Andy Jassy, the CEO of Amazon, told analysts that if Amazon’s chip business were a standalone company, it would generate more than $25 billion a year. And sales are growing at a triple-digit rate. And remember, Amazon is one of Nvidia’s largest customers.
And as if that wasn’t enough, we’re learning that Nvidia itself is raising prices by more than 15%. Exactly in the same week that it reports earnings.
For years, the picture was very simple. Nvidia makes the chips, everyone else buys them. End of story.
But now the customers are sitting at the same table and building their own. Amazon, with the $25 billion we mentioned. Google, which has been working with Broadcom for years and has just struck a deal with Marvell as well, even gave it the right to purchase $12.2 billion worth of shares.
"And why two suppliers?" you might ask. Because none of these companies wants to depend on a single one. Not on Nvidia, and not even on their own partner.
And pay attention to something else. This year, for the first time, Google will launch two versions of its own chips, its TPUs, at the same time. One will be specifically designed for fast responses. And more importantly? It has started providing them to customers outside its own cloud. In other words, Google is no longer making chips only for itself. It is selling them. Just like Nvidia.
The same thing is happening with AI companies. OpenAI is preparing its own chip, called "Jalapeno," together with Broadcom, and it is expected to enter data centers during the year. Initially, it will handle just one task, inference. And the company's semiconductor chief made it clear: this is the beginning of a series of moves aimed at giving OpenAI full control over its entire infrastructure. It is even considering providing its chips to other businesses.
And Anthropic? It has already signed contracts worth tens of billions to use chips from Google, Amazon, and AMD. And it has approached SK Hynix for memory, which means one thing. It is preparing its own chip as well.
Meta, meanwhile, has signed deals with virtually everyone in the industry. Even Qualcomm, which entered the game more recently.
Obviously, not everything is perfect. Because in practice, we are still at the beginning. Over the past few months, the amount of money Meta has spent with Nvidia has increased. And Amazon, the "chip empire" we talked about, continues to send roughly a quarter of its capital expenditures to Nvidia.
And this is where things get even more interesting. Because the very nature of what data centers do has changed.
Until recently, all the money was going into training. In other words, "feeding" the model with data so it can learn. Now we are moving toward inference. That is the stage where the model responds to us, the users, every time we type something.
And here Nvidia is no longer the only option. Because to respond quickly to a prompt, you don't necessarily need the most powerful chip in the world. You need the chip that is best suited for the task.
The result? Right now, there are 150 companies working on more than 200 different chip designs. Yes, you read that correctly. Two hundred. Cerebras held the largest IPO in the semiconductor industry's history in May. AMD doubled its data center revenue to $6.7 billion.
So you might ask... is Nvidia at risk?
Not yet. It controls roughly 90% of the market and has a 75% profit margin, twenty percentage points higher than the next-largest player. And in December, it paid roughly $20 billion for Groq's technology and employees. Precisely for inference.
In simple terms, nobody is really a competitor to Nvidia unless they reach at least one-tenth of its revenue.
And this is where the most unexpected part of the story comes in.
Nvidia's biggest customers were informed that the prices of servers containing its chips are going up by more than 15%. We're talking about systems equipped with Vera Rubin and Grace Blackwell, its flagship platforms. The price increases will take effect at the beginning of next year.
And why is this happening?
Because of memory. DRAM.
Put very simply, an AI chip without memory next to it doesn't do anything. And essentially only three companies on the planet produce this memory: Samsung, SK Hynix, and Micron. They are increasing production, but they still cannot keep up with demand. So they are raising prices.
And here's the striking part. The most profitable and most powerful company in the industry couldn't absorb the higher costs. It passed them on to its customers. Microsoft, Google, and Oracle. In fact, Nvidia also raised prices on gaming graphics cards.
It's not the only one. Apple and Qualcomm have already said that they are being forced to charge more because of chip shortages.
Of course, there's a big "but." All of this is happening in an industry that is already dealing with project delays, a shortage of workers, and more difficult access to capital. So the expansion of data centers has just become more expensive and more complicated.
