Transcriptly

The 8-Year Journey to Building a $4B AI Company | Synthesia, Victor Riparbelli

EO13:58Added Aug 31, 2026

He started the UK’s largest generative AI video startup — before ChatGPT even existed. Victor Riparbelli, CEO & Co-founder of Synthesia, breaks down why going viral isn’t the goal — and how to know if you’re building something “fun and

Watch on YouTube →
Contributed by 刘嘉琪

Transcript

Transcript format

  1. Intro

  2. 00:00So DIA is the world's leading AI video
  3. 00:02platform. We crossed 100 million AR. We
  4. 00:04launched the first product to a million
  5. 00:05AR. It took us 4 months. We then went
  6. 00:07from 1 million AR to 3 million AR in one
  7. 00:09year. But we basically tripled revenue
  8. 00:11from there. Tripled again and then we
  9. 00:12doubled and doubled. Are we doing this
  10. 00:14because it's fun and new? Are we doing
  11. 00:15this because it's actually useful for
  12. 00:17the customer? People are not trying to
  13. 00:18like use a model to do something.
  14. 00:20They're trying to solve a problem that
  15. 00:21they have. Right? In our case, the
  16. 00:23problem is making a video for something.
  17. 00:25I think a lot of AI apps are feeling
  18. 00:26this today where you may go viral, have
  19. 00:28a lot of people who are interested in
  20. 00:29your product and they come to you and
  21. 00:30play around with it. But it's really
  22. 00:32important to understand who are coming
  23. 00:34because they have a real use case that's
  24. 00:36recurring and who's coming because it's
  25. 00:37a cool fun thing that they want to try
  26. 00:39out. Do we actually think that they look
  27. 00:41real enough to work or is it a gimmick?
  28. 00:43Sometimes you'll be the hottest company
  29. 00:45in town and everybody will want to talk
  30. 00:46to you. Other times you'll be deemed as
  31. 00:48the incumbent that nobody cares about or
  32. 00:51like you're not doing something that's
  33. 00:52fun or kind of in the current sort of
  34. 00:53hype cycle. And no matter like which of
  35. 00:55those cycles you're in, I think the most
  36. 00:57important thing is really just like
  37. 01:02Victor Roelli and I'm one of the
  38. 01:03co-founders and the CEO of Synthesia.
  39. 01:05Sadesia is the world's leading AI video
  40. 01:08platform for the enterprise and for
  41. 01:09businesses. At its core, really what we
  42. 01:11help our customers do to land whatever
  43. 01:13message they have for their customers,
  44. 01:15their employees, their partners with the
  45. 01:17most engagement and effectiveness
  46. 01:19possible. So we've built a platform that
  47. 01:21enables anyone who knows how to use a
  48. 01:22PowerPoint to create a video. We're
  49. 01:24really what we're all about, right, is
  50. 01:25communicating in video, which is the
  51. 01:27most effective way to deliver a message.
  52. 01:29Today, we crossed 100 million AR. We've
  53. 01:31raised $330 million in venture capital.
  54. 01:33Commercially, we work with more than 70%
  55. 01:35of the Fortune 100. Today, we have more
  56. 01:37than 65,000 customers all around the
  57. 01:39world. People are not trying to like use
  58. 01:41a model to do something. They're trying
  59. 01:42to solve a problem that they have,
  60. 01:43right? In our case, the problem is
  61. 01:45making a video for something.
  62. 01:49We started the company in 2017. The

  63. Why I Spent 8 Years Betting on a Crazy Idea

  64. 01:51first three years of the company, it
  65. 01:52wasn't very fun. and the tech didn't
  66. 01:54really work and it was it was a kind of
  67. 01:55an uphill battle. So this was like the
  68. 01:57first time the world really saw a neural
  69. 01:59network generate video at a pretty high
  70. 02:01quality fully automatically right if you
  71. 02:03saw the output of that face tof face
  72. 02:04video and you would have to do that in a
  73. 02:06visual effect studio you would be
  74. 02:07talking like millions of dollars to
  75. 02:08create a clip like this and this thing
  76. 02:10actually did it in real time and when I
  77. 02:11saw that for the first time my mind
  78. 02:13started running I felt like this going
  79. 02:14to change everything about how we create
  80. 02:16video but it's also going to change
  81. 02:17everything about video itself as a
  82. 02:19medium but I felt like it was worth
  83. 02:21exploring starting a company around this
  84. 02:23technology when we see new technologies
  85. 02:25the first thing we do always imagine all
  86. 02:27the 1,000 ways it can go wrong, right?
  87. 02:29Cars, smartphones, AI, video generation,
  88. 02:32music generation. That's always what we
  89. 02:33default to. But it just turns out most
  90. 02:35of the time that technologies end up
  91. 02:37being used mostly for good things by
  92. 02:38good people. And there's a subset of
  93. 02:40people use it for bad things. And I
  94. 02:42think being optimistic about technology,
  95. 02:44especially in the early stages of it, I
  96. 02:46think is actually a huge arbitrage in
  97. 02:48itself. I mean, essentially we went out
  98. 02:49right and we we told people look in 10
  99. 02:51years you're going to be able to make a
  100. 02:52Hollywood film from your laptop just
  101. 02:54with your imagination. you won't need
  102. 02:55anything else than that and it's got to
  103. 02:56be more or less free to do it. And a lot
  104. 02:58of people just, you know, thought that
  105. 03:00there was a crazy idea. It was just it
  106. 03:02was in the AI winter. You know, there
  107. 03:03had been a lot of like disillusionment
  108. 03:05like a bunch of companies funding in the
  109. 03:07years before we started the company had
  110. 03:09all kind of like turned out to not
  111. 03:11really work because technology was just
  112. 03:13too immature. We got turned down by like
  113. 03:15100 investors, something like that until
  114. 03:17we eventually sent a cold email to Mark
  115. 03:19Cuban. He was probably the only person
  116. 03:21we met who totally shared our vision of
  117. 03:23the future. So he wasn't evaluating the
  118. 03:24vision. He was bought in on that. He was
  119. 03:26more kind of evaluating if he thought
  120. 03:27the team was worth funding and if you
  121. 03:28think we had a shot at at building
  122. 03:30something real. You need one person to
  123. 03:31say yes. That's all it takes. And
  124. 03:33generally it's easier to try and find
  125. 03:35people who already share kind of like
  126. 03:37your vision of of what's going to happen
  127. 03:38in the world as opposed to trying to
  128. 03:40convince them of it. Especially if
  129. 03:41you're not doing something that's in the
  130. 03:43kind of current hype cycle. If in your
  131. 03:45gut you truly really feel that you're on
  132. 03:48to something and that someone if it's
  133. 03:50not you someone is going to build a big
  134. 03:52company around your the product that
  135. 03:53you're doing you should continue
  136. 03:55throughout the entire journey of
  137. 03:56synthesis I've never a day in my life
  138. 03:58doubted that someone is going to build a
  139. 03:59big company around this technology I
  140. 04:01could doubt myself sometimes I could
  141. 04:02doubt by execution I could be afraid
  142. 04:04that I was too early for where we were
  143. 04:06but I never doubted that this technology
  144. 04:08was going to be extremely valuable for
  145. 04:09the world and someone was going to build
  146. 04:11a company around

  147. Are You Building a Painkiller or a Vitamin?

  148. 04:15The hard thing is knowing when to
  149. 04:17continue and when to pivot, right? We
  150. 04:19started off with a million dollars,
  151. 04:21right? Which is is not a lot of money
  152. 04:22when you're building a company like what
  153. 04:23we were trying to build. And so we had
  154. 04:25to figure out like what do we think we
  155. 04:26can realistically both build and sell
  156. 04:28before having to raise the next round,
  157. 04:30right? Which we would have to show some
  158. 04:31kind of progress. And the first thing we
  159. 04:33we we stumbled upon was this idea of AI
  160. 04:35dubbing, which is give us a real video
  161. 04:37shot with an actual camera. we can
  162. 04:38translate it for you by replacing the
  163. 04:40voice over with a different language but
  164. 04:41then also reanimating the face so it
  165. 04:44looks like uh the actor speaking a
  166. 04:45different language which will be useful
  167. 04:46if you're making a Hollywood film and
  168. 04:48you want to distribute that all over the
  169. 04:49world if you can make it look like it
  170. 04:51was recorded in French Italian that
  171. 04:52would be good for those audiences and
  172. 04:54that value proposition was actually good
  173. 04:55enough like that is definitely valuable
  174. 04:57when you can do it but the technology
  175. 04:59was very early it only worked if you
  176. 05:01look straight at the camera it took two
  177. 05:02PhDs you know like 10 days to make a
  178. 05:0430-se secondond clip um it didn't scale
  179. 05:07at all the quality threshold was
  180. 05:09insanely high, right? Cuz all these
  181. 05:11video people, they wouldn't accept
  182. 05:12anything but like the very best in terms
  183. 05:14of of the quality and it was just clear
  184. 05:16to us that it was a vitamin. It wasn't a
  185. 05:18painkiller, right? Like people liked
  186. 05:19that. They thought it was cool. Um but
  187. 05:21if we disappeared, no one is going to
  188. 05:23yell and scream, you know? They are, you
  189. 05:24know, we'll go back to translating the
  190. 05:26way we did things before. So we went
  191. 05:28back to the drawing board and we of
  192. 05:29course talked to lots of people during
  193. 05:30this period. We kind of realized that
  194. 05:32rather than focusing on the people who
  195. 05:34are already making video all the time,
  196. 05:36there was a huge group of people,
  197. 05:37billions of people probably that weren't
  198. 05:39making videos at all. And these people
  199. 05:41are desperate to make video. And they
  200. 05:42were basically telling us, hey, I want
  201. 05:43to deliver this message to someone. And
  202. 05:44I know I should be doing a video or
  203. 05:46audio because that's what people want to
  204. 05:48consume. I just don't know where to
  205. 05:50start. I don't know how to use a camera.
  206. 05:51My boss won't give me any budget to
  207. 05:53record it. And I know that there's so
  208. 05:55many problems after I've actually shot
  209. 05:57the content even though I get to make a
  210. 05:58video because you can't edit a video
  211. 06:00after you've shot it. And there's all
  212. 06:01these problems of of video, right? And
  213. 06:03then we we spoke more and more to these
  214. 06:05people and ultimately learned a lot what
  215. 06:07these people wanted to do was like
  216. 06:08corporate videos. And corporate videos
  217. 06:09are a lot easier than Steven Spielberg
  218. 06:12filmed, right? And so if we could just
  219. 06:14essentially help our customers with kind
  220. 06:15of like the talking head style footage,
  221. 06:17that would be pretty big. And that was a
  222. 06:19much smaller domain than make any kind
  223. 06:21of Hollywood film you can imagine,
  224. 06:22right? So we built this avatar
  225. 06:23technology that became one of the
  226. 06:24central of the product today. That's
  227. 06:26when the company started to really take
  228. 06:27off very quickly. We had a lot of uh
  229. 06:30conviction on a technology very early.
  230. 06:32It was very unclear what the right first
  231. 06:34use case was. You know, something that
  232. 06:36was both valuable and you could
  233. 06:38realistically build. But I think we
  234. 06:39found our way there. You could argue
  235. 06:41that we should just have built the
  236. 06:42avatar technology from day one. But back
  237. 06:45that day one, none of us had any idea
  238. 06:49how we would potentially do that, right?
  239. 06:51the research evolved over the years. We
  240. 06:53learned the technology better and better
  241. 06:54and so I think product market fit is one
  242. 06:56of those kind of things which is is much
  243. 06:58more art than science and I think we're
  244. 07:00very much led by by deep understanding
  245. 07:03of the customers that we're serving and
  246. 07:05of course deep insight into into the
  247. 07:06technologies. It it ultimately turned
  248. 07:08out the right ways.

  249. Fun and Cool Won't Keep Users Coming Back

  250. 07:12Once we launched the product and we had
  251. 07:14the first avatar MVP out there, we went
  252. 07:16viral very quickly. And so we had like
  253. 07:18lots of people coming onto the website,
  254. 07:20lots of people making kind of free demo
  255. 07:22videos of avatars by typing in things
  256. 07:24because it was fun and cool and they
  257. 07:26want to show all their friends and it's
  258. 07:27amazing. The first time you see it,
  259. 07:29you're kind of mind blown, right? But a
  260. 07:30lot of these people did not have any
  261. 07:31real use case and we could see that in
  262. 07:33the retention, right? They weren't
  263. 07:34coming back. They were just making a few
  264. 07:35videos, showing it to people, not making
  265. 07:37any more videos. And so that wasn't
  266. 07:38really like a sustainable revenue stream
  267. 07:40for us. But there was a small group of
  268. 07:42people and they kept using the product
  269. 07:43you know they kept coming back and uh we
  270. 07:45we we kind of went into that group and
  271. 07:47that spent a lot of time speaking to
  272. 07:49those folks and that's where we
  273. 07:50uncovered sort of interesting value
  274. 07:52equation right where these people were
  275. 07:54not comparing our videos to real videos
  276. 07:56or camera videos they're comparing our
  277. 07:58videos to text documents and when that's
  278. 07:59the comparison then the quality
  279. 08:01threshold looks very different right the
  280. 08:03feature request that they needed look
  281. 08:05very different so it was an organic
  282. 08:07thing I think of of letting yourself be
  283. 08:09guided by customers but also applying a
  284. 08:11lot of like critical thinking. The
  285. 08:13critical thinking is like difficult when
  286. 08:14you have so much inbound interest. I
  287. 08:16think a lot of AI apps are feeling this
  288. 08:17today where you may go viral have a lot
  289. 08:19of people who are interested in your
  290. 08:20product and that come to you and play
  291. 08:21around with it. But it's really
  292. 08:23important to understand who are coming
  293. 08:24because they have a real use case that's
  294. 08:27recurring and who's coming because it's
  295. 08:28a cool fun thing that they want to try
  296. 08:30out. Early days of 2021, we have
  297. 08:32customers who will scream and yell if we
  298. 08:34shut down the product tomorrow. That's
  299. 08:35always I think a good asset test of like
  300. 08:37do you have something real or not or do
  301. 08:38people kind of mildly care but not
  302. 08:40really. We all realize that there are
  303. 08:42lots of unanswered questions for sure
  304. 08:44but there clearly is something right
  305. 08:46there's a kernel of something that is uh
  306. 08:49that is huge and I think it was kind of
  307. 08:50double exciting in some ways because it
  308. 08:52felt like it was our little secret and
  309. 08:53it was for a couple of years it wasn't
  310. 08:55very obvious from the outside that
  311. 08:56anyone had any real use for this tech.
  312. 08:58That was a fun couple of years where we
  313. 08:59felt like we know something the rest of
  314. 09:01the world doesn't and and we're
  315. 09:02capitalizing on it.

  316. The Vacuum Effect of Product-Led Growth

  317. 09:05We figured very early on actually in
  318. 09:072021 that we wanted to focus in the
  319. 09:08enterprise because that was clearly
  320. 09:10where there was the most value to drive
  321. 09:11and that was clearly where we could you
  322. 09:12know get to the kind of biggest contract
  323. 09:14size. But it was also very obvious to us
  324. 09:16that the growth channels and the way to
  325. 09:17get this in the hands as many people as
  326. 09:19possible was not through enterprise
  327. 09:20sales. It was through just giving people
  328. 09:22access to the product, letting them play
  329. 09:23around with it, let them find their own
  330. 09:25use case, letting them qualify
  331. 09:26themselves before they spoke to the
  332. 09:28sales team. What was hard is when you're
  333. 09:30building both bottoms up motion and
  334. 09:32you're trying to build a tops down
  335. 09:33motion. Those things conflict in many
  336. 09:35different ways. As a startup, you have
  337. 09:36limited resource. And a lot of things
  338. 09:38you build for a big company will have
  339. 09:40very little impact on a small company,
  340. 09:42right? Like if you build admin tooling
  341. 09:44uh for like a CIO, a 50% company is not
  342. 09:47going to care about that. So they won't
  343. 09:48get much value out of it. On the other
  344. 09:49hand, if you build tools for like small
  345. 09:51businesses, a lot of that will not have
  346. 09:53value for someone who is um is really
  347. 09:55big business, right? There's always like
  348. 09:56these trade-offs that you make. But I
  349. 09:58think when you look at the last, you
  350. 09:59know, 10-15 years, some of the companies
  351. 10:01that have had like the best and most
  352. 10:02impressive runs are companies that have
  353. 10:04managed to combine the kind of vacuum
  354. 10:06effect of PLG where when you really get
  355. 10:08it working, right, it's just this vortex
  356. 10:10that just like sucks in traffic and
  357. 10:13convert it into both paying customers
  358. 10:14and credit cards, but also spits out,
  359. 10:16you know, the world's best enterprise
  360. 10:17leads with 1010 worldclass sales team
  361. 10:20that knows how to handhel customers,
  362. 10:22take them through a sales process, make
  363. 10:23them successful, and do all these
  364. 10:25different things. And it's interesting
  365. 10:26because it's not just in the expression
  366. 10:27of the product and the product roadmap.
  367. 10:29A lot of people tend to think that, but
  368. 10:30it's also very much in just like the
  369. 10:32people you hire, right? Some people
  370. 10:34really love like the self-service world
  371. 10:36and that's what they're passionate
  372. 10:37about. Other people really love
  373. 10:38enterprise sales. And so marrying like
  374. 10:40those two cultures together is
  375. 10:42definitely a challenge, but um I think
  376. 10:43we've in broadstrokes done pretty well
  377. 10:45on it.
  378. 10:48I think a mistake people often make

  379. Find the Bit-Off Underdogs

  380. 10:50especially people who maybe have already
  381. 10:53had a job in like big tech or like a
  382. 10:54successful companies they go out and
  383. 10:56then they try to hire people from
  384. 10:57Google, Meta, OpenAI like you know
  385. 11:00whatever kind of is the hot companies of
  386. 11:01our time and they go out they hey I have
  387. 11:03this cool startup idea why don't you
  388. 11:05join me I'll pay you not as much but
  389. 11:07I'll try and pay you almost roughly as
  390. 11:08much etc. And of course this can work.
  391. 11:10There are very smart people who work
  392. 11:11there who also wants to the startup
  393. 11:12hustle. But often times you just can't
  394. 11:14get them straight up or if you do get
  395. 11:16them and they are used to a very
  396. 11:18different salary, a very different
  397. 11:20lifestyle uh working in a big tech
  398. 11:21company as opposed to being in a
  399. 11:22startup. So that's actually often not
  400. 11:25the best path. So I think that the
  401. 11:26better way of doing it, which is what we
  402. 11:28did, is you try and find people that are
  403. 11:30a bit off. You know, they're not like
  404. 11:31the most obvious people to join your
  405. 11:33startup, but there's something you like
  406. 11:34about them. That could be that they have
  407. 11:36just a lot of hustle, a lot of grit
  408. 11:38engineers that maybe haven't worked at
  409. 11:40like a tier one or tier two company, but
  410. 11:42they are great open source project that
  411. 11:44they're managing and updating. Like
  412. 11:45there are a lot of other signals you can
  413. 11:46find good people, right? And if you can
  414. 11:48find these people who are truly hungry
  415. 11:51and wants to work at a startup and wants
  416. 11:53to put in the work, you gather a great
  417. 11:55group of those and they will very often
  418. 11:57outperform the 10 people who used to
  419. 11:59work in big tech uh with their comfy
  420. 12:01lifestyles. people you can take and we
  421. 12:03help shape give them opportunities to do
  422. 12:05something that's your only option in the
  423. 12:06early days finding that like racktac
  424. 12:08group of underdogs that have the
  425. 12:09collective energy to to go up against
  426. 12:12the competitors and incumbents I think
  427. 12:14that there is nothing that prepares you
  428. 12:16to really start a business like starting
  429. 12:18a business like spending 4 years in
  430. 12:20Mckenzie is not going to make you a
  431. 12:21better entrepreneur make you have a
  432. 12:22little bit more network and maybe
  433. 12:23slightly easy to raise money but if you
  434. 12:25actually truly if you know deep in your
  435. 12:27heart that you want to build a company
  436. 12:28and you are a builder I think you just
  437. 12:30get started as quickly as you And don't
  438. 12:32be afraid of starting too early. The the
  439. 12:34doom scenario is is rarely as bad as
  440. 12:36people kind of think it is or make it
  441. 12:38out to be. And most big companies really
  442. 12:40value people who have tried to do
  443. 12:42something themselves even if they
  444. 12:43failed. I do think that we will see less
  445. 12:46and less text in the world. I don't
  446. 12:47think we will be consuming most of our
  447. 12:49information in text. I think there'll be
  448. 12:50still be a place for text. You know, we
  449. 12:52still go to the theater, we still listen
  450. 12:53to vinyl. Like there'll there'll
  451. 12:54definitely be a place for text to exist.
  452. 12:56But I think in in the world of
  453. 12:57information sharing, training knowledge,
  454. 12:59etc. I do think that we will move into a
  455. 13:01world which we consume almost everything
  456. 13:03as video and audio and maybe VR and AR
  457. 13:05um in in 10 20 years with those
  458. 13:07technologies. I do think that's going to
  459. 13:08happen like and I don't think it's going
  460. 13:10to be as bad as people think it is. A
  461. 13:11lot of people are very like averse to
  462. 13:13the idea that people wouldn't
  463. 13:14necessarily be using tech as much as as
  464. 13:16they are today. But I think as humans
  465. 13:18it's very clear that we always been on
  466. 13:19this trajectory towards more interactive
  467. 13:22and more visual and more auditory
  468. 13:24content that reminds us more of like
  469. 13:26actually being in the real world, right?