Together AI Revenue and Business Model: How Does Together AI Make Money?

Together AI makes money by renting out AI computing infrastructure and charging businesses to train, fine-tune, and run AI models on it. It is not a foundation model company in the way OpenAI or Anthropic are. It is closer to a specialized cloud provider that has built its entire business around one bet: that open-source AI models, run efficiently on well-optimized GPU infrastructure, can beat closed models on cost without giving up much on quality.

The company has grown fast. Together AI closed an $800 million Series C round on July 1, 2026, at an $8.3 billion post-money valuation, led by Aramco Ventures with participation from Nvidia, Vista Equity Partners, General Catalyst, and others. Sacra, an independent research firm that tracks private company financials, estimates Together AI crossed $1 billion in annualized revenue in February 2026, up from roughly $618 million at the end of 2025.

Those two numbers, funding and revenue, get mixed up constantly in headlines. This article keeps them separate, explains exactly how Together AI earns money, and works through what is confirmed fact versus analyst estimate.

What Is Together AI?

Together AI is a San Francisco company, legally incorporated as Together Computer Inc., founded in June 2022. Its four co-founders came from a mix of research and startup backgrounds: Vipul Ved Prakash (CEO, previously founder of Topsy, which Apple acquired, and Cloudmark), Ce Zhang (CTO, formerly a systems researcher at ETH Zurich), Chris Ré (Stanford professor and co-founder of Snorkel AI), and Percy Liang (director of Stanford's Center for Research on Foundation Models). Tri Dao, known for his work on FlashAttention, later joined as a founder and chief scientist.

The company started with a research-heavy pitch: build a decentralized cloud that pools spare data center compute so AI training does not stay locked inside a handful of hyperscalers. That idea shows up in the company's early seed round, a $20 million raise led by Lux Capital. Over time the business shifted from that research framing toward a more conventional (and more monetizable) product: a managed cloud platform where developers and enterprises can access GPUs, run open-source models through an API, fine-tune those models on their own data, and rent dedicated infrastructure.

Today, Together AI describes itself as an "AI Native Cloud". In practice, it hosts more than 200 open-weight models, including Llama, DeepSeek, Qwen, Mistral, and Kimi, behind a single OpenAI-compatible API and a broader developer platform for building on those models. It also sells raw GPU capacity by the hour and offers fine-tuning services for teams that want to customize a model on proprietary data. For a closer look at how the company got here, we've covered Together AI's founding story and early growth in more depth.

How Does Together AI Make Money?

The simplest way to describe it: customers pay Together AI to access the computing infrastructure and AI capabilities needed to train, fine-tune, deploy, and run open AI models.

A startup building a customer support chatbot might pay per token every time its app calls an open-source model through Together's API. A larger company running a high-volume production workload might instead rent a dedicated GPU cluster and pay by the hour, whether they use every GPU cycle or not. A team that wants a model customized on its own data pays for fine-tuning, priced by the number of tokens processed during training, and then pays again to host that custom model.

According to Sacra's analysis, Together AI's revenue splits across two main lines: per-token API usage, which accounts for roughly 30 to 40 percent of revenue and scales directly with how much inference traffic customers send through the platform, and GPU server rentals, which make up the larger remaining share. Below is a closer look at each product line.

AI Inference

Inference is what happens every time a deployed model actually answers a prompt, generates code, or produces an image. Together AI charges for this on a per-token basis through its serverless API, and pricing varies a lot by model.

Based on pricing data compiled by third-party trackers referencing Together's own pricing pages, such as eesel AI's pricing guide and checkthat.ai's rate tracker, serverless inference generally runs from around $0.03 to as much as $9 per million tokens, depending on the model's size and capability. This token pricing model is what most self-serve model deployment on Together AI looks like in practice: no upfront infrastructure setup, just a per-call rate. A widely used mid-size model like Llama 3.3 70B has been priced at roughly $0.88 per million tokens for both input and output, while a large reasoning model such as DeepSeek-R1 can run several dollars per million tokens. Input and output tokens are often priced separately, since generating a response typically costs more computationally than reading a prompt.

This model works well as a revenue stream because it is naturally recurring. A customer does not buy inference once. Every user interaction with an AI product built on Together's infrastructure generates another small charge, and at scale those charges add up into meaningful monthly revenue. Together also offers a batch inference option at roughly a 50 percent discount for workloads that are not latency sensitive, which helps it win high-volume customers who would otherwise look for a cheaper alternative.

GPU and AI Compute Infrastructure

Underneath the API sits the actual hardware business. Together AI operates and rents Nvidia GPU clusters, ranging from small instant clusters to reserved deployments spanning thousands of interconnected GPUs, for customers who want to train their own models or run their own inference stack rather than use Together's managed API.

This is the "neocloud" side of the business, a term used for companies that specialize in AI-specific cloud infrastructure rather than general-purpose computing like AWS or Azure. On-demand GPU cluster pricing has been reported in the range of roughly $3 to $8 per GPU-hour depending on the chip generation, with H100s at the lower end and newer Blackwell-based B200 chips at the higher end. Reserved commitments of several months bring the hourly rate down, similar to how cloud providers discount long-term compute contracts.

The economics here depend heavily on utilization. An idle GPU still costs money in depreciation, power, and data center overhead, whether or not a customer is running workloads on it. That is why Together AI pushes customers toward reserved and dedicated capacity: it locks in predictable revenue and helps the company plan its own hardware purchases and lease commitments more efficiently.

Model Hosting

Model hosting sits between pure API access and raw GPU rental. Customers can deploy a specific open-weight model, including their own fine-tuned version, onto dedicated infrastructure that only they use, rather than sharing capacity with other customers on the serverless API.

Together AI has published dedicated endpoint pricing by GPU-hour, with rates that have been reported around $4 to $12 per hour depending on the hardware tier, separate from its shared-tenant serverless pricing. This tier tends to appeal to companies running production applications that need consistent latency and cannot tolerate the variability of a shared, multi-tenant endpoint. Revenue here comes from the hourly hosting charge itself, regardless of exact token volume, which makes it a more predictable line item for Together AI than pure serverless usage.

Fine-Tuning

Fine-tuning is how a company adapts an existing open-source model to its own data, tone, or task, instead of training a model from scratch. Together AI monetizes this by charging per token processed during the fine-tuning run.

Reported pricing structures show fine-tuning starting under $0.50 per million training tokens for smaller models, and scaling up for larger base models and more intensive training methods. The company supports several approaches: LoRA (a lightweight method that updates only a small portion of a model's parameters and tends to be cheaper), full-parameter fine-tuning (which touches every weight in the model and costs more), and DPO, or direct preference optimization, a technique used to align model behavior with human preferences.

The total bill for a fine-tuning job is generally a function of training dataset size, the number of training epochs, and any validation runs, multiplied by the per-token rate for that model size and method. Critically, the fine-tuning charge is not the end of the revenue relationship. Once a model is fine-tuned, it typically needs to be hosted somewhere, which triggers a second, ongoing charge through dedicated endpoints. That two-step structure, pay to train, then pay to host, is what makes fine-tuning a meaningful monetization layer on top of straightforward inference.

Enterprise AI Infrastructure

Together AI launched a dedicated enterprise platform in 2024, aimed at companies that need production-grade reliability, security controls, and predictable capacity for LLM inference rather than shared, pay-as-you-go access. Confirmed enterprise customers named in company and press materials include Cursor, whose own growth story we've covered separately, along with Cognition and Decagon. The company has said it serves thousands of paying customers overall, many of whom reach Together's models and AI APIs through the same OpenAI-compatible endpoint used for self-serve accounts.

Enterprise contracts typically bundle several of the products above (dedicated inference, reserved GPU clusters, fine-tuning, and support) into longer-term commitments. This matters for Together AI's business model because it smooths out revenue. A single enterprise customer signing a multi-month or multi-year infrastructure commitment is worth more, and more stable, than the same amount of usage spread across many small self-serve accounts. Security, data residency, and compliance requirements are a big part of why enterprises choose a dedicated deployment over shared serverless access; for a broader look at how organizations are approaching these requirements, see our guide to AI governance. It is worth being precise here: publicly available information confirms the existence of named enterprise customers and an enterprise platform, but Together AI has not disclosed exact enterprise contract values or what share of total revenue enterprise deals represent. Any specific breakdown beyond that is analyst estimation, not confirmed company disclosure.

AI Agents and Production Workloads

This part is more forward-looking than confirmed. AI agents, meaning software that uses a model to plan, call tools, and take multi-step actions with limited human input, tend to generate far more model calls than a simple chatbot. A single agent task, including the kind of agents increasingly used in business intelligence workflows, might involve dozens of reasoning steps, tool calls, and retries, each one hitting an inference endpoint.

Analysts, including Sacra, point to this as a structural tailwind for inference-focused infrastructure providers like Together AI, since more agentic software in production should translate into more token consumption over time. Together AI has also partnered with Meta's PyTorch team to build an open-source reinforcement learning framework aimed at agentic AI training, which signals the company sees this as a growth area worth investing in early. That said, this is analysis of a plausible future demand driver, not a confirmed, quantified revenue stream Together AI has broken out today.

The Revenue Model in Simple Terms

It helps to think of the money flow as a straight line:

Customer → Together AI platform → GPU/AI infrastructure → Model training or inference → Usage-based billing

A customer signs up and picks a product: an API call, a dedicated GPU cluster, or a fine-tuning job. That request runs on Together AI's underlying GPU infrastructure, whether owned, leased, or accessed through a cloud partner. The infrastructure either trains a custom model or serves live inference. Together AI then bills based on actual usage: tokens processed, GPU-hours consumed, or a combination of both for fine-tuned, hosted models.

Revenue only shows up at the last step, and it only grows if usage grows. That is different from a typical software subscription, where a customer might pay the same flat fee whether they use the product once or a thousand times a day.

Together AI Revenue: How Much Does the Company Make?

This is the part where precision matters most, because Together AI is a private company and does not publish audited financial statements. Every revenue figure below is either a third-party estimate or a company-reported bookings figure, not an audited number, and the two are different things.

A few things need explaining here. First, "annualized revenue" is not the same as revenue actually recognized over a full year. It is typically calculated by taking a recent, short period of actual revenue (a month or a quarter) and multiplying it out to a yearly figure. If a company's most recent month was unusually strong, the annualized number can overstate what a full year will actually look like. It is a real and commonly used way to track fast-growing companies, but it is an estimate, not an audited annual revenue statement.

Second, "bookings" is a different concept from revenue entirely. Bookings represent the total value of contracts signed, including work that has not been delivered or billed yet. Together AI's reported $1.15 billion in annual bookings as of its most recent quarter reflects contracted business, not revenue that has already been recognized on the company's books. A large multi-year infrastructure contract can inflate bookings well beyond what shows up as revenue in any single year.

Third, some sources report meaningfully different numbers from Sacra's estimates. One data provider, Latka, cited a $44 million revenue figure tied to 2024, which is far below Sacra's trajectory for the same period. The gap likely comes down to methodology (what counts as revenue, what time window is used, and what data each firm has access to), not a factual contradiction. Readers should treat any single figure as an estimate from a specific source at a specific point in time, not a settled number.

Together AI Revenue Growth

Despite the uncertainty in exact figures, the growth trend across independent sources is consistent and steep.

That is roughly a 10x increase in about 14 months, according to these estimates. Growth at that pace is unusual even by AI industry standards, and it lines up with the story investors were telling when they backed the Series C: open-source model usage across the industry has been climbing quickly, and Together AI has positioned itself as one of the default places to run that traffic.

Together AI Funding and Valuation

Together AI has raised capital across several rounds since 2022. Here is what is publicly documented, with dates and sources.

Together AI's valuation moved from roughly $1.25 billion in March 2024 to $3.3 billion in February 2025 to $8.3 billion in July 2026, a climb of more than 6x in about 28 months, according to the Series C announcement covered by Business Wire. Cumulative funding prior to the Series C has been reported at approximately $1.33 billion by Sacra and by Tracxn; including the $800 million Series C, total funding raised to date is roughly $2.1 billion, though exact cumulative totals vary slightly across trackers depending on what earlier rounds they count.

It is worth stating plainly, because it is easy to conflate: funding is not revenue, and valuation is not revenue. Funding is cash investors put into the company in exchange for equity. Valuation is investors' collective bet on what the company could be worth in the future, based on growth rate, market size, and comparable companies, not a measurement of current sales. A company can raise a huge round and post modest revenue, or generate strong revenue and still struggle to raise money if investor sentiment shifts. Together AI's $8.3 billion valuation reflects what a specific group of investors were willing to pay for a stake in the company on July 1, 2026. It does not tell you, by itself, what the company actually earned that year.

As part of the Series C, Together AI also secured commitments for more than 500 megawatts of additional compute capacity, financed separately by new investors rather than out of the equity round itself (TechTimes, Sacra, July 2026). The company has said it expects its infrastructure footprint to grow roughly 50-fold over five years from that base, which signals just how capital-intensive this business is expected to remain.

How Together AI's Business Model Works

Together AI's business model is usage-based infrastructure, not subscription software. Every dollar of revenue traces back to someone actually consuming compute: tokens processed, GPU-hours rented, or training runs completed.

That has real consequences for how the company grows and how risky its revenue is. A traditional SaaS company with annual contracts has revenue that is fairly predictable months in advance. Together AI's revenue moves with usage, which can grow explosively when demand for open-source AI is strong, but can also soften quickly if a large customer cuts back workloads or shifts traffic to a competitor. That is one reason enterprise contracts and reserved capacity commitments matter so much to the company's strategy: they convert some of that variable, usage-based revenue into something closer to guaranteed, contracted revenue.

The other structural feature worth understanding is that Together AI sits between hardware suppliers and end customers. It buys or leases GPU capacity (increasingly from partners like IBM Cloud, discussed below, in addition to its own data center footprint), pays for power, networking, and engineering to keep that hardware running efficiently, and then resells access to that capacity, packaged with software, at a markup. The company's profitability depends on the spread between what it costs to run that infrastructure and what customers are willing to pay for access to it.

The IBM Partnership: Revenue, Infrastructure Expansion, or Both?

On August 11, 2026, IBM and Together AI announced a multi-year, $240 million agreement under which IBM will deploy a large-scale Nvidia HGX B300 GPU cluster on IBM Cloud, expected to be available in the first quarter of 2027. Reuters covered the deal the same day. It is worth reading the structure of this deal carefully, because headlines can make it sound like a $240 million revenue win for Together AI when the actual arrangement is closer to the opposite.

Based on the companies' own statements, IBM is the one deploying and operating the infrastructure, using Nvidia hardware and Spectrum-X networking, on IBM Cloud. Together AI is the party that will use that cluster to serve inference for open-source models. In other words, this looks like a compute supply agreement in which Together AI is expanding where it can access GPU capacity, with IBM acting as an infrastructure partner and cloud host, not as a customer paying Together AI $240 million for services rendered.

That distinction matters for anyone trying to model Together AI's revenue. The deal should be read as:

• Infrastructure expansion: it adds a new, large source of GPU capacity for Together AI to sell against, supporting future revenue growth rather than representing revenue itself.
• Strategic partnership: it gives Together AI access to IBM's enterprise sales relationships and Red Hat's open-source enterprise base, which could help win the kind of large, sticky enterprise contracts the company has been chasing.
• Not confirmed current revenue for Together AI: nothing in the public announcements indicates IBM is paying Together AI $240 million, or any specific sum, for services. If anything, Together AI (or its investors) appear to be the ones financing or committing to use this new capacity.

Analysts covering the deal noted it "provides revenue visibility through 2027" for Together AI in the sense that having guaranteed access to a large, purpose-built inference cluster removes a capacity bottleneck that could otherwise cap growth, but that is a supply-side benefit, not a booked sale. Readers should treat the $240 million figure as a capital and infrastructure commitment tied to the partnership, not as a line item to add to Together AI's revenue.

Why Together AI's Revenue Could Grow

Several factors, some already playing out and some more speculative, support continued revenue growth:

1. Rising demand for open-source models. Together AI and other sources point to usage of open-weight models roughly tripling industry-wide over the past year, driven by cost-conscious developers and enterprises, citing usage data from AI gateway OpenRouter (TechCrunch, July 2026).
2. Inference demand from AI agents. As discussed above, agentic software tends to generate far more model calls per task than simple chat interfaces, which could meaningfully increase token volume over time. This is a forward-looking driver, not yet a quantified revenue line.
3. Enterprise adoption. Named customers like Cursor, Cognition, and Decagon, plus the IBM partnership's enterprise sales reach, point toward a growing enterprise pipeline. Enterprise deals tend to be larger and stickier than self-serve usage.
4. Cost-driven migration away from proprietary models. Some reporting has cited enterprise customers cutting inference costs significantly by switching workloads to open models on platforms like Together AI, though exact savings figures vary by workload and should be treated as illustrative rather than universal.
5. Geographic expansion. Together AI brought infrastructure online in Sweden in September 2025 to reduce latency and support data residency requirements for European customers, which opens up enterprise deals in regions with strict data rules.

Items 1, 3, and 5 above are already happening. Item 2 is a reasonable and widely discussed hypothesis, not a confirmed current revenue driver. Item 4 depends heavily on the specific workload and model being compared.

What Could Hurt Together AI's Revenue?

The same forces that make this business attractive also make it fragile in specific ways.

• GPU costs and depreciation. Nvidia hardware is expensive, depreciates over a few years, and needs constant reinvestment to stay competitive as new chip generations (like Blackwell) come out.
• Heavy dependence on Nvidia. Together AI's entire infrastructure stack, and much of the AI industry's, runs on Nvidia GPUs. Supply constraints, pricing changes, or a shift toward custom silicon from hyperscalers could squeeze margins.
• Price competition and inference "price wars." Inference pricing across the industry has trended downward as competitors like Fireworks AI, Baseten, Groq, and hyperscaler-native options (AWS Bedrock, Google Vertex AI, Azure OpenAI) compete for the same workloads. Lower prices can be good for adoption but bad for margins if not offset by higher volume.
• Model commoditization. As more labs release capable open-weight models, the specific model a customer chooses matters less, which shifts competition toward whoever can serve any given model fastest and cheapest. That is a tough, low-margin game to win consistently.
• Hyperscaler competition. AWS, Google Cloud, and Microsoft Azure all offer their own managed access to open models, and they have far larger balance sheets to subsidize aggressive pricing if they choose to compete directly for this segment.
• Capital intensity. Building and expanding data center capacity, even through partners like IBM, requires enormous ongoing capital. If growth in paying usage does not keep pace with capacity expansion, utilization (and margins) can suffer.
• Customer concentration. Public reporting has not disclosed what share of Together AI's revenue comes from its largest customers. If a small number of large accounts represent an outsized share of revenue, losing even one could meaningfully affect the business. This is a general risk worth naming for any infrastructure company at this stage, not a confirmed fact specific to Together AI.

None of these risks are unique to Together AI. They apply broadly across the AI neocloud category, and different companies in that category are exposed to them to different degrees based on their contract mix, hardware ownership versus leasing, and customer base. For a wider view of the capital and demand cycle driving this whole sector, see our breakdown of the AI infrastructure super cycle and its risks.

Is Together AI Profitable?

There is no publicly confirmed answer to this question, and any claim that Together AI is definitively profitable or unprofitable should be treated with skepticism, because the company has not released audited financial statements.

What can be said with more confidence, based on how similar businesses in this category operate, is that infrastructure-heavy AI companies typically carry thinner gross margins than pure software businesses. For comparison, Fireworks AI, a competitor in the same inference infrastructure space, has reported gross margins around 50 percent, below the 70-plus percent margins common in subscription software, largely because GPU infrastructure costs sit directly in the cost of revenue (Sacra). Fireworks has said it is targeting 60 percent gross margins through better GPU utilization.

If Together AI's cost structure looks similar, which is a reasonable but unconfirmed assumption given the comparable business model, its path to sustainable profitability likely depends on the same levers: keeping GPU utilization high, negotiating favorable hardware pricing and depreciation schedules, shifting mix toward higher-margin dedicated and enterprise contracts, and achieving enough scale that fixed infrastructure and engineering costs are spread across a larger revenue base.

Revenue growth by itself does not equal profitability. A company can 10x its revenue while still losing money on every dollar of infrastructure it deploys, if utilization is low or pricing is too aggressive. Investors backing the Series C were clearly betting on Together AI's growth trajectory and market position, not on a demonstrated, audited profit history, since that kind of history simply is not public for this company at this stage.

Together AI vs OpenAI and Other AI Infrastructure Companies

Together AI's economics look fundamentally different from a company like OpenAI or Anthropic, even though all three sell access to AI models.

OpenAI and Anthropic monetize their own proprietary models directly. Their entire value proposition is that their model is better than anyone else's, and they price accordingly, often at a premium. Together AI does not build a flagship proprietary model to sell (though it does run an internal research lab and has released open research). Instead, it monetizes the infrastructure and tooling layer that sits underneath whichever model a customer wants to use, whether that is Llama, DeepSeek, Qwen, or something else.

That makes Together AI's economics closer to a hybrid of AI infrastructure, developer platform, and cloud services than to a conventional AI subscription business. It is more comparable, structurally, to CoreWeave (a public neocloud that rents Nvidia GPU capacity) or to Fireworks AI and Baseten (inference specialists), than to OpenAI's ChatGPT business. Compared to AWS or Google Cloud, Together AI is far smaller and more specialized, focused specifically on AI workloads rather than general-purpose computing, which lets it optimize more narrowly for inference performance but also leaves it more exposed if the hyperscalers decide to compete aggressively on price in this specific niche.

The Future of Together AI's Business Model

Together AI's next phase looks like it will be defined by scale and enterprise depth rather than a change in the underlying model. The company has committed to expanding compute capacity roughly 50-fold over five years, added a large new infrastructure partner in IBM, and is pushing further into agentic AI infrastructure through its work with Meta's PyTorch team on reinforcement learning tooling.

The core question for the next few years is whether usage growth can keep pace with that infrastructure buildout. Revenue in this business model is a direct function of how much compute customers actually consume. If open-source AI adoption and agentic workloads continue expanding as they have over the past year, Together AI has positioned itself to capture a meaningful share of that demand. If growth slows, or if hyperscalers and well-funded competitors compress pricing faster than Together AI can grow volume, the company's aggressive infrastructure buildout could become a drag on margins rather than a growth engine. Both outcomes are plausible from where things stand in August 2026, and neither is guaranteed.

Frequently Asked Questions

1. How does Together AI make money? Together AI makes money by charging customers to access AI computing infrastructure. Revenue comes from per-token charges for AI inference, hourly charges for renting GPU clusters and dedicated hosting, and per-token charges for fine-tuning models on customer data.

2. What is Together AI's revenue? Together AI has not publicly disclosed audited revenue. Independent research firm Sacra estimates the company reached approximately $1 billion in annualized revenue in February 2026, up from about $618 million at the end of 2025. These are third-party estimates, not company-reported audited figures.

3. What is Together AI's revenue in 2026? As of the most recent available estimate, Sacra placed Together AI's annualized revenue at roughly $1 billion in February 2026. Separately, the company has said its annual bookings, meaning contracted business, topped $1.15 billion as of the second quarter of 2026. Bookings and revenue are not the same measurement.

4. What is Together AI's business model? Together AI operates a usage-based AI infrastructure business. Customers pay for what they consume: tokens processed through inference, GPU-hours for dedicated or clustered compute, and tokens processed during fine-tuning. It does not primarily sell subscriptions or its own proprietary model.

5. What is Together AI's valuation? Together AI was valued at $8.3 billion following its $800 million Series C funding round, which closed on July 1, 2026, led by Aramco Ventures. That is up from a $3.3 billion valuation in its February 2025 Series B round.

6. How much funding has Together AI raised? Together AI has raised funding across a $20 million seed round, a $102.5 million Series A in 2023, a $305 million Series B in February 2025, and an $800 million Series C in July 2026. Cumulative funding is approximately $2.1 billion, based on figures reported by Sacra and Tracxn.

7. Is Together AI profitable? There is no public confirmation that Together AI is profitable or unprofitable. As a private company, it has not released audited financial statements. Infrastructure-heavy AI companies typically carry thinner gross margins than software businesses due to GPU and data center costs.

8. Who are Together AI's customers? Together AI's publicly named customers include Cursor, Cognition, and Decagon. The company has said it serves thousands of paying customers overall and, according to a company-cited figure from August 2026, more than one million developers use its platform in some capacity.

9. Does Together AI make money from AI inference? Yes. Inference, meaning the process of running a live AI model to generate a response, is one of Together AI's two primary revenue lines. According to Sacra, per-token API usage accounts for roughly 30 to 40 percent of the company's revenue.

10. Does Together AI rent GPUs? Yes. Together AI rents Nvidia GPU capacity by the hour, both as on-demand clusters and as longer-term reserved commitments, for customers who want to run their own training or inference workloads rather than use Together's managed API.

11. How does Together AI compare with OpenAI? Together AI hosts and serves open-source models built by other labs and charges for infrastructure access. OpenAI builds and sells access to its own proprietary models. OpenAI's estimated annualized revenue (around $25 billion as of February 2026, per Sacra) is far larger than Together AI's, reflecting the premium pricing and scale of proprietary frontier models compared to open-model infrastructure services.

12. Is Together AI a cloud provider? Together AI operates as a specialized AI cloud provider, sometimes called an AI neocloud. It focuses specifically on GPU infrastructure, model inference, and fine-tuning, rather than offering the broad range of general-purpose cloud services that AWS, Google Cloud, or Microsoft Azure provide.

13. What does Together AI charge for fine-tuning? Fine-tuning is billed per million tokens processed during training, with reported starting rates under $0.50 per million tokens for smaller models. Pricing scales up based on model size and method, such as LoRA versus full-parameter fine-tuning, and hosting the resulting fine-tuned model incurs a separate, ongoing hourly charge.

14. Why is Together AI's business model important? Together AI's model matters because it represents an alternative economic path in AI: infrastructure and tooling for open models, rather than a proprietary model sold directly to consumers or businesses. Its growth is often used as a signal for how quickly enterprises are adopting open-source AI as a lower-cost alternative to closed frontier models.

15. What is the future of Together AI's revenue? Together AI's future revenue will likely depend on continued growth in inference demand, expanding enterprise adoption, GPU utilization rates, and how AI agent workloads develop, since agents tend to generate more model calls than simple chat applications. These are current growth drivers and reasonable expectations, not guaranteed outcomes.

Final Verdict

Together AI makes money primarily by providing infrastructure and services that businesses and developers use to train, fine-tune, deploy, and run open-source AI models. Its revenue comes from usage: tokens processed through inference, GPU-hours rented for dedicated or clustered compute, and tokens processed during fine-tuning jobs, not from selling a proprietary model or a flat subscription fee.

The company's growth numbers, an estimated jump from roughly $100 million to $1 billion in annualized revenue in about 14 months, and a valuation climb from $1.25 billion to $8.3 billion in under two and a half years, are genuinely striking, even accounting for the fact that most of these figures are estimates rather than audited disclosures. But growth alone does not settle the harder question of whether this business model produces durable profits.

That answer will depend on things that are still playing out: how much GPU utilization Together AI can sustain as it expands capacity 50-fold, how much pricing pressure it faces from hyperscalers and other inference specialists, how sticky its enterprise contracts prove to be, and whether AI agents really do drive the kind of inference demand analysts expect. Together AI has built a real, fast-growing business on a clear thesis: that open-source AI, run well, can be cheaper than closed AI without giving up much quality. Whether that thesis translates into a profitable, long-term infrastructure company is still an open question, and one that will likely be answered over the next few years of GPU capacity, customer contracts, and pricing decisions, not by the size of its next funding round.