Signals Inbox·July 19, 2026·AI Infrastructure

Is Together AI really worth $8.3B today?

Together AI probably is worth $8.3 billion today. Its estimated revenue already supports the price, but investors are betting that a business still dominated by GPU rentals can become a higher-margin software platform.

Q1What happened to Together AI’s valuation?

Together AI has jumped from a $3.3 billion valuation to $8.3 billion in 16 months, a huge move even by current AI standards. On July 1, the company announced an $800 million Series C led by Aramco Ventures, with Vista Equity Partners, General Catalyst, NVIDIA, Emergence Capital, March Capital, Pegatron and others joining the round.

The pace is the striking part. Together AI was valued at $1.25 billion after a $106 million round in March 2024, then reached $3.3 billion in February 2025. Its valuation has multiplied by 6.6 in roughly 28 months. The company was founded in 2022, so it reached the latest mark after only four years.

In simple terms, the new investors paid $800 million for around 9.6% of the post-money company. Their bet is that Together AI can own a valuable layer between open AI models, advanced chips and the applications consuming billions of model calls. The price only works if the company has already reached serious commercial scale.

Together AI’s valuation growth

Q2How much revenue is Together AI actually making now?

Together AI appears to be near $1 billion in annualized revenue, which makes the valuation far less crazy than the headline suggests. Sacra’s latest estimate reached that level in February, up from about $618 million at the end of 2025. Together AI itself has only disclosed annual bookings above $1 billion.

The difference between revenue and bookings is crucial. Revenue measures what the company has earned or is currently earning. Bookings can include spending committed for future periods, especially when customers reserve GPU capacity. We therefore use the Sacra estimate as the main working figure and keep a wider range around it.

Using Sacra’s estimate, Together AI is valued at 8.3 times revenue. If only 70% of bookings converts into annual revenue, the multiple reaches 10.3 times. A 50% conversion rate pushes it to 14.4 times. Around eight times looks sensible for this growth rate. Above fourteen times, the price starts looking stretched.

Together AI’s implied revenue multiples

Q3Are Together AI’s bookings as solid as they sound?

Together AI’s bookings show real demand, but they leave several important financial questions unanswered. The company has chosen to publish bookings instead of recognized revenue, ARR, net revenue retention or the amount of revenue locked in for the next twelve months.

The same $1.15 billion headline can support a multiple anywhere from 7.2 times to more than 14 times. The missing details are contract length, cancellation rights, utilization commitments and customer concentration. A one-year commitment from hundreds of recurring customers is worth much more than a handful of multiyear GPU reservations from cash-burning AI startups.

We should still take the figure seriously. Large AI customers often reserve capacity because they cannot risk running short of GPUs during a product launch. Bookings can give Together AI useful visibility. We just cannot tell how strong that visibility is yet.

Q4Is Together AI expensive compared with public cloud companies?

Together AI looks expensive beside CoreWeave and cheap beside Snowflake, which is roughly where this hybrid business should sit. Using current equity values and the latest annualized revenue figures, CoreWeave trades near 4.8 times revenue, DigitalOcean near 12 times and Snowflake near 16.8 times.

The gaps reflect very different economics. CoreWeave grew first-quarter revenue by 112%, yet its filing also showed $536 million of quarterly interest expense and a $740 million net loss. Snowflake grew more slowly, at 33%, while carrying a much more software-heavy model, 126% net revenue retention and 779 customers spending more than $1 million a year. DigitalOcean sits between the two, with profitable cloud infrastructure and AI customer ARR growing 221%.

Together AI’s multiple lands closer to infrastructure than software. That feels reasonable today because Sacra estimates that GPU rentals still produce most of its revenue. A larger share of APIs, fine-tuning and managed deployment would support a higher price.

Together AI compared with public cloud and software companies

Q5Is Together AI cheap next to Fireworks AI and Baseten?

Together AI is currently much cheaper than its closest private competitors. Fireworks has raised at a $17.5 billion valuation after reporting more than $1 billion in ARR, implying a multiple near 17.5 times. Baseten’s latest round valued it at as much as $13 billion, while Sacra estimates roughly $600 million of annualized revenue, producing a multiple near 21.7 times.

Fireworks and Baseten have reasons to trade higher. Fireworks emphasizes customized models built around customer data and workflows, which can create deeper switching costs. Baseten sources capacity across more than 20 cloud providers and sells a managed deployment layer that may require less owned infrastructure. Together AI still gets a larger estimated share of revenue from GPU rentals.

Even allowing for those differences, the gap is hard to ignore. Fireworks trades at more than twice Together AI’s estimated revenue multiple, while Baseten trades at more than 2.5 times. Against the current private-market frenzy, Together AI looks underpriced.

Q6Is Together AI growing really fast these days?

Together AI is growing very fast. Revenue, customers and capacity have all jumped. In its February 2025 financing materials, the company reported sixfold ARR growth, twentyfold customer growth and more than 450,000 developers using its platform.

Sacra later estimated a 62% increase in annualized revenue between the end of 2025 and February. The periods and definitions differ, so rolling everything into one neat growth rate would be misleading. Still, the acceleration is broad rather than dependent on one metric.

Infrastructure has expanded alongside demand. Together AI said in its latest financing announcement that investors had separately committed more than 500 megawatts of compute capacity. The company had disclosed 200 megawatts and a planned 36,000-GPU Blackwell cluster at the previous round. The power commitment has grown by at least 2.5 times since then.

Q7Are companies actually using Together AI in production?

Together AI already runs production workloads where latency, reliability and cost are easy to measure. Decagon says Together AI cut its inference cost per conversational turn sixfold while keeping p95 model latency below 400 milliseconds for its voice agents.

Other customer examples cover very different workloads. Yutori reports browser agents running twice as fast and four to five times cheaper than comparable frontier-model alternatives, with a 99.9% uptime commitment. Hedra says it cut infrastructure costs by 60%, tripled inference speed and scaled its video workload by 300 times. The Washington Post processes 1.79 billion input tokens per month on Together AI with response times around two seconds.

That is much stronger evidence than a logo wall. It shows repeated gains across voice, browser agents, video and publishing. Together AI obviously selected its best stories, so retention and concentration data are still needed before treating those results as typical.

Q8Is open-source AI demand big enough for Together AI right now?

Together AI is selling into a market that is already large and moving quickly. A McKinsey survey of more than 700 technology leaders and senior developers across 41 countries found that over half of organizations already used open-source AI across data, models and developer tools. Seventy-six percent expected usage to increase.

Adoption is strongest among the customers Together AI wants. McKinsey found that 72% of technology companies used open-source AI models, compared with 63% across the full survey. Respondents also associated open tools with lower implementation costs in 60% of cases and lower maintenance costs in 46%.

The hyperscalers confirm the demand through their own product decisions. AWS has added more open-weight models from DeepSeek, MiniMax, Kimi, Qwen, Google, NVIDIA and OpenAI. Microsoft Foundry advertises more than 1,900 models, while Google’s Model Garden offers more than 200. Open models are now a standard cloud category. Together AI has a large market, but it shares that market with companies possessing much greater distribution.

Q9Is Together AI mainly selling software or renting GPUs?

Together AI still looks more like an AI cloud than a software company. Sacra estimates that usage-based model APIs account for around 30% to 40% of revenue, leaving GPU server rentals as the larger business today.

That mix explains how the company can grow revenue so quickly. A large dedicated cluster contract can add tens or hundreds of millions of dollars in annual spending. It also brings capital commitments, lower gross margins and pressure to keep expensive GPUs busy. Serverless inference, fine-tuning, evaluations and deployment tools can scale with fewer physical assets.

Together AI is adding those higher-margin layers around the compute. Its platform now covers inference, GPU clusters, fine-tuning, evaluations, code sandboxes and custom model deployment. The valuation becomes much safer once those services produce most of the gross profit. Investors are paying ahead of that shift.

Q10What can Together AI do faster than AWS, Microsoft and Google?

Together AI’s best edge today is the speed at which its researchers turn low-level AI systems work into production gains. At its recent AI Native conference, the company said FlashAttention-4 ran 2.7 times faster than a Triton implementation and 1.3 times faster than cuDNN on NVIDIA Blackwell.

Customer workloads make the advantage more concrete. Together AI reduced one voice model’s latency from 281 milliseconds to 77 milliseconds through a custom megakernel, improving unit economics by 7.2 times. Its together.compile system sped up the generation of 200 Hedra video frames by 25%, while another image benchmark finished 41% faster than torch.compile.

AWS, Microsoft and Google offer broader catalogs, bundled security and existing enterprise contracts. Together AI can win by shipping open models quickly, tuning every layer of inference and working directly with customers whose economics depend on milliseconds. It has to keep doing it. Hyperscalers can copy successful features.

Q11Can Fireworks AI or Baseten steal Together AI’s customers?

Yes. Together AI has no easy protection against Fireworks AI or Baseten taking major accounts. All three sell optimized inference, custom model deployment and access to scarce GPU capacity, and sophisticated customers can spread workloads across several providers.

Fireworks has become especially dangerous. It powers custom models for companies including Uber, Shopify, Revolut and Doximity, and Microsoft has added Fireworks as an option inside Foundry. That gives Fireworks enterprise distribution through a platform Together AI also competes against. Baseten offers a multi-cloud layer across more than 20 providers and already shares prominent customers, including Cursor, with other inference platforms.

Together AI’s wider stack can help. It combines dedicated clusters, serverless APIs, fine-tuning, research optimization and development environments in one place. But breadth only counts when customers build workflows they do not want to move. Price and raw model access will not lock them in.

Q12Will cheaper AI models squeeze Together AI’s prices?

Together AI should expect token prices to keep falling. Open-model providers release new models constantly, GPUs improve every generation and competitors pass part of those efficiency gains to customers. Together AI’s own catalog already lists capable coding models at well below one dollar per million input tokens.

Lower prices can still produce a larger business when usage rises faster. Browser agents may call a model dozens of times during one task. Coding agents keep generating, testing and correcting code. Voice agents run throughout a conversation. Yutori’s customer case shows the pattern: one browser task creates a recurring sequence of long-context inference calls rather than one chatbot answer.

Pricing power is the problem. Together AI says open models can reduce production costs by six to twenty times. Every saving that wins a customer today becomes the baseline for the next renewal. The company needs more usage, better hardware utilization and higher-value software to outrun falling prices.

Q13Does NVIDIA have too much power over Together AI?

Together AI remains heavily dependent on NVIDIA for its current growth. NVIDIA has invested in the company, supplies the Blackwell systems behind its largest clusters and provides the architecture targeted by several of Together AI’s best-known research projects.

The 36,000-GPU cluster announced with Hypertec, the Blackwell deployment and FlashAttention-4 all tie Together AI closely to NVIDIA hardware. The relationship improves access to new chips and gives customers confidence that Together AI can support the latest systems. It also leaves a powerful supplier capturing a large share of the economics.

Together AI can reduce the risk by supporting more accelerators and earning more revenue from software that works across hardware. There is little public evidence of that diversification so far. NVIDIA’s backing lowers near-term supply risk, while the longer-term margin risk remains substantial.

Q14Can Together AI ever make software-like margins?

Together AI can improve margins considerably. Software-like margins still look far away. Sacra estimates a gross margin around 45%, healthy for AI infrastructure and weak for premium software.

The public comparison shows how punishing this business can become. CoreWeave produced $2.08 billion of first-quarter revenue and $1.16 billion of adjusted EBITDA, then reported a $740 million net loss after depreciation, operating costs and $536 million of interest expense. Together AI is smaller and may carry less debt, but its planned capacity expansion creates the same basic problem: hardware must be financed before customers fully use it.

Research can improve the equation. Faster kernels increase tokens processed per GPU, better compilers reduce engineering work and fine-tuning or evaluation tools add revenue without requiring a proportional increase in hardware. We would want gross margin above 55%, with APIs and managed services producing most of the gross profit, before valuing Together AI like a true software platform.

Q15How much revenue would make Together AI’s $8.3 billion valuation look normal?

Together AI needs between $830 million and $1.66 billion of annual revenue to support its valuation under realistic infrastructure multiples. The lower end assumes a strong ten times multiple, while the upper end assumes the market eventually values the company at five times revenue like a mature AI cloud.

The current outside estimate already clears the ten times threshold and sits only 19% below the revenue required at seven times. A five times future multiple would require another 66% of growth. Recent momentum makes that reachable. A slowdown paired with infrastructure-level margins would push the valuation underwater.

Revenue required to support an $8.3 billion valuation

Q16What has to go right for Together AI to justify $8.3 billion?

Together AI can justify $8.3 billion by improving the quality of its revenue. The company already appears large enough on sales. Now APIs, fine-tuning and managed inference need to outgrow dedicated GPU rentals.

Bookings must convert reliably, customer concentration must stay manageable and recurring API usage must grow faster than cluster rentals. Together AI also has to maintain a clear performance advantage as AWS, Microsoft, Google and specialist rivals add similar models and deployment tools.

If Together AI reaches several billion dollars of revenue, gross margins above 55% and strong retention, it could be worth far more than the latest round. Its current traction makes that path credible.

Q17What could make Together AI worth much less?

Together AI’s valuation falls apart if its disclosed commercial scale is dominated by long contracts, a few customers or low-margin GPU rentals. A revenue base closer to $600 million would place the company near 14 times sales, rich for infrastructure with an estimated 45% gross margin.

Competition could cut the multiple quickly. Hyperscalers can bundle models with storage, databases, security and existing cloud discounts. Specialist inference rivals can move faster, while large customers can self-host once their workloads become predictable enough to justify an internal platform team.

The compute commitment is where the downside gets ugly. More than 500 megawatts can support enormous growth, and underused capacity can destroy the economics just as quickly. Together AI becomes much less valuable if revenue growth slows before those clusters fill.

Q18Is Together AI really worth $8.3 billion today?

Together AI probably is worth $8.3 billion today. That judgment leans heavily on Sacra’s revenue estimate because management has disclosed little beyond bookings. At the estimated run rate, the company trades below its closest private peers and between public AI infrastructure and software companies.

Several facts push us toward yes. Together AI has grown from a $1.25 billion valuation to its current mark in a little over two years, secured major compute commitments, won demanding production workloads and built research that has delivered measurable speed and cost improvements. Open-source AI adoption is broad enough to support a large independent provider.

Four things keep us from calling it cheap. GPU rentals still appear to generate most of the revenue, gross margin is estimated around 45%, the company depends heavily on one chip supplier and management has published bookings without the retention or concentration data needed to assess their quality.

We land slightly on the yes side. Together AI looks fairly valued when Sacra’s estimate is close to reality and its software share keeps rising. Revenue near $600 million or margins stuck at infrastructure levels would make the current price too high. Investors are paying for Together AI to become a durable control layer for open-model production. It has shown enough progress to keep that outcome believable.