Dashverse founders discuss pivoting to AI video, targeting power users, and scaling production budgets
Summary
Dashverse founders explain their pivot from comics and short dramas to AI video, citing a massive total addressable market of $500B-$1T. They contrast their strategy of targeting 'power users' with competitors like Higgsfield who focus on 'tourist users' in advertising. The founders report rapid client growth, with one account scaling from $10,000 to $500,000 in billing over five months. They acknowledge that while current margins are 20%, they aim for 80% steady-state margins, though they expect open-source models to eventually erode this advantage. Key adoption barriers identified include slow generation speeds and the need for a stronger 'aha moment' in user experience.
Key takeaways
- 1 talk more about that. And you know, uh Samyar, from you give us an example where I think you and I were talking separately. You have this forward-deployed artist model, right? Where I think you had mentioned you're getting a lot of input from these artists, which goes back to the engineers, which makes the models, which makes the harness stronger, which again, you know, so that's that loop.
Unidentified speaker (speaker_1) · 30:34
Episode sections
- 0:00 – 0:51
Intro: Where does value accrue in AI video?
Speakers discuss the potential for billion-dollar revenue with high margins in AI video. They debate whether Hollywood is finished and assert that creative storytellers are gaining power while traditional production budgets become less critical for quality.
- 0:51 – 3:11
The Dashverse journey: comics → short drama → AI video
The founders explain their evolution from comics to short dramas to AI video, driven by the expanding Total Addressable Market (TAM). They note that while they started with comics due to early image model capabilities, the rapid improvement of video models prompted a pivot to a larger market.
- 3:11 – 4:02
Why they killed a growing business to chase video
The company made the difficult decision to shut down a profitable short drama business to pursue the AI video market, citing the 'curse of building an AI' where goals constantly move. They chose long-form video to solve complex problems rather than just social/ads, leading to the development of their agentic platform.
- 4:02 – 30:34
Betting on long-form
The founders claim their platform, Frameo, is now the number one agentic video platform globally on benchmarks. They discuss the massive TAM for AI video ($500B-$1T) and the importance of reducing 'cost per show' to make AI production economically viable compared to human actors.
- 30:34 – 34:02
Forward-Deployed Artist Model and Feedback Loop
Speaker_2 explains the 'forward-deployed artist' model where artists work directly with clients to solve specific problems, feeding insights back to engineers to optimize the product. This creates a flywheel where client workflows, such as generating multiple variations for selection, inform new product skills.
- 34:02 – 39:15
Higgsfield vs. Dashverse: Advertising vs. Entertainment Strategy
Speaker_0 contrasts Dashverse's focus on entertainment and long-form content with Higgsfield's focus on advertising and short-form social content. Speaker_0 argues that while Higgsfield grew faster by targeting 'tourist' users, Dashverse targets 'power users' who spend more time and money, comparing the dynamic to Lovable vs. Cursor.
- 39:15 – 44:00
Client Growth and Production Budgets
Speaker_0 reveals that their largest client expanded billing from $10,000 in July to $500,000 by October, with expectations to reach close to $1 million the following month. Speaker_0 also notes that traditional studios like Netflix and Bollywood have significant production budgets that will allow them to adapt to AI video production.
- 44:00 – 46:17
Margins, Adoption, and Open Source Threats
Speaker_2 states that Dashverse currently maintains 20% margins but aims for 80% steady-state margins, prioritizing adoption over profit in the next 12-18 months. Speaker_2 acknowledges that open-source video models will eventually erode margins, similar to the coding AI market, making the current high-margin period temporary.
- 46:17 – 51:26
Generation Speed and Personal Agents
Speaker_2 identifies the lack of an 'aha moment' and slow generation speeds (30 seconds to 2 minutes for a 30-second clip) as the primary constraints, comparing the current state to the dial-up internet buffering era. Speaker_0 predicts that personal AI agents will drive mass adoption of AI, potentially leading to personalized video content consumption.
Claims worth verifying
One billion in twelve months is possible for top-line revenue with 80% margins.
Unidentified speaker (speaker_0) ·0:00
How to verify: Check Dashverse financial filings or press releases for revenue and margin data post-2026-08-17. For recent events, verify 2026-08-17 through 2026-11-17. Use another period only when the speaker explicitly names it.
Primary sourceDashverse started in 2023.
Unidentified speaker (speaker_1) ·1:34
How to verify: Check Dashverse company registration records or press releases from 2023. For recent events, verify 2026-08-17 through 2026-11-17. Use another period only when the speaker explicitly names it.
Frameo ranked #1 on Physion Labs' Arc v1.0 Movie benchmark.
Unidentified speaker (UNKNOWN) ·5:12
How to verify: Check Physion Labs' official website or GitHub repository for the Arc v1.0 benchmark results. For recent events, verify 2026-08-17 through 2026-11-17. Use another period only when the speaker explicitly names it.
The AI video market TAM is between $500 billion and a trillion dollars.
Unidentified speaker (UNKNOWN) ·5:25
How to verify: Review market research reports from firms like Gartner, IDC, or McKinsey on AI video generation market size. For recent events, verify 2026-08-17 through 2026-11-17. Use another period only when the speaker explicitly names it.
Hicksfield has raised to a billion-dollar-plus revenue run rate.
Unidentified speaker (speaker_0) ·14:03
How to verify: Check Hicksfield's press releases, funding announcements, or financial disclosures. For recent events, verify 2026-08-17 through 2026-11-17. Use another period only when the speaker explicitly names it.
Google uses its own model for Flow.
Unidentified speaker (speaker_0) ·17:54
How to verify: Check Google's official documentation or press releases regarding the Google Flow product and underlying models. For recent events, verify 2026-08-17 through 2026-11-17. Use another period only when the speaker explicitly names it.
Dashverse's Frameo ranked #1 on Physion Labs' Arc v1.0 Movie benchmark, described as a global first from India.
Unidentified speaker (UNKNOWN) ·30:34
How to verify: Check the official Physion Labs website or Arc v1.0 benchmark leaderboard for the ranking of Frameo/Dashverse. For recent events, verify 2026-08-17 through 2026-11-17. Use another period only when the speaker explicitly names it.
A client scaled from $10k to $500k/month in five months (July to October).
Unidentified speaker (UNKNOWN) ·39:30
How to verify: Verify via Dashverse's public case studies, press releases, or direct confirmation from the company's investor relations or founder communications. For recent events, verify 2026-08-17 through 2026-11-17. Use another period only when the speaker explicitly names it.
Dashverse currently maintains 20% margins but aims for 80% steady-state margins.
Unidentified speaker (speaker_2) ·44:40
How to verify: Check Dashverse's financial disclosures if public, or verify through founder statements in subsequent earnings calls or investor updates. For recent events, verify 2026-08-17 through 2026-11-17. Use another period only when the speaker explicitly names it.
Generation time for a 30-second clip is 30 seconds to 2 minutes.
Unidentified speaker (speaker_0) ·47:25
How to verify: Test the Frameo application or similar AI video generation tools to measure real-time generation latency for a 30-second clip. For recent events, verify 2026-08-17 through 2026-11-17. Use another period only when the speaker explicitly names it.
One Piece season 2 retention dropped to 35% of users due to the two-year gap between seasons.
Unidentified speaker (UNKNOWN) ·40:27
How to verify: Search for industry reports or articles citing Netflix or streaming analytics firms (e.g., Parrot Analytics, Nielsen) regarding One Piece season 2 retention metrics. For recent events, verify 2026-08-17 through 2026-11-17. Use another period only when the speaker explicitly names it.
AI usage is currently at 2-3% global penetration.
Unidentified speaker (speaker_0) ·49:26
How to verify: Consult global AI adoption surveys from firms like Gartner, McKinsey, or Pew Research Center for current user penetration rates. For recent events, verify 2026-08-17 through 2026-11-17. Use another period only when the speaker explicitly names it.
Higgsfield targets the advertising market and grew faster than Dashverse.
Unidentified speaker (speaker_0) ·34:02
How to verify: Review Higgsfield's public marketing materials, press releases, or funding announcements highlighting their focus on advertising and growth metrics. For recent events, verify 2026-08-17 through 2026-11-17. Use another period only when the speaker explicitly names it.
Open-source video models are catching up to closed-source models, similar to the coding AI market.
Unidentified speaker (speaker_2) ·44:00
How to verify: Monitor open-source video model releases (e.g., from Stability AI, Mistral, or community models on Hugging Face) and compare their performance benchmarks against closed-source models like Sora or Runway. For recent events, verify 2026-08-17 through 2026-11-17. Use another period only when the speaker explicitly names it.
Anthropic offers 90% API margins.
Unidentified speaker (UNKNOWN) ·45:13
How to verify: Check Anthropic's API pricing page and calculate margins based on compute costs and API prices, or refer to public financial analyses of Anthropic's unit economics. For recent events, verify 2026-08-17 through 2026-11-17. Use another period only when the speaker explicitly names it.
Primary sourceDashverse ranked #1 on Physion Labs' Arc v1.0 Movie benchmark.
Unidentified speaker (UNKNOWN) ·30:34
How to verify: Visit the Physion Labs website and view the Arc v1.0 Movie benchmark results to confirm the ranking. For recent events, verify 2026-08-17 through 2026-11-17. Use another period only when the speaker explicitly names it.
Do your own research
- Verify Frameo's claim of being the number one agentic video platform on benchmarks by checking independent leaderboards or third-party evaluations.
- Analyze Dashverse's revenue concentration risk by determining if the $500k client represents a disproportionate share of total revenue.
- Compare Dashverse's 'power user' strategy against Higgsfield's growth metrics and customer acquisition costs to assess the validity of the 'tourist vs. power user' revenue model.
- Research the timeline and capabilities of open-source video models (e.g., Sora, Kling, Runway) to evaluate the threat to Dashverse's margin expansion plans.
- Review Dashverse's customer churn rates and expansion revenue to validate the sustainability of the reported billing growth.
- Investigate the technical limitations of current AI video generation regarding consistency and control, beyond just speed, to understand the 'aha moment' barrier.
- Check for any disclosed conflicts of interest or equity stakes held by the podcast hosts in Dashverse or related AI video companies.
Full transcriptexpand
- 0:00
- Unidentified speaker (speaker_0): One billion in twelve months is possible. Right. So billion dollars of top line, 80% margins, huh?
- 0:05
- Unidentified speaker (speaker_1): We are recording this. Both
- 0:06
- Unidentified speaker (UNKNOWN): will
- 0:06
- Unidentified speaker (speaker_0): not happen at 10 months.
- 0:08
- Unidentified speaker (speaker_0): Both will not happen at the same time. That I'm 100% sure.
- 0:11
- Unidentified speaker (speaker_1): What's the future of creative jobs in this world the way you guys see it? And is is Hollywood like cooked? Is it like what happens?
- 0:19
- Unidentified speaker (speaker_2): There is this uh movie like where they say writer baby. Like this is the world where it is true. The writers who are actually making this the stories had, or the creatives had the least power. I think this is when they get the most power.
- 0:33
- Unidentified speaker (speaker_0): Even a bad movie with a lot of budget could do something great. That will go away. A good story can never find a budget. All these things used to happen.
- 0:56
- Unidentified speaker (UNKNOWN): Hi, today on
- 0:57
- Unidentified speaker (speaker_3): Intelligent Indians, Tarun
- 0:59
- Unidentified speaker (UNKNOWN): and
- 0:59
- Unidentified speaker (speaker_3): I are chatting with Sanidhya and Samidib, co-founders at Dashworth, to talk everything AI video. It is clear that AI video will generate multi-billion dollar uh multiple companies, at least we hope so. Today Tarun
- 1:11
- Unidentified speaker (UNKNOWN): and
- 1:11
- Unidentified speaker (speaker_3): I will ask these two intelligent gentlemen
- 1:13
- Unidentified speaker (UNKNOWN): where does value
- 1:14
- Unidentified speaker (speaker_3): accrue
- 1:14
- Unidentified speaker (UNKNOWN): in that? uh entire
- 1:15
- Unidentified speaker (speaker_3): market.
- 1:17
- Unidentified speaker (speaker_3): So welcome Tarun, Saradea, Swami Deep. Uh great to be doing this uh after a very long time. I think uh the journey has been pretty long. Uh
- 1:27
- Unidentified speaker (speaker_0): yeah.
- 1:27
- Unidentified speaker (speaker_3): Why why why don't you talk to us a bit more about it? And Tarun, it'll be good to hear about uh how the journey started. Yeah,
- 1:34
- Unidentified speaker (speaker_1): firstly welcome guys. Thank you so much for doing this. Uh Sanadia Sanadya Swamiadeep Swamyadeep, it's uh a pleasure to have you, and I know we've been planning to do this for a while, so long over to you. Uh I'll actually start with I think our journey with uh Dashtune, then Dashverse, Frameo started in 2023. I think we were uh privileged to be sort of early partners with you
- 1:53
- Unidentified speaker (speaker_1): guys. Uh and have also seen the evolution of both the the way the company has evolved from a focus on comic creation to micro drama to AI video and now the latest thesis, which is much more of a AI services play around video, and so we'll talk about that. Uh, but
- 2:13
- Unidentified speaker (speaker_1): you know, walk us through the journey and I guess from your vantage point how it's played out.
- 2:18
- Unidentified speaker (UNKNOWN): Last
- 2:18
- Unidentified speaker (speaker_1): three years.
- 2:18
- Unidentified speaker (speaker_0): Yeah, absolutely. It's it's been crazy three years.
- 2:21
- Unidentified speaker (UNKNOWN): Uh
- 2:22
- Unidentified speaker (speaker_0): every year we have picked up uh a kind of a new uh target. Uh and trying to big up bigger and bigger TAM as much as possible. We started with comics because
- 2:35
- Unidentified speaker (speaker_0): image models was the only thing that existed in that world.
- 2:41
- Unidentified speaker (speaker_0): And we as a company, you know, we always named our company Dashwars, Uh, but we have been known by different names because of the products uh that we have launched over time. Um the core core thesis and the core reason of uh all of us coming together to to build this company was uh the content infrastructure
- 3:01
- Unidentified speaker (speaker_0): layer is going to change. From cameras and mobile cameras, it will become more like uh you know devices uh which uh which are laptops or mobile phones and stuff like that. And in that technological shift and in that kind of a creative economy, where can we plug in into? Because the biggest opportunity lies when this
- 3:20
- Unidentified speaker (speaker_0): transition is happening. So that has been an idea. Now comics or short dramas or today, framework for largely all kinds of formats is just a journey around the core thing, and how we get to build framework. which exists today has been because
- 3:40
- Unidentified speaker (speaker_0): of that journey. So comics was uh the target segment which we thought uh we could disrupt early on because image models were really good uh back then. Uh and we build a lot around it. Uh in fact, all our foundations are built on uh building dash tune uh as a company in in tech or distribution uh
- 4:00
- Unidentified speaker (speaker_0): anywhere. Uh
- 4:02
- Unidentified speaker (UNKNOWN): but
- 4:02
- Unidentified speaker (speaker_0): then you know we couldn't imagine like video models will come so early and get so good.
- 4:09
- Unidentified speaker (speaker_0): And it's what a hundred X TAM than what we were targeting. And I've heard for all from all the other intelligent people that go for the biggest stamp if you can. So we decided to take a hard call, and you you guys were there in in that call as well, right? That there is a business. which is growing and which is doing good and
- 4:29
- Unidentified speaker (speaker_0): we are deciding to just stop it and go for the next big thing.
- 4:34
- Unidentified speaker (speaker_0): That is a curse of building an AI. The goal post is constantly moving so either you adapt or perish. I believe we took the right call to get into video.
- 4:44
- Unidentified speaker (speaker_0): We picked an index on one format again uh, like the last time let's go for short drama right because if I if even build a video tool today, the models are uh good enough, but not that good that anybody is going to use it for much, apart from maybe socials or ads. I mean, that was the option. We went for a longer format video and like solving more
- 5:04
- Unidentified speaker (speaker_0): complex problems, and uh that is how our internal tool, framework, uh, evolved over time.
- 5:12
- Unidentified speaker (UNKNOWN): And
- 5:12
- Unidentified speaker (speaker_0): uh now framework FrameU is like uh number one. agentic video platform in the world, at least in terms of benchmark, if not revenue. We we hope to chase revenue in the coming five to six months as well.
- 5:25
- Unidentified speaker (UNKNOWN): But
- 5:25
- Unidentified speaker (speaker_0): yes, something beautiful has come out of that journey, which we think is the biggest possible dam that you can build your company in. right? It could be anywhere between $500 billion or a trillion dollar as well. And we want to capture that market in a meaningful way.
- 5:44
- Unidentified speaker (speaker_1): Let me pull in Swamy Deep. So Samya, I know when uh we were talking to San India, he had left your name in stealth. He had said there's somebody who's really good at AI and tech, and he's he's gonna be joining us. So, firstly, very, very glad you joined. I know you're really looked upon in the ecosystem in terms of being at the frontier of what's happening
- 6:04
- Unidentified speaker (speaker_1): in AI.
- 6:06
- Unidentified speaker (speaker_1): Talk to us a little bit about this benchmark from with Fission Labs. It was around AI video. Talk to us about what does this benchmark really mean for the company? What does it mean for India and you know AI? And what's the I guess importance of you know uh uh why this benchmark matters
- 6:26
- Unidentified speaker (speaker_1): even?
- 6:26
- Unidentified speaker (speaker_2): Technically the benchmark is just uh
- 6:30
- Unidentified speaker (speaker_2): one proof of how good the product has become and that to only one slice of the product which is the harness. Uh but what it
- 6:39
- Unidentified speaker (UNKNOWN): at
- 6:39
- Unidentified speaker (speaker_2): least for us what it means is that like something we've been building for like practically three years now is now globally best. And that is something very exciting and it was very uh um
- 6:53
- Unidentified speaker (speaker_2): like exciting for all of us that we've been able to achieve this and uh like and and I don't think anyone in the in India right now is has built from India a global benchmark and being at number one. So that that's something we are very proud of. But
- 7:09
- Unidentified speaker (UNKNOWN): from uh a
- 7:09
- Unidentified speaker (speaker_3): actually on this benchmark the good the funny thing is that uh the former number one is also an Indian company. And that's always something to for all of us to be proud of. So I would correct you there.
- 7:20
- Unidentified speaker (speaker_2): That's correct, that's correct. Yeah, the former number one is also an Indian company, and like and it's actually a lot of fun uh when that happened because I actually got to know about the benchmark because of them. So like a lot of uh banter between uh me and Sankit on that. So a lot of fun. On what it means for uh the company and the product, I think for the company, I think it's just one
- 7:40
- Unidentified speaker (speaker_2): axis of how we can sell better. That's practically it. And it also having a competitive benchmark also keeps us on a on a good treadmill that we will keep improving ourselves because competitors will keep coming and uh hopefully we'll be at the top most of the time.
- 7:58
- Unidentified speaker (speaker_1): Just in uh I know Akash wanted to get into a few more deeper details, but just for I guess benefit of those listening in, if you are to break down that benchmark into narrative consistency, cinematic sort of whatever, like
- 8:11
- Unidentified speaker (speaker_1): why, how did what allowed Framework to be right on top?
- 8:15
- Unidentified speaker (speaker_3): I think you you guys mentioned 13 out of 16 dimensions.
- 8:19
- Unidentified speaker (speaker_1): So what
- 8:20
- Unidentified speaker (UNKNOWN): are those
- 8:20
- Unidentified speaker (speaker_1): dimensions? Why? What have you all done? Like, what's the sort of secret source that allowed you all to get all the way on the top? And let me throw a hard question at you: how durable is this? Because there are, I I think between number one, number two, I was seeing between August and October already the gap is narrowing. So uh, does it make you nervous? Does that mean that this is not durable? Like,
- 8:40
- Unidentified speaker (speaker_1): talk about that.
- 8:41
- Unidentified speaker (speaker_2): Sure. So, how that benchmark happened? I think so. One of the things that we have internally always been measuring because we were we have been making short dramas, is next. net cost of creating an output content. We have never been a token maxing uh uh we because we've always monetized
- 9:01
- Unidentified speaker (speaker_2): outputs and not tokens, like we've never optimized for token using, we've optimized for net output. So, what that meant for us as a product journey is that we have always tried to make sure you get the highest quality with limited human effort
- 9:17
- Unidentified speaker (UNKNOWN): at
- 9:17
- Unidentified speaker (speaker_2): minimum possible time. So all of our iterations went behind optimizing this net number of cost per show, is what you used to call it. What that meant is that all the AI issues of consistency,
- 9:31
- Unidentified speaker (speaker_2): quality,
- 9:33
- Unidentified speaker (speaker_2): consistency of character, visual continuity, and all of this, we had to make it into our harness so that we can get this number down over time. And in fact, in fact, like existence, it was an existential question for us that we need to reduce the cost per show, otherwise, this it makes no sense to even make short drama with
- 9:53
- Unidentified speaker (speaker_2): AI, right? And especially sitting in India, because our competition was Indian uh cost of actors and not American cost of actors. So that pushed us to the limit. limits as to that we need to make sure that max handling of these AI nitty gritties are through the product, through the harness, through the through
- 10:13
- Unidentified speaker (speaker_2): technology, and humans only come for a QC or a taste bar check. So this is what we optimized for. And that optimization led us that we have to solve for coherent. Once the script is ready, you cannot go. away from the script once the switches narrative coherence for a long time. Like at least the the benchmark is for one minute,
- 10:33
- Unidentified speaker (speaker_2): but we were doing it for one hour, right? Second is uh continuity, which includes character, visuals, clothes, locations, and like and we used to monetize that content. So people used to curse us if there was not enough continuity. You would see like comments in AI videos generally, right? And we were always at that you cannot we cannot be
- 10:53
- Unidentified speaker (speaker_2): bad at that for as best as we can. And then there was production quality, like we have to show something grand if you're showing it to the user, right? And it has to look really, really beautiful, really, really amazing to look at. Otherwise, no one's gonna pay money for that. And we were always monetizing the content. So those things were like what these bet this benchmark actually surfaced and
- 11:13
- Unidentified speaker (speaker_2): gave words to. We were doing it internally. These are not the terms we were using. These are the terms that they gave us, gave it to. We were just looking at key, how do you make this content this well so that it quantizes well for the user? So we were chasing the same thing with an end-to-end cost, and it just turned out that the benchmark was measuring on the same axis, and
- 11:33
- Unidentified speaker (speaker_2): which is how we got to number one. On the
- 11:38
- Unidentified speaker (speaker_2): piece of around durability, I am sure that more people will come and it will become a hard fight. But the interesting part is that, like, as more and more people use Frameo, and we have the feedback loop that goes on, we have our harness that keeps getting learning from the user, and then we have
- 11:58
- Unidentified speaker (speaker_2): more access of use cases that open up, we it will it'll compound. compound as we go. So I'm not as scared on the durability part. It's I'm sure there will be a time we'll be number two, then we'll fight back to number one. But that's the journey of any benchmarking, right?
- 12:12
- Unidentified speaker (speaker_3): But I think the scary bit is fine, we're not scared about that. But it's something that Tarun and I worry a lot about is uh,
- 12:19
- Unidentified speaker (speaker_3): you know, how it's all great, benchmark is commendable, maybe it's durable, but how does it really capture value in the AI? video market? And Sanadev would love to hear from you that you know at one end you have uh the bottom of the layer, which is infra servers, GPUs, everything. Then you have foundation models,
- 12:40
- Unidentified speaker (speaker_3): and then somewhere in between it is, you know, the likes of Fremio, Hicksfield, Nvidio, everyone, right? And on the top obviously you have the IP owners and the platforms. Historically value has accrued to the top.
- 12:53
- Unidentified speaker (speaker_3): Seems like now the production means are shifting to compute. So there is value accrual at the bottom. Hicks will have raised to a billion dollar plus revenue run rate as claimed. Where does the slare and where does frame your capture value?
- 13:06
- Unidentified speaker (speaker_0): Yeah, it's it's it's a super hard question, first of all, right? Because what does the value chain look like in the future is very hard to predict across. right? Even you could say that for model apps do they accrually value, or the open source commoditizes it and people switch, right?
- 13:23
- Unidentified speaker (speaker_0): Or the applied AI layer actually captures any value. So we I think are a part of applied AI layer, right? What benchmark immediately did for us was just one thing. People started believing. Especially, I would say. large production houses and enterprise kind of a customer.
- 13:43
- Unidentified speaker (UNKNOWN): And
- 13:43
- Unidentified speaker (speaker_0): I was so surprised that all the meetings I've done in last one month, when I'm showing them the output that you can get, the basic question is that no, this cannot be this is some hybrid workflow or you have short or something. So the awareness is so low, right? Extremely low. And that is what I think the big play. players tells
- 14:03
- Unidentified speaker (speaker_0): about the global TAM of intelligence as well. That what is the percentile of the population who is actually using it? Even the chat GPT or Gemini. So same holds for content market. That the penetration is so low that billion dollar revenue seems like peanuts, where Hicksfield has. It has to be super easy, and that is what Hicksfield
- 14:23
- Unidentified speaker (speaker_0): has proven. right? That in 181 seconds easier I'm faster than cursor as well. Right. Uh that that's just the size of time.
- 14:30
- Unidentified speaker (UNKNOWN): Now,
- 14:31
- Unidentified speaker (speaker_0): how does uh this monetize is that you get people to start trying out, right? Especially the big studios getting into AI who are against AI. Right? When they try it, and when they see what we have seen in short drama or what China has seen in short drama, their appetite will not be linear increase, right? They will not do first the movie and
- 14:51
- Unidentified speaker (speaker_0): then do the second one. You know, after first they will do 10, after 10, they'll do 100. And because the production time will be so compressed, at least if the script is ready, right, which is the creative part, which like A cannot solve. You will see a very J-Curve happening in terms of consumption, and that should translate into revenue. That's our hope. In the next
- 15:10
- Unidentified speaker (speaker_0): 12 months, it should be a very non-linear uh expansion of revenue.
- 15:15
- Unidentified speaker (speaker_0): And benchmark maybe was a medium for us to convince them that you know, okay, we are good, now you work with us, and then we show them the output, uh, which is also the capability of actually foundation model. I mean, like people are not using it, that's why they don't know, but of course, ByteDance is doing a great work with CDAS because of which we are able to build a application
- 15:35
- Unidentified speaker (speaker_0): layer on top of it. Without that, we couldn't do anything. Right. So I think the value is going to see very meaningfully and extremely fast. I mean, just wait and watch. So
- 15:46
- Unidentified speaker (UNKNOWN): I want
- 15:46
- Unidentified speaker (speaker_1): to interject there. So OpenAI, Google, Byte Dance, these will keep getting better. Right.
- 15:52
- Unidentified speaker (UNKNOWN): Over time.
- 15:52
- Unidentified speaker (speaker_0): Yes.
- 15:53
- Unidentified speaker (speaker_1): So part of the question is the framework harness that you guys have built. Over time, does it keep getting better? Or as the underlying models get better, framework actually becomes less relevant? So question part uh sort of first part. Second one is just to make it again real and a little bit more
- 16:13
- Unidentified speaker (speaker_1): tangible, what can a filmmaker do today with framework where which is more than what they could do by just stitching together some of these tools? Right? And so, given maybe with the your example. example, talk about that.
- 16:28
- Unidentified speaker (speaker_3): And just to add to it, uh not just models, right? Uh BatDance has Capcut, the Google has Flow, they are also building these systems, right?
- 16:38
- Unidentified speaker (UNKNOWN): They're going beyond just
- 16:39
- Unidentified speaker (speaker_3): being a model. Yeah. How how do you guys uh
- 16:41
- Unidentified speaker (UNKNOWN): build something
- 16:42
- Unidentified speaker (speaker_3): defensible?
- 16:42
- Unidentified speaker (speaker_0): GPT is building dots to compete with the T see that's gonna happen uh in all uh very big TAMs like right? like like these. Uh all of this is going to happen.
- 16:53
- Unidentified speaker (UNKNOWN): Um
- 16:54
- Unidentified speaker (speaker_0): but as of today, and in the in the journey we are, uh at least I could say next 18 months. I see that all of this improvement in foundation model complements, right? Because what the end consumer wants and uh where wherewith with we are at AI, uh there is a huge gap, right? Uh maybe the
- 17:14
- Unidentified speaker (speaker_0): consumer doesn't know. but at the end of the day, what do you want? Okay, I have a crazy idea, I have a lot of creativity. I go to some ABC tool and can I make a Netflix quality movie out of it? Right. In very simple terms. Is it possible to be even on framework? You'll have to do a lot of work to get a Netflix quality movie, even if you have a great script,
- 17:34
- Unidentified speaker (speaker_0): right? You have to do a much less work on framework Frameo versus uh other tool, right, which is not built for uh such long-form content. And that delta itself is also very huge. If you don't use any tool, uh you say okay, uh model API is also available, right? Or I will go to um Google Flow to do it, which has a inferior model because
- 17:54
- Unidentified speaker (speaker_0): Ctan says a better model, Google won't, uh Google uses their own model, right? Then it will become even harder, and then the your final product will also be worse than that.
- 18:03
- Unidentified speaker (speaker_0): So
- 18:05
- Unidentified speaker (speaker_0): Fremeo is giving a lot of value today. Can it be cannibalized by a foundation? I I will say definitely they should try to. If Google has a much better video model than Seed Dance, they should try to cannibalize, right? But there is also open source which is happening
- 18:25
- Unidentified speaker (speaker_0): in the world. Uh, There is also a lot of data that I'm accruing today, right? How how I achieved a movie benchmark when I was building for short drama, right? Because it was storytelling and a lot of things could translate, and then there is a lot of secret sauce into building the product when uh real consumers are using it and it doesn't have a lot of noise.
- 18:45
- Unidentified speaker (speaker_0): If we continue to do that, right, we continue to make uh movies which are commercially successful, then we'll have a an intelligence that we could productize and give that product to a lot of creative people who doesn't understand maybe lighting so well. But the model itself, the tool itself ensures that the lighting will be very good. So there are
- 19:05
- Unidentified speaker (speaker_0): a lot of nuances into something like content, and end of the day, building it in this industry, we are a bit comfortable because it's not maths, right? So maths. is easily to copy, uh right, uh easily to be cannibalized by the model here because uh the verification loops are really sharp, right? You know, you know, like hundred percent you can
- 19:24
- Unidentified speaker (speaker_0): verify. Uh, but in content, uh, there is a taste element,
- 19:29
- Unidentified speaker (UNKNOWN): uh,
- 19:29
- Unidentified speaker (speaker_0): there is a creative element which uh is different for different users. So if you get the users today and they build their taste on the product, it is very hard for them to. get the same output by just exporting a.md file to another tool. It becomes really hard because you know the contextual engineering is too sophisticated for a tool like
- 19:49
- Unidentified speaker (speaker_0): Framework compared to other kind of harnesses for different use cases.
- 19:54
- Unidentified speaker (speaker_0): So yeah, your next question was sorry.
- 19:58
- Unidentified speaker (speaker_1): The second one was: give us an example of what can a filmmaker do uh on film. Frameio, which is perhaps you know a lot more than what they could do by just stitching together a bunch of these sort of you know AI tools.
- 20:09
- Unidentified speaker (speaker_2): Sure, so the idea of when we build framework is like we always wanted to close the gap between what you think and what comes out, right? Like and that gap is what in historically, like in movies it is that you have to shoot and then you will see the finally the director takes
- 20:29
- Unidentified speaker (speaker_2): sees it the editor is cut here you see it on the product. But in the generative AI world in any other tool you need to understand so many tools. You need to understand the prompting of it. You need to understand the prompting of a music music then sound then SFX bring them together in a place and then do that entire setup and then you get to see the final output and then you decide if okay this is
- 20:49
- Unidentified speaker (speaker_2): good or not. But what framework is trying to do is that all of this is obfuscated with our agent. You don't care about this. You care about your story, you care about how you want to tell your story, you make hundred variations for a scene, see how they feel like, and then you decide what goes out.
- 21:05
- Unidentified speaker (speaker_2): All of that thing is completely removed from you. You don't need to care about what model gets used, you don't need to care about like how a model gets used, all of these are not things as stories. a storyteller cares about.
- 21:16
- Unidentified speaker (speaker_0): You can put your budgets. Like you
- 21:17
- Unidentified speaker (UNKNOWN): want
- 21:17
- Unidentified speaker (speaker_0): to use a low mode, that means
- 21:19
- Unidentified speaker (UNKNOWN): that
- 21:19
- Unidentified speaker (speaker_0): uh essentially you don't want to spend so much time, but because maybe you're doing just rough brainstorming, right? You don't want uh 4K output over
- 21:28
- Unidentified speaker (UNKNOWN): there,
- 21:28
- Unidentified speaker (speaker_0): you just want 48 TP. That simple modes we have given, but not like you have to select everything or understand everything or select a model itself, right? That uh okay, this model I should use because it is uh cheap and I want to do only pre-visualization, I'm not making a final product. All these are productized things, right? How do you give it
- 21:48
- Unidentified speaker (speaker_0): to the end consumer?
- 21:49
- Unidentified speaker (UNKNOWN): So
- 21:50
- Unidentified speaker (speaker_3): but let's say all of this works out, and framework is great, it's durable, uh, you sustain the advantage, you don't get uh muddled up with the foundation models building something on the app layer, given. But historically
- 22:03
- Unidentified speaker (speaker_3): uh if you look at it when uh let's say digital photos did the same thing for uh you know photography, right, value went to the top of the funnel platform. Value went to the aggregator which solved for curation, distribution, identity, right? It did not accrue really meaningfully or in
- 22:23
- Unidentified speaker (speaker_3): proportion uh to the production system, right? Why would it happen this time?
- 22:30
- Unidentified speaker (UNKNOWN): I
- 22:30
- Unidentified speaker (speaker_0): would say still maybe distributor will have a bigger value. Right. Maybe. Right. What do you mean by distributor? Is there a distributor like who releases uh content in theaters or the distributor is YouTube over here?
- 22:43
- Unidentified speaker (UNKNOWN): Everyone, anyone who sits last
- 22:45
- Unidentified speaker (speaker_0): mile Smile in front of the consumer. consult. So uh if if it's a 500 billion dollar market, let's say entertainment market. distributor take a 50% chunk of it, but having users of the creative community at your platform, if you can consolidate, so YouTube never consolidated the creation, people were
- 23:04
- Unidentified speaker (speaker_0): using hundred kinds of tools, like different kinds of cameras, maybe Sony, maybe something else, right?
- 23:12
- Unidentified speaker (speaker_0): But over here, you will have that creative. community consolidating into one place. That will give you a different kind of insight where distribution doesn't become a mode. The moat is which production system can predict the next hit.
- 23:27
- Unidentified speaker (speaker_0): Right. Who can understand what's the taste of because the intelligence will also flow between creators who are creating over here. So the intelligence is building over here. Right now, the intelligence was on distribution. because RECO algorithms were the closest thing to AI intelligence that you could suit. Now the intelligence is on creation side.
- 23:47
- Unidentified speaker (speaker_0): Right. What kind of visual identity users like? I have a story, right? And I have an idea that you know this should go visually like this. But framework will recommend no. Why don't you try this visual identity? Right. So a lot of intelligence is going to accrue in the production. there, which was not true for an Adobe or camera doesn't
- 24:07
- Unidentified speaker (speaker_0): know any intelligence, right? It's just how you use it.
- 24:09
- Unidentified speaker (UNKNOWN): But
- 24:09
- Unidentified speaker (speaker_2): historically, we never had a chance to have all of your production purely digitized. This was never a this is never there in the world before. Now is the time that you can see that. And not only when I say all of the production, it means from your whatever writing script to the final output. to seeing what works on whatever
- 24:29
- Unidentified speaker (speaker_2): distribution channel and then closing the loop across creators. This was not possible before. So, like the which is why it's a very hard question to say where value will accrue, but this entire paradigm is new today, and what we believe very strongly is that the distribution
- 24:49
- Unidentified speaker (speaker_2): works because the creator. works, right? If we can accrue the value of the creator, likely there will be a lot of value created. Will it be larger or smaller than distribution? Very hard to say today.
- 25:01
- Unidentified speaker (speaker_1): So give me an example where I think, and maybe again would love to hear both your views. If I look at the future of AI entertainment, uh, how far are we from a full-length full length hundred percent AI movie that makes it to sort of you know
- 25:21
- Unidentified speaker (speaker_1): mainstream sort of uh consumption?
- 25:23
- Unidentified speaker (speaker_2): So I argue I will argue actually key we are there.
- 25:27
- Unidentified speaker (UNKNOWN): Are we weeks,
- 25:27
- Unidentified speaker (speaker_1): months, years?
- 25:28
- Unidentified speaker (speaker_2): I would argue we are there today, it's just adoption has its own speed. Like I don't think technology is blocking that. It's the adoption of the of the best directors and creators, which is blocking that today, is my argument. If there are certain chinks in technology, next six months, let's put it that. That
- 25:48
- Unidentified speaker (speaker_2): that is how fast it's
- 25:49
- Unidentified speaker (speaker_0): I'll I'll I'll interject. See why agents are so important and why we are trying being so bullish about building this agentic, right?
- 25:57
- Unidentified speaker (speaker_0): The models have capability, right? Even if you talk to a researcher in enthropic. hypothetically, right? They have released a very capable model, but they don't know what all this model can do. They don't know everything. They get to know from Twitter. Okay, it can do this only. They know maybe top ten use cases. Same is with video model. The models
- 26:17
- Unidentified speaker (speaker_0): are out there, but even the lab doesn't know what all it can do. It's the usage who comes to know what all is possible. Is AI movie possible or not? Now we are sitting at the intersection of it.
- 26:29
- Unidentified speaker (speaker_1): You're seeing people actually using it.
- 26:30
- Unidentified speaker (speaker_0): Yeah because see director doesn't know how to best utilize the model right he has to learn something called prompt engineering. What agent has done is remove that prompt engineering thing out of the contest. It happened this year. Correct. Right. Now I can make an agent who knows how to prompt
- 26:46
- Unidentified speaker (UNKNOWN): the
- 26:47
- Unidentified speaker (speaker_0): video model and also understand what director is really trying to do. because he's a creative guy and he's giving creative input. This is the value of agent to be a layer between the model and if you directly use API, I give you a window of 4,000 characters. You won't even know how to prompt 4,000 characters. But you might know how to tell my story. You can say in natural
- 27:07
- Unidentified speaker (speaker_0): language, right? And this agent has to understand that how to get the best output out of this model. That is a gap. The day people start using. it, which will be two months or three months, because we have given started giving and a lot of adult tools are also doing it, you will see a sudden adoption curve.
- 27:22
- Unidentified speaker (speaker_1): Wha what's the future of creative jobs in this world the way you guys see it? And is is Hollywood like cooked? Is it like what happens?
- 27:30
- Unidentified speaker (speaker_2): I think it's far more exciting because now, like, the real creative storyteller has all the power. There is this uh movie.
- 27:39
- Unidentified speaker (speaker_2): like where they say writer Baby. Like this is the world where it is true. Till late, though all the value in the creative industry used to be taken by these production crews, pr people who got us to pay the money, and the actors and everything. But the writers who are actually making the stories had or the creatives
- 27:59
- Unidentified speaker (speaker_2): were had the least power. I think this is when they get the most power. right? And of course, actors will have their own value because they bring distribution at the end of the day. But true stories that were not getting told before will start getting told before. I will get my thousand sniper dramas that I want to watch everywhere with because no one would fund thousand sniper dramas. Now I don't need
- 28:19
- Unidentified speaker (speaker_2): people don't need that. So it's a far more exciting thing for uh creators for sure. What about Hollywood? I think any
- 28:30
- Unidentified speaker (speaker_2): gating is getting broken by AI anyway, or at least getting smaller by AI because things are getting decentralized. And that will happen. And that's that that's always been the case in the
- 28:41
- Unidentified speaker (UNKNOWN): world.
- 28:41
- Unidentified speaker (speaker_3): But in that case, a harder question, right? Let's say all of this happens.
- 28:46
- Unidentified speaker (speaker_3): Does that not reduce the value of the final asset? So you're you're entering a world of infinite content, infinite production. You have gone through the journey yourself of seeing this with dash reels, right? What does this mean in terms of
- 29:02
- Unidentified speaker (speaker_3): the actual value of this creative IP?
- 29:05
- Unidentified speaker (speaker_0): See, can it be economical for one person to create uh content for another single person consumption? right Is the ultimate question. right or can I create content for myself just to consume would I be inclined right can that happen you have
- 29:25
- Unidentified speaker (speaker_0): enough
- 29:26
- Unidentified speaker (UNKNOWN): uh
- 29:26
- Unidentified speaker (speaker_0): you know green shoots to say the the speed of uh generation on video model will be close to uh you know buffering happens on internet
- 29:38
- Unidentified speaker (UNKNOWN): right
- 29:39
- Unidentified speaker (speaker_0): and uh if model has enough intelligence to see everything that was created in past and I want something that was created in past right similarity like he said he wants to just for a sniper movie uh do some adjacent play right but just create it for me
- 29:58
- Unidentified speaker (UNKNOWN): yeah
- 29:58
- Unidentified speaker (speaker_0): if you talk about a black pill moment that can happen
- 30:01
- Unidentified speaker (UNKNOWN): right
- 30:02
- Unidentified speaker (speaker_0): so
- 30:02
- Unidentified speaker (speaker_1): absolutely
- 30:03
- Unidentified speaker (speaker_1): historically, there's been a strong correlation between production budgets and the storytelling output.
- 30:10
- Unidentified speaker (speaker_0): Right.
- 30:11
- Unidentified speaker (speaker_1): Right? In your mind or the quality of the storytelling, right?
- 30:16
- Unidentified speaker (speaker_1): Does AI fundamentally change that equation?
- 30:20
- Unidentified speaker (UNKNOWN): Uh is
- 30:22
- Unidentified speaker (speaker_1): is it just lower cost? Is it just faster throughput? Is it just new? possibilities of AI video entertainment? Like what does this unlock?
- 30:32
- Unidentified speaker (speaker_0): It is all of it.
- 30:34
- Unidentified speaker (UNKNOWN): So talk
- 30:34
- Unidentified speaker (speaker_1): talk more about that. And you know, uh Samyar, from you give us an example where I think you and I were talking separately. You have this forward-deployed artist model, right? Where I think you had mentioned you're getting a lot of input from these artists, which goes back to the engineers, which makes the models, which makes the harness stronger, which again, you know, so that's that loop. So talk about that, and I guess what's what's possible? Like,
- 30:54
- Unidentified speaker (speaker_1): give us one real example of a capability that was unlocked because of this feedback loop, and hence what does it mean for the future of AI entertainment?
- 31:05
- Unidentified speaker (speaker_2): Sure. So
- 31:07
- Unidentified speaker (speaker_2): the reason to do a power-deployed artist is that uh firstly, we have been really strong with short trauma, drama, right? And the problem, pro like we've been golden. good and we know that the product is the absolute best for short drama. But while we open up new uh kinds of videos, because videos have some insane number of kinds also, like how do you build depth there? So the
- 31:27
- Unidentified speaker (speaker_2): way we decided we're gonna go there and build depth as well as it is great for the client also, is that we'll send a forward deployed artist, they will work with the client to solve the problem and then uh they will uh uh and then we will learn from the product and we will optimize the product, right? So like I'll explain one very creative uh use
- 31:47
- Unidentified speaker (speaker_2): client, right? The client wants to teach some really complicated uh complicated thing, like I'm not mentioning what.
- 31:57
- Unidentified speaker (speaker_2): The users, right? And they don't care about price. They care about you have to be technically correct. you have to be,
- 32:04
- Unidentified speaker (speaker_2): but I want to choose some like I will generate some seven things before I choose which one is the right one, and then decide what goes into the final. So this was their core workflow that we arrived at. Now our product was always optimized for like okay, you do this and you get the first shot, it
- 32:21
- Unidentified speaker (UNKNOWN): will
- 32:21
- Unidentified speaker (speaker_2): be the best first shot. You iterate and then you get there. But this is a different thing. So the forward deployed artist. came and told us this. The way we look at it is that our harness powers a certain set of workflows or skills that we call it. And we build a skill, built a skill with the which understood the technical nuance using a very powerful model, plus
- 32:41
- Unidentified speaker (speaker_2): gave the output in many variations that the creator can
- 32:46
- Unidentified speaker (UNKNOWN): then
- 32:46
- Unidentified speaker (speaker_2): could choose, and then that was the sk skill or a workflow they want to. follow. That I'm gonna make some like end decisions every time between seven or six options and keep selecting them, and that is what they want to follow. So this is a very simple nuance that a client wanted, and that's a very fair ask because and that is possible today. In a like as we were discussing, right? In
- 33:06
- Unidentified speaker (speaker_2): a in a if he was live action shooting this thing, he won't have those options.
- 33:10
- Unidentified speaker (UNKNOWN): You
- 33:10
- Unidentified speaker (speaker_1): can't shoot seven scenes seven
- 33:11
- Unidentified speaker (UNKNOWN): different ways and just not actually should
- 33:14
- Unidentified speaker (speaker_2): do three three cameras.
- 33:15
- Unidentified speaker (speaker_2): You can't even change camera angles midway, like
- 33:18
- Unidentified speaker (UNKNOWN): because your
- 33:18
- Unidentified speaker (speaker_2): rigs are all fixed. So all of these things just got done with a skill and like between in seven days they they went from like a like numbers like like a one K customer to a five K customer, right? That's that's that's in seven days. And that's that's the and also because they're gonna spend more, right? Because that's the imagination that opens up when you're not
- 33:38
- Unidentified speaker (speaker_2): limited. by these
- 33:40
- Unidentified speaker (UNKNOWN): constraints
- 33:41
- Unidentified speaker (speaker_2): of the real world.
- 33:42
- Unidentified speaker (speaker_0): But so this feedback loop is great, right? And you have this forward deployed artist loop, and Tarun spoke about the full flywheel. But it seems like this flywheel is probably far, at least in you know, super accelerative for people like Higgsfield because they are pointing themselves predominantly towards the advertising
- 34:02
- Unidentified speaker (speaker_0): market, right? That's the majority of the users.
- 34:04
- Unidentified speaker (speaker_3): Social industry.
- 34:05
- Unidentified speaker (UNKNOWN): You guys
- 34:05
- Unidentified speaker (speaker_0): have chosen to be more entertainment, film, production, short drama oriented, right?
- 34:12
- Unidentified speaker (speaker_0): And it seems like the difference is Higgs field has gone to a billion dollar much faster.
- 34:18
- Unidentified speaker (speaker_0): You are pretty far from it.
- 34:21
- Unidentified speaker (speaker_0): How do you think of making those choices? What does this mean for the business in the long run? Why pay pick uh entertainment and not point yourself towards the digital advertising market? Yeah, I mean that I think A that was a choice uh that happened when we started the company, right? Because we wanted to play in uh
- 34:41
- Unidentified speaker (speaker_0): entertainment space, right? We were never building an ads company.
- 34:45
- Unidentified speaker (UNKNOWN): Um somehow
- 34:47
- Unidentified speaker (speaker_0): Xfield was also building in short dema, demo, as far as I remember. Uh but they
- 34:52
- Unidentified speaker (speaker_0): kind of pivoted into a market which was faster and more ready. And so ads market, I believe, is the readiest, right? And social, so definitely is the first entry point where it started with the AI's law, but it was very obvious that you know social and vertical platforms will be the first one to explode.
- 35:10
- Unidentified speaker (UNKNOWN): But
- 35:10
- Unidentified speaker (speaker_0): yeah, I mean it wasn't a choice that we deliberated, but now we have landed here where we have an edge on something which in terms of TAM is also pretty huge, maybe bigger than maybe 2 to 3x bigger than because it is a huge uh consumption product, right? Very long format.
- 35:29
- Unidentified speaker (UNKNOWN): Uh now
- 35:31
- Unidentified speaker (speaker_0): can will we grow as fast as them? Uh time built in. Right. Uh but I I I feel uh the way the models will. improve, right? Because the base layer is very good.
- 35:42
- Unidentified speaker (UNKNOWN): I
- 35:42
- Unidentified speaker (speaker_0): think it they took 18 months, maybe we'll take 12. Right. So it can't
- 35:47
- Unidentified speaker (UNKNOWN): all
- 35:48
- Unidentified speaker (speaker_0): claim. Yeah. That can happen. Yeah. Right. Because see, I can see the output. I can see. What I have to figure out is GTM and how to convince users to come and try it out. And the economics should make sense, right? But if I figure that out, it's going to be faster. than ads. Because you know, end of the day, people on Higgs field are mostly making thirty second to one
- 36:08
- Unidentified speaker (speaker_0): second, one minute kind of an output. Here every user will start with the at least ten minute YouTube video. Right. So it should be much more nonlinear in my sense.
- 36:17
- Unidentified speaker (speaker_3): So talk to us about what's happening on that front.
- 36:21
- Unidentified speaker (speaker_2): At it is the difference is that are you making it for a power user or are you making it for a tourist? I think
- 36:30
- Unidentified speaker (UNKNOWN): the
- 36:30
- Unidentified speaker (speaker_2): Hicksfield also gets a growth spurt because they go for a tourist which we're gonna try a lot and then they will grow really fast because you can reach them reach them really quickly. I I am not making the 12 months claim because it's what the the power user we are going behind is a much more serious user. They need some coddling, hand
- 36:50
- Unidentified speaker (speaker_2): holding. to understand what they're coming for, but once they are there, they're gonna spend an insane amount of dollars on it. To give you some examples, it's the difference between a lovable and a cursor/slash a cloud code, right? A lovable goes behind a user who's gonna use the product once or twice a week, and they're gonna make
- 37:10
- Unidentified speaker (speaker_2): one website, one small app, and then they're gonna come back and play with it and do something. But that's not a cloud code or a cursor user. I'm gonna I'm a cursor cloud code user. I'm gonna stay on that product all day, 12 hours a day. And the amount of tokens I spend is far more than anyone spends on uh lovable for sure.
- 37:27
- Unidentified speaker (speaker_3): So actually that analogy is very interesting, right? Uh if we go back to uh the coding or STLC world, right? You have lovables, you have cursors, and then you have uh companies like Cognition who are more enterprise, right? And if you think of the B2B world of AI video in a similar fashion, where yes, you're saying
- 37:46
- Unidentified speaker (speaker_3): cursor, right? But you know, Tarun asked you earlier about is Hollywood cooked, uh, you spoke about uh independent creators a lot, market bifurcates, right, broadly. You'll have one end, you'll have what you guys have earlier uh told us about saying content five. factories, where these big studios, short drama companies become
- 38:06
- Unidentified speaker (speaker_3): fast through port, lot of volume.
- 38:09
- Unidentified speaker (UNKNOWN): And
- 38:09
- Unidentified speaker (speaker_3): you have the independent ones, right? How do you think of go to market in that case?
- 38:13
- Unidentified speaker (speaker_0): Yeah, so uh the simplest answer is uh the current production budgets, right? And that is why I won't say Hollywood is cooked, right? Because uh see for even this year they have billions of dollars of production, even Bollywood has, right? If they will transition today, because Creator doesn't have that kind of budget, so they will
- 38:33
- Unidentified speaker (speaker_0): be able to maintain what they can't figure out is will the formats itself explode going forward? Right? There could be new consumption formats. They cannot predict it. That would be some like you know, 25-year-old predicting it or doing something about it. But other than that, they have a very fair chance to re-adaption. readapt and they have to do nothing. They have a
- 38:53
- Unidentified speaker (speaker_0): production budget. That production budget, 20% of it.
- 38:56
- Unidentified speaker (UNKNOWN): So are
- 38:56
- Unidentified speaker (speaker_3): you going after that or are you going after the cursor and individuals?
- 38:59
- Unidentified speaker (speaker_0): Yeah, so we are going after them in the beginning and we have just started GTM two months from now. Right. We see a lot of win and a lot of uh faster movement Moment than we expected it to be. And if it continues at that pace, that is the best point to do it.
- 39:15
- Unidentified speaker (speaker_0): So
- 39:15
- Unidentified speaker (speaker_1): your largest client today, and don't need to take names, how much would they have expanded in terms of you know billings in the last what about three, four months versus and where do you expect that to be, say 12 months from now? I
- 39:29
- Unidentified speaker (UNKNOWN): mean
- 39:29
- Unidentified speaker (speaker_0): largest.
- 39:30
- Unidentified speaker (UNKNOWN): Like if you're
- 39:30
- Unidentified speaker (speaker_1): too hazard or guess.
- 39:31
- Unidentified speaker (UNKNOWN): Yeah,
- 39:31
- Unidentified speaker (speaker_0): largest client uh started in July with $10,000 billing.
- 39:36
- Unidentified speaker (UNKNOWN): Uh
- 39:36
- Unidentified speaker (speaker_0): expanded
- 39:37
- Unidentified speaker (UNKNOWN): to $500. October now,
- 39:38
- Unidentified speaker (speaker_1): so about what? about four or five
- 39:39
- Unidentified speaker (UNKNOWN): months?
- 39:39
- Unidentified speaker (speaker_1): 500k.
- 39:40
- Unidentified speaker (UNKNOWN): So they
- 39:41
- Unidentified speaker (speaker_1): want from 10k to 500k.
- 39:44
- Unidentified speaker (speaker_1): And where do you think they will be 12 months from now?
- 39:46
- Unidentified speaker (speaker_0): I
- 39:46
- Unidentified speaker (UNKNOWN): think next month
- 39:47
- Unidentified speaker (speaker_0): they should move to close to a million. And beyond that, I'm also because you know, even for them, they are thinking about it on the go.
- 39:56
- Unidentified speaker (UNKNOWN): If
- 39:57
- Unidentified speaker (speaker_0): you're able to monetize,
- 39:59
- Unidentified speaker (UNKNOWN): if you
- 39:59
- Unidentified speaker (speaker_0): you're are able to produce so fast, right, how to like even model that, right? So uh yeah, we will we'll see how it happens. Uh what I've uh read in articles about uh Netflix of the world is that they see a very low retention rate on their seasons too if they are two years ahead,
- 40:19
- Unidentified speaker (speaker_0): right? I think it happened for One Piece, which had so much of traction, but by the time season two came, the excitement got lost,
- 40:27
- Unidentified speaker (UNKNOWN): and
- 40:27
- Unidentified speaker (speaker_0): only 35% of the user. So I think it this is what consumer wants, and what consumer wants is what the platform will try to achieve. Yes.
- 40:35
- Unidentified speaker (speaker_1): And
- 40:35
- Unidentified speaker (speaker_0): it is possible with A
- 40:36
- Unidentified speaker (UNKNOWN): today.
- 40:37
- Unidentified speaker (speaker_0): So I think the fundamentally everything will change. They will release the next season within three months or six months now, and the technology is available to do it. How fast can it accelerate? Because you know, if Netflix goes full berserk, let's
- 40:50
- Unidentified speaker (UNKNOWN): say
- 40:50
- Unidentified speaker (speaker_0): to tomorrow on AI, what would be their pipeline of season twos that they want to do in the same production budget? Right, if their production budget is uh what five billion dollars, right, can they create the next season of everything that is there on an app, which even the long tail users consume? What does this mean for companies like us who who can maybe partner
- 41:10
- Unidentified speaker (speaker_0): with them?
- 41:10
- Unidentified speaker (speaker_1): Yeah,
- 41:11
- Unidentified speaker (speaker_0): right. How fast will we scale? I can't say one billion in 12 months. is possible.
- 41:16
- Unidentified speaker (speaker_1): Question.
- 41:18
- Unidentified speaker (speaker_1): If I take like a little bit of a sort of you know, gazing into the future, crystal ball gazing.
- 41:24
- Unidentified speaker (speaker_0): Yeah.
- 41:26
- Unidentified speaker (UNKNOWN): You know,
- 41:27
- Unidentified speaker (speaker_1): in the startup world, we talk about single employee unicorn company, right? Of course. Um
- 41:34
- Unidentified speaker (speaker_1): And
- 41:35
- Unidentified speaker (speaker_1): while we aren't there yet, we're I think starting to push the box. boundary on how many
- 41:40
- Unidentified speaker (UNKNOWN): employees,
- 41:41
- Unidentified speaker (UNKNOWN): five, ten
- 41:42
- Unidentified speaker (speaker_1): people, twelve people, whatever, right? Instinct is an example recently.
- 41:46
- Unidentified speaker (speaker_1): If I were to take that into the entertainment world, today, you know, big sort of uh ticket sort of releases are largely controlled by the very large studios globally. Yeah, right.
- 41:59
- Unidentified speaker (UNKNOWN): Uh
- 41:59
- Unidentified speaker (speaker_1): do you see the equivalent of that playing out where there's a very small creative team, two, three, four people, that today now has access to the same capability
- 42:09
- Unidentified speaker (speaker_1): using tools such as Framework? And do you see that kind of how does what does that change to the landscape of overall?
- 42:16
- Unidentified speaker (speaker_0): See, as long
- 42:17
- Unidentified speaker (UNKNOWN): as
- 42:17
- Unidentified speaker (speaker_0): distribution is not very gated, which in a way is not. Like, I mean, end of the day YouTube exists, right?
- 42:24
- Unidentified speaker (UNKNOWN): Uh Um yeah,
- 42:25
- Unidentified speaker (speaker_0): what has fundamentally changed is that even a bad movie with a lot budget could do something great, right? That will go away. Right. A good story can never find a budget. All these things used to happen, right? Uh I think that has fundamentally already changed, right? It's just a matter of time we as consumers start
- 42:45
- Unidentified speaker (speaker_0): experiencing it because people are making today, right? And you're you are going to see a flood of content.
- 42:51
- Unidentified speaker (speaker_0): and uh you know open distribution platform will compete very aggressively with the Hollywood and or Bollywood or the movie industry, and movie industry will also not set behind. I don't think so. The people in the industry today are so uh backdated. Uh, if they were planning to produce 10, they would also produce 100
- 43:11
- Unidentified speaker (speaker_0): to compete in this new world, and wherever the equilibrium reaches. it would be great for the consumer at least who is watching content. Both the people they are not fighting with each other. Actually, they are coming to a point which is great for consumer, and I feel everybody's making money. I mean, Bollywood will be a much bigger industry in the next five years.
- 43:30
- Unidentified speaker (speaker_3): On that making money, you know, and almost we are coming to an end.
- 43:35
- Unidentified speaker (UNKNOWN): You
- 43:35
- Unidentified speaker (speaker_3): know, if I was to draw a parallel from the past, you could think. of let's say the gaming industry where gaming engines became that production operating system, right? But they pretty much have the software margins. Seems like in your case, when you are doing this production OS of sorts for this industry now in AI Video, you're in that
- 43:55
- Unidentified speaker (speaker_3): same token flow of uh to some extent. Do the margins look similar?
- 44:00
- Unidentified speaker (speaker_2): But that's temporary.
- 44:01
- Unidentified speaker (UNKNOWN): Like
- 44:02
- Unidentified speaker (speaker_2): open source video. will solve that for us anyway. Like you can sell first models, you can open like I we already know LLM in the coding world, open source is catching up to closed source for 80% of coding, if not 100%.
- 44:16
- Unidentified speaker (speaker_2): We already have a video model out there which is number two at least, if not one, and soon they are updating and they'll get to close to two, and then the fine fine-tunes tunes will start. happening, the post-training will start happening. I think margins are a solvable problem for sure now in the current world with the
- 44:33
- Unidentified speaker (UNKNOWN): So what are
- 44:34
- Unidentified speaker (speaker_1): margins today again to the extent you're comfortable sharing? What are margins today? Where do you expect them in steady state?
- 44:40
- Unidentified speaker (speaker_2): So right now we have taken a very conscious call to keep them at 20%, but we believe we'll get to like an eighty percent type of. thing. So
- 44:46
- Unidentified speaker (UNKNOWN): you see
- 44:46
- Unidentified speaker (speaker_1): soft. So billion dollars of top line 80% eighty percent margins, huh? We are recording this. Both are not will happening. not
- 44:50
- Unidentified speaker (speaker_0): happen twelve
- 44:51
- Unidentified speaker (UNKNOWN): months.
- 44:52
- Unidentified speaker (speaker_0): Both will not happen at the same time.
- 44:55
- Unidentified speaker (speaker_2): Yeah, yeah. Because uh like I think we are at an inflection curve where first we need to give a taste of AI video to the world, for which we will have to first drop our margins.
- 45:05
- Unidentified speaker (UNKNOWN): So
- 45:05
- Unidentified speaker (speaker_1): you're focused actually on driving more adoption of the
- 45:08
- Unidentified speaker (UNKNOWN): next twelve months we should be. Like
- 45:10
- Unidentified speaker (speaker_2): it makes no sense.
- 45:11
- Unidentified speaker (UNKNOWN): This time is
- 45:11
- Unidentified speaker (speaker_0): for that. So
- 45:12
- Unidentified speaker (UNKNOWN): you
- 45:12
- Unidentified speaker (speaker_1): will prioritize that versus worse.
- 45:13
- Unidentified speaker (UNKNOWN): You
- 45:13
- Unidentified speaker (speaker_2): have to have a $200 cloud code subscription to get the 90% API margins of Anthropic.
- 45:19
- Unidentified speaker (speaker_1): Of course.
- 45:19
- Unidentified speaker (speaker_2): Right. So like I think the next 12 to 18 months is for that subscription first. Then you get to the enterprise margins. While we will keep having certain enterprises, but I think that is where the meat will become soon. Like like to get the adoption to be driven, someone has to spend that money. Yeah.
- 45:38
- Unidentified speaker (speaker_1): See,
- 45:38
- Unidentified speaker (speaker_0): with the kind of KPS that has happened in this industry, like intelligence uh industry, we think uh already the foundation has been set for applications and companies to make margins in the future. Yeah. Right. Somebody had to invest money and they have. It was not us, we couldn't do it. Somebody has invested money, right? And the margins will happen and might be
- 45:58
- Unidentified speaker (speaker_0): more exciting than SaaS, traditional SARS.
- 46:00
- Unidentified speaker (speaker_1): Final question for me, I know we're out of time. I'll let uh sort of cash give his closing comments. What worries you? What's your biggest constraint today? I mean, leave capital aside, but what is the biggest constraint? Like what do you what do you need to solve that you still haven't solved yet for this to really sort of achieve its potential?
- 46:17
- Unidentified speaker (speaker_2): I think
- 46:20
- Unidentified speaker (UNKNOWN): any
- 46:21
- Unidentified speaker (speaker_2): AI video product yet that hasn't reached that
- 46:26
- Unidentified speaker (speaker_2): a high. aha quick aha moment that okay I try it and then I get hooked to it. Like Hicksfield has done the best job there, but they've done it for like a fifteen second, thirty second. But no one has been able to get there. And that is something at least
- 46:41
- Unidentified speaker (UNKNOWN): at
- 46:41
- Unidentified speaker (speaker_2): least I'm spending the next three months on because and I don't think it's a I don't think it's a tech tech problem. It's it's a very, very product in UX problem as to that. Like you want to give the aha moment while giving the user control. And that balance is something that we have to find. And that's something that that's my core next to three months.
- 47:01
- Unidentified speaker (speaker_0): Faster generation is the key to uh explode this industry. Like today, input to output time for a video generation is much more high.
- 47:10
- Unidentified speaker (UNKNOWN): Uh
- 47:10
- Unidentified speaker (speaker_0): which is uh
- 47:12
- Unidentified speaker (UNKNOWN): How
- 47:12
- Unidentified speaker (speaker_1): long does it take? Let's say I want to do a 30-minute thirty minute sort of you know, short feature film or whatever it is. How long does it take versus obviously doing it physically would have taken months or whatever it
- 47:22
- Unidentified speaker (UNKNOWN): is?
- 47:22
- Unidentified speaker (speaker_1): But how long does it take and what's the proportionate cost difference?
- 47:25
- Unidentified speaker (speaker_0): I would just break it down to one clip, let's say a thirty-second clip, the generation time could be thirty second to two minutes. Okay. Depending on the complexity of the input. Yeah. Like references and all. Right. For penetration, that is very high time. Today, you know, you're in a short video economy, instant gratification, right? But if it renders in two seconds,
- 47:45
- Unidentified speaker (speaker_0): three seconds, that will supercharge the entire ecosystem of content creation. Right. And that is the biggest bottle. That is why people don't try. People give only five minutes and they like make a decision K it's not working
- 47:59
- Unidentified speaker (UNKNOWN): anymore,
- 47:59
- Unidentified speaker (speaker_0): right? Because this red just just imagine pre-buffering pre buffering era of you. internet. We are there. Once you something like that.
- 48:06
- Unidentified speaker (UNKNOWN): You were in the
- 48:06
- Unidentified speaker (speaker_1): dial up days even today, and you think once we get to broadband,
- 48:10
- Unidentified speaker (speaker_0): you two used to buffer, right? And now it is like instantaneous and look at how the consumption has grown. Same thing has to happen here.
- 48:17
- Unidentified speaker (speaker_3): Got it. No, so I think super interesting cha uh ch conversation. Great for Taro and I. We are super excited to hear the predictions of uh billion dollar run rates in twelve 12 months. and uh 70-80 percent margins, let's say not in 12 and 25.
- 48:30
- Unidentified speaker (speaker_0): I mean it's it's it's the possibility depends on our execution, how good it is.
- 48:34
- Unidentified speaker (speaker_3): And uh we we are super excited. Uh hopefully all those predictions come true. But parting note, each one of you individually, one prediction besides your own company and Femio uh about AI. Uh where do we uh where one big prediction on what changes, what happens. in the next 12 months. Either of you can go first. Yeah,
- 48:54
- Unidentified speaker (speaker_0): I'm thinking.
- 48:56
- Unidentified speaker (UNKNOWN): I
- 48:57
- Unidentified speaker (speaker_2): think I've like the re one of my core tenets of why we are doing this is every company will have to become a content company. I think we know this. Every company has to become a media company, and uh yeah, we are here to help them do that really well.
- 49:12
- Unidentified speaker (speaker_0): I I uh am very bullish about. personal agents.
- 49:16
- Unidentified speaker (UNKNOWN): Uh
- 49:16
- Unidentified speaker (speaker_0): right, might also some way impact our product, how we design it in the future. And uh I think this is one thing, it's it's very cost inefficient right
- 49:25
- Unidentified speaker (UNKNOWN): now,
- 49:26
- Unidentified speaker (speaker_0): but something like this will drive penetration. Like you know, Facebook took a six to eight years for our parents to start using it, right? And today also AI is not used by a larger population at a word level, like standard. status 2%, 3%.
- 49:40
- Unidentified speaker (speaker_1): Absolutely.
- 49:40
- Unidentified speaker (speaker_0): That to 50 will happen very suddenly because of something like personal agent, maybe through devices or maybe within an app.
- 49:47
- Unidentified speaker (speaker_2): Actually, in order for time calculations, we don't even think that we might start consuming most content through videos that are made only for me.
- 49:55
- Unidentified speaker (UNKNOWN): Yeah.
- 49:56
- Unidentified speaker (speaker_2): And personal agents are one way to tap into that market.
- 50:01
- Unidentified speaker (speaker_3): And as a last note, I'm also gonna going ask to the Tharo same question. One big prediction. What what's your big prediction in the AI world for the next 12 months?
- 50:10
- Unidentified speaker (speaker_1): I think
- 50:12
- Unidentified speaker (speaker_1): what excites me is one of the questions I was asking, which is I know we talk about it in our sort of startup world, but I genuinely think the small creator who today just didn't have access, didn't have tools, didn't have distribution, didn't know how to you know bring budgets to support their creative inputs. I can see a
- 50:32
- Unidentified speaker (speaker_1): mainstream, you know, full-length feature film being created by a small creator in a tier two, tier three city in India and that getting global recognition. And I don't think that days far away. Hopefully, you know, you guys are the ones that enable that.
- 50:45
- Unidentified speaker (speaker_2): Yeah.
- 50:45
- Unidentified speaker (UNKNOWN): Hopefully.
- 50:46
- Unidentified speaker (speaker_3): On that note, thank you. Clearly you you heard all of us being extremely optimistic about the video market. Uh highly ambitious, high agency folks going after it from India, and that is what we need uh to do and we try and bring those stories out through the Intelligent Indian series with uh uh what we do with founders.
- 51:06
- Unidentified speaker (speaker_3): Thank you and we hope you enjoyed it.
- 51:08
- Unidentified speaker (speaker_1): Thank you.
- 51:09
- Unidentified speaker (UNKNOWN): Thank
- 51:09
- Unidentified speaker (speaker_0): you so much.
- 51:09
- Unidentified speaker (speaker_2): Thank you.
- 51:10
- Unidentified speaker (speaker_0): You're watching Z Z47 four seven moments, the podcast by India's oldest
- 51:13
- Unidentified speaker (speaker_3): venture capital firm Z. Z47. If you like this episode, consider hitting the subscribe button and the bell icon so you never miss an update from Z47 Moments.
More from Z47 Moments
272: Masterclass on building a D2C Business - By Blissclub Founder | Minu Margeret | Unstarted
Episode discussion: Bliss Club founder Minu Margeret details single-SKU strategy, proprietary fabric innovation, and finance-first D2C metrics
Speakers’ claims: Minu Margeret explains launching Bliss Club with a single hero product (black leggings) to perfect fit and iterate rapidly, achieving over a crore in monthly revenue. She contrasts this with industry norms, citing return rates under 10% versus the typical 30-40%, driven by deep consumer feedback loops and proprietary fabric development for Indian climates. Margeret emphasizes a finance-first approach, prioritizing unit economics and supply chain consolidation over rapid SKU expansion. The discussion covers the challenges of fundraising for women founders and the strategic decision to remain a solo founder for faster decision-making.
271: How to build an AI startup from scratch in 2026 | Rohan Joshi | Stellon Labs
Episode discussion: Rohan Joshi on building Stellon Labs, pivoting from consumer to developer market, and AI voice inflection points
Speakers’ claims: Rohan Joshi, founder of Stellon Labs, discusses his journey from Meta to launching an AI research lab focused on CPU-efficient open-source voice models. He explains how initial pricing experiments revealed that non-technical users were not the target audience for privacy-focused tools, leading to a pivot toward developers. Joshi highlights the virality of open-source distribution and argues that AI voice models have reached a threshold of naturalness suitable for enterprise applications like customer support and collections. He also outlines future inflection points for voice AI, including bidirectional interfaces and real-time video streaming.