Together AI's valuation reached $8.3 billion in July 2026, up from $20 million raised four years earlier by a team of academic researchers who thought enterprises would rather rent GPUs by the hour than build their own data centers. They were right often enough to raise five rounds in four years.
Together AI is a "neocloud" — infrastructure that sits between raw Nvidia chips and the companies that want to train or run AI models without owning the hardware. This deep dive covers what the company actually sells, its full funding history, the agentic compiler it says is already optimizing production inference, its push into Saudi Arabia, the competitive field, and what is and isn't known about an eventual IPO.
Company Snapshot
Figures from Together AI's own funding announcements (via TechCrunch, SiliconANGLE, PYMNTS, Yahoo Finance) and Sacra research, as of November 2026. Together AI is private and does not publish audited financials.
Founders
Vipul Ved Prakash (CEO), Ce Zhang (CTO), Percy Liang
HQ
San Francisco, CA
Chief Scientist
Tri Dao, creator of FlashAttention
Headcount
Not disclosed; third-party trackers estimate roughly 250-430 in 2026
Key investors
Aramco Ventures, Vista Equity Partners, General Catalyst, Kleiner Perkins, Lux Capital, Nvidia
Named customers
Cursor, Salesforce, Decagon, Zoom, Datadog, ElevenLabs, IBM, Google DeepMind
What Together AI Actually Does
Together AI was founded in 2022 in San Francisco by Vipul Ved Prakash, Ce Zhang, and Stanford professor Percy Liang, with Tri Dao joining as chief scientist in July 2023. The pitch: companies that want to train or run open-source AI models shouldn't have to buy and operate their own Nvidia GPU clusters. Together AI rents that compute instead, layering optimized inference software on top so customers get both the hardware and the engineering needed to run models efficiently.
The company doesn't build its own models as a primary business. It hosts other organizations' open-weight models — Meta's Llama, DeepSeek, Alibaba's Qwen, Moonshot AI's Kimi, Zhipu's GLM, and others — and sells access to them, betting that enterprises increasingly want the cost and control benefits of open weights without the complexity of self-hosting. Usage of open-source models across the industry roughly tripled over the year leading into its Series C, Together AI said, citing OpenRouter data.
How Together AI Makes Money
GPU cluster rentals (billed hourly)
On-demand Nvidia GPU access priced per hour: $3.99 for an H100 up to $9.99 for a B300, per together.ai/pricing. Reserved 91-180 day commitments cut the H100 rate by about 20%. This is the larger of Together AI's two core revenue lines, per Sacra's estimated revenue mix.
Serverless per-token inference
Customers call a hosted open model and pay per million tokens processed, from $0.14 for a small model like Llama 3 8B Instruct Lite up to $15.00 for a larger model like Kimi K3 — the product most comparable to calling the OpenAI or Anthropic APIs directly.
Dedicated Inference
A single GPU reserved for one customer's model, running $5.49/hour for an H100 and $8.99/hour for a B200 — a middle tier between shared serverless inference and running a full self-managed cluster.
Together Post-Training (launched Oct 19, 2026)
A managed service for customizing open models with supervised fine-tuning or reinforcement learning (GRPO, DPO), billed per GPU-minute at an introductory $9/hour for B200 usage. Early customers cited include Deel and Collinear AI.
Every revenue line is usage-metered rather than a flat subscription or per-seat license. That is the core mechanic of the neocloud category: a customer's bill scales with compute consumed, which is why revenue has tracked the broader AI compute buildout as closely as it has. For the full pricing tables and how that mix breaks down, see our dedicated breakdown of how Together AI makes money.
Funding History
Together AI has raised roughly $1.33 billion across five disclosed rounds in a little over three years — and the Saudi Aramco connection runs deeper than the Series C headline suggests: Prosperity7 Ventures, an Aramco-linked fund, co-led the Series B a year and a half before Aramco Ventures itself led the Series C outright.
| Round | Date | Amount | Lead Investor(s) | Valuation |
|---|---|---|---|---|
| Seed | May 15, 2023 | $20M | Lux Capital | Undisclosed |
| Series A | Nov 29, 2023 | $102.5M | Kleiner Perkins | Undisclosed |
| Series A ext. | Mar 13, 2024 | $106M | Salesforce Ventures | $1.25B |
| Series B | Feb 20, 2025 | $305M | General Catalyst, Prosperity7 | $3.3B |
| Series C | Jul 1, 2026 | $800M | Aramco Ventures | $8.3B |
Sources: TechCrunch (seed, Series A), SiliconANGLE (Series A extension), PYMNTS (Series B), and Yahoo Finance (Series C), compiled November 2026.
Together AI: Valuation Milestones, 2023-2026
Valuation climbed from $1.25B at the March 2024 extension to $8.3B at the Series C 28 months later — a 6.6x increase, with the biggest single jump (2.5x) coming in the last 16 months
SiliconANGLE, PYMNTS, and Yahoo Finance reporting on Together AI's funding rounds, compiled November 2026
Data
| Item | Value |
|---|---|
| Seed ('23) | $20M raised |
| Series A ('23) | $102.5M raised |
| Series A ext. ('24) | $1.25B valuation |
| Series B ('25) | $3.3B valuation |
| Series C ('26) | $8.3B valuation |
Product Portfolio
GPU Clusters
On-demand and reserved Nvidia GPU access (H100, H200, B200, B300) for training, fine-tuning, and dedicated serving, billed hourly.
Inference API
Serverless, per-token access to dozens of open-weight models, plus a Batch API priced 50% below real-time rates for non-latency-sensitive workloads.
Together Post-Training
Managed reinforcement-learning and fine-tuning infrastructure, launched October 19, 2026, with dedicated GPU allocations and multi-LoRA adapter serving.
Together Compiler
An agentic system, launched November 4, 2026, that continually rewrites and tunes inference-engine configurations for whatever model and traffic pattern a customer is running (detailed below).
Deep Research & agents tooling
APIs and reference stacks for retrieval-augmented generation and multi-step agent workflows built on Together-hosted open models.
Humain data center (Saudi Arabia)
A 250-megawatt facility under construction with HUMAIN, expected to roughly triple Together AI's existing data center capacity once online.
The Together Compiler: Letting Agents Rewrite the GPU Code
Most of the coverage Together AI gets is about valuation. The more interesting engineering story is what the company shipped on November 4, 2026: Together Compiler, a system that uses AI agents to continually rewrite and test the low-level configuration of its inference engines, rather than hand-tuning them once and leaving them static.
Together AI describes it, in its own launch post, as the first "truly continual" optimization loop for inference: multiple agents propose candidate optimizations in parallel, a persistent test engine evaluates each one in seconds instead of restarting from scratch (cutting a 467-second startup down to about 9 seconds), and a human still has to approve any change before it deploys. The company reports up to 2.6x higher throughput across frontier open models (Kimi K3, GLM 5.3, DeepSeek V4.1-Flash) achieved in under 24 hours of agent search, and says the system set a state-of-the-art mark on SemiAnalysis's InferenceX low-latency benchmark.
Together AI says the Compiler is "serving trillions of tokens today" in production, not just in benchmarks — a claim that fits the company's broader scale: at the time of its August 2026 IBM partnership (below), Together AI said its platform was processing roughly 400 trillion tokens a month. According to the company, if the throughput gains hold up across more of that volume, it is a direct lever on gross margin in a business where GPU time is the single largest cost.
Revenue and Key Metrics
End-2025 ARR
~$618M
Sacra estimate
Feb 2026 ARR
~$1B
Sacra estimate
Q2 2026 Bookings
$1.15B
company-disclosed
Monthly tokens served
~400T
as of Aug 2026
Together AI does not publish audited financials. Its own disclosed figure is annual bookings, not revenue: the company said bookings crossed $1.15 billion in the quarter before its July 2026 Series C announcement. Sacra, a research firm that tracks private-company financials, separately estimates annualized revenue reached roughly $1 billion by February 2026, up 375.4% year over year from about $618 million at the end of 2025, with a gross margin around 45% — consistent with a business still paying third parties for a meaningful share of its GPU capacity.
The company's largest disclosed enterprise contract is a IBM deal signed August 11, 2026: $240 million over multiple years to build a dedicated inference cluster on IBM Cloud running Nvidia HGX B300 systems, targeted for availability in the first quarter of 2027. Separately, the Nvidia-powered Humain partnership in Saudi Arabia, announced August 31, 2026, projects more than $5 billion in first-year gross annualized revenue once its 250-megawatt facility is built — a target for a data center that does not yet exist, not booked revenue.
Competitive Landscape
Together AI's closest head-to-head rivals are Fireworks AI and Baseten — both sell the same usage-metered, GPU-time-or-per-token access to open and custom models. CoreWeave competes one layer down, as a raw GPU cloud rather than a model-serving API, though the lines are blurring as all four add adjacent products.
| Attribute | Together AI | Fireworks AI | Baseten | CoreWeave |
|---|---|---|---|---|
| Founded | 2022 | 2022 | 2019 | 2017 (as Atlantic Crypto) |
| Last priced valuation | $8.3B (Jul 2026) | $17.5B (Jul 2026) | $13B (Jun 2026) | Public company |
| Total funding/capital raised | ~$1.33B | ~$2.25B | ~$2.3B (per Series F) | N/A (public) |
| Est. annualized revenue | ~$1B (Sacra) | >$1B (reported) | ~$600M (reported) | $1.9B+ (2025 public revenue) |
| Primary pricing model | Per GPU-hour + per token | Per token + per GPU-hour | Per GPU-minute | Per GPU-hour (raw compute) |
| Headline customer concentration risk | Not publicly disclosed | Not publicly disclosed | Not publicly disclosed | Microsoft ~62% of 2024 revenue |
How we counted: each company's own funding announcement or most recent reported figure as of November 2026; CoreWeave figures are from its public 2025 financial filings. Revenue figures for the three private companies are third-party estimates (Sacra and other research firms), not audited disclosures.
Neocloud Valuations: Together AI vs Fireworks AI vs Baseten (2026)
Each company's funding announcement, compiled November 2026
Baseten has since drawn reports of talks at higher, unpriced marks ($19-30B); those have not closed as of November 2026 and are excluded here.
Leadership Team
Vipul Ved Prakash — Co-Founder & CEO
Previously founded Topsy, a social-media search and analytics company Apple acquired in 2013 for a reported $200 million-plus. Prakash is Together AI's public face on funding and partnership announcements, including the July 2026 Series C and the August 2026 Humain deal.
Ce Zhang — Co-Founder & CTO
An associate professor at ETH Zürich and the University of Chicago before co-founding Together AI, Zhang leads the company's systems and infrastructure engineering.
Percy Liang — Co-Founder
A Stanford computer science professor known for foundation-model research, Liang brings the academic research pedigree that shapes Together AI's open-source positioning and its research publications.
Tri Dao — Chief Scientist
Joined as chief scientist in July 2023 alongside the release of FlashAttention-2. Dao, also an assistant professor at Princeton, is the lead author of the original FlashAttention algorithm — the memory-efficient attention mechanism now used across most modern AI inference stacks, including later versions released through Together AI's research arm.
Bull Case / Bear Case
The Bull Case
- +Valuation has climbed 6.6x in 28 months, tracking disclosed bookings growth rather than a speculative market-wide re-rating.
- +Open-weight model usage is reported to have roughly tripled industry-wide in the year before the Series C — exactly the trend Together AI's business is built to capture.
- +The Together Compiler's agentic kernel optimization is a real technical moat if the throughput gains generalize: a direct lever on gross margin in a GPU-time-constrained business.
- +The Humain partnership diversifies compute supply geographically and brings an energy-linked sovereign investor (Aramco) deeper into the cap table across two consecutive rounds.
- +A roster of enterprise logos (Cursor, Salesforce, IBM, Google DeepMind) signals the business has moved beyond developer-tool early adopters into large commercial contracts.
The Bear Case
- -The headline $1.15B figure is bookings, not revenue — it can include multi-year commitments not yet recognized, and no audited revenue number has ever been published.
- -Together AI is priced below both Fireworks AI ($17.5B) and Baseten ($13B) on a last-round basis, despite comparable or larger disclosed top-line figures — the market may be pricing in a weaker moat.
- -The $5 billion Humain revenue figure is a first-year projection for a data center that doesn't exist yet, not a contracted or booked number.
- -GPU-rental and per-token economics are commodity-adjacent: pricing is public and comparable across Together AI, Fireworks, Baseten, and the hyperscalers, which caps pricing power if any single player's compute costs rise relative to peers.
- -Headcount, margins, and customer concentration are not disclosed by the company, which makes independent verification of the Sacra revenue estimate and gross margin figure impossible from outside sources.
Where We Could Be Wrong
The ~$1 billion ARR estimate and the 45% gross margin figure both come from Sacra, a third-party research firm, not from Together AI itself, and have not been independently verified.
Bookings and revenue are not the same thing — treating the $1.15 billion figure as realized revenue would overstate the business's current scale.
The Together Compiler's reported speedups (up to 2.6x) are company-published benchmark results; independent, third-party verification of those exact figures was not found.
IPO Outlook
Together AI's IPO timeline is not publicly disclosed. As of November 9, 2026, no S-1 has been filed with the SEC, no underwriter or bank mandate has been reported, and Prakash has made no on-record comment about a target listing date. The closest thing to a market signal is Vista Equity Partners' participation in the Series C — a firm known for taking pre-IPO growth positions in other companies — but that is an observer's inference about the cap table, not a statement from Together AI.
Five funding rounds in roughly three and a half years, each closing faster and larger than the last, is itself a data point: it suggests a company still finding private capital abundant enough that it hasn't needed to test public markets. Peers in the same category — Databricks and CoreWeave among them — offer the closest comparable timelines for when a well-funded AI infrastructure company eventually lists, but neither path maps directly onto Together AI's own, undisclosed plans.
Bottom line: Together AI turned a $20 million seed round into an $8.3 billion business in just over three years by renting GPUs and metering inference for a market betting on open-weight models over closed ones. The Together Compiler and the Humain partnership are real bets on defensibility — one technical, one geographic — in a category where pricing is otherwise commodity-adjacent. The $1.15 billion bookings figure and the Sacra revenue estimate both point to real growth, but neither is an audited number, and the company trades at a lower multiple-implied valuation than its two closest rivals despite comparable scale. No IPO signal exists yet; that is itself the most honest thing that can be said about the timeline.
All financial figures are based on publicly reported data, company announcements, and credible media reporting as of November 2026. Together AI is a private company and does not publicly disclose audited financials or headcount; the Sacra revenue and gross-margin estimates have not been independently confirmed by the company.
Latest from the Pulse
Get VC data most people never see
— free to subscribe
Trace's notes on venture, AI and startups, a few times a week. Join 5,000+ subscribers. No spam.
No spam, unsubscribe anytime. Sent via Substack · Privacy · or subscribe on Substack ↗