Google and Amazon may still need NVIDIA, even as they build their own AI chips

Alphabet’s Google and Amazon have made no secret of their desire to run more of their AI workloads on custom hardware. Google markets Tensor Processing Units (TPUs), purpose-built accelerators for machine learning, while Amazon promotes Trainium, its own AI chip designed for training at scale. Together, the two companies represent the clearest challenge to NVIDIA’s dominance in data center graphics processing units and AI compute platforms.

Yet a recent market analysis argued that, in practice, both hyperscalers will keep buying large amounts of NVIDIA’s systems. The core of the case is not that custom chips are inferior. It is that moving entire AI operations off a mature vendor stack is slower than chipmakers and buyers often expect, especially when supply constraints and software compatibility become the main friction points.

For Google, building and deploying TPUs can reduce dependence over time, but it also requires continuous work to keep models, training pipelines, compilers, and performance-tuning aligned with the latest workloads. For Amazon, Trainium is meant to offer similar leverage. The challenge is that even when a company can train efficiently on its own silicon, production systems often involve multiple components across research, training, and inference. That increases the likelihood that NVIDIA-powered infrastructure remains part of the broader workflow for a long time.

The analysis also highlights a supply bottleneck dynamic that can make substitution harder than it sounds. Custom chips may be designed to hit specific cost and performance targets, but availability, manufacturing throughput, and platform readiness determine how quickly they can absorb demand. When AI demand spikes, companies tend to prioritize getting working systems into production, which can favor suppliers with established manufacturing capacity and rapid deployment pathways.

Another factor is the “software moat” surrounding NVIDIA’s stack. NVIDIA’s platform is not just the hardware itself. It includes the tooling and libraries that help developers map AI models onto GPUs efficiently, optimize kernels, and debug performance issues. Even large buyers that build their own chips can still face a tradeoff: they may prefer to reuse proven software patterns and spend their engineering effort on model development rather than re-creating an entire optimization ecosystem for every new workload.

For investors, this creates a more complex picture than a simple head-to-head competition between custom silicon and NVIDIA accelerators. If buyers cannot fully transition at once, NVIDIA’s revenue exposure may persist through the years even if Google and Amazon gain incremental share. In other words, custom chips can reduce marginal demand for certain tasks, while still leaving plenty of compute needs that NVIDIA is positioned to serve.

Still, the uncertainty is how much of that dependence is structural and how much is temporary. Public information typically does not reveal the internal allocation of training and inference workloads across chip types at the level needed to quantify precisely how NVIDIA demand could change. The market post did not lay out detailed numbers or a timetable for migration, so the argument is best treated as a strategic interpretation rather than a confirmed forecast.