Together AI
GPU "neocloud" + open-model inference & training platform · usage-based API + dedicated-capacity rental
Seed (May 2023, $20M) and Nov 2023 Series A ($102.5M) had no disclosed valuation — omitted per honesty gate. Series C ($7.5B) is reported by The Information / private-market trackers, not company-confirmed.
Earnings, margins, COGS & capex
No public filings exist. Revenue mix (est.): ~60–70% dedicated GPU-cluster/server rental, ~30–40% serverless/dedicated inference API + fine-tuning. Gross margin (~45% est.) sits between a pure-API software business and a bare-metal GPU reseller because much capacity is historically rented (CoreWeave, Lambda) and increasingly self-operated (Hypertec 36K-GB200 co-build; Rumble HGX B300). Capex/financing intensity is high and rising — the business funds GPU buildout via equity rounds (>$1.5B raised) and multi-year customer prepays. The structural P&L question is whether API/inference (high-margin, sticky) outgrows GPU rental (low-margin, capital-hungry) fast enough to lift blended margin as it scales. No earnings cadence, beat/miss, or guidance — private.
Income statement — where each revenue dollar goes
% of revenueOf every $1 of revenue, ~55¢ is cost of goods and ~45¢ operating expense, leaving ~0¢ of operating profit.
Revenue trend
Margins
COGS structure
Not disclosed (private company).
Capex
Not disclosed.
Growth drivers
- ~3x growth to ~$1B ARR validates real demand, not a science project; the re-rate to ~$7.5B (>2x in ~14 months) is supported by the revenue, not just sentiment
- FlashAttention/spec-decoding DNA = durable per-token cost advantage; whoever serves open models cheapest and fastest wins, and Together's researchers literally wrote the primitives
- Day-0, 200+-model catalog + OpenAI-compatible API makes it the path of least resistance for the enterprise exodus from closed-API lock-in
- Mix-shift from rented GPU to self-operated capacity + higher API share is a clean margin-expansion story from ~45% upward
Bull & bear
Inference is the recurring tollbooth of the AI economy, open weights keep winning enterprise share, and Together is the research-led, model-neutral default to serve and tune them — with a kernel-level cost edge that widens as it shifts to API and owns more capacity.
- ~3x growth to ~$1B ARR validates real demand, not a science project; the re-rate to ~$7.5B (>2x in ~14 months) is supported by the revenue, not just sentiment
- FlashAttention/spec-decoding DNA = durable per-token cost advantage; whoever serves open models cheapest and fastest wins, and Together's researchers literally wrote the primitives
- Day-0, 200+-model catalog + OpenAI-compatible API makes it the path of least resistance for the enterprise exodus from closed-API lock-in
- Mix-shift from rented GPU to self-operated capacity + higher API share is a clean margin-expansion story from ~45% upward
- Marquee strategic backers (NVIDIA, Salesforce, Prosperity7) provide silicon access, distribution, and capital for the GPU treadmill
Strip out the research halo and Together is a capital-hungry GPU reseller with ~45% margins, sandwiched between hyperscalers above and well-funded neocloud peers beside it, in a market where inference price is a race to zero and GPUs are a depreciating, financeable asset.
- ~45% gross margin and a rental-heavy mix mean this trades like infrastructure, not software — the ~$7.5B mark prices in margin expansion that isn't proven
- Inference is commoditizing fast; AWS/Azure/Google can bundle and underprice, and Fireworks/Baseten/CoreWeave are equally or better capitalized
- GPU pricing is cyclical and capex never stops — equity rounds and customer prepays fund a depreciating asset base; a price-collapse cycle hits margin and the financing model at once
- Customer/GPU-supplier concentration risk (CoreWeave-style fragility) is plausible and undisclosed
- Disintermediation: open-model vendors or NVIDIA serving direct erodes the middle-layer thesis; the moat is real on speed but thin on lock-in
What it is worth
Revenue-multiple comp vs. neocloud/inference peers (no DCF — private, unaudited). Anchor: ~$7.5B reported Series C valuation on ~$1B ARR ≈ ~7.5x ARR. Cross-check vs peers: Fireworks ~$15B on ~$800M (~19x), Baseten ~$11B on ~$600M (~18x), CoreWeave (CRWV, public) trades at a lower infra multiple. Together's ~7.5x is a DISCOUNT to private inference peers — reflecting its rental-heavier, lower-margin (~45%) mix, but leaving room if margin/mix shifts and growth holds.
~$3.5–5B
if inference price competition from hyperscalers + neocloud peers compresses growth and margin, GPU-pricing cycles bite the rental book, and the multiple de-rates toward infra/CoreWeave-like levels; a down-round risk in a GPU-oversupply scenario.
~$7.5–9B
consistent with the reported Series C mark (private trackers $7.5–8.5B); ~7.5–9x a ~$1B ARR, a deserved discount to pure-API peers for the rental-heavy, ~45%-margin mix until the shift proves out.
~$12–15B
if ARR compounds past ~$1.5–2B with API/inference mix lifting gross margin toward 55–60%, Together re-rates toward Fireworks/Baseten's ~15–19x; an IPO or strategic round prices the open-model-inference scarcity premium.
SWOT
Strengths
- Research moat — the FlashAttention / FlashAttention-2/3 (Tri Dao), speculative-decoding, and RedPajama lineage converts directly into faster, cheaper per-token serving — a real cost edge, not marketing
- Largest open-model catalog (200+ models) with day-0 support for new releases (DeepSeek, Llama, Qwen, Mistral) and OpenAI-compatible API → low switching friction for builders leaving closed APIs
- Full-stack span — serverless inference, dedicated endpoints, fine-tuning, and large GPU-cluster training on one platform (NVIDIA H100/H200/B200/GB200)
- Hypergrowth at scale — ~3x to ~$1B ARR with marquee investors (General Catalyst, NVIDIA, Prosperity7, Salesforce Ventures, Kleiner Perkins, Coatue, Lux)
Weaknesses
- Modest ~45% gross margin — structurally below a pure-software business because a majority of revenue is GPU capacity, much of it rented
- Capital-intensive — must keep raising equity / prepays to fund GPU buildout, diluting and exposing the model to GPU-pricing and financing cycles
- Private-market estimates only — revenue/margin unaudited; concentration in a handful of large GPU-rental customers (CoreWeave-style fragility) is plausible but undisclosed
- Depends almost entirely on NVIDIA silicon and on the continued momentum of open-weight models vs. closed frontier labs
Opportunities
- Enterprise migration from closed APIs to fine-tuned open models for cost/control/privacy — Together is the neutral, multi-model home for that shift
- Sovereign / regional AI clouds and regulated enterprises wanting open models on dedicated, compliant capacity
- Inference is the durable, recurring layer of the AI stack as training spend concentrates in a few labs — Together is inference-weighted
- Owning more of its own data-center capacity (vs. renting) to capture the margin spread, plus reasoning-model demand driving more tokens/inference compute
Threats
- Hyperscalers (AWS Bedrock, Azure AI Foundry, Google Vertex) bundling open-model serving at scale and cutting GPU prices (AWS cut H100 ~45% mid-2025)
- Direct neocloud peers with more ARR/capital — Fireworks (~$800M ARR, $15B talks), Baseten (~$600M ARR, ~$11B), CoreWeave (public) — compressing inference prices
- GPU oversupply / price collapse cycles (H100 spot $8→$1–2→partial recovery) whipsawing rental margin
- Open-model vendors (DeepSeek, Meta, Alibaba) or NVIDIA (DGX Cloud) going direct, disintermediating the middle layer
Moats, dependencies & bottlenecks
Moats
FlashAttention (1/2/3), speculative decoding, custom CUDA kernels (Tri Dao, Chris Ré, Percy Liang lineage) → lower per-token serving cost
Largest open-model catalog (200+) with day-0 support and OpenAI-compatible API → developer default + low switching friction
Full-stack integration (inference + fine-tune + train on one platform) reduces tool sprawl for AI-native customers
Salesforce Ventures, Prosperity7 backing
research-grade home for open models (RedPajama, open datasets)
Dependencies
near-total hardware dependency and allocation priority
DeepSeek, Qwen, Mistral) — Together monetizes others' models
Rumble) and historically rented capacity (CoreWeave, Lambda)
Continued equity / customer-prepay financing to fund GPU buildout
closed frontier APIs
Advantages
- Writes the inference primitives the rest of the industry uses (FlashAttention) — credibility + cost edge
- Model-neutral, multi-vendor — not captive to one frontier lab's roadmap
- Inference-weighted revenue (recurring) vs. lumpy training-only peers
- NVIDIA-aligned: early access to GB200/B300-class capacity
Weaknesses
- ~45% gross margin caps software-like valuation until mix shifts
- Capital intensity + dilution risk from continuous GPU funding
- Unaudited private metrics; concentration risk undisclosed
- Limited lock-in — OpenAI-compatible API cuts both ways (easy to leave too)
Bottlenecks
- GPU supply, allocation, and price volatility gate revenue and margin
- Power and data-center capacity for next-gen liquid-cooled Blackwell racks
- Capital availability for the perpetual capex treadmill
- Engineering talent to sustain the kernel/inference research edge
Top signals & trends
Top signals
Reportedly closed ~Apr 2026 at ~$7.5B pre-money; formal confirmation + any secondary-mark step-up validates the re-rate
Margin expanding above ~45% on more self-operated capacity + API share is the core upside; flat/declining margin is the bear confirm
Backlog/prepay deals de-risk capex and signal enterprise demand; watch for customer concentration
Price competition from above and beside compresses the inference tollbooth
Oversupply/price-collapse cycles directly hit rental margin and the financing model
No filing or announced plan as of mid-2026; a confidential filing or growth-round would mark maturity
Trends
DeepSeek/Llama/Qwen quality closing the gap → more workloads on neutral clouds like Together
Together's own thesis behind the Series B — reasoning ↑ token/GPU demand, not ↓
Demand tailwind, but financing fragility (CoreWeave-style) is the systemic risk
AWS cut H100 ~45% mid-2025; race-to-zero on per-token price pressures the middle layer
H100 spot $8→$1–2→partial recovery→softening — margin whipsaw for rental-heavy players
Prosperity7/DAMAC/SK Telecom backers position Together for non-US sovereign demand
Ecosystem & competitor graph
Suppliers feed the company; customers pull from it. Line thickness shows the strength of each tie (supply-chain dependency, customer earnings contribution). Hover to isolate a tie.
Primary GPU supplier (H100/H200/B200/GB200/B300) + strategic investor
Historical rented-capacity supplier (also a competitor)
Data-center partner co-building the 36K-GB200 NVL72 cluster
GPU server/system OEM supplying the broader neocloud hardware stack
AI server OEM supplying GPU systems into the neocloud ecosystem
$270M multi-year HGX B300 cloud deal (Jun 2026) — flagship named customer
Generative-media AI-native customer running on Together inference
AI video generation startup — inference customer
AI video/creator tooling — inference customer
Strategic investor (Salesforce Ventures) and enterprise-AI ecosystem partner
Public GPU neocloud — larger scale, more capital-intensive, training-heavy; both a competitor and historically a capacity supplier to Together
Public full-stack AI cloud (ex-Yandex) — direct neocloud + managed inference competitor
Closest direct rival — speed-optimized open-model inference (FireAttention), ~$800M ARR, reportedly in $15B talks
Model-deployment/inference platform (Truss), ~$600M ARR, reportedly ~$11B — competes on dedicated inference
Serverless Python compute for inference/training/batch — developer-centric competitor
Hyperscaler serving open models at scale; can bundle + undercut on price
Hyperscaler open-model + OpenAI serving; distribution + pricing threat
Hyperscaler model-garden inference; TPU + open-model serving competitor
Supplier AND potential competitor — could serve open models direct, disintermediating the middle layer