Amazon’s Chip Play: Beyond Nvidia

by

Amazon, that relentlessly expanding empire of everything, recently dropped a hint about its future spending. Two hundred billion dollars, they say, earmarked for capital expenditures this year. That’s a sum so large it’s almost impossible to grasp. It’s roughly the GDP of a moderately sized European nation, and they’re planning to spend it on… well, mostly data centers. And data centers, as any reasonably informed person knows, are built on chips.

Now, everyone’s been assuming Nvidia, the current darling of the AI boom, would be the primary beneficiary. And they will be, to a degree. Amazon will be loading up on Nvidia’s GPUs, those little silicon workhorses that power so much of the digital world. But something rather interesting is happening. Amazon, it turns out, isn’t content to simply buy the future; it wants to build it. And that means chips of its own.

The $10 Billion Question

Andy Jassy, Amazon’s CEO, casually mentioned that the company’s custom chip business within Amazon Web Services is now running at over $10 billion annually. Ten billion! That’s like selling enough Kindles to outfit the entire population of Belgium. And, get this, it’s growing at a triple-digit percentage rate. Which, if my calculations are correct (and they often aren’t), means it’s doubling roughly every year. It’s enough to make a person wonder if they should have paid more attention in physics class.

Amazon’s foray into chip design isn’t just a whimsical hobby. They’ve developed Graviton CPUs, and then there are Trainium and Inferentia – custom AI accelerators for training and inference. Apparently, the Trainium2 chip was in such demand they could barely keep up. They’ve already announced Trainium3, and Jassy confidently predicts it will be fully committed by mid-2026. It’s a bit like a very popular restaurant – you need to book well in advance just to get a table.

Loading widget...

It’s not just Amazon, either. Alphabet (Google) is tinkering with Tensor Processing Units, Anthropic reportedly ordered $20 billion worth of chips, Microsoft and Meta are also getting in on the act. Everyone, it seems, wants a piece of the silicon pie. Nvidia remains a key supplier, naturally, but this growing appetite for custom silicon suggests that the landscape is shifting. Perhaps Nvidia’s most spectacular growth days are behind it, while other companies, those quietly designing their own chips, are poised to benefit.

Marvell: The Quiet Beneficiary

And that brings us to Marvell Technology. Now, Marvell isn’t a household name, unless your household happens to be a data center. But they’re a crucial partner in Amazon’s chip ambitions. They’ve been working with Amazon to design the Trainium chips, and recently solidified their partnership with a five-year agreement to supply chips to AWS data centers. It’s a bit like being the unsung hero of a blockbuster movie – you don’t get the credit, but without you, the whole thing falls apart.

However, the market hasn’t exactly been showering Marvell with accolades. There have been whispers that its position with both Amazon and Microsoft might be diminishing. Apparently, Amazon is increasingly relying on a company called AIChip for the design of its Trainium3 and Trainium4 chips, reducing Marvell’s role to a smaller part of the overall interface. It’s a bit like being relegated to the bench after being a star player.

Loading widget...

But here’s the thing: Marvell does a lot more than just custom AI accelerators. They’re a major player in networking chips, which are essential for data centers. Amazon will likely continue to rely on Marvell for interconnects, switching, and storage – the essential plumbing of the digital world. And their recent acquisition of Celestial AI should further strengthen their position in AI-focused interconnect chips. It’s a bit like having a diversified portfolio – you’re not putting all your eggs in one basket.

There were also fears that Marvell might lose its Microsoft design contract. But that doesn’t seem to be the case, at least not for now. Management is forecasting a significant increase in fiscal 2028 revenue from custom AI accelerators, driven by Microsoft’s Maia 300 chip. CEO Matt Murphy insists that nothing has changed in the company’s outlook, despite all the recent noise. It’s a bit like a seasoned captain calmly steering the ship through a storm.

So, Marvell appears to be one of the best opportunities to capitalize on the growing demand for custom silicon in hyperscale data centers. With shares trading at just 22.6 times forward earnings, it looks like an excellent opportunity for investors. It’s a bit like finding a hidden gem – you just need to know where to look.

Read More

2026-02-20 00:33