# Google, Amazon and Meta Are Out-Growing Nvidia With Their Own AI Chips > Source: > Published: 2026-08-18 21:06:29+00:00 *Custom AI chips from Google, Amazon, Microsoft and Meta are set to grow shipments 44.6% this year, nearly triple the 16.1% growth rate for Nvidia's GPUs, according to TrendForce.* Nvidia's biggest customers are becoming its biggest competitors. Google, Amazon, Microsoft and Meta together spend tens of billions of dollars a year on Nvidia's GPUs, and all four are now shipping their own AI chips in growing volume: Google's TPU, Amazon's Trainium, Microsoft's Maia and Meta's MTIA. According to TrendForce, shipments of these custom accelerators, known in the industry as ASICs, will grow 44.6% this year, compared with 16.1% for general-purpose GPUs like Nvidia's. That's a nearly three-to-one gap, and it's the first year custom silicon has meaningfully outpaced the GPU it was built to replace. First time for everything. TrendForce expects ASIC-based servers to make up 27.8% of all AI server shipments in 2026, rising to 40% by 2030. Nvidia still controls roughly 70% of the AI chip market by most industry estimates. This isn't a collapse. It's a slow leak, and it's happening at the exact companies that made Nvidia the most valuable chipmaker on earth. The economics are blunt. Nvidia charges roughly 70% gross margins on its GPUs, which is great for Nvidia and expensive for everyone buying at scale. Jensen Huang has more or less admitted the calculation his customers are making. Asked about the threat from custom chips, he said hyperscalers don't need an ASIC that beats Nvidia. It "just needs to be not more than 70% worse," he said, because that's the margin they're paying to avoid. He's also pushed back on the idea that ASICs are actually cheaper to run, arguing ASIC margins land closer to 65%, not far off Nvidia's own. Huang isn't wrong about the hard part. Building a chip that beats Nvidia is genuinely difficult. But he's also dismissed custom silicon as "a science project," which is a harder line to hold now that Broadcom is guiding to $56 billion in AI semiconductor revenue for fiscal 2026, almost entirely from designing other companies' chips. No hyperscaler is moving faster than Google's TPU program, engineered together with Broadcom. In April, Broadcom and Google signed a long-term supply deal running through 2031 that covers roughly 3.5 gigawatts of next-generation TPU capacity starting in 2027, according to CNBC. Anthropic is part of that build-out too, already running on about 1 gigawatt of Google TPUs this year. That scales to 3 gigawatts by 2027. Mizuho analysts put a number on what that's worth to Broadcom: $21 billion in 2026 revenue from the Anthropic commitments alone, $42 billion in 2027. Broadcom and Marvell design roughly 95% of the custom AI chip market between them. They profit no matter which hyperscaler wins. Amazon's Trainium is shipping too. So is Microsoft's Maia, and so is Meta's MTIA - just without a deal that size attached to any of them. And now OpenAI wants in: it's developing its first custom chip with Broadcom for a 2027 launch, with more than a gigawatt of planned capacity. Even Nvidia's most GPU-hungry customer wants an exit ramp. [Nvidia is now bankrolling the company that buys its chips and the numbers are getting hard to ignore](https://startupfortune.com/nvidia-is-now-bankrolling-the-company-that-buys-its-chips-and-the-numbers-are-getting-hard-to-ignore/) Nvidia is in talks to guarantee up to $250 billion for OpenAI's data center buildout in Ohio, with separate discussions covering $350 billion more in chip financing. The deal reignites circular-financing concerns as Nvidia's total AI deal exposure crossed $750 billion last week. - [nvidia financing openai infrastructure deals](https://startupfortune.com/nvidia-is-now-bankrolling-the-company-that-buys-its-chips-and-the-numbers-are-getting-hard-to-ignore/) - [gpu supplier funding data center buyers](https://startupfortune.com/nvidia-is-now-bankrolling-the-company-that-buys-its-chips-and-the-numbers-are-getting-hard-to-ignore/) ## Is this actually denting Nvidia yet Not yet, at least not in the numbers. Nvidia posted $81.6 billion in revenue for its most recent quarter, reported in May, up 85% year over year, with data center revenue alone hitting $39.1 billion, up 69%. Huang says Nvidia has $1 trillion in committed orders through 2027 and that the company is "going to be short" on supply, not sitting on GPUs nobody wants. Every chip built on TSMC's 3-nanometer process, custom or otherwise, is competing for capacity that's already running flat out, with demand outstripping supply by roughly three to one. So this isn't Nvidia losing customers. It's Nvidia's customers diversifying their supply while still buying everything Nvidia can make - a real distinction, for now. But shipment growth running three times faster than Nvidia's own doesn't stay a rounding error forever, and 27.8% of the AI server market is not a science project anymore. It's a second industry, built inside the first one, by the four companies with the most reason to want out. **Also read:** [Etched's AI chip valuation doubles to $21 billion in a single month](https://startupfortune.com/etcheds-ai-chip-valuation-doubles-to-21-billion-in-a-single-month/) • [Anthropic Shows AI Agents Can Infect Each Other With a Self-Spreading Goal](https://startupfortune.com/anthropic-shows-ai-agents-can-infect-each-other-with-a-self-spreading-goal/) • [Claude AI Suffers Widespread Outage Across All Its Models on August 18](https://startupfortune.com/claude-ai-suffers-widespread-outage-across-all-its-models-on-august-18/)