Should Nvidia Worry About Annapurna, AWS’s Secretive Chipmaker?

Ten years ago, Amazon paid big bucks for an unknown Israeli chip startup. Today, custom silicon maker Annapurna makes the specialized AI processors that power AWS, the cash cow underpinning Amazon’s entire business model.

The firm – which still operates as an independent unit – is so vital to growth that analysts call its processors the secret sauce behind AWS. What’s so special about the company, and how has it displaced Nvidia inside one of the digital economy’s cornerstone firms?

We dig into the story of Big Tech’s most successful 70s-style vertical integration.

Table of Contents Table of Contents

1. Key Takeaways
2. Competing for Scale
3. What Is Annapurna?
4. The Brains of AWS
5. The Bottom Line
6. FAQs
7. References

Show Full Guide

Table of Contents

1. Key Takeaways
2. Competing for Scale
3. What Is Annapurna?
Show Full Guide
4. The Brains of AWS
5. The Bottom Line
6. FAQs
7. References

Competing for Scale

The AI market’s rush for gold has kicked off a trillion-dollar arms race between the titans of tech. Amazon, Alphabet, and Microsoft are all investing gargantuan sums on specialized processors that can handle generative AI’s heavy, and growing, compute requirements. All three firms are building custom chips of their own and working to reduce their reliance on Nvidia. No company is more committed to that goal than Amazon.

The cloud and e-commerce giant has a $100 billion capex plan in place for new AWS infrastructure designed to meet the demands of AI systems.

It calls for the creation of a massive supercomputer built on an ultracluster of advanced processors from Annapurna. By creating its own custom silicon, Amazon reckons it can deliver faster and more cost-effective AI cloud services.

A company that made its name selling books and CDs is now focused on machine learning and microchips. That pivot was made possible by a canny startup acquisition made over a decade ago.

What Is Annapurna?

Annapurna bills itself as a ‘specialist in what’s never been done.’ From a careers page inside the Amazon website, it touts a chip product suite that consumes ‘60% less energy than comparable instances’ (combinations of CPU, memory, storage, and networking capacity), and 600+ instance types ‘for virtually every business need’.

The firm was founded in 2011 by Israeli entrepreneur Nafea Bshara. At that time, tech was obsessed with touchscreens, smartphones, mobile apps, and tablets. He and his team of chip industry veterans went against the grain and stuck with what they knew best: IT infrastructure and the processors that power it.

Named for a mountain in Nepal that’s notoriously difficult to scale, Annapurna seems to have timed its debut perfectly – just as the spotlight was moving from mobile and consumer gadgets to cloud software and services.

Though backed by big names like ARM and VC firm Walden International, the early-stage firm operated in the shadows, launching in stealth mode to avoid scrutiny by larger competitors. A fateful meeting with Amazon’s lead engineer led to a full-scale buyout offer from then AWS chief Andrew Jassy (now Amazon CEO), a few years later.

Amazon’s plans for the cloud were still developing and it had only started thinking about building its own cloud-focused chips, hoping they might give AWS more pricing flexibility with customers, as well as control over a key part of its tech supply chain.

The Brains of AWS

From labs in Tel Aviv, Toronto, and Austin, Texas, Annapurna designs and manufactures AI-accelerating processors, including the flagship Trainium2, which enables high-compute AI model training at reduced power consumption and lower cost.

Engineered for efficiency and performance, AWS says features like enhanced heat management and fewer internal components combine to boost Trainium2’s computational heft. These improvements cater specifically to the training of machine learning models, positioning the Trainium2 chip as a formidable contender against Nvidia’s offerings.

Analysts say the processors have the capabilities to speed-up model training times over first generation Tranium, accelerating new LLM rollouts while delivering up better energy efficiency.

In a review published in December 2024, researchers at SemiAnalysis said that Trainium2 offers a performance uplift over first-generation Tranium and “a competitive custom silicon with training and inference on the chip, system, and at the software compiler/frameworks level.”

While not a ringing endorsement on its own, some of the world’s biggest AI firms have been impressed enough to put the platform to the test. Apple and OpenAI competitor Anthropic are both reportedly trialling Trainium2 for model training, while Apple has moved to AWS services powered by Annapurna’s Graviton and Inferentia chips.

Emily Webber speaks to @JonKrohnLearns about her work at Amazon Web Services, from its Annapurna Labs-developed Nitro System, a foundational technology that can enhance securities and performance in the cloud and how Trainium2 became AWS’ most powerful AI chip with four times the… pic.twitter.com/nMR7c5jbbD

— SuperDataScience (@superdatasci) April 22, 2025

The Bottom Line

Amazon has been criticized for falling behind big tech rivals, repeatedly delaying the launch of a generative AI version of its Alexa chatbot while Google and Apple unveiled increasingly smarter AI digital assistants.

But critics missed the fact that the real AI action has been going on behind the scenes for years. Annapurna has been quietly speeding up Amazon’s corporate transformation, from mail-order paperbacks to building AI-accelerating chips in the cloud.

FAQs