“By 2024, 60% of the data used for the development of AI and analytics projects will be synthetically generated.” This is a prediction from Gartner that you will find in almost every single article, deck or press release related to synthetic data.
We are repeating this quote here despite its ubiquity because it says a lot about the total addressable market of synthetic data.
Let’s unpack: First, describing synthetic data that is “synthetically generated” may seem tautologic, but it is also quite clear: We are talking about data that is artificial/fake and created, rather than gathered in the real world.
Next, there’s the core of the prediction — that synthetic data will be used in the development of most AI and analytics projects. Since such projects are on the rise, the correlation is that the market for synthetic data is also set to grow.
Last but not least is the time horizon. In our startup world, 2024 is almost today, and people at Gartner already have a longer-term prediction: Some of its team published a piece of research “Forget About Your Real Data — Synthetic Data Is the Future of AI.”
“The future of AI” is the kind of promise that investors like to hear, so it’s no surprise that checks have been flowing into synthetic data startups.
In 2022 alone, MOSTLY AI raised a $25 million Series B round led by Molten Ventures; Datagen landed a $50 million Series B led by Scale Venture Partners, and Synthesis AI pocketed a $17 million Series A.
Synthetic data startups that have raised significant amounts of funding already serve a wide range of sectors, from banking and healthcare to transportation and retail. But they expect use cases to keep on expanding, both inside new sectors as well as those where synthetic data is already common.
To understand what’s happening, but also what’s coming if synthetic data does get more broadly adopted, we talked to various CEOs and VCs over the last few months. We learned about the two main categories of synthetic data companies, which sectors they address, how to size the market and more.
The tip of the iceberg
Quiet Capital’s founding partner, Astasia Myers, is one of the investors bullish about synthetic data and its applications. She declined to disclose whether she invested in this space, but said that “there’s a lot to be excited about in the synthetic data world.”
Why the enthusiasm? “Because it gives teams faster access to data in a secure way at a lower cost,” she told TechCrunch.
“We can simply say that the TAM of synthetic data and the TAM of data will converge.” Ofir Zuk (Chakon)
Access to large troves of data has become critical for machine learning teams, and real data is often not up to the task, for different reasons. This is the gap that synthetic data startups are hoping to fill.
There are two main contexts in which these startups focus: structured data and unstructured data. The former refers to the kind of datasets that sit in tables and spreadsheets, while the latter points toward what we could call media files, such as audio, text and visual data.
“It makes sense to distinguish between structured and unstructured synthetic data companies,” Myers said, “because the synthetic data type is applied to different use cases and therefore different buyers.”
According to MOSTLY AI CEO Tobias Hann, most of the demand for structured synthetic data comes from banking, insurance, telecommunications and healthcare companies.
These four highly regulated sectors are attracted by the possibility of plentiful — yet privacy-preserving — data. Whether synthetic data can deliver this or not is still somewhat controversial, but several companies think so, as do their investors.
Recently funded companies in this structured data vertical include MOSTLY AI; Tonic.ai, which raised a $35 million Series B funding round; and Gretel AI, which closed a $50 million Series B round last October. A fuller list and market map can be found in this Medium post by synthetic data advocate Elise Devaux, whose employer Statice is also a competitor.
As captured by Devaux, the unstructured synthetic data side of the market is represented by a whole other range of companies, such as the above-mentioned Datagen and Synthesis AI. Some, such as Parallel Domain, appeared a few years ago, while others, such as Scale AI, entered the space more recently. But they have one thing in common: They have no need to envy their structured data peers when it comes to attracting funding or clients.
Visual synthetic data, for instance, has a variety of use cases. According to Datagen CEO Ofir Zuk (Chakon), four of these are accelerating faster than others: AR/VR/metaverse, in-cabin automotive and automotive in general, smart conferencing and home security.
However, Datagen is also making sure its algorithms and technology are domain agnostic to ensure it will be ready when synthetic data usage takes off in other sectors, such as retail and robotics. And there’s little reason to doubt that it will be the case.
An ongoing democratization
“Deep learning with synthetic data will democratize the tech industry,” general partner at LDV Capital Evan Nisselson predicted in a 2018 TechCrunch guest column.
By democratization, Nisselson meant leveling the playing field. By using synthetic data, startups would be able to do applied machine learning without the type of big data that only large tech companies had at that point.
Nisselson’s prediction both held up and didn’t. Synthetic data helped underdogs in their David-Goliath fight. But now, Meta and the like want to have their cake and eat it, too. Nisselson acknowledges as much.
In 2018, “a lot of the people at big companies said: ‘Evan, we have more data than we need, we don’t need to make synthetic data.'” The only exception was for addressing edge cases in applications such as autonomous vehicles.
“But many things have changed, and I think more and more [big companies] will leverage synthetic media,” Nisselson said.
Synthetic data has selling points that can appeal to companies of all sizes. “It’s a faster and sometimes cheaper way to train systems,” Nisselson said. This is in part because the data is itself generated with AI, for instance via generative adversarial networks (GANs). But it is also because costly steps such as annotation and labeling could potentially be skipped.
If one buys into the bullish hypothesis on synthetic data becoming the main form of training data, it becomes easy to calculate its total addressable market (TAM), Zuk said. “We can simply say that the TAM of synthetic data and the TAM of data will converge.”
This market could even be bigger than expected if new use cases are unlocked within companies. “We are just starting to see the early innings of synthetic data’s role in organizations,” Myers said.
It is not clear yet how this democratization within organizations will happen, but a low-code or no-code approach will likely help. MOSTLY AI’s strategy, for example, is very much focused on unlocking business value for clients that include Fortune 100 banks and insurers, as well as telcos. Because of this, the startup created a platform that can not only be used by data scientists, but also by software engineers and quality testers.
What’s next?
Some applications of synthetic data are still novel for the general public, but already too late for a VC firm like LDV, whose thesis is to invest as early as possible in teams leveraging cutting-edge technology.
This makes Nisselson interesting to talk to, keeping that bias in mind. For instance, he describes synthetic data applied to transportation or to robotics — such as training robots to pick up items in warehouses or factory lines — as “crowded” already.
Use cases highlighted by Nisselson also included training autonomous retail, like AiFi does; marketing, advertising, and e-learning, like his portfolio company Synthesia, which uses synthetic avatars for corporate training; gaming, including the metaverse; and what he calls “pure content creation.”
Myers too highlighted the use of synthetic data in content creation. “Synthetic data is most affiliated with deepfakes that replace one person’s likeness or voice with synthetic versions,” she said, but “these technologies can also be applied to commercial applications.”
One thing is for sure: With startups blossoming and use cases aplenty, some of these companies will eventually buy each other or get acquired by tech giants. Last October, even before rebranding as Meta, Facebook already quietly acquired AI.Reverie. With other large companies partnering with synthetic data startups more or less openly, and considering the current venture capital climate, we would be surprised if some of these didn’t lead to M&As in the near future. It’s something that we will definitely be tracking.