Roblox is Deploying its Own Generative AI Models and Infrastructure at Lower Cost

Cost, data privacy, and regulation will give open-source models a boost in 2024

Roblox CEO David Baszucki commented during a recent earnings call about the company’s plans for generative AI. Some of the most instructive comments were made in response to a question asked by a BTIG analyst:

We think experiences, for example, a fashion design experience is going to migrate all the way from using a virtual sewing machine and scissors to purely based text prompts. And ultimately, we'll enter an age on Roblox where anyone can make their avatar or clothing 100% through text prompts. So if I or you wanted to build a piece of clothing, we could describe it. We're going to see that created in real time.

We think for developers, this is going to accelerate quality up and down the stack. …We will see people who before didn't expect to be creators making immersive 3D experiences. And we'll see the ones that are created by developers become more rich and dynamic. We may ultimately even see experiences that are dynamically personalized for each individual player. So it's really early, but we think it's a really exciting frontier for 3D creation.

Game platforms genuinely believe that generative AI will enable game designers to create game code using text prompts, software developers will use the same text prompt features to generate 3D characters, objects, and images, and everyone else will use both. This is the future they are actively building.

Baszucki also highlighted how generative AI could be used within a game experience with the fashion design example. If you consider that generative AI may be adopted for both game development and to support real-time in-game features, the obvious conclusion is that gaming may become a large market for AI inference jobs.

Running its Own Infrastructure

In prepared remarks, Baszucki also commented on the company’s intent to run its own generative AI infrastructure.

There are many, many, many areas that we're already live on the platform with ML and AI stacks and more to come. I want to highlight that we have 70 machine learning training stacks right now. We have trained, for example, our own translation model with one billion parameters. This is the model that helps auto-translate all experiences when a creator makes them. And the range of verticals that we have live right now, we've mentioned material generation and code generation have shipped.

We mentioned that we're making really impressive gains in both quality and cost throughout our safety systems on all types of assets, which is live.

We've mentioned that we're building our own model internally and running our own inference on voice, safety, internally on search—we're live—and we have a lot of 3D generative coming as we move to creation everywhere and our avatar project.

…

We believe long term that on our platform because we can run infrastructure, inference on our own infrastructure, there's an amazing opportunity to run inference all over the platform and run it at extremely low cost.

Roblox is a very large gaming platform. Baszucki said the company is approaching five billion monthly interaction hours on the platform. That turns out to be a lot of potential training data for AI models, and also a lot of potential generative AI inference cost.

The comments suggest Roblox does not plan to pay OpenAI or any of its many foundation model competitors for training or inference. The economics will be far stronger if Roblox can efficiently use its own infrastructure.

Our personalized recommendations right now, 100% of these are running on our own infrastructure, running inference, and doing it really cheaply at scale. Also, all of our safety pipelines, image, audio, voice, text, and 3D are running on our own platform, doing inference on our own platform at significant efficiencies of cost and driving quality there.

The Market for Customized Foundation Models

While most companies will find it more cost-effective and convenient to use a proprietary generative AI foundation model from third-party developers, some of the largest organizations will deploy their own to make the economics feasible.

Training and inference costs are net new expenses that many companies will soon have to incorporate into product pricing. That cost impact may be lower if they deploy their own models and don’t have to pay a third party for access.

There are also reasons beyond economics that are compelling a number of large companies to adopt and run open-source models on their own infrastructure. Data privacy and security are among the biggest motivations. This is particularly true for regulated industries. They simply don’t want or cannot have their data—user input or system output—traveling through another company’s servers. There are compliance considerations as well as other legal and policy requirements that will make third-party models hard to adopt.

A large portion of Roblox’s user base is children. That means privacy and regulatory considerations need to be considered alongside cost impacts.

When you consider how the generative AI market and even your own generative AI selection criteria, cost, privacy, and security will be important variables. These will also be key reasons why the open-source generative AI segment will see tremendous activity in 2024.