https://www.computerweekly.com/news/366627053/Enterprise-AI-adoption-moving-beyond-experimentation
For Moe Abdula, Google Cloud’s vice-president of customer engineering for Asia-Pacific, the past year has seen a “drastic difference” in the artificial intelligence (AI) landscape, with customer conversations shifting from experimentation to tangible business value and return on investment (ROI).
“A year ago, people were saying, ‘everybody’s experimenting, but when are we going to get to production?’ Nobody’s asking that now,” he said in a recent interview with Computer Weekly. “We’re starting to see people build ROI and thinking about building AI by default.”
In doing so, the spotlight often falls on the latest AI models, but Abdula noted the importance of the underlying infrastructure that powers them. He pointed to two key pillars for Google: its partnership with Nvidia and the parallel development of its own specialised hardware.
At Google Cloud’s Next 2025 conference in April, the company announced it is bringing its Gemini models to on-premise, air-gapped environments with Nvidia Blackwell systems running on Google Distributed Cloud. The move is aimed at addressing the needs of sectors with stringent data sovereignty and security requirements, such as government, defence and regulated industries like healthcare and finance.
At the same time, Google has developed its own tensor processing unit (TPU) architecture, which Abdula dubbed as the “backbone of AI” for Google. While graphics processing units (GPUs) have traditionally been optimised for training AI models, TPUs are designed for inference – the process of using a trained model to make predictions – which is where the bulk of the cost and computational load lies for large-scale AI services.
The rapid adoption of AI is also forcing a rethink of underlying infrastructure standards, including Kubernetes. The open-source container orchestration platform, which has become the de facto standard for modern applications, now faces a juncture as it adapts to the unique demands of AI workloads.
“Do we sustain a dual architectural model, which is pods and so on, and APIs [application programming interfaces] that allow you to interface at the resource level for something a little bit lighter and more connected with the resource managers of AI systems? Or do we simply let go of the whole pod?” said Abdula, adding that Google does not have a formal position yet, as it monitors developments in the open-source community.
In the interim, Google Cloud is enhancing its Google Kubernetes Engine to better support AI. This includes introducing code libraries that simplify the deployment of workloads to TPUs instead of GPUs. Additionally, Google has added more granular resource management controls to help organisations efficiently share expensive GPU resources among different teams.
As infrastructure evolves and AI models proliferate, Abdula positioned Google’s Vertex AI as the central platform to help customers manage the escalating complexity. Vertex AI lets organisations access and manage models from Google, open-source providers and other commercial labs, simplifying everything from commercial agreements to model migration.
For enterprises, a key challenge is managing the rapid pace of model updates, such as the transition from Gemini 1.5 to 2.5. Vertex AI has tools to ease this process, including an evaluation service to ensure quality parity between model versions and even the ability to make a new model behave like its predecessor. For users struggling with prompt compatibility, Abdula said Google has “introduced a toolkit that allows you to do the migration of prompts”.
Abdula also touched on the emergence of agentic AI systems – AI that can reason, plan and automate complex workflows by interacting with data and other systems. “This is perhaps one of the areas that everybody is seeking to understand how to work through,” said Abdula.
Google Cloud was one of the first in the market to provide agentic AI development tools – its Vertex AI Agent Builder and the open-source Agent Development Kit (ADK) aim to simplify the creation of agents that can connect to any data source and interact with other agents.
While the productivity gains are promising, AI agents can bring governance challenges. A primary concern is how to manage the permissions and access rights of these agents as they become more integrated into core business processes.
“The idea of being able to deliver an agent is actually not that sophisticated or complex,” said Abdula. The real evolution, he said, is moving from simple task automation to agents that have reasoning capabilities with the right integrations and controls around what they’re authorised to do or not do.
To address this, Google has developed Agentspace, a platform that serves as both a gallery for discovering and reusing agents, and a governance framework where each agent has a defined set of governance rules. “The governance rules will say, ‘Can I inherit Moe’s credentials and act as Moe? Or are there specific restrictions?’” said Abdula. These rules can dictate whether an agent is fully autonomous or requires human-in-the-loop approval for certain actions.
However, as companies begin to deploy legions of these digital workers, a key question arises: how can IT and security teams manage their permissions from a single, unified location, just as they do for human employees through identity and access management (IAM) systems?
When asked if enterprises can manage agent permissions in the broader Google Cloud IAM console alongside human users, Abdula acknowledged that it could be a potential feature. “That has not yet been integrated in a seamless way and it’s a great feature, but today you have to kind of do it in the construct of Agentspace,” he said.
For Abdula, who relocated to Singapore two years ago, seeing AI technology translate into local success stories is particularly rewarding. He cited the work by enterprises like DBS Bank and Prudential as proof of the “world-pioneering” innovation happening in the region.
DBS, for instance, is using AI to enhance customer service and provide hyper-personalised financial guidance to customers. Prudential is also partnering with Google Cloud to use its specialised MedLM models to analyse medical documents in a bid to speed up insurance claims processing and reduce errors.
This enterprise adoption is bolstered by strong government support. Abdula pointed out the AI Cloud Takeoff programme, a joint initiative with Digital Industry Singapore to help up to 300 local companies establish their own AI centres of excellence. It builds on the success of the AI Trailblazers programme, which has already helped scores of organisations build generative AI products.
“The epiphany hit me,” said Abdula, reflecting on his career shift from engineering, where he worked on Docker containers, among other technologies, to a customer-facing role.
“We bring all these amazing advancements, but are we enabling people to take advantage of them?” he asked.
With a maturing ecosystem and AI adoption shifting from experimentation to ROI-driven implementation, it seems the answer for many enterprises is increasingly “yes”.
02 Jul 2025