The sentence that explains Synthesia is not about artificial intelligence. It is about two PhDs, 10 days and 30 seconds of video. In the company's early years, that was roughly what a bespoke dubbed clip could require. The result looked futuristic: a person seemed to speak a language they had never recorded. The process looked like a consultancy with an expensive science project in the back room.
It could still make a splash. In 2018, a BBC presenter appeared to speak Spanish, Mandarin and Hindi. The following year, a Malaria Must Die campaign gave David Beckham nine languages and more than 700 million digital impressions. The work made synthetic video legible to the public. It did not make production repeatable.
Victor Riparbelli, Steffen Tjerrild and computer-vision professors Matthias Niessner and Lourdes Agapito had founded Synthesia in London in 2017. For three years they sold dubbing and translation work to film, advertising and media customers. Riparbelli now describes that product with the chill of a founder who has seen the invoice: mildly successful, not very big, not scalable.
The customers arrived carrying slide decks
Around 2021, strange inbound requests began to pile up. People inside large companies wanted to make training and onboarding videos. They did not know how to operate a camera. They could not win budget for a conventional shoot. Often, what they needed was simply a talking head explaining a process clearly. This was no red-carpet market, but it recurred every time a company hired people, changed software, launched a product or rewrote a policy.
Synthesia turned the service into Studio, a browser application. A user types or pastes a script, picks a licensed avatar and synthetic voice, adds slides or a screen recording, then generates the result. A custom avatar can reproduce an employee's likeness with recorded consent. Translation, dubbing and a multilingual player handle localization. Brand kits, comments, workspaces, analytics, SCORM export, an API, SSO and audit controls turn a clever rendering trick into something a corporate department can actually govern.
That last row is the important one. HeyGen, Colossyan, D-ID, DeepBrain AI and Hour One can all put synthetic presenters on screen. Loom can capture a walkthrough. Runway, Veo, Sora and Firefly chase more cinematic generation. Synthesia has chosen the less glamorous middle: repeatable business communication, where permissions, updates and dozens of language versions matter as much as the face.
The cost is easy to quote. The savings need context.
Synthesia runs a familiar SaaS ladder. A Basic plan is free. Public monthly pricing observed in 2026 listed Starter at $29 and Creator at $89; annual billing discounts both. Enterprise pricing is negotiated and adds unlimited minutes, the full avatar library, centralized workspaces, brand controls, SSO and higher API limits. Each regeneration consumes usage, so a sloppy review cycle can still have a bill.
The useful numbers come from the jobs, not the price card. Berlitz says it made more than 1,700 language-course micro-videos in six weeks, cutting production time 70% and costs 66%. Avetta says its support-training program saved 1,300 hours over two years; new agents reached proficiency 20% faster. KONE reports 30% faster production and says pre-learning can save a day on multi-day technical courses. These are customer case-study figures, not controlled comparisons, but they reveal where the economics can work: many short videos, frequent revisions, broad distribution and expensive coordination.
A $100 million workflow, disguised as a talking face
By April 2025, Synthesia said annual recurring revenue had passed $100 million. It counted more than 65,000 businesses, with thousands of enterprise customers using the platform daily for training, product marketing, support and sales enablement. Adoption later rose above 90% of the Fortune 100 by the company's count. Revenue follows subscriptions, with enterprise contracts historically supplying the majority.
Capital followed. A $90 million Series C made Synthesia a unicorn in 2023. A $180 million Series D in January 2025 valued it at $2.1 billion. Twelve months later GV led a $200 million Series E at $4 billion, joined by NVentures, Accel, Kleiner Perkins, NEA and others. Adobe Ventures also made a strategic investment in 2025. Publicly announced rounds add to more than $530 million; the January 2026 deal included a Nasdaq-arranged secondary sale so employees could turn some paper wealth into cash.
The company's expertise now spans video generation, speech synthesis, lip synchronization, computer vision and applied model research. Yet its defensible work may be closer to systems design: get a document into video, keep it on-brand, route it through review, publish it to an LMS, measure it and update every language version when paragraph four changes.
The trust department is part of the product
Synthetic media comes with a problem that normal slide software does not: the output can make a real-looking person say something they never said. Synthesia requires recorded consent to build personal avatars, moderates content before generation and uses automated checks plus human review. Its acceptable-use rules restrict impersonation, election manipulation, explicit deepfakes and several high-risk applications. The company says 10% of its team will always work on safety and ethics, and it was the first generative AI company certified to the ISO 42001 management standard.
This is also market positioning. Large employers care about provenance, privacy, content ownership and whether a tool can be switched off across every workspace. Synthesia's work with the Partnership on AI, the Content Authenticity Initiative and external red teams gives procurement officers something more durable than a promise to behave. Its Shutterstock research license puts a similar idea on the input side: use selected, licensed footage to investigate better avatar models.
There is friction by design. A script may enter manual review. A stock avatar cannot endorse every product or appear in every paid advertisement. Political content faces stricter identity and account rules. Those limits annoy anyone expecting an unrestricted generator. They are also part of why a bank, pharmaceutical company or government department might permit the tool in the first place.
Now the training video wants an answer
Synthesia 2.0 expanded creation into screen recording, document-to-video assistance, managed translations and interactive players. Synthesia 3.0 pushes further. Express models improve gesture and vocal delivery. Dubbing 2.0 aims for more natural translation and lip sync. A conversational Assistant plans an on-brand video through chat. Video Agents and Roleplays let a learner speak to an avatar, rehearse a customer call and receive feedback instead of merely pressing play.
This is a larger claim than cheaper production. The product wants to sit between corporate knowledge and employee performance. It can make training active, personalized and measurable. It can also make workers wonder who reads the transcript, whether the coach is right and whether practice scores will leak into employment decisions. Synthesia's own policy says outputs must not be used as a basis for employment decisions. The feature succeeds only if managers honor that boundary and learners trust the room.
What the next founder can steal
The portable lesson is not “add an avatar.” It is to notice who is asking with urgency. Synthesia's film customers liked the magic; enterprise employees repeatedly needed a video and lacked the budget or skill to make one. The company built for that desperation, packaged a service as software and added the dull controls that made repeated use possible. Internally, Riparbelli calls the principle “utility over novelty.” Projects that look impressive but fail to solve a recurring problem get shut down.
- Count unsolicited requests, not applause for a demo.
- Choose work that repeats and needs revision.
- Turn expert labor into a customer-operated workflow.
- Build permissions and review before enterprise sales demand them.
- Measure hours removed from the process, not content generated.
- Keep a human editor responsible for meaning and accuracy.
Where the camera still wins
Synthesia is least convincing when the person is the evidence. A founder apology, a patient testimony, a documentary interview or a performance that depends on tiny emotional choices should probably involve the person and a camera. The platform is also a poor shortcut for weak writing. A polished avatar cannot rescue an overlong compliance script; it can merely pronounce every dreary sentence in 160 languages.
The business case weakens for one-off hero films, small audiences, infrequent updates or teams that already produce video cheaply. Regulated advice and safety-critical instruction require native-speaker review, subject-matter approval and careful disclosure. Conversational agents add latency, hallucination and surveillance questions that static video avoids. Under those conditions, the old shoot may be slower but more honest.
Synthesia found its market by accepting that limitation. It did not need to replace cinema. It needed to replace the awkward moment when a training manager looks at 80 pages of updated policy, six regional teams and no production budget. The first act was making Beckham speak nine languages. The durable act was letting everyone else fix sentence 37 without calling him back.
Company and customer metrics are the latest publicly reported figures available through September 2026. Customer savings are self-reported case-study results and will vary with volume, workflow and review requirements.