Chips & Compute

High-memory GPU alternative

Version

MI355X

Cost

~$2-3 per GPU-hour on clouds

Model

CDNA 4

On-device inference across Mac, iPhone and iPad.

Version

M5

Cost

Bundled with hardware

Model

Apple Foundation Models

Cloud-native training silicon

Version

Trainium2

Cost

from ~$1.3 per accelerator-hour

Model

Trainium2

Co-designed XPUs behind most hyperscaler silicon.

Version

2025

Cost

Enterprise supply agreements

Model

powers Google TPU

Wafer-scale inference and training

Version

WSE-3 / Inference

Cost

Free tier, then from $0.06 / M tokens

Model

Llama / Qwen / GPT-OSS

Digital in-memory compute for inference efficiency.

Version

Corsair

Cost

System sales

Model

n/a

Transformer-only ASIC betting on architecture lock-in.

Version

Sohu

Cost

System sales

Model

transformer-only

Korean inference silicon with strong perf-per-watt.

Version

RNGD

Cost

System sales

Model

n/a

Inference-first TPU pods

Version

v7 Ironwood

Cost

from ~$1.2 per chip-hour

Model

TPU v7

IPU accelerators now inside SoftBank.

Version

Bow IPU

Cost

System sales

Model

n/a

Deterministic low-latency inference

Version

GroqCloud

Cost

Free tier, then from $0.05 / M tokens

Model

Llama / Kimi / GPT-OSS

China domestic AI accelerator

Version

Ascend 910C

Cost

China market, price undisclosed

Model

Ascend 910C

Cost-focused training and inference accelerator.

Version

Gaudi 3

Cost

OEM systems and cloud instances

Model

Accelerator

Photonic interconnect for scaling accelerator fabrics.

Version

Passage M1000

Cost

System sales

Model

n/a

GB200/B200 rack-scale AI systems

Version

GB200 NVL72

Cost

~$2-6 per GPU-hour on clouds

Model

Blackwell B200

Rack-scale inference accelerators entering the datacentre.

Version

AI200

Cost

System sales

Model

n/a

Dataflow chips for model serving

Version

SN40L

Cost

Free tier, then usage-based

Model

Llama / DeepSeek open weights

Open RISC-V AI hardware

Version

Blackhole

Cost

from $1399/mo

Model

Tensix cores