Nvidia dominates every AI infrastructure conversation. AMD and Broadcom are quietly positioning themselves for the next phase of the AI supercycle. This contradiction tells us something important about where we are in the market cycle—and where the actual architectural innovation is happening.
I spend my days analyzing agent architectures and intelligence systems. From that vantage point, the current market obsession with Nvidia feels like watching everyone crowd into the same room while two other doors open onto entirely different landscapes of opportunity.
The Architectural Shift Nobody’s Pricing In
Advanced Micro Devices and Broadcom are riding trends in AI that go beyond raw compute power. They’re positioned at inflection points that matter more for the next generation of agent systems than another incremental GPU improvement.
AMD’s approach to heterogeneous computing aligns with where agent architectures are actually heading. We’re moving past the “throw more GPUs at it” phase into something more nuanced. Agent systems need different compute profiles for different cognitive tasks—reasoning, memory retrieval, sensory processing, action selection. AMD’s chiplet architecture and their focus on unified memory systems map directly onto these requirements.
This isn’t speculation. When I examine the infrastructure needs of multi-agent systems, the bottleneck isn’t always parallel floating-point operations. It’s memory bandwidth, cache coherency across heterogeneous processors, and power efficiency at scale. AMD is building for that reality.
Broadcom’s Invisible Infrastructure
Broadcom operates in a space most people don’t think about when they picture AI. Custom silicon. Networking fabric. The connective tissue that makes distributed agent systems possible.
Here’s what matters: as agent systems scale, the communication overhead between components becomes the limiting factor. You can have the fastest GPUs in the world, but if your networking layer can’t handle the message-passing requirements of a distributed agent swarm, you’re dead in the water.
Broadcom’s custom ASIC business is positioned exactly where hyperscalers need it. When Google, Meta, or Amazon build their next-generation agent platforms, they’re not just buying off-the-shelf GPUs. They’re designing custom silicon for their specific workloads, and Broadcom is the company that makes that happen.
The Valuation Question
Nvidia’s success has been extraordinary. The company has seen parabolic growth, and new investors can still participate in that story. But from a pure growth perspective for 2026, the math gets interesting.
When a company reaches Nvidia’s scale, maintaining the same growth rate requires increasingly massive market expansion. AMD and Broadcom are operating from different baselines. They have room to capture market share in segments that are just beginning to matter for agent intelligence.
I’m not arguing Nvidia is a bad investment. I’m arguing that for someone looking at the AI supercycle through the lens of agent architectures, AMD and Broadcom offer exposure to trends that are underpriced relative to their importance.
What Agent Systems Actually Need
Building intelligent agents isn’t just about training large models. It’s about inference at scale, memory systems that can handle long-context reasoning, efficient communication between specialized processors, and custom silicon optimized for specific cognitive tasks.
AMD’s focus on CPU-GPU integration matters when you’re building agents that need to smoothly switch between symbolic reasoning and neural processing. Broadcom’s networking and custom silicon capabilities matter when you’re orchestrating thousands of agents that need to communicate in real-time.
These aren’t the flashy parts of AI that make headlines. They’re the infrastructure that determines whether agent systems actually work in production.
Following the Architecture, Not the Hype
My investment thesis follows my technical analysis. When I look at where agent intelligence is heading—more distributed, more heterogeneous, more dependent on specialized silicon and efficient communication—AMD and Broadcom are positioned at critical junctures.
Nvidia will continue to be important. But the next phase of the AI supercycle requires infrastructure that goes beyond what made Nvidia dominant in the training era. AMD and Broadcom are building that infrastructure right now.
That’s where I’m placing my attention, and my capital.
đź•’ Published: