OpenAI gpt-oss

gpt-oss-120b hardware requirements

Fits on one GB300 NVL72 GPU 288GB at INT4 / FP4. For 1,000 tokens/s at 32k context you need 2 x H100 SXM5 80GB at $3.66/hr or 4 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $2.00/hr.

Parameters

117B

Active per token

5.1B

mixture of experts

KV cache per token

72 KB

16-bit cache

Native precision

MXFP4

Rent Model card

Top options for gpt-oss-120b

The strongest part each vendor offers for gpt-oss-120b at INT4 / FP4 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.

1 x GB300 NVL72 GPU 288GB 1 x Instinct MI355X OAM 1 x Ascend 950DT 1 x Gaudi 3 128GB 1 x TPU v7 192GB 4 x Inferentia2

Memory fills at: 1 x GB300 NVL72 GPU 288GB 174 users, 1 x Instinct MI355X OAM 174 users, 1 x Ascend 950DT 59 users, 1 x Gaudi 3 128GB 46 users, 1 x TPU v7 192GB 97 users, 4 x Inferentia2 44 users. Roofline estimates at INT4 / FP4.

Bare minimum

Fewest GPUs

1 x GB300 NVL72 GPU 288GB

INT4 / FP4, 4,096 tokens, one stream

Tokens/s: 2.0k
Per hour: $12.42
Per million tokens: $1.75
Memory used: 24%

Rent on DataCrunch

Cheapest per hour

1 x RTX PRO 6000 Blackwell Max-Q Workstation Edition

INT4 / FP4, 4,096 tokens, one stream, live price

Tokens/s: 438
Per hour: $0.50
Per million tokens: $0.32
Memory used: 72%

Rent on RunPod

Cheapest per million tokens

2 x RTX PRO 6000 Blackwell Max-Q Workstation Edition

INT4 / FP4, 32,768 tokens, 32 streams

Tokens/s: 868
Per hour: $1.00
Per million tokens: $0.32
Memory used: 57%

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; 68 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	1	25 %	6.3k tok/s	216 tok/s	—	—	—	Specs →
2	Rubin SXM	1	38 %	5.9k tok/s	204 tok/s	—	—	—	Specs →
3	TPU 8t announced	1	50 %	~1.8k tok/s	60.4 tok/s	—	—	—	Specs →
4	TPU 8i announced	1	38 %	~2.2k tok/s	79.6 tok/s	—	—	—	Specs →
5	Instinct MI355X OAM	1	38 %	2.1k tok/s	74.0 tok/s	$5.99	$0.79	1.4 kW	Rent →
6	GB300 NVL72 GPU 288GB	1	38 %	2.2k tok/s	74.0 tok/s	$12.42	$1.58	—	Rent →
7	GB200 NVL72 GPU 186GB	1	58 %	2.1k tok/s	74.0 tok/s	$10.50	$1.39	—	Rent →
8	Instinct MI350X OAM	1	38 %	2.1k tok/s	74.0 tok/s	$6.16	$0.82	1.0 kW	Rent →
9	TPU v7 192GB	1	57 %	~1.5k tok/s	68.3 tok/s	—	—	—	Specs →
10	Instinct MI430X OAM announced	1	25 %	~4.9k tok/s	216 tok/s	—	—	—	Specs →
11	B300 SXM 262GB	1	41 %	2.2k tok/s	74.0 tok/s	$4.99	$0.64	1.4 kW	Rent →
12	B200 SXM 180GB	1	60 %	2.0k tok/s	71.3 tok/s	$4.09	$0.57	1.0 kW	Rent →
13	Ascend 970 announced	1	38 %	3.3k tok/s	133 tok/s	—	—	—	Specs →
14	B100 SXM 192GB	1	57 %	2.0k tok/s	74.0 tok/s	—	—	0.7 kW	Specs →
15	Jalapeño announced	1	50 %	~3.9k tok/s	143 tok/s	—	—	0.7 kW	Specs →
16	Instinct MI325X OAM	1	42 %	~1.0k tok/s	55.5 tok/s	$3.80	$1.03	1.0 kW	Rent →
17	Instinct MI300X 192GB	1	57 %	~955 tok/s	49.1 tok/s	$2.39	$0.69	0.8 kW	Rent →
18	BR100	2	86 %	~722 tok/s	28.1 tok/s	—	—	1.1 kW	Specs →
19	Ascend 960 announced	1	38 %	2.1k tok/s	88.9 tok/s	—	—	—	Specs →
20	H200 SXM 141GB	1	77 %	~804 tok/s	44.4 tok/s	$2.48	$0.86	0.7 kW	Rent →
21	H100 SXM5 80GB	2	69 %	~1.2k tok/s	58.9 tok/s	$3.66	$0.83	1.4 kW	Rent →
22	TPU v6e 32GB	4	88 %	~1.4k tok/s	54.6 tok/s	—	—	—	Specs →
23	Data Center GPU Max 1550 128GB	1	85 %	~599 tok/s	30.3 tok/s	—	—	0.6 kW	Specs →
24	H200 NVL 141GB	1	77 %	~744 tok/s	44.4 tok/s	$3.61	$1.35	0.6 kW	Rent →
25	H100 NVL 94GB	2	58 %	~1.3k tok/s	68.6 tok/s	$6.38	$1.41	0.8 kW	Rent →
26	Ascend 910C	1	85 %	~580 tok/s	29.6 tok/s	—	—	0.6 kW	Specs →
27	H100 PCIe 80GB	2	69 %	~793 tok/s	35.2 tok/s	$5.00	$1.75	0.7 kW	Rent →
28	H800 PCIe 80GB	2	69 %	~793 tok/s	35.2 tok/s	—	—	0.7 kW	Specs →
29	Ascend 950DT	1	76 %	900 tok/s	37.0 tok/s	—	—	—	Specs →
30	RTX PRO 6000 Blackwell Workstation Edition	2	57 %	884 tok/s	31.5 tok/s	$3.60	$1.13	1.2 kW	Rent →
31	Ascend 950PR	1	85 %	421 tok/s	14.8 tok/s	—	—	0.9 kW	Specs →
32	MTT S5000	2	69 %	~597 tok/s	28.1 tok/s	—	—	—	Specs →
33	RTX PRO 6000 Blackwell Server Edition	2	57 %	798 tok/s	28.1 tok/s	$1.38	$0.48	1.2 kW	Rent →
34	Gaudi 3 128GB	1	85 %	~812 tok/s	34.2 tok/s	—	—	0.9 kW	Specs →
35	TPU v5p 95GB	2	58 %	~796 tok/s	48.6 tok/s	—	—	0.9 kW	Specs →
36	RTX PRO 6000 Blackwell Max-Q Workstation Edition	2	57 %	868 tok/s	31.5 tok/s	$1.00	$0.32	0.6 kW	Rent →
37	Cloud AI 100 Ultra	1	85 %	~135 tok/s	5.1 tok/s	—	—	0.1 kW	Specs →
38	Gaudi 2 96GB	2	57 %	~726 tok/s	43.1 tok/s	—	—	1.2 kW	Specs →
39	Atlas 350	1	98 %	363 tok/s	13.0 tok/s	—	—	0.6 kW	Specs →
40	Data Center GPU Max 1100 48GB	4	58 %	~895 tok/s	40.9 tok/s	—	—	1.2 kW	Specs →
41	GeForce RTX 5090 32GB	4	88 %	1.6k tok/s	59.7 tok/s	$4.76	$0.81	2.3 kW	Rent →
42	Ascend 910B	2	86 %	~551 tok/s	28.1 tok/s	—	—	0.8 kW	Specs →
43	MI250X 128GB	1	85 %	~412 tok/s	30.3 tok/s	—	—	0.6 kW	Specs →
44	RTX 6000 Ada 48GB	4	58 %	~722 tok/s	32.0 tok/s	$3.96	$1.52	1.2 kW	Rent →
45	Instinct MI250 128GB	1	85 %	~399 tok/s	30.3 tok/s	—	—	0.5 kW	Specs →
46	L40S 48GB	4	58 %	~668 tok/s	28.8 tok/s	$3.04	$1.26	1.4 kW	Rent →
47	A800 PCIe 80GB	2	69 %	~549 tok/s	34.0 tok/s	—	—	0.6 kW	Specs →
48	A800 SXM4	2	69 %	~563 tok/s	35.9 tok/s	—	—	0.8 kW	Specs →
49	A100 PCIe 80GB	2	69 %	~549 tok/s	34.0 tok/s	$2.70	$1.37	0.6 kW	Rent →
50	A100 SXM4 80GB	2	69 %	~563 tok/s	35.9 tok/s	$2.70	$1.33	0.8 kW	Rent →
51	A100 PCIe 40GB	4	70 %	~926 tok/s	51.8 tok/s	—	—	1.0 kW	Specs →
52	A100 SXM4 40GB	4	70 %	~926 tok/s	51.8 tok/s	$6.24	$1.87	1.6 kW	Rent →
53	GeForce RTX 5090 D V2 24GB	8	59 %	~2.3k tok/s	84.6 tok/s	—	—	4.6 kW	Specs →
54	RTX 5880 Ada 48GB	4	58 %	~661 tok/s	32.0 tok/s	—	—	1.1 kW	Specs →
55	TPU v4 32GB	4	88 %	~756 tok/s	40.0 tok/s	—	—	—	Specs →
56	RTX 5000 Ada 32GB	4	88 %	~455 tok/s	19.2 tok/s	$3.32	$2.03	1.0 kW	Rent →
57	RTX PRO 5000 Blackwell 48GB	4	58 %	~1.2k tok/s	44.8 tok/s	$3.84	$0.90	1.2 kW	Rent →
58	RTX PRO 5000 Blackwell 72GB	2	76 %	~627 tok/s	23.6 tok/s	—	—	0.6 kW	Specs →
59	RTX PRO 4500 Blackwell Server Edition	4	88 %	736 tok/s	26.7 tok/s	—	—	0.7 kW	Specs →
60	RTX PRO 4500 Blackwell 32GB	4	88 %	813 tok/s	29.9 tok/s	$2.88	$0.98	0.8 kW	Rent →
61	Radeon AI PRO R9700 32GB	4	88 %	~446 tok/s	21.3 tok/s	—	—	1.2 kW	Specs →
62	Inferentia2	4	88 %	~519 tok/s	27.3 tok/s	—	—	—	Specs →
63	Trainium	4	88 %	~519 tok/s	27.3 tok/s	—	—	—	Specs →
64	Instinct MI100 32GB	4	88 %	~421 tok/s	40.9 tok/s	—	—	1.2 kW	Specs →
65	Arc Pro B70 32GB	4	88 %	~425 tok/s	20.3 tok/s	—	—	0.9 kW	Specs →
66	L40 48GB	4	58 %	~524 tok/s	28.8 tok/s	$3.28	$1.74	1.2 kW	Rent →
67	Instinct MI210 PCIe	2	86 %	~379 tok/s	28.8 tok/s	—	—	0.6 kW	Specs →
68	GeForce RTX 4090 24GB	8	59 %	~1.0k tok/s	63.4 tok/s	$3.20	$0.86	3.6 kW	Rent →
69	A30 24GB	8	59 %	~991 tok/s	58.7 tok/s	—	—	1.3 kW	Specs →
70	RTX 4500 Ada 24GB	8	59 %	~607 tok/s	27.2 tok/s	—	—	1.7 kW	Specs →
71	RTX A6000 48GB	4	58 %	~458 tok/s	25.6 tok/s	$2.00	$1.21	1.2 kW	Rent →
72	A40 48GB	4	58 %	~427 tok/s	23.2 tok/s	$2.36	$1.53	1.2 kW	Rent →
73	H20 141GB HBM3e	1	77 %	~224 tok/s	37.0 tok/s	—	—	0.4 kW	Specs →
74	H20 96GB	2	57 %	~426 tok/s	70.3 tok/s	—	—	0.8 kW	Specs →
75	RTX PRO 4000 Blackwell 24GB	8	59 %	~1.1k tok/s	42.3 tok/s	$4.56	$1.11	1.2 kW	Rent →
76	BI-V100 Tiangai 100	4	88 %	~558 tok/s	40.0 tok/s	—	—	1.0 kW	Specs →
77	Atlas 300I Duo 48GB	4	58 %	~298 tok/s	13.6 tok/s	—	—	0.6 kW	Specs →
78	Atlas 300I Duo 96GB	2	57 %	~157 tok/s	7.2 tok/s	—	—	0.3 kW	Specs →
79	RTX A5500 24GB	8	59 %	~817 tok/s	48.3 tok/s	—	—	1.8 kW	Specs →
80	Wormhole n300s	8	59 %	~683 tok/s	36.3 tok/s	—	—	2.4 kW	Specs →
81	V100S PCIe 32GB	4	88 %	~507 tok/s	37.8 tok/s	$3.52	$1.93	1.0 kW	Rent →
82	ATOM-Max	2	86 %	~254 tok/s	18.0 tok/s	—	—	0.7 kW	Specs →
83	A10 24GB	8	59 %	~686 tok/s	37.8 tok/s	$8.81	$3.57	1.2 kW	Rent →
84	Radeon PRO W7900 Dual Slot 48GB	4	58 %	~436 tok/s	28.8 tok/s	—	—	1.2 kW	Specs →
85	Radeon PRO W7900 48GB	4	58 %	~436 tok/s	28.8 tok/s	—	—	1.2 kW	Specs →
86	Radeon RX 7900 XTX 24GB	8	59 %	~863 tok/s	60.4 tok/s	—	—	2.8 kW	Specs →
87	L4 24GB	8	59 %	~434 tok/s	18.9 tok/s	$4.72	$3.02	0.6 kW	Rent →
88	L20 48GB	4	58 %	~429 tok/s	28.8 tok/s	—	—	1.1 kW	Specs →
89	Radeon AI PRO R9600D 32GB	4	88 %	~337 tok/s	21.3 tok/s	—	—	0.6 kW	Specs →
90	Arc Pro B65 32GB	4	88 %	~328 tok/s	20.3 tok/s	—	—	0.8 kW	Specs →
91	Arc Pro B60 24GB	8	59 %	~530 tok/s	28.7 tok/s	—	—	1.6 kW	Specs →
92	L2 24GB	8	59 %	~405 tok/s	18.9 tok/s	—	—	—	Specs →
93	RTX PRO 4000 Blackwell SFF 24GB	8	59 %	~739 tok/s	27.2 tok/s	—	—	0.6 kW	Specs →
94	MLU370-X8	4	58 %	~325 tok/s	20.5 tok/s	—	—	1.0 kW	Specs →
95	Radeon PRO W7800 48GB	4	58 %	~366 tok/s	28.8 tok/s	—	—	1.0 kW	Specs →
96	Radeon PRO W7800 32GB	4	88 %	~305 tok/s	19.2 tok/s	—	—	1.0 kW	Specs →
97	GeForce RTX 3090 24GB	8	59 %	~610 tok/s	58.9 tok/s	$4.00	$1.82	2.8 kW	Rent →
98	A10G 24GB	8	59 %	~512 tok/s	37.8 tok/s	$8.05	$4.37	2.4 kW	Rent →
99	Atlas 300V Pro	4	58 %	~149 tok/s	6.8 tok/s	—	—	—	Specs →
100	Atlas 300I Pro	8	59 %	~282 tok/s	12.9 tok/s	—	—	—	Specs →
101	Atlas 300V	8	59 %	~250 tok/s	12.9 tok/s	—	—	0.6 kW	Specs →
102	GeForce RTX 3090 Ti 24GB	8	59 %	~406 tok/s	63.4 tok/s	$3.68	$2.52	3.6 kW	Rent →
103	Radeon PRO W6800 32GB	4	88 %	~167 tok/s	17.1 tok/s	—	—	1.0 kW	Specs →
104	RTX A5000 24GB	8	59 %	~287 tok/s	48.3 tok/s	$2.16	$2.09	1.8 kW	Rent →
105	Radeon PRO V710 28GB	8	50 %	~253 tok/s	28.2 tok/s	—	—	1.3 kW	Specs →

By vendor

At INT4 / FP4 the NVIDIA pick is 2 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at about 868 tokens/s for $1.00/hr; the AMD pick is 1 x Instinct MI300X 192GB at about 955 tokens/s for $2.39/hr. Per rental dollar NVIDIA delivers 2.17x the tokens of AMD here.

NVIDIA 51 parts fit

Best value: 2 x RTX PRO 6000 Blackwell Max-Q Workstation Edition, 868 tok/s for $1.00/hr

Fewest GPUs: 1 x Rubin SXM, 5.9k tok/s

AMD 18 parts fit

Best value: 1 x Instinct MI300X 192GB, 955 tok/s for $2.39/hr

Fewest GPUs: 1 x Instinct MI455X, 6.3k tok/s

Huawei 10 parts fit

Fewest GPUs: 1 x Ascend 950DT, 900 tok/s

Intel 7 parts fit

Fewest GPUs: 1 x Gaudi 3 128GB, 812 tok/s

Google 4 parts fit

Fewest GPUs: 1 x TPU v7 192GB, 1.5k tok/s

AWS 2 parts fit

Fewest GPUs: 4 x Inferentia2, 519 tok/s

Other 7 parts fit

Fewest GPUs: 1 x Cloud AI 100 Ultra, 135 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	1,000 x TP1	0	804k tok/s	$2,480	700 kW	$0.86
B	100 x GB200 NVL72 GPU 186GB	100 x TP1	0	210k tok/s	$1,050	—	$1.39

Size gpt-oss-120b 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 gpt-oss-120b need?

The weights take 65.7 GB at the native INT4 / FP4 precision (234 GB at BF16, 117 GB at FP8, 65.7 GB at INT4). Each concurrent stream adds 72 KB of KV cache per token: 0.2 GB at 4,096 tokens and 1.2 GB at 32,768 tokens.

What is the bare minimum to run gpt-oss-120b?

1 x GB300 NVL72 GPU 288GB at INT4 / FP4 with 4,096 tokens of context and one stream. The cheapest live rental for that footprint is 1 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $0.50 per hour.

How many H100s do you need to run gpt-oss-120b?

one H100 at INT4 / FP4 for 4,096 tokens of context and one stream. For 32,768 tokens and 32 concurrent streams the tensor-parallel group is 2 H100s, and 1,000 tokens per second takes 2 H100s across 1 replicas.

How much does gpt-oss-120b cost per million tokens on an H100?

About $0.83 per million output tokens at INT4 / FP4, 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 2 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $0.32.

What precision does gpt-oss-120b ship in?

The published checkpoint is MXFP4. That 4-bit format is what the lab validated, so the INT4 column is the native one; BF16 figures describe a dequantised copy nobody would deploy.

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. Active parameters per the OpenAI model card. Half the layers attend over a 128-token sliding window, so the cache grows at half the rate the head count suggests.