Vertex AI

Vertex AI is an offering from Google that companies can rent over the internet. It allows the development of programs that learn from examples and then independently solve tasks. Google provides the computers that such programs need for this. It also supplies ready-made language programs that can write texts or answer questions. Anyone using Vertex AI therefore doesn’t buy hardware but pays for usage based on consumption. Such rental offerings on the internet are called cloud services, and Vertex AI is the part of that which handles artificial intelligence.

Why companies don’t build it themselves

Modern AI models require special graphics chips, known as GPUs. A single such chip quickly costs several tens of thousands of euros. Training a large model requires hundreds of them, often for weeks. For most companies, this would be a pointless investment, because the chips would just sit around afterward.

Vertex AI spreads these costs across many customers. A bank that wants to build a chatbot for its customers rents computing time for a few hours. Payment is only for what was actually consumed. This significantly lowers the barrier to entry and is the main reason for the success of such platforms.

For Google itself, Vertex AI is an important business area. The company competes directly with Amazon and Microsoft, which operate similar platforms with Bedrock and Azure AI Foundry. In quarterly reports from Alphabet, Google’s parent company, the growth of the cloud division therefore regularly appears as a separate metric.

From dataset to running application

Vertex AI covers the entire journey of an AI project. At the beginning there is data, such as thousands of customer inquiries from support. This data is stored and sorted within the platform. Then one selects a model that should learn from it.

For the selection, there is a catalog called Model Garden. It contains Google’s own Gemini models, but also freely available models from other providers. One can use such a model unchanged or retrain it with one’s own data. This retraining is called fine-tuning and makes a general model more accurate for a specific task.

Once the model is finished, it is deployed as an endpoint. An endpoint is a fixed internet address to which another program sends questions and receives answers back. Vertex AI then ensures that enough computers are running when many requests come in simultaneously. Additionally, the platform logs how good the answers are. This way, an operator notices when the model gets worse over time because the real data has changed.

Vertex AI in products and headlines

As a private individual, one never uses Vertex AI directly. Yet one constantly encounters the result. When an insurance company offers a chat that answers contract questions, a cloud platform like this often runs behind it. Search functions in online shops or automatic translations on company websites also often arise this way.

In business news, the name usually comes up in connection with major contracts. Headlines like “Corporation X to rely on Google’s Vertex AI going forward” are indicators of how the market is distributed among the three major cloud providers. A common misconception here is confusing Vertex AI with Gemini. Gemini is a single model, whereas Vertex AI is the environment in which this and many other models are operated.