2026.09.22 | TB | Cloud | Together AI Expert

Primer • 11 Sep 2026

Expert Former Senior Executive in Sales & Marketing

Sidharth Nair (Analyst) — Third Bridge Sector Analyst

Agenda

• Over 18 years' experience in the software space
• Over one year of experience at Together AI (left within the last year)

Introduction

Analyst: Please could you state I agree or I disagree to the following statement: You understand the definition of material non-public information and agree not to disclose any such information, or any other information which is confidential, during the Primer Interview today.

Expert: I agree.

Questions

Could you describe what Together AI does and the products and services it provides?

Expert: The company portrays itself as being an AI-native cloud. What that means is cloud infrastructure and managed services to facilitate the training. That is both pre-training and post-training, as well as inference of both open-weight and custom AI models. Even beyond that, I would say that there's a broader vision to provide various kinds of services that would be more or less essential for the purpose of developing and operating AI agents and systems, chatbots, what have you, multimodal kinds of AI. I think that would also fit into some of the vision. That's the broad set of capabilities.

What would you say is the value proposition of Together AI's AI-native cloud? What problem is the company looking to solve?

Expert: I mean, at the core, AI workloads are incredibly different than the CPU-based workloads that have been most associated with the hyperscale cloud infrastructure providers, Amazon, Azure, to an extent Google. These workloads are obviously very, very compute-intensive and training and inference themselves are quite different from each other and certainly extremely different than traditional internet services. Given the fact that compute and power are now proven to be very serious limiting factors on AI, the progress of AI, if you will, making most effective use of the computing and power that is available is obviously essential. Together, it's just to solve that as best as possible.

What is the full range of industries and business sizes that could use Together AI's services? What might be the company's TAM?

Expert: We should expect an evolution in this. There are different segments that I would say the company has already catered to especially well. The company calling itself "an AI-native cloud," from my perspective, there's some good and bad in that. The AI-native venture-backed start-ups, ones that are training models at times and doing the inference, that's the ideal customer profile, number one. I will also flag that Together AI has, at this point, I'd be surprised if it's not more than 1 million individuals sign up for the platform. There are lots of individual developers out there, many with entrepreneurial aspirations and basically all of them now using AI models for the purpose of coding. The coding agent use case is a very, very significant one for the industry. Together AI certainly has a role to play as far as providing inference to potentially individuals or software development teams. That's another category in and of itself, the individual prosumer.

The other sets of businesses that exist where Together AI has not exactly had massive uptake yet. There's obviously GPU clouds, CoreWeave, Nebius, Lambda. CoreWeave especially has emerged as a hyperscaler owning and operating hundreds of thousands of GPUs. Together has had a different operating model than that. I think it has somewhat similar aspirations as far as scale is concerned. What I'm getting at is Together, they're not among those that are signing these multi-billion-dollar contracts to build data centres and sell that capacity to Microsoft or Meta or what have you. It could potentially get into that at some point, but it's not doing that.

Could you profile Together AI's typical target customer across industries and business sizes? You mentioned the company's ICP [ideal customer profile] being AI start-ups. What might be its revenue split from AI start-ups compared to individual researchers? Might it be 41-50% or 51-60%?

Expert: The revenue is extremely slanted towards businesses. There's a time when the fuzzy line between when an individual or a small team suddenly become one of those businesses that are spending a lot of money? 91-100% of the revenue is coming from customers that would be considered businesses. It's just the funnel, there are two sides to it. You certainly have a strong inbound product-led growth, but at the same time, you also have a lot happening at the executive level to put together some very large deals. As far as the AI-native customers that are particularly strong fits for the platform, I would further segment that by saying that there's a set of companies that have had a need to procure multi-million-dollar GPU clusters in order to train, do some reasonably extensive training, but not at the scale of OpenAI, but maybe it's 1,000 GPUs or what have you. Then, they train their models and then they need a place to run those models. In those scenarios, Together AI is a particularly strong fit because it is that full AI life cycle.

Analyst: How might the 91-100% revenue be split between GPU clusters and AI-native companies?

Expert: There are bands here. The companies that are doing training, they are often spending millions to procure these GPU clusters. Companies that are basically inference-only, occasionally, they will spend millions, if they're among the AI-natives that are already basically at scale as far as them having customers, like a Cursor or a Decagon. Those are doing millions of dollars in inference already. Then, there are lesser-known companies that are maybe in the hundreds of thousands or tens of thousands or you get down to the real seed-stage start-ups. You're looking at thousands of dollars or something, an individual spending USD 20. That's how to break it down.

Analyst: What might Together AI's sweet spot be in size, average revenue or employee count?

Expert: As far as the average revenue per customer, I'm sure there's the mean and then there's the median. The mean is certainly going to be dragged to a pretty different value than the median. The number of paying customers, just anyone with a credit card, I don't know how many thousands that would be at this point. The median is probably in the tens of thousands. The mean, I wouldn't be surprised if it's 10 times that. I will also flag that I haven't articulated yet, there's a handful of very, very large customers, probably with USD 100m-plus contracts. I believe that the information reported that, basically, Nvidia is supposedly investing in Together AI and renting capacity back from Together AI.

Could you mention any customer names that are either on Together AI's website or in the public domain?

Expert: Obviously, Cursor has been one of the more prominent ones. I mentioned Decagon.

What are the key regions Together AI derives its revenue from?

Expert: Certainly, the company is headquartered in San Francisco and has strong ties within the community of start-ups based here in the Bay Area and certainly has raised funds from venture capitalists and has built a professional network and introductions. Probably most of the top companies in the AI space, I think, probably have had interactions with Together AI at this point. The Bay Area, number one and New York City, really, number two. The company has set up some operations in Europe, more recently has announced. I forget the name of the partners in India, but India is an interesting market. I haven't seen a company do a heck of a lot in other parts of APAC, but the other region where there's a lot happening is Saudi Arabia. Prosperity7, which is very closely tied in with, why am I forgetting, what they exactly call themselves, the fund in Saudi Arabia. Together has announced that it's working with Humain and will have a very, very substantial amount of compute in the region.

How much of the company's revenue might come from the US, Europe, India and the Middle East, in percentage ranges?

Expert: I honestly think almost all of it is happening. I don't know about the Humain arrangement exactly and then how things are being accounted for and so on. The actual customers, at least among the businesses, I think, they're almost all in the US. Probably a handful of exceptions and then the Humain financing, the accounting for that being a question mark right now. [So, 91-100% from the US, 0-10% each from the other three regions].

What are Together AI's core revenue-generating product segments?

Expert: Certainly, there's an evolution in the percentage split here. The company, as an AI cloud platform, has within it GPU clusters, so that's just the GPU infrastructure as a service product line, with GPUs being rented by the hour, more or less. Then, different kinds of inference product lines, so serverless model APIs and dedicated endpoints and so on. The three product lines of clusters and the different kinds of endpoints, that's going to make up right there fairly 91-100% of the revenue. If you add up GPU clusters and serverless inference and dedicated endpoint inference, you're probably looking at 91-100%.

How would you split Together AI's 91-100% revenue across GPU clusters, serverless inference and dedicated endpoints?

Expert: These are the percentages that are undoubtedly continuing to shift. I would say that earlier in Together's history, the GPU clusters, the revenue from that on a percentage basis, absolutely dwarfed the inference revenue. Probably in August '24, it was probably 91-100% GPU cluster revenue, with some of that on the back of a couple of large USD 10m-plus deals. Now, the inference business has grown and every product line is growing significantly. It continues to be the case that sometimes these GPUs are sold as clusters and a transaction, like if Nvidia's renting back GPU clusters. That's a big-ticket item there. I would speculate that the current percentage breakdown is probably more like, say, 61-70% of the revenue from clusters, whereas, say, another 21-30% from the dedicated endpoints product line and the remainder [0-10%] is from serverless inference.

What is the typical revenue model for each segment? Might it be reoccurring for GPU clusters, given you mentioned it's by the hour? Is it recurring [subscription], reoccurring [usage-based] or transactional [one-time] for dedicated endpoints?

Expert: The clusters, the dedicated endpoints, those two product lines are oftentimes sold under contracts, particularly if you're selling to a customer that's spending hundreds of thousands or millions of dollars. You're typically on structures there. As you go smaller, certainly, Together AI has built its product to enable self-service as much as possible. It definitely wants to encourage customers to help themselves and continues to make their products more elastic. Just a couple of days ago, the company introduced pre-emptible GPU clusters to offer discounts if you're willing to relinquish your compute.

Analyst: What might be the revenue model for serverless inference? Is it reoccurring by the hour or more recurring?

Expert: This is something that has evolved as well. Together, historically, the serverless product was not built to be relied upon at scale. It is only fairly recently that the company has completely revamped it with more of a provision throughput capability and more reliability guarantees. Now, they're positioning it as something that can be relied upon. I'm sure it's growing a lot. I said earlier that dedicated endpoints are 21-30%. It could be wrong. That might be overcounting it now because the serverless is potentially much more. The structures for serverless are more variable.

Analyst: Are the contracts typically annual or multi-year?

Expert: When it comes to inference, those tend to be a lot more fluid. When it comes to GPU clusters, that's when you typically have, depending on the scale, the bigger you go, the longer-term contract. Inference is that is the workload that is so dependent on your customer's product actually getting traction. It's volatile. Whereas a GPU cluster that's often put towards training is a much more predictable purchase in the contract structures. I have to follow that.

Analyst: What is the payment typically based on? Is it more consumption-based or variable based on transactional usage?

Expert: The inference is much more of a consumption-based product. Companies that are at scale, like Cursor, would have enough predictability into their business, into their customer adoption, that they're oftentimes comfortable making commitments. They may still insist on having some cancellation terms and things, if things are to change. At least the customers with scale can make commitments.

Analyst: Might the revenue model be recurring for GPU clusters and dedicated endpoints, given you mentioned they are more predictable? Might it be reoccurring for serverless inference?

Expert: The thing I'll flag is that, at least when it comes to GPU clusters, clusters that are put towards training, a lot of those are by design temporary or at least not necessarily recurring. Training jobs, at least historically, are meant to finish. Together AI, the fact that its customer base is so geared towards these AI-native VC-backed start-ups, those companies are not necessarily proven to be long-term businesses yet. The clusters are not necessarily recurring, but the compute is often just resold to someone else, if that's okay. [So, reoccurring for all of them].

Could you provide a use case for each segment? How do customers use the GPU clusters segment?

Expert: Another company customer, like Cartesia, is less well-known than ElevenLabs, which is also a customer, but building amazing text-to-speech models and therefore relying on substantial amounts of compute to train those models and then to serve those models at scale. That would be an example or Hedra, which has historically trained video generation models. Cursor is even doing different kinds of training with the clusters.

Analyst: How do the clusters help in the training process?

Expert: It is the essential infrastructure. If you're doing a pre-training run or post-training, you need the GPUs. That's what these are.

Analyst: Does the company typically manufacture and source GPUs or have third-party companies along the value chain?

Expert: The GPU clusters are Nvidia chips. Together, as far as how it goes about getting those systems, it has a supply strategy that is either renting from some bare metal GPU clouds or otherwise directing the build-out of GPU clusters and working with some data centre partners to stand those up.

Analyst: What might be a use case for dedicated endpoints? Could you mention an example of a customer?

Expert: That would be for a company that has a model. It could be a model that there's a custom model that maybe they did the pre-training. Maybe they did some amount of post-training or it could just be an open-weight model exactly as is available on Hugging Face. They're operating an application with enough scale and sustained use that they want. It could be as little as, say, 8 GPUs, but more likely tens or hundreds or potentially even thousands, but oftentimes more hundreds of GPUs to power high-performance, reliable, cost-efficient inference. As I said earlier, historically, the serverless endpoints product was flaky, multi-tenant, noisy neighbours, not exactly reliable. The dedicated endpoint product was basically with Together and some other of its competitors have similar things. That was the solution of, like, "Okay. You want to do inference. You really want to count on it. You want to scale it. That's what you use." That's what dedicated endpoints have been for.

Analyst: What might be a use case for serverless inference?

Expert: There's the old serverless and then there's the new serverless. You can think of that as really serving two different kinds of audiences now. Continues to be the case for individual developers or people just starting a company and just trying to get going with it. That doesn't require any contracts. It's a paper token. It's a great fit. Plugs into coding agents. There are alternatives to Claude Code, like OpenCode and Kilo Code. That plugs in great. [We're] building stuff with these days. Then, if you're just building an app, you can use that or if you're, say, an Ollama and you're effectively reselling inference capacity as an API, it lends itself all to that form factor as well. Really, anything where the inference utilisation is volatile. You have different times of day, different times of week, different times of month. Wherever the inference is volatile, the serverless lends itself particularly well for that.

Which of Together AI's three solutions do you consider to be the fastest-growing and potentially driving the largest revenue share in future?

Expert: I honestly am very bullish about the serverless form factor. Honestly, [earlier] at Together, [we were] always a bit puzzled as to why there was seemingly so much emphasis on a dedicated endpoints product line. Shouldn't the serverless product just work the way people would expect? It seems that, "Okay, the company finally has made it work," so they claimed. That's the form factor that makes open-weight inference a truly viable alternative to the GPT API, to the Claude API. I think that across the industry, particularly as coding agents have now proven to be so remarkably capable, I'm fairly certain across the industry that the inference of open-weight models, the revenue growth there, the past six months, has been astounding and at this point is capped not by demand, but by compute and infrastructure. Whoever can stand this stuff up at scale has a very, very large market opportunity.

What do you estimate Together AI's NRR [net revenue retention] rate to be – less than 80%, 80-90%, 90-100%, 100-110%, 110-120% or above 120%? What percentage of the company's revenue has been retained from its existing customer base?

Expert: I'm understanding correctly. It's a weird dynamic here, these customers are growing. However you were to calculate this exactly, the reality is the customers are growing at times very, very quickly.

Analyst: Might the company's NRR be plus-120%?

Expert: I would think so. NRR calculation is basically your customer's revenue and how fast that's growing, at times it's growing very, very quickly. A Decagon or a Cursor revenue or what have you, those are all going up very fast and Together is providing infrastructure to these companies. I'm not sure of the figure.

What do you think are some of the key drivers of Together AI's NRR? Might it be a land-and-expand model?

Expert: It is land-and-expand. The companies that have signed on, there's certainly within Together a set of known accounts that are very deliberately sought after and oftentimes ultimately landed and then the expansion. Some of that is happening organically because the customer is growing. Then, other times, it's happening because the customer is satisfied and pleased with performance. They want to consume more because they find the products to be very strong. A lot of the growth here is just like the industry is growing here.

Could you specify the industry that best describes Together AI's core operations?

Expert: It's providing infrastructure. AI cloud infrastructure and services, enabling the development of agents and AI-native applications.

What has been the AI cloud infrastructure space's AAGR over the past three years – less than 0%, 0-5%, 6-10%, 11-20% or more than 20%?

Expert: The entire industry of AI cloud infrastructure is, I think, growing. Doubling, tripling, what have you. CoreWeave's revenue is, they went on an absolute hockey stick from 2022 on. The growth rate is following basically, I don't know, the number of GPUs out there.

Analyst: Might the space's AAGR be plus-20%?

Expert: Yes. I think that's certainly undercounting.

What do you think have been the 1-2 key drivers of the AI cloud infrastructure space's historical growth?

Expert: I think there are multiple kinds of demand spikes that have played out and are likely to continue to play out. These scaling laws are one indicator. I mean, there was demand for GPUs and therefore the cloud infrastructure companies that had them, that originally followed the scaling law of pre-training, which is like you keep throwing more compute at the problem, increase your compute in order of magnitude and your model gets a heck of a lot better. That alone created massive demand for infrastructure. As the models therefore prove capable, then you started using them, not just training them, but actually doing inference and building consumer applications, ChatGPT, of course. Now that the models have proven capable of doing astounding professional work that would have taken humans a routine to do in minutes, but used to take a team weeks of coding and writing and creating all sorts of digital documents and presentations. We're just continuing to see phenomenal spikes in demand here.

What do you think the space's AAGR will be over the next three years? Do you see it at plus-20%?

Expert: Yes. Absolutely, [plus-20%]. The demand here is insatiable. There are obviously some people who have adopted AI faster than others. Of course, engineers are leading the way here. I think we're still very early in putting this infrastructure to work.

What might be the space's future growth drivers? Do you see any emerging trend in the space that might create new or more demand?

Expert: Actually putting agents to work is still so early. As I said, the engineers are all doing this as much as they possibly can right now and everyone's maxing out there a lot of the amount of tokens and so on. Beyond engineers, you have a small set of early adopters of capabilities, like Claude Cowork and ChatGPT Work. There are newer kinds of personal agents, like, just the other day, Meta announced it's, what they're calling Muse. It's a personal agent. [Some] are using a new agent called Instinct that is looking at all your e-mails and calendar. It's scary, but texting you various useful things. Agents are still in their very early adoption right now. As people figure out what this stuff can do, the demand is just going to keep going through the roof.

What do you estimate the adoption rate to be for AI cloud infrastructure solutions across the addressable market – 0-10%, 10-20%, 20-30% or higher?

Expert: I think that there's a set of companies that are obviously building with AI and language models and so on. A lot of them so far have used either the Claude API or the OpenAI API. I don't think that they necessarily need to continue doing that. A lot of them can potentially switch to open-weight models and therefore hopefully land on a company like Together AI's platform. That switch is still potentially very early, where there's also just a lot of earliness, as I said, is like the actual consumer and white-collar worker adoption of the very emerging products and services that all require AI infrastructure, that all require inference. That remains very early. There's potentially still some big waves ahead. If there ever is a large-scale shift from the Claude API, OpenAI API to open models, Together AI is a beneficiary of that and then this inevitable complete reinvention of what it means to work and, in a lot of ways, what it means to even live, that's also coming.

Analyst: Might adoption rate be the same for generic inference engines and more inference-optimised AI clouds? Might they have higher adoption rates?

Expert: Yes. I think so. When it comes to Together in this industry, one of the things to be paying attention to is who has a path to scale. Together has, in some public statements, claimed that it will have, I don't know if it said the date, 500MW of capacity. I honestly don't recall seeing Fireworks or Baseten or some of the other kinds of inference companies exactly make that claim. I wouldn't be surprised if they're going to get there, too. There are other companies that maybe have very strong inference technology, but don't have an obvious path to that level of scale.

Analyst: What might be the space's adoption rate in a percentage range – 40-50% or 50-60%?

Expert: 100%-plus. All this is, if I'm understanding, when you say 11-20% of it, it's 1.2 times what you did the year before.

Analyst: What percentage of AI-native companies has adopted AI cloud infrastructure in the space?

Expert: I would say that 90-100% have adopted some form of AI infrastructure or language model, but how many of them have adopted an open-weight model? Maybe that's 10-20%.

To what degree do you think the industry is correlated to economic cyclicality? Is it counter-cyclical, very cyclical, somewhat cyclical, somewhat resilient or resilient?

Expert: From my perspective, the broader macroeconomic environment has absolutely nothing to do with how this is going to unfold. Anything in the broader, whatever, policy world, there are policies that will potentially regulate AI. That can have some impact, but interest rates and the like are irrelevant.

Based on public information, what impactful tech innovations should we anticipate in the AI cloud infrastructure space over the next 3-5 years? Could you mention one example?

Expert: As far as innovations are concerned, what's happening now is that AI is beginning all the more to improve itself. It's beginning all the more to evolve in ways beyond human control. The infrastructure is likely to further enable those developments. I may see some things in the infrastructure world that can attempt to hopefully control this stuff better. AI, so far, the history of the past few years is that models are still updated periodically, but where things are going to be probably reasonably quickly is that the AI is going to learn much more as you and I do continuously. I think there's more infrastructure work that needs to happen to make that really function. We should also expect a lot more multi-modality, more infrastructure. Obviously, there are world models, some companies that are really specialising in that. We should all anticipate at some point in the next, I don't know how many years, but having this avatar, not the movie, but being able to talk fluidly with a digital being and having that, in effect, be as good as when you're on a Zoom talking to somebody you can see. That's still unfolding.

Are there any downside risks you believe might materially impact market growth over the next three years? What might hurt the AI cloud infrastructure space?

Expert: Certainly, alarms are sounding. The Anthropic researcher just a few days ago was very publicly resigning and saying this stuff, "There are not adequate controls." The regulatory response, we'll probably see some more efforts, both by governments as well as some in the industry to try to slow this all down and that's you slow it down and that slows down the infrastructure segment. My opinion is this is all an inevitable thing. Any slowdown by governments, by the industry, will slow infrastructure growth.

Analyst: What might be an example of any regulation that might cause a slowdown?

Expert: There are certainly things happening, much more conversations about, like, "Don't build that data centre." That's a thing. The other thing that's also still not really resolved is that a lot of these open-weight models, most intelligent ones, are made in China. There's the Kimi K3 launch, I don't know if it was two months ago now. That definitely sparked a lot of discussion about, "Should Chinese models be banned?" Shortly after, there was an open letter that Microsoft and Nvidia and basically every company in the industry other than OpenAI and Anthropic signed, supporting open-weight AI. We could still see some restrictions on the use of Chinese models.

How fragmented do you consider the AI cloud infrastructure industry to be? Is the majority of market share captured by 2-5 strong players? Is it a monopoly? Is it fragmented or highly fragmented?

Expert: There are a lot of companies in the AI cloud and inference world. GPUs, they are quite distributed. There are a lot of companies here. OpenAI and Anthropic, they have paths towards, I forget if it was to try to have 31-40% of all AI compute or something like that.

Analyst: Might the space be fragmented or highly fragmented?

Expert: It's fragmented, but then there are still some ultimate companies with huge, huge supply or at least OpenAI and Anthropic, they're not clouds themselves per se, but there is, call it, significant customer concentration. That's a real thing.

Who are Together AI's top two, most direct competitors?

Expert: Fireworks and Baseten are most similar.

What are Together AI, Fireworks AI and Baseten's market shares, in percentage ranges? Alternatively, could you rank them?

Expert: I would say Fireworks at number one for inference. Together and Baseten, I think, are about tied right now. I'll go with Together, number two.

What was each player's revenue for the last financial year – USD 0-50m, USD 50m-100m, USD 100m-200m, USD 200m-300m or more?

Expert: I think the ARR [Together AI] recently went public with was USD 1.1bn-1.2bn. I'm sure it's got to be more than that right now. It's announced some pretty interesting deals lately. [Fireworks AI] probably USD 900m-1bn. Together has had more revenue because it's been in the GPU cluster business that Fireworks has not been in. [Baseten], probably USD 900m-1bn.

What might be each player's revenue growth rate – less than 0%, 0%, 0-5%, 6-10%, 11-20% or more than 20%?

Expert: Definitely well over 20% [for all of them].

What has been each player's growth driver for the last financial year? Has Together AI's growth been organic, acquisition-based or mixed?

Expert: It's organic. It's made some acquisitions of other companies, but the core revenue is from products it's built. [Fireworks and Baseten], organic.

In your opinion, what is the average EBITDA margin for the players – less than 0%, 0-10% or higher, up to more than 60%? How profitable do you think the players are?

Expert: 21-30% for sure.

Would you say each player's EBITDA is in line with, above or below industry?

Expert: [Together AI], just a little bit above. [Fireworks], I would say on industry. [Baseten], on industry.

How do you think the players differentiate themselves, starting with Together AI?

Expert: Together, the fact that it has the GPU cluster product and then it's more of a full AI life cycle cloud, that's something that Baseten and Fireworks are not quite.

Analyst: What might be Together AI's other differentiators based on approach, breadth of offerings or pricing model?

Expert: I think Together AI is perhaps a bit more hands-on as far as helping companies just train models, bringing researchers into engagements. That's one thing.

Analyst: What is Fireworks' differentiating factor?

Expert: I think that they have been probably more focused on inference. At times, they have been first to really optimise model performance on some new, emerging models. Together is not necessarily far behind, but Fireworks at times has had just a bit of a first-mover advantage in some areas.

Analyst: What is Baseten's differentiating factor?

Expert: Honestly, I think that they have really done an excellent job cross-functionally in some ways that Together and Fireworks have not. In particular, I'm going to complement their marketing because it's the best in the industry. They have a very strong product, but I honestly feel like their marketing has allowed them to really get in this race as something being about as good as any other company.

What do you think are customers' top three purchasing criteria? How do you think they rank by importance?

Expert: Number one is, do you have a capacity? Number two is, can you run a model with performance characteristics that meet the requirement? Number three is, basically, is your service reliable enough to be dependent upon in a production service?

How important do you consider price to be for customers in the space? Is the market commoditised, somewhat commoditised or not commoditised?

Expert: The price is maybe number two or number three [criterion], actually. Basically, for a customer to buy, you have to check off all of these requirements. Your pricing, your performance needs to be in an acceptable range. Reliability needs to be in an acceptable range. That's how decisions are made.

Analyst: Might capacity be the number one purchasing criterion, performance two, price three and reliability four?

Expert: I think that's fair.

Analyst: Would you say the space is commoditised or somewhat commoditised?

Expert: I would say it's somewhat commoditised with the appearance of being highly commoditised.

What is the industry average annual churn rate for AI cloud infrastructure software solutions – 0-5%, 6-10%, 11-15%, 16-20%, 21-25%, 26-30% or above 30%?

Expert: If churn is, what per cent of customers churn, then I'm going to guess maybe 21-25%. Potentially so. Just because it's a very volatile segment, but any churn is overcome by the fact that the customers who do not churn, some of them are growing exponentially.

How do you think Together AI's churn rate compares to the industry average? Is it in line with, better or worse than the industry?

Expert: I think it's in line with the industry.

What do you think has been the key driver of Together AI's growth over the past couple of years? What led the company to capture more market share historically?

Expert: As far as what has allowed the company to break out, I think that the company has established a reputation in, say, 2023-24 as being a platform for inference of these open models and also for putting out some research and data sets and the like. It established itself at a very critical moment in time and built a community, built a following that it has continued to grow. That 2023-24 time period was a really critical one.

How fast do you think Together AI could grow annually over the next three years? Might it be as fast as the industry at plus-20% or slower than the average?

Expert: I think faster than that, for sure. It's in a good position right now, [so plus-20%].

What could be the most interesting way to grow Together AI? How would you improve its organic growth profile if you were running the company?

Expert: Honestly, where the company, I think, has yet to truly execute is really motivating, switching from Claude APIs and GPT APIs onto open models on Together. I feel like it's partially executed that, but there remains a significant opportunity there. As well, the company can consider more vertical integration. It's been a cloud platform, but there are alternative agent harnesses to Claude Code, to ChatGPT. There are things like OpenCode out there. I would certainly welcome Together getting more into some of these adjacent categories that are already consuming meaningful inference, but potentially have a lot more exponential growth ahead.

How would you improve Together AI's margin and operational efficiency?

Expert: Where the company is, I would say, making some headway or efforts. There will be more efficiencies, I would say, as Together manages to drive more control of the infrastructure. As Together manages to take more control of its infrastructure and work with partners on more favourable financial terms, like the margin efficiency when in that 2023-24 time period where it was very dependent on renting GPUs from some other clouds, some of the margin structures there were not necessarily as favourable as what I would expect going forward. I think that it's building those relationships to procure infrastructure and stand it up and operate it and have that be on more favourable terms going forward.

Based on publicly announced acquisitions, how would you describe Together AI's M&A strategy? Has the company carried out any M&A in the past?

Expert: Yes. They have. There was a company called CodeSandbox that Together acquired probably about a year and a half ago. That acquisition was even more forward-looking than [we] understood it to be at the time. Agents are now writing code and then running it. It seems that this Sandbox for running agent-generated code is like its own product category in its own right and it's growing very, very quickly. That was one. Then, Together also acquired, I forget the name of the start-up, but another company that actually, maybe they weren't even doing voice AI, but maybe that's just what they started doing, Together. There have been at least a few. I don't know if there's been anything beyond those two. I'm certainly hoping to see Together make more acquisitions and get into some of these adjacent categories.

Could it make sense for Together AI to acquire, merge with or be acquired by other players in or outside the space?

Expert: It's certainly possible. In fact, I'm speculating as to who. Honestly, the one real dark horse that I would love to see happen is that a good portion of Together's founding team had been running Siri at Apple, obviously. Any day we're all anticipating a completely new Siri. I believe I'm disappointed because I don't think it's using Together AI under there, but I feel like it should be. That would be one. I don't really think it's going to happen, but in some ways, gosh, that would make so much sense. Google would be an interesting one. I do feel like Google has obviously fallen behind OpenAI and Anthropic in the model race and could potentially use a bit of a shot in the arm to better position the company going forward. I could imagine a scenario in which Google did. Of course, Nvidia is another. What is an acquisition anyway these days? That's another weird messed up thing about the way things seem to be operating right now. We were going to hire everybody but not actually buy the company, I don't like this model. Nvidia could or potentially do a merger with something like CoreWeave. There's some complementary technology there. My honest expectation, though, is Together should be on a trajectory as a standalone public company.

What would you say are the 1-2 biggest risks Together AI faces?

Expert: The potential ban of Chinese open-weight models is a risk. Then, I think, the fact that inference is perceived as a commodity. There's also open-source inference, vLLM and now there's a company called Inferact that has raised meaningful funding to further build that. Open-source inference is a thing and a risk of the GPU infrastructure being so decentralised, so distributed. Talking about Together having a path towards 500MW. If Together, for whatever reason, can't execute on that, if for some reason they're not able to get to the level of scale, then that's a real risk.

Is there anything the market does not properly understand about Together AI? Are there any common misconceptions?

Expert: I'm sure there are people who don't quite understand what the company does because the company doesn't always explain itself super well. I think there's that.