Global cloud giants are building their own AI-focused chips to rein in costs and gain more control. Still, NVIDIA dominates GPUs by a mile.
These custom silicon projects don’t really replace GPUs—they complement them. It’s a layered approach that’s changing data-center architectures, market share, and maybe even the whole growth story for the industry.
AI data-center chip landscape and NVIDIA’s enduring prominence
Cloud providers are racing to develop their own accelerators. Yet NVIDIA still commands well over 80% of the GPU market, and most analysts think that’s not changing soon.
The GPU market sits at about $36 billion now. Some estimates say it could skyrocket to $811 billion by 2035, with data-center revenue just climbing year after year.
Chips from Google, AWS, and Microsoft are starting to shape the economics of inference and specialized workloads. But they haven’t knocked GPUs off their pedestal as the go-to for big model training.
In reality, the industry’s leaning into a layered architecture. GPUs handle the heavy parallel computation, while CPUs and other accelerators juggle and optimize tasks across the data center.
This shift is fueling demand for a wider variety of chip types. Nobody’s ditching NVIDIA GPUs—they’re just adding more options to the mix.
It’s turning into a more collaborative ecosystem. GPUs are the backbone, while cloud-specific silicon steps in to cut costs and boost efficiency where it counts.
Google’s TPU strategy and the CUDA ecosystem
Google’s TPU 8t and 8i chips are built to run Google’s own models and select partner workloads with high efficiency, mainly for inference. They’re priced to beat some NVIDIA options in certain inference tasks, really zeroing in on cost-per-inference gains.
But here’s the thing: Google still leans on NVIDIA’s CUDA ecosystem for a lot of its workflows. Even as they push their own silicon, they can’t quite escape the gravity of established software stacks and tooling.
- In-house accelerators target inference and high-volume, specialized tasks, delivering cost-control and throughput gains where they matter most.
- Full displacement of GPUs remains unlikely, given the entrenched CUDA tooling and the scale of training workloads that GPUs handle efficiently.
AWS: a broad custom silicon portfolio to complement, not supplant NVIDIA
AWS has rolled out a whole lineup of in-house chips: Graviton CPUs, Trainium for training, and Inferentia for inference. TSMC manufactures these chips, and AWS positions them as add-ons, not replacements, for NVIDIA, AMD, and Intel hardware.
AWS aims to optimize specific workloads at scale. They want to lower costs per inference or training run and get a little more leverage with cloud customers.
- Inference and training specialization aim to boost efficiency for high-volume workloads.
- Manufacturing at TSMC keeps AWS at the cutting edge with solid supply lines.
- These chips slot into existing data centers as accelerators alongside GPUs.
Microsoft: Maia, Azure Cobalt, and the systems approach
Microsoft’s Maia accelerator and Azure Cobalt CPU follow a similar, systems-focused playbook. They’re all about efficiency for specific tasks, not about pushing GPUs out of the picture for big language models.
Microsoft believes specialized accelerators can really crank up throughput and energy efficiency for certain jobs. But GPUs still drive the broader training and model development pipeline.
- Specialized accelerators get used where they make the biggest impact, while GPUs stay central for model development.
- Azure Cobalt and Maia show how a multi-chip ecosystem can optimize different parts of the AI workflow.
Data-center architecture and market outlook
The AI data-center keeps getting more layered and interoperable every year. GPUs tackle large-scale parallel computation. CPUs and other accelerators step in to orchestrate workloads, juggle memory, and speed up non-parallel tasks.
TrendForce points out that the CPU-to-GPU ratio could move toward parity. That would mean growing demand for a mix of chip types, but it doesn’t really mean GPUs will lose importance.
Market dynamics to watch include:
- NVIDIA’s market position remains resilient against the rise of cloud-specific silicon. Its dominant GPU stack and broad software ecosystem keep it ahead.
- AMD is still the main GPU rival. Still, it seems unlikely that custom silicon from cloud vendors will knock NVIDIA off its central spot anytime soon.
- The ecosystem keeps shifting into a mixed, collaborative model. GPUs, CPUs, and accelerators all coexist to get the best AI performance possible.
Here is the source article for this story: Will the AI Boom Upend the Chip Industry?