the competitive calculus changes. But for now, the market is pricing in the more benign scenario: a growing AI compute market where Nvidia captures the high-end training segment and custom silicon serves cost-sensitive inference workloads.
This article is for informational purposes only and does not constitute investment advice.
