Together AI

TL;DR. Together AI is a San Francisco-based AI infrastructure cloud for building, fine-tuning, and deploying open-source generative models. Founded in 2022, it provides serverless and dedicated inference, GPU clusters, and fine-tuning across 200+ models via an OpenAI-compatible API. It targets ML engineers, AI-native startups, and enterprises running open-weight models at scale.

Overview

Together AI (legally Together Computer Inc.) is a San Francisco-based company providing a full-stack cloud platform for building, training, fine-tuning, and deploying generative AI models, with a focus on open-source model infrastructure. The company positions itself as "the AI Native Cloud," purpose-built for AI engineers and researchers, offering serverless and dedicated inference, GPU compute clusters, fine-tuning, and agentic tooling across over 200 open-source models spanning chat, image, audio, vision, code, and embeddings.

Founding and Leadership. Together AI was founded in June 2022 by Vipul Ved Prakash (CEO), Ce Zhang (CTO), Chris Ré, Percy Liang, and Tri Dao (Chief Scientist). Prakash previously co-founded Tap-root Systems and was an early Twitter engineer. Liang and Ré are Stanford computer science professors; Dao is the creator of FlashAttention. The company is headquartered at 251 Rhode Island Street, San Francisco, CA.

Funding. Together AI has raised approximately $533.5 million in total funding. In March 2024, a Salesforce Ventures-led round valued the company at $1.25 billion. In February 2025, Together AI closed a $305 million Series B led by General Catalyst and co-led by Prosperity7 Ventures, bringing the valuation to $3.3 billion. Key investors include NVIDIA, Kleiner Perkins, Lux Capital, Coatue, and Salesforce Ventures. As of mid-2025, reports indicated the company was seeking to raise up to $1 billion at a potential $7.5 billion valuation.

Scale and Customers. The platform serves over 450,000 developers and enterprises including Salesforce, Zoom, SK Telecom, The Washington Post, and Cognition. Forbes lists approximately 350 employees; PitchBook reports 335. The company has secured 200 MW of power capacity and is deploying NVIDIA Blackwell GPU clusters across multiple North American data centers.

Research and Product. The company's research lab has produced innovations including FlashAttention-3, Mixture of Agents, Medusa, and Sequoia. In 2024, Together AI launched the Together Enterprise Platform, acquired CodeSandbox for built-in code interpretation, and became available on AWS Marketplace.

Key links: Website | About | Blog | Documentation | LinkedIn

Products & Features

Together AI operates a full-stack "AI Native Cloud" platform spanning three product lines: Inference, Fine-Tuning, and GPU Clusters. The company offers 200+ open-source and third-party models across text, image, video, code, and audio modalities through a single OpenAI-compatible API (api.together.xyz/v1).

Inference is offered in four deployment modes. Serverless Inference provides pay-per-token access with no provisioning, supporting models from Meta Llama, Qwen, DeepSeek, Kimi, GLM, MiniMax, Google Gemma, NVIDIA Nemotron, and others. Batch Inference processes up to 30 billion tokens per model asynchronously at up to 50% lower cost. Dedicated Model Inference reserves isolated compute backed by the Together inference engine, and Dedicated Container Inference lets customers run their own engine on managed infrastructure, targeting generative media (video, audio, image) workloads. Together AI claims up to 2.75x faster serverless inference than competing providers, driven by in-house research including the ATLAS speculative decoding system and intelligent quantization.

Fine-Tuning supports both LoRA (lightweight, fast iteration) and full fine-tuning of models up to 100B+ parameters, including DeepSeek-V3 and Qwen3-235B. Users can fine-tune any Hugging Face model without format conversion. Recent upgrades added support for longer contexts (2-4x at no extra cost via the UPipe attention chunking technique), vision model training on raw image data, and tool/function-calling dataset integration. Advanced capabilities include speculative decoding, FP8/NVFP4 quantization, and PyTorch-based reinforcement learning.

Together GPU Clusters provide self-serve access to NVIDIA H100, H200, B200, and GB200 NVL72 hardware, scaling from 8 to 4,000+ GPUs with InfiniBand networking. The Together Kernel Collection, built by Chief Scientist Tri Dao (creator of FlashAttention), delivers up to 90% faster training on B200 versus H100. Clusters support Managed Kubernetes and Slurm orchestration, with pre-built Grafana observability. Storage options include Weka and VAST parallel filesystems with zero egress fees. Together AI is an NVIDIA Cloud Partner operating in 25+ cities across the US, Europe, and Asia.

Pricing is usage-based with no minimum commitment for serverless. Serverless text model pricing ranges from $0.03 per 1M input tokens (LFM2-24B) to $1.74 per 1M input tokens (DeepSeek V4 Pro), with cached input pricing significantly lower. GPU cluster pricing is per-GPU-per-hour: H100 at $5.49/hr on-demand ($3.99/hr reserved), H200 at $6.79/hr ($4.55/hr reserved), and B200 at $9.95/hr ($9.09/hr reserved). Image generation ranges from $0.0006 to $0.134 per image, video from $0.18 to $1.60 per video, and TTS from $10 to $65 per 1M characters. There is no free trial; a minimum $5 credit purchase is required. Platform access is via web console, CLI, SDK, API, and Terraform, with SOC 2 Type II and ISO 27001:2022 compliance.

Security & Compliance

Together AI holds SOC 2 Type 2 certification and complies with HIPAA and GDPR, positioning its platform for regulated-industry deployments including healthcare and life sciences. The SOC 2 Type 2 examination was conducted by an independent auditor over several months, validating access management, data encryption, incident response, and change management controls.

Data handling and privacy. Together AI supports zero data retention (ZDR) by default: it does not store user inputs or outputs. Temporary caching may be used for performance unless explicitly disabled. Data sharing for training other models is opt-in and disabled by default, configurable at both account and organization scope. For passthrough (third-party) models, Together forwards prompts to the upstream provider under that provider's data policy, controllable via a separate organization-level toggle.

Encryption and infrastructure. All data is encrypted in transit and at rest. The platform employs network segmentation, continuous monitoring, and automated threat detection. Third-party models hosted on Together (e.g., DeepSeek, Qwen, Mistral) run on Together's own North American infrastructure; the original model authors receive no user requests or API traffic from these deployments.

Enterprise security features. Together AI enforces multi-factor authentication (MFA) and role-based access controls (RBAC) for access to sensitive environments. The Together Enterprise Platform supports deployment on Together Cloud, within a customer VPC, or on-premise, keeping data within the customer's firewall. Private networking and VPC-based deployments are available for customers with data-residency or regulatory requirements.

Compliance program. Together AI adheres to HIPAA requirements including BAAs, audit logging, and encryption. The dedicated security team conducts regular vulnerability assessments, penetration testing, and code reviews, and maintains an incident response plan. No publicly known security breaches or incidents have been reported as of this writing.

Model safety. Together hosts models at full precision without distillation, forced system prompts, or additional censorship layers, meaning the version called is the version the model author published.

User Feedback & Adoption

Together AI has limited public review volume but strong enterprise adoption, with a 4-star rating on AWS Marketplace (6 reviews, sourced from G2) and a 4.8-star G2 rating across a small number of verified reviews. The platform's review footprint is notably thin for a company of its scale, reflecting its developer-centric, API-first go-to-market model rather than a traditional B2B SaaS review cycle.

Review Platforms and Ratings G2 lists Together AI at 4.8 stars from 4 verified reviews, while the AWS Marketplace page aggregates 6 external reviews (also from G2) at a 4-star average, with 83% rated 4-star or above and one 1-star review. Trustpilot reviews are sparse and polarized, ranging from praise for responsive support and transparency to sharp complaints about prepaid billing practices and lack of pre-purchase performance transparency. No TrustRadius or Capterra presence was identified.

Common Praise Themes Users consistently highlight inference speed (one reviewer reported approximately 400 tokens/second in production), the breadth of open-source model access (200+ models via API), and the elimination of GPU infrastructure management overhead. A G2 reviewer noted the ability to "prototype an idea in the afternoon as opposed to spending a week just setting up the environment." Enterprise customers like Zomato and The Washington Post cite cost savings versus closed-source providers and reliable low-latency performance at scale.

Common Complaint Themes The most prominent complaint is pricing unpredictability: the per-token, per-megapixel, and per-GPU-hour billing structure makes monthly costs difficult to forecast. One G2 reviewer described "bait and switch" tactics after a serverless instance was deprecated without clear pre-purchase notification, followed by difficulty obtaining a refund on a prepaid balance. Other recurring themes include thin documentation in places, a steep learning curve for non-developers, and the platform being unsuitable for teams without ML expertise.

Notable Customers and Case Studies Together AI's customer base includes AI-native companies (Cursor, Decagon, Cartesia, Hedra, ElevenLabs, Dippy AI) and enterprise adopters (Salesforce, Zoom, Zomato, The Washington Post). Zomato's case study is the most detailed: migrating from GPT-4 to Together AI's optimized Llama models yielded a 2x improvement in customer satisfaction, 75% reduction in response times, and scaling to over 1,000 messages per minute. The Washington Post's CTO praised Together AI for delivering "optimized performance at scale, and at a lower cost than closed-source providers." Cursor partnered with Together AI for real-time, low-latency inference at production scale.

Adoption Signals Together AI describes itself as powering "millions of AI-native companies and enterprises." The company was valued at $3.3 billion as of 2025. Its developer-focused GTM strategy, emphasizing free-tier access and frictionless API adoption, has driven organic growth without a traditional enterprise sales motion. The primary barrier to broader adoption remains the technical sophistication required: the platform targets ML engineers and developers, not business users seeking plug-and-play AI solutions.

Monetization & Business Model

Together AI operates on a pure consumption-based model: customers pay per token for serverless API inference, per GPU-hour for dedicated compute, and per unit for image, video, and audio generation. There are no subscription tiers, no setup fees, and no minimum commitments for standard usage.

Serverless inference pricing ranges from $0.03 to $4.50 per million tokens depending on the model, with input and output priced separately. For example, Llama 3.3 70B costs $1.04/1M tokens (input and output), while smaller models like gpt-oss-20B start at $0.05/1M. Dedicated GPU endpoints are billed hourly: HGX H100 clusters range from $1.76 to $2.39 per GPU-hour, HGX H200 from $3.15 to $3.79, and HGX B200 from $4.00 to $5.50, with pricing dependent on commitment length. Fine-tuning is billed per million training tokens, and image generation ranges from $0.002 to $0.134 per image.

As of July 2025, Together AI discontinued its free trial credits (previously $25 for new signups) and now requires a minimum $5 prepaid credit purchase to access the platform. A startup accelerator program offers free credits for qualifying early-stage companies, and an enterprise tier provides dedicated support, compliance features (SOC 2, HIPAA), and data residency. The platform is also available through AWS Marketplace for enterprise procurement.

Together AI has raised approximately $1.33B in total funding. Its $305M Series B (February 2025) valued the company at $3.3B, led by General Catalyst and co-led by Prosperity7. In July 2026, an $800M Series C led by Aramco Ventures raised the valuation to $8.3B. Sacra estimates Together AI reached approximately $1B in annualized revenue by February 2026, up from roughly $618M at the end of 2025, with annual bookings exceeding $1.15B as of Q2 2026. Gross margins are estimated at around 45%, with the larger revenue share (60-70%) coming from GPU server rentals for training and fine-tuning rather than per-token API calls. The company serves over 450,000 developers and enterprise customers including Salesforce, Zoom, and The Washington Post.

Leadership & Team

Together AI was founded in June 2022 by Vipul Ved Prakash, Ce Zhang, Chris Ré, Tri Dao, and Percy Liang. The company is headquartered in San Francisco and operates as Together Computer Inc.

The five founders bring a blend of deep academic research (Stanford, ETH Zurich, Princeton, UChicago) and entrepreneurial operating experience. Prakash is the third company he has founded, after Cloudmark and Topsy Labs.

The company has scaled rapidly, growing from roughly 130 employees at the end of 2024 to approximately 300-385 by mid-2026, according to PitchBook and Tracxn estimates. The leadership team spans engineering, infrastructure, product, revenue, and security functions.

Together AI's stated values emphasize open and responsible development, "doing more with less," and model stewardship for societal benefit, reflecting the founders' open-source research roots.

Target Audience & Use Cases

Together AI serves AI-native startups, enterprises, and individual developers building production applications with open-source models. The platform has grown from roughly 150,000 developers in late 2024 to over 450,000 developers, AI-native companies, and global enterprises by early 2025, and now serves thousands of paying customers with annual bookings exceeding $1.15 billion as of mid-2025.

Primary personas. The core audience comprises ML engineers, AI researchers, and backend developers who need API access to open-source models (Llama, DeepSeek, Mixtral) without managing GPU infrastructure. Enterprise AI/ML platform teams use the Together Enterprise Platform to run inference in their own VPC or on-premise with full data control. AI-native startup founders and engineering teams rely on the platform for cost-efficient, high-throughput inference at scale.

Top use cases. (1) Production inference of open-source LLMs via OpenAI-compatible APIs, with customers like Cursor using Together for real-time, low-latency inference in its AI code editor. (2) Fine-tuning and custom model training on managed GPU clusters (NVIDIA H100, H200, Blackwell). (3) Retrieval-augmented generation (RAG) and customer support automation, with Zomato deploying AI chatbots at food-delivery scale and The Washington Post using the platform for AI-forward publishing. (4) Generative media workloads including text-to-video and image generation, exemplified by Pika Labs and Krea. (5) Cybersecurity and threat detection, with Nexusflow as a named customer.

Industry verticals. Technology and SaaS, media and publishing, food delivery and e-commerce, cybersecurity, and generative media. Named enterprise customers include Salesforce, Zoom, SK Telecom, and Zomato; AI-native customers include Cursor, Pika Labs, Arcee AI, Cartesia, Hedra, Cognition, and Krea.

Company size sweet spot. The platform's self-serve, pay-per-token pricing makes it accessible to individual developers and early-stage startups, while dedicated GPU clusters and the Enterprise Platform (VPC/on-premise deployment, unlimited rate limits, dedicated support) address mid-market and large enterprise needs. The company is headquartered in San Francisco with North American data center infrastructure, though its customer base and investor syndicate (including Aramco Ventures and SK Telecom) reflect a global footprint.

Tags & Categories

Together AI is an AI infrastructure platform providing serverless and dedicated inference, fine-tuning, GPU compute clusters, and code sandboxing for open-source generative AI models. The company positions itself as "The AI Native Cloud," targeting developers and enterprises that want to deploy, fine-tune, and run open-weight models (such as DeepSeek, Llama, Qwen, and Mistral) without managing underlying GPU infrastructure.

Primary Category: AI Infrastructure Categories: AI Infrastructure, AI DevOps Tags: ai infrastructure, llm inference, open-source models, serverless inference, gpu clusters, fine-tuning, openai-compatible api, generative ai, model deployment, developers, enterprise ai, nvidia gpus

The platform supports over 200 open-source models across text, image, video, code, and audio modalities, accessible through a single OpenAI-compatible API. Key products include Serverless Inference, Batch Inference, Dedicated Model Inference, Dedicated Container Inference, Accelerated Compute (Instant GPU Clusters), Code Sandbox, Managed Storage, and Fine-Tuning. Performance is driven by proprietary research innovations including the ATLAS speculative decoding engine, FlashAttention, and Together Kernel Collection. The platform holds SOC 2 Type II and ISO 27001:2022 certifications. Competitors include Fireworks AI, Groq, Baseten, Replicate, OpenRouter, Hugging Face, Anyscale, and Amazon Bedrock. Together AI counts Cursor, Zoom, and Salesforce among its customers.

Impact & Recommendations

Together AI has established itself as the leading production inference platform for open-weight models, with annual bookings exceeding $1.15 billion as of Q2 2026 and a $8.3 billion valuation following its $800M Series C. The company's trajectory places it at the center of the open-source AI shift: open-weight model usage on its platform tripled over the twelve months preceding the round, and Sacra estimates annualized revenue reached approximately $1 billion by February 2026, up from roughly $618M at the end of 2025.

Growth signals. Together AI serves thousands of paying customers including Cursor, Cognition, Decagon, ElevenLabs, and Suno. Customer outcomes validate the economic thesis: Decagon reduced inference costs sixfold after migrating from closed models, with co-founder Ashwin Sreenivas citing costs of one-fifth to one-seventh of proprietary alternatives. The company has secured over 500 MW of compute capacity commitments from investors and is deploying NVIDIA Blackwell GPU clusters in its own data centers across Maryland, Memphis, and Sweden.

Competitive moat. Together AI differentiates through a full-stack, research-driven approach. Its team co-authored FlashAttention (now at version 4 for NVIDIA Blackwell), released the RedPajama dataset (30+ trillion tokens, downloaded over 1.2 million times), and contributed Mixture of Agents, Mamba, FlexGen, and DeepCoder to the open-source community. This research pipeline feeds directly into proprietary optimizations including Together Megakernel, together.compile, ThunderAgent, and ATLAS-2, creating a compounding performance advantage that pure infrastructure resellers cannot match.

Competitive landscape. Among dedicated inference platforms, Fireworks AI is the closest peer on throughput and model catalog breadth, while DeepInfra competes aggressively on price. Replicate serves a different niche (diverse model types, prototyping). GPU cloud providers CoreWeave and Lambda Labs overlap on raw compute rentals but lack Together's inference software stack. Groq differentiates on ultra-low latency via custom silicon but supports a narrower model range. Together AI's combination of kernel-level optimization, fine-tuning tooling, and owned infrastructure gives it a structural advantage over reseller-dependent competitors.

Risk factors. The company faces GPU supply dependency on NVIDIA, margin pressure as hyperscalers (AWS, Google Cloud, Azure) expand their own managed open-model inference offerings, and potential commoditization as inference APIs converge on OpenAI-compatible standards with limited switching costs. Capital-intensive infrastructure ownership (~45% gross margins currently) requires sustained scale to justify.

ICP fit. Together AI is ideally suited for AI-native companies and engineering teams running production workloads on open-weight models at scale, particularly those needing fine-tuning, custom model hosting, or cost optimization beyond what closed-model APIs offer.

Outlook. Strong. The convergence of open-weight model quality parity, escalating closed-model inference costs, and Together's research-to-production flywheel positions the company as a primary beneficiary of the shift toward open AI infrastructure.

Chiri Analysis

Chiri Score: 87/100

Verdict

Best for:

• AI-native startups and engineering teams running production inference on open-weight models at scale

• ML engineers who need OpenAI-compatible API access without managing GPU infrastructure

• Teams fine-tuning large open-source models (up to 100B+ parameters) via LoRA or full training

• Enterprises requiring VPC or on-premise deployment with SOC 2, HIPAA, and zero data retention

• Companies seeking cost reduction versus closed-model APIs (customers report 5-7x lower costs)

• Organizations training on managed NVIDIA H100/H200/B200 GPU clusters

Not for:

• Business users seeking plug-and-play, no-code AI solutions

• Teams without ML or backend engineering expertise

• Buyers needing highly predictable, fixed monthly cost forecasting

• Organizations requiring only closed frontier models (GPT, Claude) rather than open-weight models

• Companies wanting a free trial before committing (minimum $5 prepaid credit required)

Head-to-head

Frequently Asked Questions

Is Together AI SOC 2 compliant?

Yes. Together AI holds SOC 2 Type 2 certification validated by an independent multi-month audit, along with ISO 27001:2022 certification. It also complies with HIPAA, including data encryption, audit logging, and business associate agreements (BAAs).

How much does Together AI cost?

Together AI uses pure consumption-based pricing with no subscription tiers or minimums. Serverless inference ranges from $0.03 to $4.50 per million tokens (input and output priced separately); Llama 3.3 70B costs $1.04/1M tokens. Dedicated GPUs run $1.76-$5.50 per GPU-hour depending on hardware and commitment. A $5 minimum prepaid credit is required.

Who are Together AI's competitors?

Together AI's direct competitors include Fireworks AI, Groq, Baseten, Replicate, OpenRouter, DeepInfra, Anyscale, and Amazon Bedrock. GPU cloud providers CoreWeave and Lambda Labs overlap on raw compute rentals.

Is Together AI good for enterprise?

Yes. Together AI serves enterprises including Salesforce, Zoom, SK Telecom, and The Washington Post, offering VPC and on-premise deployment, dedicated support, SOC 2 and HIPAA compliance, and AWS Marketplace procurement. Its developer-first design suits teams with ML expertise more than non-technical business users.

What models does Together AI support?

Together AI provides access to over 200 open-source and third-party models across text, image, video, code, and audio modalities, including Llama, DeepSeek, Qwen, and Mistral, all through a single OpenAI-compatible API.

Does Together AI store my data?

No. Together AI supports zero data retention (ZDR) by default and does not store inputs or outputs. Data sharing for training other models is opt-in and disabled by default. All data is encrypted in transit and at rest.

Can I fine-tune models on Together AI?

Yes. Together AI supports both LoRA and full fine-tuning of models up to 100B+ parameters, including DeepSeek-V3 and Qwen3-235B. Users can fine-tune any open-source model from Hugging Face Hub without format conversion.

How fast is Together AI's inference?

Together AI claims up to 2.75x faster serverless inference than competing providers, driven by proprietary research including the ATLAS speculative decoding system and the Together Kernel Collection. One G2 reviewer reported approximately 400 tokens per second in production.

---

Reviewed by Chiri Atlas Research Desk (AI Tooling Analyst) on 2026-07-05.