Amazon just opened the doors to its secretive chip development lab in Austin, Texas, and what’s inside explains why Anthropic, OpenAI, and even Apple are buying in. TechCrunch AI got an exclusive tour of the facility where Amazon designs Trainium, the custom silicon that’s quietly becoming a real threat to Nvidia’s dominance.
The timing isn’t random. This tour came right after AWS CEO Andy Jassy announced a massive $50 billion deal with OpenAI, making AWS the exclusive provider for OpenAI’s new AI agent builder, Frontier. As part of the agreement, Amazon committed to supplying OpenAI with 2 gigawatts of Trainium computing capacity. That’s a staggering promise, considering Anthropic and Amazon’s own Bedrock service are already consuming Trainium chips faster than Amazon can make them.
The Numbers Behind the Silicon
Here’s what stands out: there are now 1.4 million Trainium chips deployed across three generations. Anthropic’s Claude alone runs on over 1 million Trainium2 chips, according to TechCrunch AI. The chip originally focused on model training, but it’s shifted to handle inference too, which is where the real bottleneck sits right now.
Trainium2 already handles the majority of inference traffic on Amazon’s Bedrock service, the platform enterprise customers use to build AI applications. Lab director Kristopher King didn’t mince words about demand: “Our customer base is just expanding as fast as we can get capacity out there.” He even compared Bedrock’s potential to EC2, AWS’s foundational compute service. That’s a bold claim, but the trajectory supports it.
Why This Matters for Nvidia’s Grip
Amazon says the latest Trainium3 chips, running on new Trn3 UltraServers, cost up to 50% less than classic cloud servers for comparable performance. Combined with new Neuron switches that let every Trainium3 chip communicate with every other chip in a mesh configuration, the result is significantly lower latency.
“That’s why Trainium3 is breaking all kinds of records,” director of engineering Mark Carroll told TechCrunch AI, “particularly in price per power.” When you’re processing trillions of tokens daily, those savings compound fast.
The switching cost problem, traditionally Nvidia’s biggest moat, is getting smaller too. Amazon’s team says Trainium now supports PyTorch natively. Carroll described the migration effort as “basically a one-line change, and then recompile, and then run on Trainium.” That’s a direct shot at Nvidia’s CUDA lock-in.
Apple’s Quiet Endorsement
One detail worth noting: Apple publicly praised this chip team back in 2024, according to TechCrunch AI. Apple’s director of AI described how the company used Graviton, the team’s ARM-based server CPU, and gave a nod to both Inferentia and Trainium. For a company as secretive as Apple, that’s a meaningful endorsement.
What’s Next
Amazon isn’t stopping at silicon. The team designs the full stack: servers, networking, Nitro virtualization hardware, and liquid cooling systems. AWS also announced a partnership with Cerebras Systems this month, integrating Cerebras inference chips on Trainium-powered servers for what Amazon promises will be ultra-low-latency AI performance.
This is the classic Amazon playbook: find what people want, build a cheaper in-house version, and scale relentlessly. The difference now is the stakes. With AI infrastructure demand outpacing supply across the industry, Amazon’s ability to produce its own chips gives it leverage that pure cloud resellers simply don’t have.
One complication worth watching: Microsoft reportedly believes OpenAI’s AWS deal may violate its own agreement with OpenAI, specifically around access to all of OpenAI’s models and tech. That tension could reshape how these partnerships evolve.
The full story, including photos from inside the lab, is available at TechCrunch AI.