All right, great. Welcome to the afternoon of the Needham & Company Virtual Tech Week conference. This afternoon, we're excited to have Veritone presenting. We have CEO Ryan Steelberg and CFO Mike Zemetra presenting, and we'll leave about 10 to 15 minutes for Q&A after about a 25-minute presentation, so if you have any questions, please submit them in the Q&A box or email them to me at jreilly@needhamco.com, and with that, Ryan, I will hand it off to you.
All right, great. Thank you, everybody. And happy to, excited to do this virtual week. I think we're going to be presenting in person here the first part of the year as well. So for those of you, hoping to see you in person at some point in the future. I just want to give a few moments. Obviously, we're going to be talking about some future guidance. We did disclose just recently our Q3 earnings, and we did provide some guidance for Q4 as well as for the entirety of 2025. But again, we will be touching on, again, some future guidance. So again, please keep this information confidential. I'm sorry, not confidential, but just please execute and follow all rules and guidance.
Okay, I'm going to start off with a little bit of background on Veritone and what our current and active product and service offerings are in the AI space today. And then I want to spend the majority of the time today to talk about what we announced this morning, which we call the Veritone Data Refinery, which is a really exciting expansion of our business that really we believe is going to be a cornerstone in the future AI training for large language models and multimodal models going forward. But let me start by saying, excuse me, that we started this business in 2014 as an AI, pure AI company.
That statement alone is pretty interesting, that we're running the companies that really sort of saw this opportunity arising, not just the large language models and the Gen AI models, but we were really obsessed with what cognitive science and cognitive AI could do to the landscape. That would be like transcription or object detection or face detection and what that could do to the landscape as a whole. I come from a pretty rich background, as do a few of the other founders in the ad tech space. I started my first company in 1994 right out of school. It was called AdForce. So we've had a very successful career building large distributed advertising technology platforms, taking a couple of companies public. I sold our previous business to Google, and I headed up all of Google's offline ad efforts, which ironically include a lot of audio and video.
It was kind of really the kernel that was kind of impressed and sort of sunk in me about that there was a major need for AI to start to ingest and understand audio and video at scale. So let me say that again. So we saw that there was a huge demand because of the explosion of audio and video, which we'll put in the category as unstructured data, that we needed to build software, a solution to be able to ingest these huge corpuses of audio and video and leverage AI to understand what's happening inside it. And that was the impetus. Veritone is kind of a play on the words of truth in the signal, veritas tone signal, truth in the signal. Today, we've been public since 2017.
As we exit the year here, we're forecasted to have right around 485 employees, and we have over 3,200 software and software solution customers today. So what is Veritone? Veritone is a leading enterprise AI company. We've kind of kept that moniker, and we believe we're one of few that can say that for almost 10 years. And we primarily, I would say our true specialty is unstructured data in ingestion analysis and harnessing and understanding unstructured data. But in that class, audio and video, we believe we're the best in the business for managing and understanding and activating and activating upon audio and video at scale. What we sell is software SaaS-based solutions as well as enterprise-class like API access to our platform solutions as well.
More recently, just because of the demand and need for a lot of our customers to actually work on preparing their data, we did launch a consulting services arm just earlier this year in coordination with AWS to assist our companies along their AI journey. But first and foremost, we are a product software-based company. Our flagship product and platform we're just going to talk about later is called aiWARE. And we primarily service enterprise-class customers really in two major verticals, commercial and public sector. And here's just a quick logo representation of some of the marquee leading companies that license our AI-based solutions and software. And I'd like to say is we are mission-critical. Many of our customers have been customers now and so in AI for our AI products and services for several years, and we have a very high retention rate.
So I'm going to go through that in more detail. We also have a very exciting and fast-growing public sector business, and that's really broken down by selling AI-based software and solutions to state and local law enforcement as well as FedCiv like the DOJ, but also now more recently the Department of Defense with many different agencies that we're currently working with, including the Defense Logistics Agency. The problems that we solved, and I don't want to, this is more of a high-level framework, but the biggest opportunity, and it's going to be a major opportunity for the foreseeable future, is just the ever-exponential growth of unstructured data and how important it is to ingest and harness and take advantage and leverage that unstructured data. And that's what we do very, very well.
I would say it's probably put us up against anybody and truly goes to our legacy and our experience and frankly the scale that aiWARE is now processing unstructured data today. The way we built aiWARE, there's a lot of different AI models and components out there. aiWARE literally is a one-stop shop end-to-end so customers can come to us with their data. We can build the workflows. We can build the data piping. We can tap into hundreds of pre-built AI models, but also train and tune new models, so in effect, we are their wilderness guide, right? One-stop shop for really all their aiWARE needs, I mean, their AI needs and we design our platform that we can get customers on board with applications or services for as little as a few thousand dollars a year.
But at the same time, that same platform, that same customer can scale to a multi-million dollar a year contract and scale of PB of data in terms of scale. I apologize. I'm recovering from COVID, so I'm going to have unfortunately a little few intervals through the presentation where I'm going to have to have some water and drink a little bit, but I appreciate you guys bearing with me. And then lastly, speed. You hear sometimes that a lot of the AI implementations are like proof of concepts, but what's really the speed and time and understanding the expense to go to full production at scale? We've been doing this for 10 years, and as I've sort of articulated, we can onboard fast and we can scale fast.
Just put in some magnitude, we've processed from paid customers tens of millions of hours of audio and video over the last several years. So we can handle some of the largest, most dynamic, most complex data payloads. How we do it? I'm not going to go into this in too much detail. We got a lot of information on our website, but our core platform is aiWARE. This has been built and improved and iterated upon over the years. But again, it truly is your end-to-end stack from data ingest to standing up the data lake. We have vector databases. We have a full orchestration layer to be able to run and invoke different AI models, both cognitive and large language and multimodal models, I'll call Gen AI models, but also a very robust application level as well. So this is not just ingestion and preparing the data.
We have a full workflow and application layer to actually build on top of it and in effect, turn it into real value. We have tens of thousands of daily active users, our customers and users who use our software every day. So this is not just, again, a platform where it's just the data science team or the CIO, CTO department. This is a platform that's been tried and true that does touch and can be deployed as an application to service end users, people like tagging audio clips, police investigators across the United States. So we're just showing you that's the kind of the flexibility, extensibility of the platform. The last thing to point out is we are an open platform. We can ingest and sort of interrogate and understand and work with really any type of data set.
We are open also in the context of what AI models we can use. We do build and run our own proprietary AI models, which we've built and trained ourselves, but we also have hundreds of third-party models fully deployed on aiWARE available, including a lot of the large language models, Anthropic, Claude, and the like, and so how we do it in action, just think of the sheer explosion of all these different data sets or data silos that enterprises may currently have or they're producing every day: video files, images, text, social media, audio files, IoT sensor data, emails. All that gets ingested into aiWARE. We create that data lake, and then we have a robust and orchestrated cognitive and generative AI layer that makes sense of it all. Sometimes it's just trying to find the needle in the haystack.
Other times it's preparing that data to activate and monetize down the road. From just a scale and scope of our business today, in Q3, we just recently reported earnings of around $22 million in GAAP revenue. About $14.7 million of that is pure software-based revenues. At the end of September, we had about $46 million of total cash and cash equivalents. We have over 3,200 customers just in our software inside. We have a managed services division that uses a lot of our software as well, where we have frankly thousands of more customers. But to be clear, this KPI of 3,200 plus customers is software customers. At the end of Q3, we had about $63 million of annual recurring revenue, and our gross retention rate for Q3 was well over 90%, meaning we keep our customers. This is not novelty. This is not testing AI solutions.
These are groups that we've proven time and again that once we onboard customers, we keep them and we can grow them. In terms of guidance for the balance of the year, we guided to do between $92 million and $92.5 million and $93 million of GAAP revenue for the fiscal 2024 with a range of a non-GAAP loss between $36.5-$37.5 on a non-GAAP net loss basis. We did provide guide for 2025, and we are guiding to have a very, very strong year. This is after a couple of years of, I'd say, doing a lot of cleanup and sort of getting back to our core focus. We do have by far the largest pipeline that we've ever had. That's also now expanding into the public sector and the DOD.
And so we are projecting, again, to do up to 30%+ in year-over-year forecasted revenue growth and up to 45%+ in forecasted growth in our improvement in non-GAAP net loss. There's a lot of work to do, but to be clear, a couple of points about that is the pipeline is there, qualified pipeline. And we believe we already have made the investments and we have the products in good standing, meaning we don't need to go deliver a net new, brand new product to realize these revenues. So I think we're by far in the best shape to get this to a very strong grower for 2025 and through into 2026 and beyond. Just a quick shot of kind of what we do, a little bit more detail for the commercial side of our business. And for commercial, let's just pick the category of media and entertainment.
We have hundreds of marquee leading customers in media and entertainment that use our AI-based software solutions every day. They range from the ESPNs to the iHearts to the CNBCs, the Masters Golf Tournament. But fundamentally, for all these customers, they have lots of data. They have lots of unstructured data. Some have old archives. Some are producing content every day, like a CBS News. We ingest all that content for them. We stand up a data lake. We run the AI cognition. So we turn it into, I'll say, an intelligent data lake. And then we have a full application layer that turns that into value. So for the context of ESPN, we help them sift through the hundreds and thousands and millions of hours of content that they're exposed to, trying to interrogate that content to understand and cull it down to produce SportsCenter, right?
We also help their advertising and sponsorship department find all the logos and brand mentions through their programming so they can articulate to their advertisers how their campaigns are doing, what's the value and efficacy of those campaigns. And we also support their research department, providing them the insights of and do correlations between when this type of programming or hosts are airing, is that driving ratings? Is that driving extended video usage and viewership on YouTube, for example? So the use cases for our end customers all vary a little bit, but fundamentally, they all get the value. And collectively, it's become very efficient because of the scale of data ingest, data lake, intelligent data lake, AI cognitive processing, and application and workflow development. And they're some of the best and brightest.
We recently announced a major deal expansion on a multi-year basis with the NCAA that we are kind of the system of record and the licensing partner of record for their entire archive that goes back decades. That's one of many. So it's been a very exciting, very mission-critical part of our business. Public sector. It's a newer side of our business. This business, our focus on the public sector, started several years ago through inbound demand. Today, we have a very robust product offering of AI-based products and solutions that we're selling into state and local law enforcement as well as the public sector. And I'll just give you another example as I kind of did with ESPN. But for state and local law enforcement, they have a data problem. Their world now is not just about sworn officers and peace officers and homicide detectives.
They have to deal with tonnage and exponential growth of data from body cameras, from dash cams, from security cameras, drone footage, you name it. And so for them to try to thwart crime or investigate crime, the speed of their ability to ingest these complex, diverse data sets into a common data structure and then apply AI and cognitive processing around that to make sense of it all and then eventually turn it into an application so an investigator or a homicide detective can use is what we've done. So think of it like it's a Microsoft Office suite of products for law enforcement, which includes our applications like redaction, programmatic redaction. You don't think of it intuitively, but when you're producing all this audio and video footage of potential suspects, we have to release that by law to the public.
Historically, we're talking millions of hours that people manually sitting there going frame by frame trying to find and obfuscate faces and PII information. Our redaction product automates that with brilliant speed and efficiency. We have a product called Track, which allows you to track people of interest and cars across cameras without certain biometric markers such as face. In many jurisdictions and states and counties, you cannot use facial recognition right in the context of investigation unless you have a court order, etc. Track allows you to find hypothetically that person who's wearing the red sweatshirt and a backpack across cameras. We like to say it's our Jason Bourne, right? And that's one of many. So we are selling those powerful, efficient driving solutions. I kind of kid. They're DOGE ready, Department of Government Efficiency ready that is bringing real value to police agencies.
We now have almost 400 police agencies and sheriff departments across the United States that are using our products and services. More recently, we are now starting to sell those exact same products and solutions into the FedCiv, the DOJ, as well as the DOD. The one recent deal that was publicly disclosed was the DLA, the Department of Logistics, sorry, the Defense Logistics Agency. The beauty is the exact same product suite. So it's not like we're building net new applications or AI-based solutions. We're selling the exact same solutions into these larger agencies. Collectively, our pipeline for public sector between SLED and Fed is the largest in the company. To be clear, these are investments that we started to make years ago. The Fed Space grows, as many of you may know, very slow. It's hard to predict, but it's a long game.
You do have to wrap the right products and services. You have to have the right contract vehicles. But we're there, and 2025 will be our breakout year. And points of activity that you should look at as an investor is what we've been able to disclose with the $50 million blanket purchase agreement with the DOJ. What you should look at is the announcement of the DLA, which is a multi-million-dollar deal with the Defense Logistics Agency. More to come, a lot more to come, but we're very excited about this sector. Okay. That's the quick backstory. Just from a time check, I want to quickly just talk about what we announced today, which today is mostly going to impact and be a great growth driver for us on our commercial side. But we're calling it the Veritone Data Refinery or VDR.
As I touched on earlier, and let's take CBS News as an example. We have historically already been licensing technology and services to them where we're ingesting their content, we're indexing it, we're organizing it, and we're providing them great value from that workflow ecosystem. And we're helping them license their content for revenue that shows up in movies and on SportsCenter and stuff like that. So that's been a great engine for us, and we're continuing to grow that. However, a radical new use case has been presented to us over the last few years that we're in a prime and a very unique situation to capitalize and leverage, which is not just acting upon these large audio video data sets with AI for intelligence or prediction, but it's now using those audio and video assets as an input function to improve or train the AI.
So let's pick two examples: large language models and multimodal models. Large language models, as many of you are now familiar with, they were trained on their underlying transformer type of architecture on the readily available text-based data that's available on the public web, right? That's the super majority of the content. However, and if you shift over to multimodal models, right, they also need to train on audio, video, and images. We are actually running into a major resource gap that there's not enough new, differentiated, and clean publicly available free data. So it's why you're starting to see and read about some of the hyperscalers or some of the larger multimodal and large language model companies having to license content from walled garden sources or proprietary IP sources. Now you can kind of see what the introduction of VDR is.
We are sitting on one of the largest corpuses and libraries, library if you want to call that, already indexed and organized of premium audio and video. And similar to how we've been organizing and structuring this audio and video for our customers' internal use and for, frankly, licensing some of that footage to show up in TV shows, we are now starting to license that premium content to third-party groups who are looking to train their large language models and multimodal models to a whole another level or even fine-tuning. So, for example, they want to know, they want to be able to train or optimize their models to really understand the movements of professional athletes running on a field, or they really want to understand how an organic, natural conversation happens in a podcast.
We're sitting on some of the largest corpuses of finely structured and, in many areas, time-correlated data sets out there. That's what VDR is setting out to do. Again, let me highlight the problem set out there that we are introducing this new functionality to go tackle. Again, there's a definitive growing concern that large language models are running out of readily available data to train on, number one. Number two is the amount of publicly accessible human-generated text, right, is not enough to power or continue to differentiate at the level of advancement that is needed, right, to advance these large language models. You guys probably heard about some of this in the press. Multimodal models, again, think of audio and visual, require video, audio, and images to train, and they're running out of quality data sets even faster than text.
Some experts predict that we are there. We're already running into the limitation, right, as early as 2026. VDR, as I touched on earlier, already represents and is going to facilitate the licensing availability of the largest corpuses of premium audio and video that is already structured. You don't have to try to piece together. Veritone is a one-stop shop. It's the same process by which we ingest into our data lakes. We use our AI cognition and aiWARE to organize that content and prepare it for AI model training and next-generation applications, and the beauty is this is organic to us. We're already doing it. We're already processing their data already for other use cases.
So this is a great organic, natural extension of the value we can bring to our existing customers on the IP side, but also an entirely new revenue and growth opportunity for the business as we're going to be licensing this content to third parties. And the dollar amounts and the size are really exciting. If you want to look to some of the market dynamics around this, I'll mention just two companies that have been doing stuff sort of [inaudible] to this. One is called Shutterstock, one of the largest stock footage and stock video footage companies out there. And this is a whole new revenue line that they activated before us, and they've been monetizing this now for a few years.
And another company that's private is called Scale AI, which is one of the groups that really got into training data preparation and human annotation to prepare that data. It's a very large opportunity. We're very excited about it. And we do think that we're in a very unique and proprietary area, again, because of our unique understanding and relationships and, in certain areas, contracts and representation of these premium IP audio and video data sets. I'm going to go pretty fast because I want to get to Q&A, but I think you guys get the gist. But just to show you some of the scale, today, we process over 150,000 hours a day cognitively and with AI every single day for our customers. And again, those range from large sporting events broadcasters to, in certain regards, to now public safety institutions like law enforcement agencies. Let me skip ahead.
But what we have and what we're going to be mobilizing is both new content, but also archival content. Now, again, the farther back in time you go of activating a legacy archive for potential training of an AI model, the quality of the content is important. So as, let's say, the resolution degrades and the quality of the audio degrades, as you go back in time and it's stored on old film, right, the efficacy and viability of that as a training data can be impaired. And we want to make sure we don't have to spend too much money, I'd say, to improve or clean up the data. But to be clear, we have a huge amount of readily available high, super high-quality content that is available, that is actionable, right, to really help fine-tune and train and take these models to a next level.
I'm going to skip this slide. It goes a little bit more detail. But just give you some of the flavor of the content that in entertainment, feature films, scripted series, documentaries, you guys can read it for yourself. News, we have 60+ years of our iconic news footage. Also, new content, new film. This is not just archives. This is content that's being produced every day that's flowing into our system. So meaning we want buyers to understand we're a one-stop shop, which is a very powerful moat for us, which we think against competition. And again, just more flavor of the sort of variety. NCAA, the Australian Open, Alpha Media, it ranges, again, from scripted shows, Court TV, which was the show that aired the Menendez trial. Variety, variety, but high quality. And we can index it and package it at scale very fast for our end customers.
I think I kind of touched on this, guys. I apologize for timing. But I think I hit this over the head, is that Veritone is in the right place, the right time with the right scale that VDR is in a great position to be another major revenue source for the business, but also a very new, powerful new revenue source for our customers, our IP owners, which is a great spot to be in. And we're in this trend right now. You're going from spending a fortune on general large language models, which are awesome and amazing, but you're seeing a transition or growth of finer-tuned or narrow domain-specific large language models or domain-specific multimodal models. And that's where we're going to be a great partner for this ecosystem, fueling it for the future. Just scale, it's pretty interesting, right?
GPT-4 was trained on less than 1 PB of text-based data. JPMorgan's, right, where they created a kind of finely-tuned version, it's 150 PB, right? Just the CBS News, if they wanted to mobilize their entire database, and I put 10+ here, it's over 50. There's a tonnage of amazing high-quality audio and video content that's going to be made available to these models. So we're super excited about where this is taking us. And I think with that, I'll turn it back to Josh. I appreciate it for some Q&A, and then we can close up.
Awesome. All right. Well, let's get started with some questions here. So just as a reminder, if anybody has any questions, they can enter them in the Q&A box, and I can read those off here. Maybe just starting off with the VDR opportunity, what is the how does the end customer for that product ultimately, is it the same as your current set of customers? And how do you think about the business model structuring in terms of what is pricing going to what parameters just broadly is pricing based on? And maybe just touch on how does that compare to how the pricing for some of your other products works.
Yep. So answer question one is we do think that there will be some overlap customers that we're selling products and services to. As kind of we disclosed, we do sell certain products and services to large hyperscalers today. But for the most part, these are net new customers. So different model-based customers, some of the bigger names that you know, right, that are building some of the largest large language models. But really, anybody who is making a decision to build and train their own LLMs or multimodal models are in the market for customers. So again, we think it's both an upsell and expansion to our existing customer base as well as an expansion to net new customers.
How the market is kind of being set that typically for these premium libraries for audio and video, the cost basis is typically defined in terms of hours or minutes of footage with or without sidecar data files. So we're talking tonnage. So thousands of hours of audio or video with or without, let's say, time-correlated metadata like transcription or objects, et cetera. But think of it, it's based upon time. And that's somewhat consistent with the very mature model of licensing footage. When you license footage for your commercial or TV show, it's typically priced on a minute or hourly basis. We're seeing consistency of that in terms of brokering it in terms of minutes or hours of audio and video for training AI models. The revenue model does vary a little bit.
Some groups want to do all you can eat,[cannery] , and they come up with a kind of a negotiated upfront price, which is, again, somewhat consistent in the licensing world. We have certain customers that write, let's just say, a large annual check that's kind of all you can eat, or it's very à la carte. For us, this will be, in effect, a revenue share for the most part between Veritone and our IP owners of the content that we're ingesting and managing for them, and so we will get, I'd say, a healthy percentage of the gross dollar, so when a third party licensing this footage or this content for AI-based and model training purposes, Veritone will be able to capture a large percentage of that revenue basis on a revenue share basis. I hope that answered the question.
Yeah, that's super helpful. So how do you determine what large language model to use when you're creating one of these AI applications? And just more broadly, how do you kind of determine what use cases or workflows to ultimately kind of create? And I'm thinking more from the public sector side of the business, where I know you already have five and you're building more.
So it's a great question. First, in the public sector, they really need to understand where these models and data is ultimately going to be deployed. Unlike commercial, public sector has a different level of scrutiny as it relates to security, right, and data protection. Because of that point, let's say it has to go into a private cloud or, in some instances, a network-isolated environment, by default, what I just stated would eliminate, again, if you're trying to license a third-party model, you may not be able to invoke it through a public API and start using that large language model and start to tune it. You may have to find one that you can actually deploy in your own tenant, in your own environment, and train it yourself.
So first, again, specifically answering the question to public sector: you need to understand ultimately even before will this model as a baseline work for the end use case? You need to understand where you're going to be deploying this. And once you make that decision, then you could say, "Okay, do we want it to create a new vast-based model from scratch on open source and we're going to do it ourselves? Or we're going to license Claude from Anthropic and we're going to fine-tune it and deploy that in our private cloud," which you can. So, meaning of the availability now of an ever-growing amount of portable and licensable LLMs and multimodal models out there, you need to understand deployment. You need to understand data integrity. Do you have the rights to move data? I mean, does data have to stay on-prem?
And then you can evaluate the efficacy of these models. There's a lot of great documentation about the effectiveness of both LLMs and multimodal models about, again, if you're trying to do a RAG type of solution or now agentic type of functionality, there's a lot of research for it. Again, we'll remain agnostic. I don't want to name any groups right now. To be clear, from a search perspective, just from a scale and cost perspective, we've built our own, right, off of open source in our training up Veritone. And I'll call this our default search for Veritone. But to be clear, we also work with a lot of customers where using AI where we've deployed Claude or another solution. So the point is you have to map out what your end solution is. What are you trying to achieve?
That will dictate or predicate what type of model or models you should be using to deliver the end solution. You have to think of a multitude of different factors: cost, deployment, data integrity, et cetera.
Got it. That's super helpful. Okay. So in terms of the revenue guidance for 2025, the range between the high end and the low end is fairly wide. Maybe you could just help investors, give us a sense of what are some of the factors that could lead the revenue guidance to the lower end of the range versus the higher end of the range?
So I think by far the biggest variable is going to be the timing and expansion of a lot of our larger Fed contracts. It's the law of small numbers that we're dealing with a little bit. Our public safety division is the smallest today, but it's got by far the fastest, largest pipeline and projected fastest growth rate. But when those contracts go from, let's say, in theory, one military base to 50+, we don't have great clarity of when those expansions roll out. So that variance by far is dictated by the timing of these. However, we're hopeful that we'll be able to communicate the bookings.
So even if we see, candidly, that there's going to be a delay of a quarter for certain types of revenue, but we see this massive opportunity that's going to be a five-year deal. We hope our investors will be excited about the longer-term opportunity, even if we land the contract, but the deployment is actually pushed out three to six months. That's the biggest variance by far. There is some exposure relative to some of our consumption business lines. We just talked about the licensing business, not the VDR business, but the licensing business for footage. That's consumption-based. That could have some ebbs and flows. As we saw this past year with Bernie coming out of a lot of the Screen Actors Guild and the writers' strike, that's normalized, obviously, but there's some more exposure to some of our consumption-based businesses.
So I think those are two of the primary factors of why we have such a large range of guide from a revenue perspective next year.
Got it. We only have time maybe for one last quick point. We got about 60 seconds left. But I did want to touch on this just because I think it's important for investors. Is now you've done a series of cost cuts, you've kind of optimized the business going forward. How are you thinking about the capital allocation for 2025 in terms of what are you thinking about in terms of cash burn relative to your cash balance today versus the relative growth rate of the business? Maybe that's a good point to wrap up on.
Yep. So if you kind of look at our midpoint, at some point, we are going to have to raise some additional capital in some form or fashion in the future so we can get to a fully funded plan. We still are confident that we're looking at a 2026 profitability run rate as we move into 2026. But based upon, and people can do the math based upon our burn, we are going to have to have not a lot. Part of our frustration, not frustration, is just what we want to calculate here is we want to take very measured approaches. We do have a big chunk of cash parked from our recent divestiture of our legacy media agency. We do have about $24.7 million parked in escrow and as an earn-out.
That's a material amount of cash relative to our current cash balance and our cash burn. So again, as we progress through the years and some of those escrows burn off and we're able to collect that cash, we just want, again, we want to be measured. I'm very aligned with our shareholders. Us and a few of the larger shareholders are very in tune with dilution. There is a lot of demand for Veritone stock, and we have a lot of demand to raise capital. But to be clear, we want to be very measured, but measured to a point where we do not want to screw up on this huge opportunity. Hopefully, everybody's really excited what I've been talking about, and our public sector business is real, and it's going to be very large.
And so therefore, we want to make sure that if we do see momentum building from a contracting perspective, bookings perspective, we may elect to raise a little bit more cash just to make sure that there's no risk on fully realizing the opportunity. But again, just be very clear, as we sit here today at our midpoint, we will be needing to raise some additional capital in the future unless there's some earlier than sort of forecasted huge breakout earlier wins, which would be great. But again, I want to be clear, and I just want to be transparent with our shareholders.
Awesome. That's awesome. And thank you, Ryan, for the presentation today. And I think we can wrap it up with that.
Thank you. Thank you, Josh. Thank you, everybody. Bye.