Amazon opens Trainium lab — the chip chosen by Anthropic, OpenAI, and Apple

Amazon held an exclusive tour of its Trainium lab for TechCrunch journalists immediately after announcing its largest investment of $50 billion in OpenAI. This is no coincidence: Trainium is precisely the chip that underpins AWS's new strategy as an AI infrastructure player. Trainium is a line of custom AI accelerators that Amazon has been developing under the AWS brand since 2020.

Trainium 2, announced in 2023, is positioned as a direct competitor to NVIDIA's H100: it is optimized for training large language models and promises comparable performance at significantly lower cost per operation. Amazon is investing billions in its own silicon foundation precisely because NVIDIA holds a dominant position in the AI hardware market — this turns its chips into a scarcity and a source of pricing power. Notably, among Trainium's clients are companies that are simultaneously competitors and partners to Amazon in the AI space.

Anthropic — a startup in which Amazon invested up to $4 billion — actively uses Trainium to train models in the Claude family. OpenAI, which just received $50 billion from Amazon, is also among the users of this platform. The most unexpected name on the list turned out to be Apple: a company that traditionally builds vertically integrated products on its own hardware, turned to AWS Trainium for some of its AI tasks.

It is precisely this combination of names — Anthropic, OpenAI, Apple — that turned the lab tour into an event. AWS is essentially telling the market: if the three most demanding AI organizations in the world when it comes to hardware quality chose Trainium, this is no accident and no marketing. Development is led by the team at Annapurna Labs — an Israeli startup acquired by Amazon in 2015.

Annapurna is the company that created the Graviton chips and the Nitro hypervisor. Trainium is the next bet by the same team, this time in the AI accelerator segment. The fundamental difference between Trainium and NVIDIA is the business model.

Trainium is available exclusively through AWS: it cannot be purchased and deployed in one's own data center. Amazon monetizes the chip through cloud compute hours, not through direct equipment sales. For clients, this is advantageous — access to powerful accelerators without capital expenditure.

For Amazon — a tie-in of the world's largest AI labs to its own infrastructure. The $50 billion investment in OpenAI is not simply a financial bet on the market leader. It is a commitment from OpenAI to use AWS as a key cloud partner for training and inference.

Trainium is the physical embodiment of this agreement. Amazon is systematically building its position as the default AI infrastructure player: through investments in leading labs, through custom silicon, and through AWS's scale. If Trainium demonstrates performance comparable to NVIDIA's at lower cost — pressure on NVIDIA's dominance in the AI segment will become significantly more real.

The fact that even Apple — the company with the most closed hardware stack in the industry — uses this chip speaks to the fact that the strategy is already working.