Remember when everyone said you had to run your inference workloads on Nvidia GPUs or you weren’t serious about AI? I spent six months last year optimizing bot architectures around that assumption, watching my cloud bills climb while I convinced myself there was no alternative.
Then I started actually looking at what Advanced Micro Devices and Broadcom are shipping.
Don’t get me wrong—Nvidia built something incredible. Their CUDA ecosystem is still the gold standard, and if you’re training massive foundation models, you’re probably not switching anytime soon. But here’s what I’ve learned building production bot systems: most of us aren’t training foundation models. We’re fine-tuning them, running inference at scale, and trying to keep costs reasonable while serving real users.
AMD Is Solving Real Bot Infrastructure Problems
AMD’s recent moves in the AI space aren’t just about matching Nvidia’s specs on paper. They’re addressing the actual pain points I hit when deploying conversational AI systems. Their MI300 series chips are showing up in more cloud providers, which means I have options when architecting multi-region bot deployments.
What matters for bot builders is inference throughput per dollar, not just raw performance. AMD is competitive here, and the gap keeps closing. When I’m running hundreds of concurrent chat sessions, the ability to spread workloads across more affordable hardware changes my unit economics completely.
The software story is improving too. ROCm isn’t CUDA yet, but it’s getting there. I’ve successfully migrated several inference pipelines without rewriting everything from scratch. That’s the threshold that matters—when switching costs drop low enough that you’ll actually consider alternatives.
Broadcom’s Quiet Dominance in Custom AI Silicon
Broadcom doesn’t get the same headlines, but they’re powering some of the most sophisticated AI deployments out there. Their custom ASIC business is where things get interesting for anyone thinking about the next five years of AI infrastructure.
Major cloud providers and tech companies are designing their own AI chips, and Broadcom is often the partner making that happen. Google’s TPUs, various custom inference accelerators—this is where Broadcom plays. As AI workloads mature and companies optimize for specific use cases, custom silicon becomes more attractive than general-purpose GPUs.
For bot builders, this matters because it shapes the infrastructure space we’ll be deploying on. The cloud services I use are increasingly running on custom chips designed for specific AI tasks. Broadcom’s position in this supply chain makes them a solid bet on where AI infrastructure is headed.
Why This Matters for Building Bots in 2026
I’m not buying Nvidia stock right now because I’m betting on a more distributed future for AI compute. The current concentration around one vendor’s hardware creates risks—supply constraints, pricing power, and architectural lock-in.
AMD and Broadcom represent different paths to the same destination: more options for running AI workloads efficiently. AMD gives us direct competition in the GPU space with improving software support. Broadcom enables the custom silicon revolution that’s already reshaping how major platforms run AI.
When I’m planning bot architectures for the next few years, I’m assuming more hardware diversity, not less. I’m assuming cloud providers will offer more chip options, not fewer. I’m assuming the cost curve for inference will keep improving as competition increases.
Both AMD and Broadcom are positioned to benefit from these trends. They’re riding the same AI wave as Nvidia, but with more room to grow and less of the hype premium already priced in.
As someone who builds bots for a living, I care less about which stock performs best and more about which companies are solving my infrastructure problems. Right now, that’s AMD and Broadcom. The fact that they might also be smarter investments is just a bonus.
đź•’ Published: