Together AI — research brief

# Together AI — research brief

**Slug:** together-ai  
**Date:** 2026-09-21 (America/Chicago)  
**Website:** https://www.together.ai  
**Industry:** AI Inference / AI Native Cloud  

## Label key
- **FACT** — primary: together.ai pricing, docs, company blog/about  
- **SECONDARY** — press / analyst (cited, not invented)  
- **INFERENCE** — reasoned from public product shape; not a private metric  

## Company snapshot
Together AI (Together Computer Inc.) positions itself as an **AI Native Cloud** for open-source and custom models: serverless inference APIs, dedicated endpoints, GPU clusters for training/high-throughput, fine-tuning, sandboxes, and managed storage (FACT — site/pricing).

Company blog (2026-07-01) announces **$800M Series C** with investors including Aramco Ventures, NVIDIA, Vista Equity, General Catalyst, Emergence Capital, Salesforce Ventures, Lux Capital, and others; plus **commitments for over 500 MW** of compute capacity capitalized independently by investors (FACT — company blog). Named customers in that post: Cognition, Decagon, Eleven Labs, Cursor, Suno (FACT claim).

## Revenue / pricing ladder (FACT — together.ai/pricing + docs)
1. **Serverless Inference** — pay per token / megapixel / audio-video second; **no minimums, no provisioning** (FACT docs serverless overview). 100+ open models; OpenAI-compatible APIs. Example chat rates (per 1M tokens, input/output): Llama 3.3 70B **$1.04/$1.04**; gpt-oss-120B **$0.15/$0.60**; GLM-5.3-Flash **$0.15/$0.50** (FACT pricing page at research time).
2. **Batch API** — up to **50% lower** cost for async workloads (FACT docs + pricing).
3. **Cached input** — select chat models bill cached prefixes at steep discount (automatic; FACT docs).
4. **Provisioned Throughput (PTUs)** — reserved capacity units; contact/estimator on pricing (FACT pricing).
5. **Dedicated Inference** — single-tenant GPU-hour: NVIDIA HGX **H100 $5.49** on-demand (**promo $3.99** until 09/30/26), **B200 $8.99**; H200/B300/GB200 contact (FACT pricing).
6. **GPU Clusters** — H100 on-demand **$3.99**/GPU-hr; reserved 7–30d **$3.69**, 31–90d **$3.45**, 91–180d **$3.19**; preemptible **$1.99**; H200/B200 higher tiers (FACT pricing).
7. **Fine-Tuning** — SFT/DPO priced per 1M tokens with per-model minimum charges (FACT pricing).
8. **Sandbox / Code Interpreter** — Code Sandbox **$0.0446**/vCPU-hr + **$0.0149**/GiB RAM-hr; Code Interpreter **$0.03**/60-min session (FACT pricing).
9. **Managed Storage** — shared filesystem **$0.16**/GiB/month (FACT pricing).

Teaching arc: **Serverless tokens → Batch/cache discounts → Dedicated GPU-hr → Clusters (on-demand→reserved) → Fine-tune + Sandbox attach → Enterprise/PTU sales**.

## Value proposition themes (FACT product)
- Open-weights economics (company claims 6x–20x lower cost vs closed frontier in production; Decagon 6x — FACT company blog claim)
- Full stack: models + inference systems + training infra + kernels (FlashAttention lineage via research culture — company narrative)
- Same API surface from serverless prototype to dedicated/reserved (FACT docs)

## Partners / ecosystem (FACT / company-stated)
- NVIDIA (investor + Cloud Partner narrative — company blog/news)
- Open-weight model labs hosted (DeepSeek, Nemotron, MiniMax, Kimi, GLM, Llama, Qwen — company blog + pricing catalog)
- Named production customers (Cognition, Decagon, Eleven Labs, Cursor, Suno — Series C post)

## What we do NOT invent
- No ARR, burn, headcount, or token-volume figures unless company-published (none used as FACT metrics beyond blog qualitative claims).
- Sacra / press valuations labeled SECONDARY if mentioned; not used as keyMetrics.

## Primary sources
- https://www.together.ai/pricing  
- https://docs.together.ai/docs/serverless/overview  
- https://docs.together.ai/docs/dedicated-endpoints/pricing  
- https://www.together.ai/blog/announcing-our-series-c  
- https://www.together.ai/brand