Why Every Big Tech Company Wants Its Own AI Chip

On July 12, SK Hynix closed the largest foreign IPO ever recorded in the United States, driven almost entirely by demand for HBM memory chips used in artificial intelligence. That same week, rumors about Apple's upcoming M7 Ultra chip suggested support for up to 1.5 TB of RAM. Two seemingly unrelated news items, but both point to the same trend: big tech companies no longer want to rely solely on Nvidia to run their AI models.

Why Buying Nvidia Chips Isn't Enough

Nvidia has dominated the market for processors that train and run AI models for years, and for good reason: its chips and the surrounding software ecosystem are hard to match. The problem is that this dominance has become a bottleneck. Demand for processing power has grown much faster than Nvidia's ability to manufacture and deliver, and companies that depend entirely on a single supplier become hostage to its pricing and waiting list.

The Real Bottleneck Is Memory, Not Just Processing

Running a large language model isn't just about computing power. The entire model, with its billions of parameters, needs to be as close to the processor as possible, and that's where HBM memory (High Bandwidth Memory) comes in. It's more expensive and harder to manufacture than standard computer memory, and demand for it has grown so much that manufacturers like SK Hynix have become as strategically important as Nvidia itself. That explains why a memory company, not a processor company, led the largest foreign IPO in American history.

Who's Already Building Their Own Chip

This trend isn't new; it's just accelerating. Google has been developing its own TPU chips for over a decade to run Gemini and other internal models. Amazon has Trainium and Inferentia chips powering part of AWS's infrastructure. Microsoft unveiled the Maia chip for its AI data centers. Even OpenAI, which doesn't manufacture hardware, has signaled interest in developing its own chip capabilities to reduce reliance on external suppliers.

Apple is taking a similar path through a different door: it already dominates its own silicon design in iPhones and Macs, and is expanding that strategy to support increasingly large AI workloads locally, as rumors around the M7 Ultra suggest. None of these companies have stopped buying Nvidia chips. The goal isn't to completely replace the supplier, but to reduce dependence and gain control over part of their own infrastructure.

What This Means for Everyday AI Users

For end users, this trend will likely show up indirectly: faster models, gradually falling per-query costs, and new features like larger context windows becoming feasible. At the same time, it creates a long-term rivalry between hardware makers and model builders, which should shape the price and availability of AI tools in the coming years.

FAQ

Does this mean Nvidia will lose the AI market?

Not in the short term. Nvidia remains the world's leading supplier of AI chips. The big tech move is about reducing dependence, not completely replacing that supply.

Why is memory as important as the processor?

Because an AI model needs to keep its parameters quickly accessible during processing. Without enough high-speed memory, even the world's fastest processor sits idle waiting for data.

Will this lower the price of AI tools?

That's one of the industry's bets. More hardware suppliers and less dependence on a single bottleneck tend to increase supply and, over time, push costs down.

Can smaller companies also do this?

Unlikely. Designing a chip from scratch requires billions of dollars in investment and years of development, which is only feasible for companies with the scale and cash of Google, Amazon, Microsoft, and Apple.