What this covers

Patrick O'Shaughnessy talks with Captions co-founders Gaurav Misra and Dwight Churchill about building an AI video company. Misra supplies most of the technical grounding. He explains how a diffusion model starts from pure noise and reveals structure step by step, guided by text, unlike next-token prediction. He says video data is far heavier and rarer than text, and that downloading their training videos alone would cost about a million dollars. Diffusion video models sit in the tens of billions of parameters, against roughly 400 billion for text, so he sees more room to scale. He also forecasts that distillation could cut denoising from about a hundred steps to a few, an efficiency gain of roughly an order of magnitude. That forecast is one of several claims flagged as contested.

The company's own model gets equal attention. A free consumer app, instrumented from day one, generates data for a paid B2B product, a pattern they compare to Facebook and Google (Freemium-to-B2B Data Model). Misra attributes the growth to Use-Case Unlock Growth: each new capability opens a use case only they can serve, and he puts their share of the space at one to five percent. He also warns that Product-Market Fit Masks Bad Decisions. As copycats arrive, he expects competition to move to workflows, pricing, and packaging.

Other turns range further. Churchill argues investors over-focus on frontier labs, are rushing to price AI by labor replacement, and may be overlooking mature SaaS companies at 80 to 90 percent margins that are open to reinvention. Misra predicts the average likeness will become worthless while trusted, widely recognized ones gain value (Scarcity Shift to Trusted Identity). The two also discuss how cheaper production could change filmmaking.

Sharpest takeaway

Gaurav Misra and Dwight Churchill argue that AI businesses built on bounded problems like video generation, where the output becomes a durable asset and the data flywheel compounds, can have normal software-like economics. Unbounded general-intelligence races lack that property.

  • Video generation is a solved rendering problem with a finite endpoint, while general intelligence is unsolved and may have no end to the model race
  • A product that generates usage data can build a proprietary data flywheel that outpaces scraped internet data
  • Inference and GPU costs are deflating, so bounded-problem companies should trend toward traditional software margins

The claims · ranked62 claims · weighted by value

This asset isn't compiled yet

You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.

0.80

Diffusion models generate video by starting from pure noise and, at each denoising step conditioned on text, progressively revealing structure such as a man in a blue shirt, unlike next-token prediction, which predicts the next word from prior context.

definitionhigh valueestablishednovelty 2/4durability 4/4· Gaurav Misra

“they actually work by starting from noise—it starts from literal noise, like the static you see on TV.”

0.73

Because video generation is a rendering problem rather than an intelligence problem, its compute needs are bounded by known rendering costs, and it should require no more resources than rendering, possibly fewer.

causalhigh valuecontestednovelty 3/4durability 3/4· Gaurav

“what will save us on the video-model side is actually the fact that it is an easier problem than the text problem.”

0.70

Video training data is much heavier, rarer, and more expensive to acquire than text or audio, and downloading the training videos alone would cost about one million dollars for their company.

factualhigh valueestablishednovelty 2/4durability 2/4· Gaurav Misra

“even if we were to just download all of our training videos, it would cost us a million dollars to download the training videos.”

0.68

Scaling models depends on hardware, architecture, and training techniques, but ultimately on sustainable data sources, and the internet's stock of text is limited and growing, so winners will be determined by who secures sustainable data.

causalhigh valuecontestednovelty 2/4durability 3/4· Gaurav Misra

“what really matters is the data. At the end of the day, a lot of companies are scraping the internet.”

0.68

Captions' model resembles Facebook and Google: a mass-market free consumer product generates data that powers a paid B2B product.

factualhigh valueestablishednovelty 2/4durability 3/4· Gaurav Misra

“It seems to me similar to the Facebook or Google business model, where you have a mass consumer free product”

0.66

Text models are in the roughly 400-billion-parameter range, while diffusion video models are in the tens of billions of parameters, and Meta's MovieGen was about 30 billion, so video has far more room to scale.

factualhigh valueestablishednovelty 2/4durability 2/4· Gaurav Misra

“text models are already in the 400-billion-parameter range... Diffusion models are still in the tens of billions”

0.65

Video inference is currently inefficient, but distillation now lets models produce output in a few diffusion steps instead of about a hundred, and efficiency could improve by roughly an order of magnitude.

forecasthigh valuecontestednovelty 3/4durability 2/4· Gaurav Misra

“We can distill models and have them work with a few steps of diffusion now... It could be a factor of at least an order of magnitude, like 10x”

0.65

When a company has strong product-market fit, growth can hide harmful decisions, so employees conclude that everything they did was right; distinguishing good actions from bad requires measuring what is actually causing growth.

causalhigh valuespeaker onlynovelty 4/4durability 4/4· Gaurav Misra

“the company was growing despite those actions.”

0.57

Capital spent on a general-intelligence model is often beaten by the next model and rendered obsolete, creating an open-ended spending race with no clear end, whereas a finished media-generation model becomes a durable asset that retains value.

causalhigh valuespeaker onlynovelty 3/4durability 3/4· Gaurav Misra

“only for that model to be beaten by the next model—and that model becomes essentially useless and obsolete”

0.57

Captions instrumented its app from day one so that every user interaction generates training data that improves the model, creating a flywheel in which a larger user base produces data for better models, which deliver better experiences and attract more users.

causalhigh valuespeaker onlynovelty 3/4durability 3/4· Gaurav Misra

“the app is an AI app. People come in, they use the app, we use the data to make the model better”

0.57

Text models already function as general intelligence that can be applied to bounded tasks such as coding, where the work amounts to translation into a higher-level language, so coding does not require AGI.

causalhigh valuespeaker onlynovelty 3/4durability 3/4· Gaurav Misra

“we’ve already created what is essentially a tool for intelligence—intelligence in a box—that you can just apply to something to solve a bounded problem.”

0.57

Early leaders can hold a moat while they stay ahead, but as more companies raise money and copy the playbook, the competition shifts to software: workflows, pricing, and packaging.

forecasthigh valuespeaker onlynovelty 3/4durability 3/4· Gaurav Misra

“it’s going to become a software race.”

0.57

Early-stage companies with unusually high margins are attractive targets for entrepreneurs to attack, and mature SaaS companies at 80-90 percent margins are losing pricing power and are open to reinvention.

normativehigh valuespeaker onlynovelty 3/4durability 3/4· Dwight Churchill

“higher-margin businesses actually sound to me like perfect attack vectors for another entrepreneur.”

0.57

Lower production costs let lower-budget creators make more complex work, and the ticket-price parity between blockbusters and small films means cost removal can premiumize certain aspects of film.

causalhigh valuespeaker onlynovelty 3/4durability 3/4· Dwight Churchill

“those high-budget films... it may even create more premium on certain aspects.”

0.57

Text-generation companies are actually pursuing the unsolved problem of intelligence, while video and audio generation are solved-problem rendering tasks that AI makes about 100 times easier and cheaper, so their business models differ fundamentally.

definitionhigh valuespeaker onlynovelty 3/4durability 3/4· Gaurav Misra

“it's, I think, a little bit less of solving an unbounded intelligence problem and a little bit more of solving actually rendering a solved problem.”

0.57

As generating likenesses becomes free, the value of an average likeness goes to zero, and value shifts to likenesses that are trusted and recognized by millions, including fabricated personas that become famous.

forecasthigh valuespeaker onlynovelty 3/4durability 3/4· Gaurav Misra

“the value of the average likeness goes to zero, because anyone can make one out of nothing... it’s going to be determined by who’s known”

0.54

Once a media generation model reaches near-perfect quality, it becomes an asset that is expensive to create but keeps generating value and does not lose value easily, after which the company behaves like a software business.

causalhigh valuespeaker onlynovelty 3/4durability 3/4· Gaurav

“it creates an asset, and there might be very soon a point where, 'Oh wow, it’s just really good'... And then it’s an asset, and after that it’s just a software company”

0.53

Captions' rapid growth comes from being the only company able to serve each newly unlocked use case, and they estimate they have only captured about one to five percent of the potential use-case space.

factualhigh valuespeaker onlynovelty 3/4durability 2/4· Gaurav Misra

“we really think that the use case unlocked so far is somewhere in the one-to-five-percent range. We’ve barely scratched the surface”

0.53

Investors focus heavily on frontier AI labs and their R&D spending, underappreciating how AI tools are changing work inside ordinary companies, where the adoption and value are happening.

normativehigh valuespeaker onlynovelty 3/4durability 2/4· Dwight Churchill

“there’s just so much attention on the large AI labs that are effectively doing what Gaurav was talking about before—solving intelligence.”

0.53

Investors are rushing toward pricing AI by labor replacement, but the CFO wants lower costs regardless of source, and the ongoing subscription may hold more pricing value than labor-based models, so the market needs more exploration.

normativehigh valuespeaker onlynovelty 3/4durability 2/4· Dwight Churchill

“people are maybe rushing into the labor argument... there might be more continuing alpha in the typical subscription.”

0.53

Captions originally had no free product, yet users still paid immediately, a result that would not have worked with earlier technology, showing that AI value lets companies skip the free-tier hurdle.

factualhigh valuespeaker onlynovelty 3/4durability 2/4· Gaurav Misra

“for a long time, we had no free product. You couldn’t even use it once without paying. People still said, 'Okay, here’s the money,' and moved on.”

0.52

GPU and compute costs are deflationary, with each new chip generation (such as H100 prices falling as H200 rolls out) making compute cheaper, so AI application businesses can reach very high margins over time.

forecasthigh valuespeaker onlynovelty 2/4durability 3/4· Dwight Churchill

“Throughout history, these prices have never gone the other way. It’s highly deflationary as the next one rolls out”

0.52

The winners of generative media will be the companies whose models consistently outperform others, which depends on data acquisition and a flywheel that keeps improving the model, while a later software race will be decided by workflows, pricing, and packaging.

normativehigh valuespeaker onlynovelty 2/4durability 3/4· Gaurav

“the winners are going to be determined by who has the best model that’s consistently outperforming everyone else. All that comes down to data acquisition, a flywheel”

0.52

The 2010s design boom, in which tools like Canva and Figma made good design and websites ubiquitous, shows that once a capability becomes a commodity it still remains valuable and required, and video is following the same path.

causalhigh valuespeaker onlynovelty 2/4durability 3/4· Gaurav Misra

“If you don’t have a good one, you might suffer, even though it’s basically a commodity now.”

0.51

AI lab researchers who cannot release their work are attracted by companies that provide compute, data, and environment, which makes recruiting and negotiation easier.

causalhigh valuespeaker onlynovelty 3/4durability 2/4· Dwight Churchill

“There are a lot of folks in AI labs who can’t release anything they’re working on”

0.50

Building AI products now allows teams to invent the entire stack from hardware drivers to user experience, and current AI products are poorly designed because they simply put models on a UI with a press-button-output interaction.

normativehigh valuespeaker onlynovelty 2/4durability 3/4· Gaurav Misra

“They’re just taking models, throwing them on a UI, and saying, “Press button, output.””

0.49

Captions generates hundreds of thousands of unwatermarked original videos daily for social networks, which need content, and so it semi-collaborates with platforms such as Instagram rather than competing with them.

factualhigh valuespeaker onlynovelty 2/4durability 2/4· Gaurav Misra

“we generate tons of content—hundreds of thousands of pieces a day—that goes to social media.”

0.49

Pricing and value will ultimately depend on what share of a given use case AI can cover, which Gaurav expects could approach 100 percent because the video problem is solvable.

forecasthigh valuespeaker onlynovelty 2/4durability 2/4· Gaurav Misra

“We’re at 5% today, but is the limit 100%? 75%? 50%? ... My guess is we can go all the way to 100%”

0.49

Captions' AI Edit currently takes style preferences, with the future goal being that users can prompt editing in natural language like instructing a human editor, eliminating the need to understand keyframes, animation curves, and timelines.

forecasthigh valuespeaker onlynovelty 2/4durability 2/4· Gaurav

“That’s where it will go in the future. Currently, it’s more style preferences that you provide to it.”

0.49

The first Captions app, which added text to videos and used speech-to-text, was built in two days and reached the top of the App Store overnight with about 600 videos created per minute, and the app was instrumented from day one to collect data for improving models.

factualhigh valuespeaker onlynovelty 2/4durability 2/4· Gaurav

“this was built in two days on a weekend, really just band-aided together. And we put it on the App Store, went to sleep. The next morning, it was top of the App Store.”

0.49

Video editing tools are commoditized, many are built offshore to minimize cost, and this made it very hard for a large company like Snap to win with a standalone video creation platform, so Captions needed an AI differentiator.

causalhigh valuespeaker onlynovelty 2/4durability 2/4· Gaurav Misra

“video is kind of a commodity—video editors are commodities. A lot of these companies are actually foreign, and that’s because we’re just trying to minimize costs”

0.49

It is too early to know whether AI products will be priced like labor, since they cannot yet fully replace operationally heavy workflows, so pricing will depend on how much of a use case the product can cover over time.

forecasthigh valuespeaker onlynovelty 2/4durability 2/4· Gaurav Misra

“I don’t know if we completely understand it yet. It’s almost too early to tell, in a way, because we can’t fully replace labor”

0.48

Captions plans to train fully on licensed data because its scale of data collection enables this, and in a saturated market customers will pay more for the guarantee and reputation of licensed models.

forecasthigh valuespeaker onlynovelty 3/4durability 2/4· Gaurav Misra

“We plan to go fully licensed—one of our main differentiators”

0.48

A bounded investment in the hundreds of millions of dollars can produce a model reasonably capable of generating anything a CGI studio can do, which is the threshold the company needs to reach.

forecasthigh valuespeaker onlynovelty 3/4durability 2/4· Gaurav Misra

“That bounded cost is probably in the hundreds of millions of dollars, but it actually gets us to a solution”

0.47

Building world-class generative models requires only a dozen or fewer people with a very specific skill set, though finding them is the main challenge, and talent shortages will grow as use cases expand because experience takes years to develop.

factualhigh valuespeaker onlynovelty 2/4durability 3/4· Gaurav Misra

“delivering world-class results can be done with maybe a dozen people or fewer”

0.47

Captions deliberately narrowed to communication-oriented talking videos for marketing, sales, education, and internal communications rather than general video or stock footage, because focusing on everything would solve none of the problems well.

normativehigh valuespeaker onlynovelty 2/4durability 3/4· Gaurav Misra

“We’ve actually niched down quite a bit on purpose, because as you said, video is huge”

0.46

Captions' real arms race is delivering far beyond what customers need today, because each release is commercialized on day zero and built into customers' actual work rather than tested first, so the competition is about inventing design patterns rather than beating rivals.

normativehigh valuespeaker onlynovelty 2/4durability 2/4· Dwight Churchill

“Whenever we release something, it’s commercialized on day zero immediately. We’re not testing it with a bunch of people”

0.46

Multiple former Snap employees have started nearly identical companies, and Gaurav reads this copying as a sign that Captions is doing the right things, while arguing that its strategy should stay driven by its own mission rather than competitor moves.

normativehigh valuespeaker onlynovelty 2/4durability 2/4· Gaurav Misra

“in a way, most people are copying us—is that it’s a great sign. It means we’re doing the right things.”

0.46

Video models struggle with human anatomy such as fingers, arms, and drinking, and Captions addresses this by training specifically on people and planning skeleton conditioning, because models that see few hands in varied configurations fail to represent normal anatomy.

causalhigh valuespeaker onlynovelty 2/4durability 2/4· Gaurav Misra

“it may not fully realize that if it hasn’t seen enough hands in all kinds of configurations doing all kinds of things.”

0.45

Gaurav argues that TikTok, not Facebook or Google, is now the copying company, and that Meta has become a good actor by open-sourcing models.

factualhigh valuespeaker onlynovelty 2/4durability 1/4· Gaurav Misra

“Facebook suddenly became the good guys—Mark Zuckerberg is a hero now for releasing all these models”

0.45

Given the pace of video model improvement, video that is close to indistinguishable from a real recording is likely about a year to a year and a half away, possibly sooner.

forecasthigh valuespeaker onlynovelty 3/4durability 1/4· Gaurav Misra

“I could easily see within a year and a half or so something coming pretty close to indistinguishable”

0.45

Consumer video app pricing has historically sat around $7.99 to $12.99 per month, but AI video tools can charge around $25 a month, and some AI consumer subscriptions reach $2,000 a month because competition is thin.

factualhigh valuespeaker onlynovelty 3/4durability 1/4· Gaurav Misra

“historically it’s somewhere around $7.99 to $12.99 per month... Some people even pay $2,000 a month for consumer subscriptions.”

0.44

Fine-tuning a bounded-problem model is much cheaper than training a foundation model from scratch, so continuous incremental training on new data will be a low-cost component of the business.

factualhigh valuespeaker onlynovelty 2/4durability 2/4· Gaurav Misra

“fine-tuning is generally cheap. It’s actually not even close to as expensive as training a foundation model from scratch.”

0.43

Models will be able to generate a specific product, such as a branded water bottle, interacting with a person within six months, using image conditioning and multiple angles of the object, because models infer unseen sides from general object knowledge.

forecasthigh valuespeaker onlynovelty 2/4durability 1/4· Gaurav Misra

“Yeah, I think that will happen within six months—guaranteed.”

0.43

The company estimates that only roughly one to five percent of possible video use cases have been unlocked, meaning most of the potential market has barely been touched.

factualhigh valuespeaker onlynovelty 2/4durability 1/4· Gaurav Misra

“we really think that the use case unlocked so far is somewhere in the one-to-five-percent range. We’ve barely scratched the surface”

0.36

Speech-to-text technology had become highly accurate, including for names and obscure terminology, but the average person did not realize how good it had become even though tech circles did.

factualestablishednovelty 2/4durability 2/4· Gaurav

“the average person actually didn’t understand how good the tech had gotten and how accurate it was with names and obscure terminology”

0.31

A company can own a generated likeness of a person who does not exist as brand IP and use it as a spokesperson, and fabricated personas like Lil Miquela show that such people can become famous.

forecastspeaker onlynovelty 2/4durability 2/4· Gaurav

“you could own a likeness as IP for your company of a person that doesn’t exist, and have them be the spokesperson of the company.”

0.30

Captions offers free consumer video recording and editing software to build a large user base and then upsells users into AI Creator and AI Edit, which mirror the recording and editing functions.

factualestablishednovelty 1/4durability 2/4· Gaurav Misra

“The play here for us is to provide a service to a large number of people, kind of a freemium business model”

0.30

Large enterprises such as L'Oréal are already deploying internal LLMs that any employee can query, which Dwight cites as evidence of broad enterprise AI adoption.

factualestablishednovelty 1/4durability 2/4· Dwight Churchill

“L’Oréal, the beauty company, has an internal GPT, basically an internal LLM of some sort, and any employee can ask any question.”

0.28

Snap missed the public-sharing model of TikTok because its private, low-abuse design, which avoided resharing, was the opposite of broadcast-to-everyone sharing.

causalspeaker onlynovelty 2/4durability 2/4· Gaurav Misra

“They kind of missed the TikTok thing because it was the exact opposite of that: it was shared to everybody.”

0.27

Captions is currently the only company training a foundation model specifically to generate talking human videos, while other companies in the space are not training foundation models.

factualspeaker onlynovelty 2/4durability 1/4· Gaurav

“We’re actually the only company training a foundation model to do this type of thing today”

0.26

Providing a skeleton as conditioning would let users specify the exact animation or dance a generated person performs, which also improves the model's understanding of human anatomy.

forecastspeaker onlynovelty 2/4durability 2/4· Gaurav

“the ability to provide a skeleton, for example: 'This is the exact animation I want to play out; this is the exact TikTok dance I want you to do.'”

0.26

At Snap, a small design team of about ten to twelve people drove many decisions, and Gaurav's career advanced because the CEO recognized people who could learn quickly.

factualspeaker onlynovelty 2/4durability 2/4· Gaurav Misra

“He had a great design team that he built, and a lot of the decision-making was driven through that design team. It was a small set of people—like 10 to 12 people on that team”

0.25

The rapid improvement from the crude 'Will Smith spaghetti' video meme to convincing output illustrates how fast generative video is advancing.

factualestablishednovelty 1/4durability 1/4· Gaurav Misra

“it went from, like, really horrible to like, 'Wow, this is actually good.'”

0.25

People are already experimenting with generative video, and companies will soon reinvent much of everyday work, so the gap between this technology and widespread workflow change is small.

forecastspeaker onlynovelty 2/4durability 1/4· Dwight Churchill

“we’re not far off at all from really reinventing a lot of how people end up doing their everyday work.”

0.24

Dwight's wife made it possible for him to take the risk of starting Captions while they already had a child, which he describes as the kindest thing anyone has done for him.

factualspeaker onlynovelty 0/4durability 4/4· Dwight Churchill

“she enabled me to take more risk.”

0.24

Snap's innovation came from its CEO's deep understanding of product and customers, who pushed ideas that others opposed and were often the company's best decisions.

factualspeaker onlynovelty 1/4durability 2/4· Gaurav Misra

““I don’t care, we’re doing it.” We did it, and it was the best thing we ever did.”

0.23

A large competitor copied Captions' App Store description, website, press release wording, and brand colors, but Gaurav believes the rival's software is mediocre and that Captions wins on product quality, while the rival relies on distribution.

factualspeaker onlynovelty 1/4durability 1/4· Gaurav Misra

“the software they create is mediocre; it just works because they have great distribution through TikTok. I think we win because we have a better product.”

0.23

If Captions achieves CGI-level capability, it would be a starting point for transformation across industries, including social networks, film and TV, and education, with long-run possibilities that are practically endless.

forecastspeaker onlynovelty 1/4durability 1/4· Gaurav

“You could imagine a social network based on something like this. You could imagine film and TV and stuff being dominated by these types of technologies.”

0.21

When a new technology is dramatically better, distribution temporarily becomes easy because users are willing to adopt rather than stick with entrenched tools.

causalspeaker onlynovelty 1/4durability 2/4· Patrick O'Shaughnessy

“when tools are just 10x better or more—100x better—distribution for a time becomes really easy.”

0.21

Building AI products now gives teams a rare chance to invent the entire stack from hardware drivers up to user experience, since the field is new enough that bugs at the NVIDIA driver and hardware levels still need handling.

factualspeaker onlynovelty 1/4durability 2/4· Gaurav Misra

“When do we get a chance in history to actually invent the entire stack from the bottom to the top, all the way down at the hardware level?”

0.18

A vast majority of Captions users are paid, AI Creator and AI Edit are roughly equally popular, and many users chain both tools to go from text prompt to fully edited video.

factualspeaker onlynovelty 1/4durability 1/4· Gaurav

“today a vast majority of our users are paid users. Between AI Creator and AI Edit, they’re both about equally popular.”