CPUs Are Back, Baby

(NVDA), (AMD), (AMAT), (ASML), (META)

Last quarter, a senior AI architect I’ve known for years quietly left his plush post at Nvidia (NVDA). Where did he land? AMD (AMD). When the talent starts migrating, especially the ones who know how the sausage gets coded, you can bet they see the shift before the market does. 

AMD just posted 36% year-over-year revenue growth with earnings of $1.17 per share, beating estimates on both fronts while the stock rallied over 100% year-to-date. Wall Street’s celebrating the GPU wins and generally missing the forest for the very expensive semiconductor trees. 

Because here’s what’s actually happening: AMD isn’t just winning the GPU arms race anymore. They’re quietly orchestrating something far more interesting, and it showed up buried in an earnings call comment that most people glossed over.

CEO Lisa Su mentioned that multiple hyperscale clients are forecasting “significant CPU build into 2026.” If you’ve been around semiconductor cycles as long as I have, and trust me, I’ve lived through enough boom-and-bust cycles to know the pattern, you know that when hyperscalers start ordering CPUs at scale after a GPU buying spree, something fundamental is shifting. 

The data center segment grew 34% quarter-over-quarter hitting $4.3 billion, which got all the headlines, but the client segment quietly ticked up 10% sequentially. That’s the signal hiding in the noise.

The AI workload evolution is creating this fascinating dynamic where you need powerful CPUs to orchestrate all those GPUs and handle the actual productive work that AI generates. I remember back in the dotcom days when companies would buy server farms without thinking about bandwidth, power, or cooling. Same mistake, different decade. 

Turns out buying a warehouse full of GPUs without the CPU infrastructure to feed them data is like owning a stable of Ferraris without bothering to pave the roads. Some lessons never get old, just more expensive.

AMD’s new Helios platform integrates their Instinct MI400 series GPUs with Venice EPYC CPUs and Pensando network interface cards into a single optimized rack system. This is their direct shot across Nvidia’s bow, basically saying “we’re competing on the entire AI infrastructure now.” 

Nvidia has been selling complete solutions for years, which is why they command those obscene margins. AMD is finally getting serious about closing that gap, and judging by the talent they’re attracting, the engineers who built Nvidia’s ecosystem think AMD’s approach might actually work.

Now, about that ROCm 7 software release. The company claims up to 4.6 times higher inference performance and 3 times better training compared to ROCm 6. 

Having spent years watching AMD struggle with software while Nvidia’s CUDA moat seemed insurmountable, seeing them actually deliver meaningful improvements is genuinely noteworthy. Nvidia’s CUDA advantage has been the real competitive barrier, not chip performance. Every percentage point AMD closes on the software side reduces switching costs for customers. 

And when senior architects start jumping ship to work on AMD’s software stack, they’re betting their careers that the gap is closable. Engineers vote with their feet, and that vote matters more than any analyst report.

Another thing that caught my attention was the client mix AMD’s attracting. Sovereign AI programs and national supercomputing initiatives are choosing AMD hardware. The Department of Energy is building their Lux AI system with AMD chips. 

When governments start betting on your technology for long-term strategic projects, that’s the kind of sticky revenue that compounds beautifully over time. I’ve watched this play out with defense contractors for decades – once you’re in, you’re really in.

The gross margin story tells you everything about AMD’s execution trajectory. They’re guiding to 54.5% for Q4, which is approaching the promised land of high-margin semiconductor economics. 

A few years ago, AMD was a scrappy underdog happy to win business on price. Now they’re commanding premium pricing while expanding share. 

I’ve seen this movie before with companies like Applied Materials (AMAT) and ASML Holding (ASML). Once you cross that margin threshold, the flywheel effect kicks in and suddenly you’re swimming in R&D budget.

The valuation conversation gets spicy here. Projections floating around suggest AMD could hit $775 per share, implying a trillion-dollar valuation predicated on EPS reaching $15 with a 25 times multiple. Aggressive? You bet. Impossible? Not if the CPU refresh cycle materializes and AI infrastructure spending continues at current rates. 

AMD’s chiplet architecture approach is proving more flexible than monolithic designs as workloads diversify, and that architectural advantage compounds over time.

The risk isn’t technical execution anymore. AMD has proven they can ship competitive products and attract the talent to build ecosystems around them. The risk is macro, plain and simple. 

If AI benefits don’t materialize fast enough and CFOs start questioning whether they’re getting returns on those billions in capital expenditures, the entire AI infrastructure buildout could hit the brakes harder than my ’69 Mustang on black ice. 

We’ve already seen Meta (META) issue debt to fund AI spending because free cash flow isn’t covering it anymore. I’ve got gray hair from learning that conviction and delusion often look identical until time passes.

What separates AMD’s story from typical hype cycles is they’re not selling dreams anymore. They’re shipping products, winning designs, and taking share in the most important technology transition of our generation. 

The CPU demand surge validates that AI infrastructure is maturing beyond the initial GPU land grab into something more sophisticated. When your revenue growth comes from multiple product lines serving different parts of the AI stack, that’s portfolio diversification at the chip level.

AMD at current prices reflects a company that’s already won significant AI share. The upside case requires believing the CPU cycle is real, the software improvements stick, and the integrated systems approach resonates with sophisticated buyers. 

Given the talent migration patterns and what infrastructure engineers are actually deploying in the field, all three seem increasingly likely. And apparently, it’s compelling enough that Nvidia’s own architects are placing their bets accordingly. 

When the rats start swimming toward the ship instead of away from it, you might want to pay attention.