Key Partners
- NVIDIA — investor + Cloud Partner narrative (FACT blog/news).
- Open-weight labs hosted — DeepSeek, Nemotron, MiniMax, Kimi, GLM, Llama, Qwen (FACT blog + pricing catalog).
- Series C investors — Aramco Ventures, Vista, GC, Emergence, Salesforce Ventures, Lux, others (FACT blog).
- 500 MW+ compute capacity commitments capitalized by investors (FACT blog).
- Named customers — Cognition, Decagon, Eleven Labs, Cursor, Suno (FACT blog claim).
- OpenAI-compatible API ecosystem for tool/SDK partners (FACT docs).
Key Activities
- Operate serverless multi-model inference APIs.
- Provision Dedicated endpoints + GPU clusters (on-demand/reserved/preemptible).
- Ship Fine-Tuning (SFT/DPO), Sandbox, Code Interpreter, Storage.
- Advance kernels/compilers and model serving (company research narrative).
- Publish pricing + docs; run Batch and cache optimizations.
- Sell PTUs / enterprise capacity for steady production.
Key Resources
- GPU inventory — H100, H200, B200, GB200/B300 paths (FACT pricing tables).
- 500 MW+ committed capacity pipeline (FACT Series C blog).
- Serverless + Dedicated + Cluster product surfaces (FACT pricing).
- OpenAI-compatible inference APIs + model catalog (FACT docs).
- Research/serving IP narrative (FlashAttention-4, Megakernel — company blog).
- Named production customer relationships (FACT blog claim).
Value Propositions
- Serverless — 100+ open models, no provisioning, pay per use (FACT docs).
- OpenAI-compatible APIs — swap providers with less rewrite (FACT docs).
- Batch up to 50% off; cached input discounts on select models (FACT).
- Dedicated single-tenant GPUs for latency/custom models (FACT pricing).
- GPU Clusters on-demand → reserved → preemptible (FACT pricing).
- Fine-Tuning + Sandbox + Storage on the same platform (FACT pricing).
- Company claim — 6x–20x lower cost vs closed frontier at scale (blog).
Customer Relationships
- Self-serve signup + public pricing cards (FACT site).
- Docs-led expansion — serverless → dedicated when busy (FACT DMI docs).
- Batch/cache education to keep cost competitive (FACT docs).
- PTU estimator + Contact sales for reserved capacity (FACT pricing).
- Startup Accelerator program narrative (company blog index).
- Reference customers named in Series C communications (FACT claim).
Channels
- Primary — together.ai + /pricing public cards.
- docs.together.ai quickstarts and model catalogs.
- OpenAI-compatible API / SDK discovery.
- Company blog — research, model library, funding narrative.
- NVIDIA Cloud Partner + marketplace surfaces (company news).
- Contact sales for PTUs, reserved clusters, enterprise.
Customer Segments
- App/agent builders on serverless open models (FACT docs ICP).
- AI-native production companies (named blog logos — FACT claim).
- Teams needing dedicated/custom model endpoints (FACT pricing).
- ML orgs training or self-serving on GPU clusters (FACT pricing).
- Residency/compliance buyers pinning dedicated regions (FACT docs).
- Products attaching Fine-Tune + Code Sandbox (FACT pricing).
Cost Structure
- GPU / power / colo COGS behind serverless + dedicated (INFERENCE).
- Cluster inventory and reserved-capacity utilization risk (INFERENCE).
- Serving-stack and kernel R&D (company research narrative).
- Fine-tune + sandbox compute COGS (INFERENCE from pricing).
- MW-scale capacity commitments with investors (FACT blog claim).
- GTM — PLG content + enterprise sales (INFERENCE).
- No public Together AI P&L line invented.
Revenue Streams
- Serverless — PAYG tokens/images/audio/video; no minimums (FACT).
- Examples — Llama 3.3 70B $1.04/$1.04; gpt-oss-120B $0.15/$0.60 per 1M (FACT pricing).
- Batch up to 50% off; cached input discounts (FACT docs).
- Dedicated — H100 $5.49/hr (promo $3.99 thru 09/30/26); B200 $8.99 (FACT).
- Clusters — H100 $3.99 OD → $3.19 reserved (91–180d); preemptible $1.99 (FACT).
- Fine-Tuning — $/1M tokens + minimums; Sandbox/Storage attach (FACT).
- PTU / Enterprise — contact sales for reserved throughput (FACT).