Image Source: Unsplash

Today’s artificial intelligence (AI) boom is driven by massive computer “factories” filled with special processors called GPUs. These machines are extremely powerful, but they use huge amounts of electricity—so much that companies are now struggling to find enough energy to keep them running.

NVIDIA (NVDA) is the clear leader in this space. About 90% of the servers built for AI use NVIDIA’s GPUs, thanks to their speed and the strong software support that comes with them. Think of NVIDIA as the “Apple” of AI hardware: everything works together smoothly and reliably.

But big tech companies like Google, Microsoft, and Meta (Facebook’s parent company) are starting to experiment with building their own chips, called ASICs (Application-Specific Integrated Circuits). These custom chips are like finely tuned race cars: they cost a lot to design, but they can be much cheaper and faster for specific jobs. ASICs can cut the cost of running AI by 20–40% compared to NVIDIA’s chips, and they don’t have to wait in line for NVIDIA’s often-sold-out hardware.

If these custom chips become widely available—and if the software that runs on them gets as good as NVIDIA’s—then by 2028, up to one-third of new AI servers could use these alternatives instead of NVIDIA.

Tough Choices for Tech Leaders

Companies must balance spending big for top performance against saving money. For now, the top companies—Google, Amazon, Microsoft, and OpenAI—still rely mostly on NVIDIA for the best results. But they’re also investing in alternatives like AMD’s chips or building their own to avoid price increases or hardware shortages.

There are two main approaches:

• NVIDIA Approach: Buy NVIDIA’s hardware and use its powerful, easy-to-use software. This is expensive but works for almost any AI job.
• Custom Approach: Use a mix of chips from AMD, custom ASICs (like Google’s TPUs), or others. This can be cheaper and more efficient for certain tasks, but is harder to set up and requires more technical know-how.

NVIDIA: The All-in-One Solution

NVIDIA offers a complete package: fast GPUs, top-notch networking, and software that makes it easy to build and run AI systems. Their latest Blackwell chips can run AI models up to 30 times faster than their previous H100 model. Plus, more than 20,000 companies work with NVIDIA, so help and support are always available.

The downside? The price. NVIDIA’s best GPUs can cost tens of thousands of dollars each, and there’s often a waiting list to get them. Relying on one company can also be risky if something changes in the market or with regulations.

The Custom (Non-NVIDIA) Approach

Instead of relying on NVIDIA, some companies choose other chips, like AMD’s MI300X or Google’s TPUs. These can be better for specific jobs—especially when running lots of AI tasks (called “inference”) rather than training new AI models.

For example, OpenAI, known for ChatGPT, started using Google’s TPUs for some of its work, even though it mainly uses NVIDIA hardware. Why? Because TPUs can be cheaper for certain jobs, and sometimes NVIDIA hardware isn’t available.

AMD’s chips can also be a good deal. Building a cluster (a group of connected computers) with AMD’s chips and networking hardware can save around 40% on costs compared to NVIDIA’s setup. Companies like Microsoft, Oracle, and some in the Middle East are already adding AMD chips to their data centers, especially for handling very large AI models that need lots of memory.

What Determines the Best Choice?

A key factor is cost per token—basically, how much it costs to process each word or piece of data with AI. AMD’s MI300X chip, for instance, can bring the cost down to about $0.0015 per token, compared to $0.002 per token with NVIDIA’s H100. That adds up to big savings at scale.

Here’s an analogy: Imagine an AI as a chef making a recipe. The chef (AI) needs fast access to ingredients (data). The faster and bigger the fridge (memory), the better the chef can work. AMD’s chips have a “big fridge” right next to the chef, saving time and energy.

What’s Next?

Until NVIDIA’s next-generation chips (with even bigger memory) are widely available—expected in late 2025—AMD is in a good spot, especially for companies that have the technical skills to make the most of its chips. Google’s TPUs are also becoming more attractive for companies that want to cut costs and run very large AI models.

But it’s not about everyone leaving NVIDIA behind. Instead, companies are starting to mix and match: using NVIDIA when they need versatility and top performance, but turning to AMD or Google for cost savings or to avoid hardware shortages.

The Big Picture

NVIDIA is still the giant in AI hardware, but things are changing. As AI models get bigger and more complex, companies care more about efficiency—how many results they can get for each dollar and each unit of energy.

Cloud providers face new challenges: higher energy bills, long waits for the latest chips, and a need for more flexibility. AMD is offering GPUs with more memory and lower energy use, and Google’s TPUs give companies another way to save money.

If AMD’s software keeps improving—and if more companies can use Google’s TPUs—the future might be less about a single hardware supplier and more about picking the right tool for each job. The AI hardware world is shifting from a one-size-fits-all model to a competitive marketplace, where cost, performance, and energy use all matter.

In summary:

NVIDIA remains the leader, but Google and AMD are catching up—giving companies more choices and helping drive down costs. The next few years will be all about finding the best balance between price, performance, and availability as AI continues to grow.

More By This Author:

Positioning For A Bond Market Hurricane
In The Eye Of The Bond Market Hurricane?
Stock Pause, Small-Cap Bets; Bond Market Trouble Ahead?