Alibaba Qwen3

Qwen3-235B-A22B hardware requirements

Needs at least 2 GPUs at BF16 (Instinct MI355X OAM). For 1,000 tokens/s at 32k context you need 16 x H100 SXM5 80GB at $29.28/hr or 24 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $12.00/hr.

Parameters

235B

Active per token

22.0B

mixture of experts

KV cache per token

188 KB

16-bit cache

Native precision

BF16

Rent Model card

Top options for Qwen3-235B-A22B

The strongest part each vendor offers for Qwen3-235B-A22B at BF16 with 32,768 tokens per user, rentable parts first. Each card is one replica, the fewest accelerators that hold the model, and the line it draws on the graph below. Select a card to plot it.

Tokens per second as users grow

Each line is one replica, the fewest accelerators that hold the model, serving more users at once. It climbs while memory bandwidth is the limit, flattens once compute is, and stops where memory runs out at 32,768 tokens per user. Hover a line to bring it forward.

4 x GB200 NVL72 GPU 186GB 2 x Instinct MI355X OAM 4 x Ascend 950DT 4 x Gaudi 3 128GB 4 x TPU v7 192GB 16 x Inferentia2

Memory fills at: 4 x GB200 NVL72 GPU 186GB 38 users, 2 x Instinct MI355X OAM 13 users, 4 x Ascend 950DT 13 users, 4 x Gaudi 3 128GB 3 users, 4 x TPU v7 192GB 42 users, 16 x Inferentia2 1 users. Roofline estimates at BF16.

Bare minimum

Fewest GPUs

2 x Instinct MI355X OAM

BF16, 4,096 tokens, one stream

Tokens/s: 251
Per hour: $11.98
Per million tokens: $13.25
Memory used: 85%

Rent on Daytona

Cheapest per hour

8 x RTX PRO 6000 Blackwell Max-Q Workstation Edition

BF16, 4,096 tokens, one stream, live price

Tokens/s: 201
Per hour: $4.00
Per million tokens: $5.54
Memory used: 64%

Rent on RunPod

Cheapest per million tokens

8 x RTX PRO 6000 Blackwell Max-Q Workstation Edition

BF16, 32,768 tokens, 32 streams

Tokens/s: 348
Per hour: $4.00
Per million tokens: $3.19
Memory used: 92%

Rent on RunPod

Configurations

Every datacenter GPU in the catalog, sized for these settings. Change precision, context and concurrency and the whole page follows. Each point is one replica: the smallest tensor-parallel group of that GPU that holds the model.

Precision Context Streams

Vendor

Priced only

Priced configurations only; 42 more fit without a live price and appear in the table. Marker size grows with the tensor-parallel width. Roofline estimates.

#									Rent
1	Instinct MI455X	2	81 %	1.2k tok/s	54.0 tok/s	—	—	—	Specs →
2	Rubin SXM	4	61 %	2.1k tok/s	96.7 tok/s	—	—	—	Specs →
3	TPU 8t announced	4	81 %	~773 tok/s	28.7 tok/s	—	—	—	Specs →
4	TPU 8i announced	4	61 %	~925 tok/s	37.8 tok/s	—	—	—	Specs →
5	Instinct MI355X OAM	4	61 %	874 tok/s	35.2 tok/s	$23.96	$7.61	5.6 kW	Rent →
6	GB300 NVL72 GPU 288GB	4	61 %	873 tok/s	35.2 tok/s	$49.68	$15.81	—	Rent →
7	GB200 NVL72 GPU 186GB	4	94 %	873 tok/s	35.2 tok/s	$42.00	$13.37	—	Rent →
8	Instinct MI350X OAM	4	61 %	857 tok/s	35.2 tok/s	$24.64	$7.99	4.0 kW	Rent →
9	TPU v7 192GB	4	91 %	805 tok/s	32.4 tok/s	—	—	—	Specs →
10	Instinct MI430X OAM announced	2	81 %	~903 tok/s	54.0 tok/s	—	—	—	Specs →
11	B300 SXM 262GB	4	67 %	852 tok/s	35.2 tok/s	$19.96	$6.51	5.6 kW	Rent →
12	B200 SXM 180GB	4	97 %	827 tok/s	33.8 tok/s	$16.36	$5.49	4.0 kW	Rent →
13	Ascend 970 announced	4	61 %	1.2k tok/s	63.3 tok/s	—	—	—	Specs →
14	B100 SXM 192GB	4	91 %	796 tok/s	35.2 tok/s	—	—	2.8 kW	Specs →
15	Jalapeño announced	4	81 %	~1.2k tok/s	67.7 tok/s	—	—	2.8 kW	Specs →
16	Instinct MI325X OAM	4	68 %	597 tok/s	26.4 tok/s	$15.20	$7.08	4.0 kW	Rent →
17	Instinct MI300X 192GB	4	91 %	546 tok/s	23.3 tok/s	$9.56	$4.87	3.0 kW	Rent →
18	BR100 2 nodes	16	69 %	657 tok/s	23.4 tok/s	—	—	8.8 kW	Specs →
19	Ascend 960 announced	4	61 %	723 tok/s	42.2 tok/s	—	—	—	Specs →
20	H200 SXM 141GB	8	62 %	887 tok/s	39.8 tok/s	$19.84	$6.21	5.6 kW	Rent →
21	H100 SXM5 80GB 2 nodes	16	55 %	1.2k tok/s	49.1 tok/s	$29.28	$6.76	11.2 kW	Rent →
22	Data Center GPU Max 1550 128GB	8	69 %	644 tok/s	27.2 tok/s	—	—	4.8 kW	Specs →
23	H200 NVL 141GB	8	62 %	~840 tok/s	39.8 tok/s	$28.91	$9.56	4.8 kW	Rent →
24	H100 NVL 94GB	8	94 %	~729 tok/s	32.4 tok/s	$25.52	$9.73	3.2 kW	Rent →
25	Ascend 910C	8	69 %	625 tok/s	26.6 tok/s	—	—	4.8 kW	Specs →
26	H100 PCIe 80GB 2 nodes	16	55 %	757 tok/s	29.3 tok/s	$40.00	$14.68	5.6 kW	Rent →
27	H800 PCIe 80GB 2 nodes	16	55 %	757 tok/s	29.3 tok/s	—	—	5.6 kW	Specs →
28	Ascend 950DT	8	61 %	640 tok/s	33.2 tok/s	—	—	—	Specs →
29	RTX PRO 6000 Blackwell Workstation Edition	8	92 %	360 tok/s	14.9 tok/s	$14.40	$11.10	4.8 kW	Rent →
30	Ascend 950PR	8	69 %	330 tok/s	13.3 tok/s	—	—	7.2 kW	Specs →
31	MTT S5000 2 nodes	16	55 %	~582 tok/s	23.4 tok/s	—	—	—	Specs →
32	RTX PRO 6000 Blackwell Server Edition	8	92 %	329 tok/s	13.3 tok/s	$5.52	$4.66	4.8 kW	Rent →
33	Gaudi 3 128GB	8	69 %	820 tok/s	30.7 tok/s	—	—	7.2 kW	Specs →
34	TPU v5p 95GB	8	93 %	476 tok/s	23.0 tok/s	—	—	3.6 kW	Specs →
35	RTX PRO 6000 Blackwell Max-Q Workstation Edition	8	92 %	348 tok/s	14.9 tok/s	$4.00	$3.19	2.4 kW	Rent →
36	Cloud AI 100 Ultra	8	69 %	~131 tok/s	4.5 tok/s	—	—	1.2 kW	Specs →
37	Gaudi 2 96GB	8	92 %	430 tok/s	20.3 tok/s	—	—	4.8 kW	Specs →
38	Atlas 350	8	78 %	287 tok/s	11.6 tok/s	—	—	4.8 kW	Specs →
39	Data Center GPU Max 1100 48GB 2 nodes	16	93 %	456 tok/s	18.0 tok/s	—	—	4.8 kW	Specs →
40	Ascend 910B 2 nodes	16	69 %	551 tok/s	23.4 tok/s	—	—	6.4 kW	Specs →
41	MI250X 128GB	8	69 %	491 tok/s	27.2 tok/s	—	—	4.5 kW	Specs →
42	RTX 6000 Ada 48GB 2 nodes	16	93 %	~364 tok/s	14.1 tok/s	$15.84	$12.10	4.8 kW	Rent →
43	Instinct MI250 128GB	8	69 %	479 tok/s	27.2 tok/s	—	—	4.0 kW	Specs →
44	L40S 48GB 2 nodes	16	93 %	333 tok/s	12.7 tok/s	$12.16	$10.14	5.6 kW	Rent →
45	A800 PCIe 80GB 2 nodes	16	55 %	582 tok/s	28.3 tok/s	—	—	4.8 kW	Specs →
46	A800 SXM4 2 nodes	16	55 %	601 tok/s	29.9 tok/s	—	—	6.4 kW	Specs →
47	A100 PCIe 80GB 2 nodes	16	55 %	582 tok/s	28.3 tok/s	$21.60	$10.32	4.8 kW	Rent →
48	A100 SXM4 80GB 2 nodes	16	55 %	601 tok/s	29.9 tok/s	$21.60	$9.98	6.4 kW	Rent →
49	RTX 5880 Ada 48GB 2 nodes	16	93 %	~343 tok/s	14.1 tok/s	—	—	4.6 kW	Specs →
50	RTX PRO 5000 Blackwell 48GB 2 nodes	16	93 %	~429 tok/s	19.7 tok/s	$15.36	$9.96	4.8 kW	Rent →
51	RTX PRO 5000 Blackwell 72GB 2 nodes	16	61 %	~429 tok/s	19.7 tok/s	—	—	4.8 kW	Specs →
52	L40 48GB 2 nodes	16	93 %	283 tok/s	12.7 tok/s	$13.12	$12.87	4.8 kW	Rent →
53	Instinct MI210 PCIe 2 nodes	16	69 %	423 tok/s	24.0 tok/s	—	—	4.8 kW	Specs →
54	RTX A6000 48GB 2 nodes	16	93 %	249 tok/s	11.2 tok/s	$8.00	$8.94	4.8 kW	Rent →
55	A40 48GB 2 nodes	16	93 %	230 tok/s	10.2 tok/s	$9.44	$11.41	4.8 kW	Rent →
56	H20 141GB HBM3e	8	62 %	309 tok/s	33.2 tok/s	—	—	3.2 kW	Specs →
57	H20 96GB	8	92 %	309 tok/s	33.2 tok/s	—	—	3.2 kW	Specs →
58	Atlas 300I Duo 48GB 2 nodes	16	93 %	151 tok/s	6.0 tok/s	—	—	2.4 kW	Specs →
59	Atlas 300I Duo 96GB	8	92 %	85.8 tok/s	3.4 tok/s	—	—	1.2 kW	Specs →
60	ATOM-Max 2 nodes	16	69 %	279 tok/s	15.0 tok/s	—	—	5.6 kW	Specs →
61	Radeon PRO W7900 Dual Slot 48GB 2 nodes	16	93 %	248 tok/s	12.7 tok/s	—	—	4.7 kW	Specs →
62	Radeon PRO W7900 48GB 2 nodes	16	93 %	248 tok/s	12.7 tok/s	—	—	4.7 kW	Specs →
63	L20 48GB 2 nodes	16	93 %	245 tok/s	12.7 tok/s	—	—	4.4 kW	Specs →
64	MLU370-X8 2 nodes	16	93 %	183 tok/s	9.0 tok/s	—	—	4.0 kW	Specs →
65	Radeon PRO W7800 48GB 2 nodes	16	93 %	217 tok/s	12.7 tok/s	—	—	4.2 kW	Specs →
66	Atlas 300V Pro 2 nodes	16	93 %	75.9 tok/s	3.0 tok/s	—	—	—	Specs →

By vendor

At BF16 the NVIDIA pick is 8 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at about 348 tokens/s for $4.00/hr; the AMD pick is 4 x Instinct MI300X 192GB at about 546 tokens/s for $9.56/hr. Per rental dollar NVIDIA delivers 1.52x the tokens of AMD here.

NVIDIA 30 parts fit

Best value: 8 x RTX PRO 6000 Blackwell Max-Q Workstation Edition, 348 tok/s for $4.00/hr

Fewest GPUs: 4 x Rubin SXM, 2.1k tok/s

AMD 11 parts fit

Best value: 4 x Instinct MI300X 192GB, 546 tok/s for $9.56/hr

Fewest GPUs: 2 x Instinct MI455X, 1.2k tok/s

Huawei 8 parts fit

Fewest GPUs: 8 x Ascend 950DT, 640 tok/s

Intel 4 parts fit

Fewest GPUs: 8 x Gaudi 3 128GB, 820 tok/s

Google 2 parts fit

Fewest GPUs: 4 x TPU v7 192GB, 805 tok/s

Other 5 parts fit

Fewest GPUs: 8 x Cloud AI 100 Ultra, 131 tok/s

Fleet what-if

Compare whole fleets at these settings: a thousand of one part against a hundred of another. Each fleet splits into replicas of its tensor-parallel width; GPUs left over sit idle.

A GPUs Part

B GPUs Part

Fleet letter	Fleet	Replicas	Idle	Tokens/s	Per hour	Power	Per M tokens
A	1,000 x H200 SXM 141GB	125 x TP8	0	111k tok/s	$2,480	700 kW	$6.21
B	100 x GB200 NVL72 GPU 186GB	25 x TP4	0	22k tok/s	$1,050	—	$13.37

Size Qwen3-235B-A22B for a tokens per second target

The calculator starts from your traffic instead of a fleet: set a target rate and read the replica count for every GPU.

Open in the LLM GPU Calculator

Frequently asked questions

How much GPU memory does Qwen3-235B-A22B need?

The weights take 470 GB at the native BF16 precision (470 GB at BF16, 235 GB at FP8, 132 GB at INT4). Each concurrent stream adds 188 KB of KV cache per token: 0.8 GB at 4,096 tokens and 6.3 GB at 32,768 tokens.

What is the bare minimum to run Qwen3-235B-A22B?

2 x Instinct MI355X OAM at BF16 with 4,096 tokens of context and one stream. The cheapest live rental for that footprint is 8 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $4.00 per hour.

How many H100s do you need to run Qwen3-235B-A22B?

8 H100s at BF16 for 4,096 tokens of context and one stream. For 32,768 tokens and 32 concurrent streams the tensor-parallel group is 16 H100s, and 1,000 tokens per second takes 16 H100s across 1 replicas.

How much does Qwen3-235B-A22B cost per million tokens on an H100?

About $6.76 per million output tokens at BF16, 32,768 tokens of context and 32 streams, using the cheapest live on-demand price of $1.83 per GPU-hour. The cheapest part per token is 8 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $3.19.

What precision does Qwen3-235B-A22B ship in?

The published checkpoint is BF16. FP8 and INT4 figures describe post-training quantisations that halve and quarter the footprint at a small accuracy cost.

All figures are roofline estimates from the model's config.json and each GPU's published memory, bandwidth and tensor-core peak, with fixed efficiency factors. Prices are the cheapest live on-demand listing per GPU when the site has one. They are a planning floor; a tuned serving stack can do better.