TBC Home →
← Back to stack

Compute

SubstituteCore: AI training, AI inferenceHigh: EV / auto, Consumer / IoT, IndustrialMedium: Robotics

AI accelerators / GPUs

Domestic contenders exist, but frontier datacenter performance still depends on software, HBM, packaging, networking, and foundry access.

The simple version

These are the chips built to run many calculations in parallel. GPUs started in graphics, but the same parallel architecture turned out to be ideal for training and running large AI models.

What the technology does

  • The chip matters, but the full product is really a system: silicon, memory, packaging, networking, software libraries, and developer tools.
  • Training needs huge clusters and fast memory. Inference can often use smaller, cheaper, or more specialized chips.
  • For AI, software compatibility can be as important as raw chip speed. Nvidia's CUDA ecosystem is the benchmark.

Market scale

  • Nvidia is the global reference point and crossed $130 billion in fiscal 2025 revenue, mostly because of datacenter AI.
  • AMD is the main merchant GPU challenger; Google TPU and Broadcom-linked custom silicon show how hyperscalers can build their own paths.
  • Chinese contenders are strategically important, but much smaller and harder to compare because many deployments are private or state-linked.
Global references
  • NvidiaUnited States · NVDA
  • AMDUnited States · AMD
  • Google TPUUnited States · GOOGL / GOOG
  • BroadcomUnited States · AVGO

China angle

Huawei Ascend is the anchor to watch because it has chips, systems, cloud channels, and political support. Biren, Moore Threads, Cambricon, and Enflame matter too, but the question is less 'can they make a chip?' and more 'can customers run real workloads at scale?'

China references

How we got here

The category moved from graphics to AI in the 2010s, then became the center of the semiconductor cycle after the generative AI boom in 2022-2024. Export controls turned Chinese AI accelerators from a nice-to-have into a national priority.

  1. 2012-2016

    Deep learning makes GPUs strategically important beyond gaming and visualization.

  2. 2017-2021

    Datacenter GPU clusters become standard for frontier AI labs and cloud providers.

  3. 2022-2024

    Generative AI demand turns accelerators, , and advanced packaging into bottlenecks.

  4. 2025+

    China's test is cluster reliability, software compatibility, and usable supply, not one-off chip demos.

Why it matters

This is the most visible layer in AI compute, but the chip alone is only part of the stack.

Main bottleneck

CUDA-like software ecosystem, supply, advanced packaging, leading-node access, and large-scale deployment proof.

Where this sector shows up

Core
AI training, AI inference
High
EV / auto, Consumer / IoT, Industrial
Medium
Robotics

What to watch

  • Real cloud deployments
  • Model compatibility
  • Cluster networking
  • Training benchmarks
  • Export-control workarounds

Chip pricing

41 chips from this sector are in the pricing dataset — domestic parts and their global references, each with dated, source-linked price observations. Latest figure shown per chip.

Full generation tables →