Nvidia Falls After Hyperscaler Reports

📉📉📉 Nvidia fell after reports from hyperscalers, even though capex increased by $15 billion.
On April 29, Meta, Amazon, Microsoft, and Google collectively raised their capex forecast for 2026. According to the old logic of AI trading, Nvidia should have soared. Instead - a sell-off. The reason is that GPUs are no longer the main bottleneck in AI infrastructure.
Mark Zuckerberg stated directly in Meta's report: "We are increasing our infrastructure CapEx forecast for this year. Most of that is due to higher component costs, particularly memory pricing." Meta's capex is rising to $125-145 billion (up from the previous range of $115-135). HBM (high-bandwidth memory) is sold out until the end of 2026, and prices are rising. The pricing power is now with SK Hynix, Samsung, and Micron, not Nvidia.
Amazon announced ~$200 billion capex for 2026. But the most interesting part is that the AWS custom chip business (Graviton + Trainium + Nitro) is already on a run-rate of $20 billion a year, growing three hundred percent year over year. Andy Jassy mentioned that Trainium will provide savings of "tens of billions of dollars" in capex annually and "several hundred basis points" of operating margin advantage over using third-party chips for inference. This is the most direct public hit to Nvidia's margins from a hyperscaler.
The picture across all four:
<ul>
<li>Meta: rolling out over 1 gigawatt of its own silicon developed with Broadcom, plus a significant volume from AMD</li>
<li>Amazon: Trainium/Graviton at a $20B run-rate, AWS grew 28% year over year (the fastest in 15 quarters)</li>
<li>Microsoft: Maia accelerator provides "30%+ improved tokens per dollar", CPU Cobalt is already in half of the data center regions</li>
<li>Google: pushing for full integration - "own the frontier models, own the silicon"</li>
</ul>
None of them said they are replacing Nvidia. Everyone said they are buying less Nvidia per unit of new compute than they would have before.
Meanwhile, the second bottleneck is growing - power. Caterpillar and GE Vernova quietly reported strongly because after memory, the next limitation is electricity. Chips can be bought, but gigawatts cannot be brought to the data center.
<a href="https://www.fool.com/earnings/call-transcripts/2026/04/29/meta-meta-q1-2026-earnings-call-transcript/">Meta Q1 2026 Earnings Call Transcript</a>
<a href="https://www.aboutamazon.com/news/company-news/amazon-ceo-andy-jassy-aws-ai-q1-2026-earnings">Amazon CEO Andy Jassy on AWS AI Q1 2026 Earnings</a>
#nvidia #ai #hbm #trainium #mark_zuckerberg