Data Center AI Accelerators

GPUs and chips for training and large-scale inference.

Free plan Reset

26 products

NVIDIA GB200 NVL72

A rack-scale exaflop AI supercomputer that acts as one giant GPU.

9.6/10 From $10.50/per hour

NVIDIA GB300 NVL72

Rack-scale Blackwell Ultra: 72 GPUs + 36 Grace CPUs as one giant accelerator

9.5/10 From $12/per hour

NVIDIA GB300 NVL72 (Blackwell Ultra)

Blackwell Ultra rack-scale system tuned for the age of AI reasoning.

9.5/10 From $12/per hour

NVIDIA B300 (Blackwell Ultra)

Blackwell Ultra single-GPU module for AI reasoning at scale

9.4/10 From $30000/one-time

NVIDIA Blackwell B200

The flagship Blackwell GPU that more than doubles Hopper's throughput.

9.4/10 From $3.50/per hour

AMD Instinct MI400 / MI455X

CDNA-Next flagship: 432GB HBM4 and 40 PFLOPS FP4

Google TPU v7 (Ironwood)

Google's 7th-gen TPU built from the ground up for the age of inference.

NVIDIA H200

The memory-upgraded Hopper GPU that remains the AI workhorse.

8.9/10 From $3.72/per hour

AMD Instinct MI355X

AMD's CDNA 4 flagship with the biggest memory and native FP4/FP6.

AWS Trainium3

Amazon Web Services

Amazon's 3nm training chip: 2.52 PFLOPS FP8, 144GB HBM3e

NVIDIA H100 Tensor Core GPU

The Hopper workhorse that powered the LLM boom, now widely available

8.8/10 From $2/per hour

Google TPU v6e (Trillium)

Google's widely available 6th-gen TPU for efficient training and inference.

8.6/10 From $1.22/per chip-hour

NVIDIA GH200 Grace Hopper Superchip

Grace CPU + Hopper GPU fused over NVLink-C2C

8.6/10 From $1.99/per hour

AWS Trainium3 (Trn3)

Amazon Web Services

Amazon's newest 3nm-class training chip, up to 4.4x faster than Trn2.

Microsoft Maia 200

Azure's second-gen custom AI accelerator, 3x FP4 of Trainium3

AWS Trainium2 (Trn2)

Amazon Web Services

Amazon's purpose-built training chip with the best price/performance on AWS.

Tenstorrent Galaxy (Blackhole)

Open RISC-V AI server: 23 PFLOPS from 32 Blackhole chips

8.4/10 From $100000/one-time

AMD Instinct MI325X

256 GB of HBM3e for memory-hungry inference at lower cost than NVIDIA.

SambaNova SN40L

SambaNova Systems

Reconfigurable Dataflow Unit with three-tier memory for trillion-param AI

Cerebras WSE-3 (CS-3)

Cerebras Systems

A single wafer-scale chip the size of a dinner plate for record inference speed.

Meta MTIA (v3 / 300-series)

Meta's RISC-V custom accelerators for ranking and GenAI

Huawei Ascend 910C

China's flagship AI accelerator: 96GB HBM, 1.8 TB/s

FuriosaAI RNGD

Power-efficient inference chip: 512 TFLOPS FP8 at 150-200W

Groq LPU

A deterministic inference engine built for the lowest latency per token.

8.0/10 From $0.05/per 1M input tokens Free plan

Tenstorrent Blackhole p150

$1,399 PCIe AI dev card with 32GB and open software

8.0/10 From $999/one-time

Intel Gaudi 3

An open, cost-effective AI accelerator priced well below NVIDIA.

7.8/10 From $15000/per card (est.)

#	Product	Vendor	Score	From	Free
	NVIDIA GB200 NVL72	NVIDIA	9.6	From $10.50/per hour
	NVIDIA GB300 NVL72	NVIDIA	9.5	From $12/per hour
	NVIDIA GB300 NVL72 (Blackwell Ultra)	NVIDIA	9.5	From $12/per hour
	NVIDIA B300 (Blackwell Ultra)	NVIDIA	9.4	From $30000/one-time
	NVIDIA Blackwell B200	NVIDIA	9.4	From $3.50/per hour
	AMD Instinct MI400 / MI455X	AMD	9.0	—
	Google TPU v7 (Ironwood)	Google	9.0	—
	NVIDIA H200	NVIDIA	8.9	From $3.72/per hour
	AMD Instinct MI355X	AMD	8.8	—
	AWS Trainium3	Amazon Web Services	8.8	—
	NVIDIA H100 Tensor Core GPU	NVIDIA	8.8	From $2/per hour
	Google TPU v6e (Trillium)	Google	8.6	From $1.22/per chip-hour
	NVIDIA GH200 Grace Hopper Superchip	NVIDIA	8.6	From $1.99/per hour
	AWS Trainium3 (Trn3)	Amazon Web Services	8.5	—
	Microsoft Maia 200	Microsoft	8.5	—
	AWS Trainium2 (Trn2)	Amazon Web Services	8.4	—
	Tenstorrent Galaxy (Blackhole)	Tenstorrent	8.4	From $100000/one-time
	AMD Instinct MI325X	AMD	8.3	—
	SambaNova SN40L	SambaNova Systems	8.3	Free	✓
	Cerebras WSE-3 (CS-3)	Cerebras Systems	8.2	Free	✓
	Meta MTIA (v3 / 300-series)	Meta	8.2	—
	Huawei Ascend 910C	Huawei	8.1	—
	FuriosaAI RNGD	FuriosaAI	8.0	—
	Groq LPU	Groq	8.0	From $0.05/per 1M input tokens	✓
	Tenstorrent Blackhole p150	Tenstorrent	8.0	From $999/one-time
	Intel Gaudi 3	Intel	7.8	From $15000/per card (est.)

Compare

Compare