What is an inference provider?

A company that runs AI models on its own or rented GPUs and sells the output by the token through an API. Open-weight models such as DeepSeek, GLM, Kimi and Qwen can be served by anyone, so dozens of providers compete on price, speed and uptime for the same model.

Inference Providers

Who serves open-weight model tokens, and how many.

Finance your GPUs

What it shows

Tokens each provider served over Oct 3 to Oct 9, 2026, through a large public LLM router. That is one channel: direct, cloud-marketplace and enterprise traffic is not included.

Who serves the most open-weight tokens

Tokens served for other labs' models; a lab serving its own model is left out of its own total.

#	Provider	HQ	Tokens, 7d	Share	Largest family	Open models
1	Together	United States	11.5T	18.4%	DeepSeek	14
2	Relace	-	9.54T	15.3%	DeepSeek	6
3	DeepInfra	United States	6.43T	10.3%	DeepSeek	69
4	inference.net	United States	4.13T	6.6%	DeepSeek	7
5	NovitaAI	United States	3.26T	5.2%	DeepSeek	51
6	Parasail	United States	2.67T	4.3%	DeepSeek	38
7	Wafer	United States	2.57T	4.1%	DeepSeek	10
8	StreamLake	China	2.53T	4.1%	DeepSeek	19
9	Fireworks	United States	2.20T	3.5%	DeepSeek	4
10	Sail Research	United States	1.71T	2.7%	DeepSeek	6
11	CoreWeave	United States	1.55T	2.5%	DeepSeek	19
12	GMICloud	United States	1.45T	2.3%	DeepSeek	26
13	DigitalOcean	United States	1.37T	2.2%	DeepSeek	16
14	AtlasCloud	United States	1.17T	1.9%	DeepSeek	23
15	Morph	United States	1.02T	1.6%	DeepSeek	5
16	Decart	United States	981B	1.6%	GLM	5
17	SiliconFlow	Singapore	893B	1.4%	GLM	40
18	Open Inference	United States	759B	1.2%	GLM	4
19	Baseten	United States	598B	1.0%	GLM	8
20	Mistral	France	576B	0.9%	GLM	11
21	Baidu Qianfan	China	500B	0.8%	DeepSeek	11
22	Alibaba Cloud Int.	Singapore	466B	0.7%	Qwen	18
23	Venice	United States	465B	0.7%	DeepSeek	38
24	Modal	United States	460B	0.7%	GLM	5
25	Makora	-	414B	0.7%	DeepSeek	5
26	Friendli	United States	350B	0.6%	GLM	7
27	Phala	United States	311B	0.5%	DeepSeek	20
28	ModelRun [by Modular]	United States	283B	0.5%	Qwen	3
29	Cohere	United States	283B	0.5%	DeepSeek	2
30	Reka AI	-	280B	0.4%	GLM	8
31	Crusoe	United States	204B	0.3%	GLM	6
32	DekaLLM	Indonesia	164B	0.3%	Qwen	10
33	Groq	United States	161B	0.3%	gpt-oss	6
34	Inceptron	Sweden	160B	0.3%	DeepSeek	6
35	NextBit	Spain	154B	0.2%	Gemma	7
36	Cloudflare	United States	114B	0.2%	GLM	18
37	Darkbloom	-	108B	0.2%	Gemma	9
38	Ionstream	United States	103B	0.2%	DeepSeek	3
39	io.net	United States	99B	0.2%	MiMo	6
40	AkashML	-	89B	0.1%	gpt-oss	7

Tokens by model family

Family	Tokens, 7d	Share	Week on week	Providers	Top provider	Lab serves
DeepSeek	48.3T	47.0%	+34%	39	Together 18%	12%
GLM	16.4T	16.0%	+20%	42	Together 16%	8%
MiMo	13.1T	12.7%	+11%	11	Xiaomi 91%	91%
Hunyuan	10.1T	9.8%	+14%	7	Tencent Cloud 97%	97%
Nemotron	7.85T	7.6%	+12%	11	NVIDIA 100%	100%
Kimi	2.22T	2.2%	+20%	33	inference.net 28%	15%
Qwen	1.47T	1.4%	+5%	31	Alibaba Cloud Int. 47%	47%
MiniMax	1.32T	1.3%	+0%	17	MiniMax 86%	86%
Gemma	882B	0.9%	+4%	19	DeepInfra 22%	n/a
gpt-oss	733B	0.7%	-5%	22	Groq 20%	n/a
Mistral	323B	0.3%	-2%	7	DeepInfra 78%	8%
Llama	117B	0.1%	-6%	12	Groq 66%	n/a

Lab serves: the share the model's own lab served through its API; n/a where it has none.

Price alone does not win share. On DeepSeek V4.1 Flash, the most-used open model, Together served 23% of tokens at $1.20 per million output tokens. The cheapest listing, Decart at $0.18, served 0.3%.

Prices and uptime. Each family page lists every provider's price and speed per model. Daily history is in the price index and the Inference Uptime Tracker; monthly totals in the Open-Weight Token Report; gateways compared in LLM routers.

Finance your GPUs

Serving tokens on your own GPUs? Tell us the cluster, the models you serve and the contracts behind them, and we will share it with partner funders that fit.

Prefer email? hello@amcompute.com

Questions

What is an inference provider?: A company that runs AI models on its own or rented GPUs and sells the output by the token through an API. Open-weight models such as DeepSeek, GLM, Kimi and Qwen can be served by anyone, so dozens of providers compete on price, speed and uptime for the same model.