Amazon’s Little Rebellion

by

One does rather tire of the predictable. Amazon, it seems, has decided to be…difficult. A rather substantial investment in infrastructure – two hundred billion, if you please – has caused a momentary flutter amongst the financial pigeons. Mr. Jassy, the Amazon CEO, assures us they are “monetizing capacity as quickly as they install it.” One hopes it’s a profitable endeavor, though frankly, one suspects they’re having a spot of fun upsetting the established order.

The truly interesting development, however, isn’t the spending itself, but what they’re spending it on. Apparently, 1.4 million of their Tranium2 chips are now humming away in their data centers, and the results, shall we say, are…noticeable. A ten-billion-dollar annual run rate, growing at a rather impudent 100%? One begins to suspect a deliberate attempt to create a bit of a stir.

Now, Amazon remains a perfectly good customer of Nvidia, naturally. But this Tranium business…it introduces a rather charming little crack in Nvidia’s otherwise impeccable facade. A crack, mind you, that could, with a bit of luck, become a rather significant fissure.

Lowering Costs – Or Simply Being Clever

Nvidia, of course, dominates the AI accelerator market, and their chips are priced accordingly. One understands the necessity of profit, naturally, but it does leave other players rather…constrained. Amazon, with its Tranium chips, is claiming a 30-40% performance-per-dollar improvement. A bold claim, certainly, but one that seems to be resonating.

Anthropic, a rather fashionable AI start-up, is now employing these Tranium2 chips to train their next-generation models. AWS’s Project Rainier, boasting 500,000 Tranium2 chips (eventually a million, naturally), is the engine driving their Claude family of models. One can’t help but admire the audacity.

And the future? Tranium3 is on its way, promising another 40% improvement. The current fleet is sold out, and Tranium3 is expected to follow suit by 2026. They’re even working on Tranium4! It’s almost…enthusiastic.

Their Graviton CPU, installed in their data centers, is also proving rather successful, generating a multi-billion-dollar revenue stream. Apparently, 90% of AWS’s top 1,000 clients are utilizing it. One wonders what Nvidia thinks of all this. Probably nothing they’d share with us, naturally.

Nvidia’s Little Problem

Amazon isn’t selling these custom chips directly, of course. But every Tranium chip installed represents one less Nvidia GPU purchased. And they aren’t alone. Alphabet has been tinkering with TPUs for years, and Microsoft is now on its second generation of Maia AI chips. The competition, you see, is rather…stirring.

These custom chips are particularly efficient at AI inference – the act of using a trained AI model. As AI usage expands, inference is likely to surpass training in terms of overall computing power. Having alternatives, therefore, is rather…sensible.

Loading widget...

Increased AI adoption depends on costs coming down, and these efficient custom chips could be the key. More powerful AI models are enabling exciting new applications, but broader adoption is hindered by the expense of running them.

Amazon and other cloud giants will continue to purchase Nvidia GPUs in vast quantities. But having alternatives provides leverage. Nvidia may find its profit margins under pressure as the market becomes more competitive. One suspects a bit of a price war is brewing, and frankly, one is rather looking forward to it. It’s all terribly…entertaining.

Read More

2026-02-09 23:02