LLM brokers · ranked
Comparison · who brokers other companies' models
Ten LLM brokers, ranked.
Gateways that route to other providers, and direct hosts of open-weight models — sized by team, valuation, funding, users, tokens and model count. Click any column header to sort; click again to reverse.
Figures checked 25 Oct 2026 from each company's own newsroom/pricing pages, live model APIs and press (TechCrunch, Reuters, NYT, Bloomberg, Forbes) — method and per-figure sources in the cards below. Loads sorted by team size, biggest company first.
| Service▲ | Team3▲ | Users4▲ | Models1▲ | Val., $B2▲ | Raised2▲ | Tokens/mo5▲ | HQ▲ | Scope▲ | Type▲ | Pay6▲ |
|---|---|---|---|---|---|---|---|---|---|---|
| Together AIserverless + dedicated | 438PitchBook, Aug 2026 | 1M+1M+ developers | 200open-weight served | $8.3BSeries C, Jul 2026 | $1.33B$800M Series C, Jul 2026 | 400Tblog, Jul 2026 | San FranciscoCA, US | open-weight only | direct host + GPU rental | per-token, no platform fee |
| Groq (GroqCloud)hosted inference, LPU heritage | 313LinkedIn, Oct 2026 (post-NVIDIA) | 6M+6M+ developers | 14curated + enterprise-quoted | $3.5Bhalved post-NVIDIA deal | $4.25B$4.25B total (trackers vary) | —trillions/week, unnumbered | Mountain ViewCA, US | open + licensed | hosts on NVIDIA capacity post-LPU | per-token + free tier |
| Fireworks AIinference cloud + FireRouter | 309LinkedIn-derived, Sep 2026 | 10K10K+ organizations | 329+ customer post-trained | $17.5B$30–40B talks reported, unclosed | $1.83B$1.5B Series D, Jul 2026 | 1,200T40T/day × 30 (derived) | Redwood CityCA, US | open + post-trained | hosts + fine-tunes + FireRouter | credits; GPUs from $8/hr |
| Hugging FaceInference Providers (routes Hub) | 250CEO-stated, Jan 2025 | —none; 20M+ requests/mo | 200docs 200+; 132 live chat | $4.5BSeries D, Aug 2023 (last known) | $395M$395M incl. $235M D | —requests only, no tokens | New YorkNY, US | open only (Hub) | routes Hub models to partners | no markup; PRO $9/mo |
| OpenRoutergateway · part of Stripe | 70trackers 54–90 diverge | 10M+10M+ own claim (TC: 8M) | 467API count, 25 Oct 2026 | $7.5BStripe price, Aug 2026 | $153M$113M B, May 2026 | 576TSep rankings; site 700T+ | New YorkNY, US | open + closed via routes | gateway to 80+ providers, no GPUs | pass-through; 5.5% credit fee |
| Nous ResearchPortal · every lab + tool gateway | 46Dealroom, Aug 2026 (40–66) | —24M agent clones ≠ Portal | 300300+ across labs | $1.5Bconfirmed Oct 7, 2026 | $158M$90M Series B, Oct 2026 | —unpublished | AustinTX, US | open + closed + own Hermes | proxies every lab + tools | $20→$22 credits, rollover caps |
| DeepInfraserverless open-model APIs | 43LinkedIn, Oct 2026 | —none; $100M ARR, Sep 2026 | 175API count, 25 Oct 2026 | —~$1B claims unsourced | $133M$133M incl. $40M debt | 95.7T22T/wk × 30.4 (derived) | Palo AltoCA, US | open only | hosts on own GPUs | per-token, no contracts |
| Replicateserverless runs + fine-tunes | 25LinkedIn, Oct 2026 | 2M+2M signups, Dec 2023 | 6,158sitemap, 25 Oct 2026 | $0.35BSeries B, Dec 2023 (dated) | $58M~$58M, none since Dec 2023 | —run counts, not tokens | San FranciscoCA, US | open + official | serverless GPU runs (Cog) | per-GPU-second |
| Novita AImodel APIs + GPU cloud | 24LinkedIn, Oct 2026 | —~7.7M HF hub reqs/mo | 200open-source catalog | —bootstrapped — none | $0M$0 — no rounds found | —unpublished | SingaporeSG + SF office | open only | HF official inference partner | per-use; GPUs $2.59/hr |
| Ollama Cloudhost · Free / Pro / Max | 14CEO Jul 2026; trackers 69–76 | 8.9M+local tool, not Cloud | 21open-weight only | —declined to disclose | $88M$15M A + $65M B | —no cloud volume published | Palo AltoCA, US | open only | direct host on NVIDIA partners | Pro $20→$60 credits (3×) |
Spot check: the same model, on two opposite brokers
| Model▲ | Ollama input6▲ | Ollama output▲ | On Pro: effective in / out▲ | OpenRouter input▲ | OpenRouter output▲ | Better input rate▲ |
|---|---|---|---|---|---|---|
| beamOR id: reflection/beam-1.0-501b-a23b | freefree on Ollama Cloud — no Pro needed | free | freeout free | $0.25 | $1.00 | Ollama |
| deepseek-v4-proOR id: deepseek/deepseek-v4-pro | $1.32−50% off-peak on Ollama | $3.96 | $0.44out $1.32 | $0.21 | $0.42 | OpenRouter |
| deepseek-v4.1-flashOR id: deepseek/deepseek-v4.1-flash | $0.30−50% off-peak on Ollama | $1.20 | $0.100out $0.40 | $0.020 | $1.00 | OpenRouter |
| glm-5.3-flashOR id: z-ai/glm-5.3-flash | $0.15same headline rate on both | $0.50 | $0.050out $0.17 | $0.15 | $0.50 | Ollama |
| gpt-oss:120bOR id: openai/gpt-oss-120b | $0.15same headline rate on both | $0.60 | $0.050out $0.20 | $0.15 | $0.60 | Ollama |
What counts as a broker
Rows are services that serve other companies' models: gateways/routers (OpenRouter, Nous Portal, HF Inference Providers — they hold no GPUs and forward to providers) and direct hosts of open-weight models (Together, Fireworks, GroqCloud, DeepInfra, Replicate, Novita, Ollama Cloud). Closed-weight-only vendors (OpenAI, Anthropic, Google) are excluded. Groq is in despite its chip era ending: NVIDIA took the LPU IP and ~90% of staff in a ~$20B deal (Dec 2025) and Groq Cloud kept serving on NVIDIA capacity. ↑
Market cap does not exist here — valuation stands in
None of the ten is public, so there is no market cap. The column shows the last post-money valuation from funding or deal news, with its date. Two entries are acquisitions/marks rather than raised rounds: OpenRouter's $7.5B is the reported Stripe price (announced 19 Aug 2026), and Fireworks' $17.5B is its Jul 2026 Series D — press reported talks at $30–40B (Oct 2026) but nothing has closed. Replicate's mark is dated Dec 2023; DeepInfra declines to disclose; Ollama's founders declined; Novita has taken no outside money. ↑
Team counts: stated vs tracked
Two rows carry company-stated counts (Hugging Face 250, CEO Jan 2025; Ollama 14, CEO to TechCrunch Jul 2026) — but trackers estimate Ollama at 69–76 and OpenRouter at 54–90, so the sort column uses the mid-range with ranges in the cells. Groq's 313 is post-NVIDIA re-staffing (LinkedIn, Oct 2026); Together's 438 is PitchBook (Aug 2026); Fireworks' trackers span 208–309. ↑
Users: every row counts something different
OpenRouter's 10M+ ("8 million" per TechCrunch, May 2026) and Groq's 6M+ developers are the platforms' own claims. Ollama's 8.9M are users of the free LOCAL tool — not Cloud subscribers. Replicate's 2M signups date to Dec 2023. Fireworks counts organizations (10K+). Hugging Face, Nous, DeepInfra and Novita publish no user counts — HF publishes request volume instead. ↑
Token volume: who publishes what
Only four rows have real numbers. Together: 400T tokens/month (company blog, Jul 2026). Fireworks: 40T tokens/day → ~1.2 quadrillion/month derived (×30). OpenRouter: 576T in Sep 2026 — aggregated from OpenRouter's own CC-BY rankings data (CrbonFree); its homepage claims 700T+, and its newsletter reported 128.4T in one week of Sep 2026. DeepInfra: 22T tokens/week → ~95.7T/month derived. Groq says “trillions of tokens every week” without a number; HF counts 20M+ requests/month; Ollama, Nous, Replicate and Novita publish nothing. T = trillion. ↑
Pricing: where fees hide
OpenRouter passes through provider list prices but takes 5.5% on credit top-ups ($0.80 min; 5% crypto) and 5% on BYOK past $25,000/mo. Hugging Face charges no markup (PRO is $9/mo with $2 credits). Stipends differ: Ollama triples $20→$60 but expires monthly; Nous pays a 10% bonus ($20→$22) with small rollover caps. Fireworks/Novita/DeepInfra/ Together are raw pay-as-you-go; Replicate mostly bills per GPU-second, not per token. ↑
Bottom line
Scale and money concentrate at the top: Fireworks carries the biggest valuation ($17.5B) and the largest measured volume (~1.2P tokens/mo, derived), Together the largest headcount and cleanest 400T/mo disclosure, and OpenRouter the widest catalog (467 models) with the most disclosed per-route data. The cheapest paths stay at the bottom of the size ranking: on the table below, Ollama Pro is 3× on glm-5.3-flash and gpt-oss and free on beam, while OpenRouter's $0.02/M DeepSeek input is the cheapest agent input price anywhere here. ↑