HARDWARE DARWINISM: SURVIVAL OF THE MOST COST-EFFICIENT

(NVDA), (AAPL), (GOOG), (MSFT), (META), (AMD), (AMZN), (ORCL)

I was touring a Google data center last month with an old contact from my Tokyo correspondent days who’s now deep in the cloud infrastructure world, when something clicked that completely changed how I think about the AI investment landscape. 

We were standing in front of rows upon rows of humming servers, and he casually mentioned, “You know, John, about half of these aren’t even running Nvidia (NVDA) chips anymore.” 

That comment hit me like a freight train because here I was, looking at the physical manifestation of a massive shift that most investors are completely missing.

For the past two years, I’ve been hearing the same narrative from every financial analyst and their grandmother: Nvidia owns AI, end of story. 

And sure, the numbers seem to back that up – roughly 90% of AI servers still use Nvidia’s GPUs, thanks to their speed and that bulletproof software ecosystem that makes everything work seamlessly together. 

Think of Nvidia as the Apple (AAPL) of AI hardware: expensive, premium, but everything just works. The problem with this thinking is that it assumes the status quo will last forever, which in technology is about as reliable as a weather forecast in tornado season.

What my data center tour revealed, and what subsequent conversations with chip designers and cloud operators have confirmed, is that we’re witnessing the early stages of a fundamental diversification in AI hardware. 

Companies like Google (GOOG), Microsoft (MSFT), and Meta (META) aren’t just experimenting with alternatives to Nvidia – they’re actively deploying them at scale. 

Google’s TPUs, AMD’s (AMD) MI300X chips, and custom ASICs are starting to carve out meaningful market share, driven by two powerful forces that make perfect business sense: cost savings and supply chain independence.

The economics are frankly staggering when you dig into the details. One contact at a major cloud shop walked me through the math, and it made my inner quant swoon.

AMD’s chips can cut operational costs by 20-40% compared to Nvidia’s top-tier hardware, while Google’s TPUs are delivering cost per token improvements that can reach $0.0015 versus $0.002 for Nvidia’s H100s. 

That might sound like pocket change, but when you’re processing billions of tokens daily, those fractions add up to millions in savings.

But here’s where it gets really interesting from an investment perspective. This goes beyond saving money. It’s about strategic risk management. 

Nvidia’s best GPUs cost tens of thousands of dollars each and often come with waiting lists longer than the line for cronuts back in 2013. 

Relying on a single supplier for something as critical as AI infrastructure is like putting all your retirement savings in one stock – it might work out great, but it’s not exactly what you’d call prudent portfolio management.

The technical details matter because they drive the investment thesis. 

AMD’s MI300X chips pack significantly more memory right next to the processor, which is like having a massive refrigerator next to your kitchen counter instead of in the garage – everything runs faster and more efficiently. 

Meanwhile, Google’s TPUs are optimized for the specific mathematical operations that AI models love most, making them incredibly efficient for certain workloads even if they’re not as versatile as Nvidia’s offerings.

What we’re seeing is a multi-tool market emerging. 

Nvidia still rules for training bleeding-edge models. But for inference – the grunt work of running AI models at scale – AMD and Google are serving up cheaper, targeted tools that get the job done.

It’s hardware Darwinism: adapt or get outcompeted. And investors who only see Nvidia are staring at the sun and missing the solar system.

The ripple effects go way beyond semis. Microsoft, Amazon (AMZN), Oracle (ORCL) – they’re all broadening their hardware stacks. 

More flexibility means better margins and less risk. AI application developers get more levers to pull. Even Nvidia benefits. After all, nothing sharpens innovation like hot breath from competitors.

My prediction, based on conversations with industry insiders and my own analysis of the cost dynamics, is that by 2028 we’ll see roughly one-third of new AI servers using non-Nvidia hardware. That doesn’t mean Nvidia is doomed – far from it. 

The AI market is growing so rapidly that even a smaller slice of a much bigger pie could still represent enormous growth. But it does mean we stop treating Nvidia like the Pope of AI and start thinking about the sector like a toolbox, where different jobs need different tools.

As we stepped out of the data center, my friend turned and said, “The future isn’t about one company owning everything. It’s about having the right tool for every job.”

Smart guy, that one. No wonder we’ve stayed friends all these years.