Standard Chartered’s AI Factory, a computing environment for running artificial intelligence (AI) workloads, has consumed about 500 billion tokens to date. That volume of AI use is part of a wider challenge. As models and applications evolve, the bank needs reliable data, controls built into development and systems it can update without disrupting services.

At a media roundtable in Kuala Lumpur on 22 September, the bank’s technology leaders described how they are embedding controls earlier in development, bringing data together across systems and expanding capacity to run AI. The work spans markets with different regulatory requirements, where faster technology change must be managed without weakening resilience or customer trust.

Alvaro Garrido, Group Chief Information Officer, said it was “not as simple as purely retrofitting AI into how we’ve been doing things traditionally in the bank”. Alongside experimentation with AI, Standard Chartered had spent substantial time fixing its technology foundations, he said. The work spans shared capabilities such as infrastructure, data and payments, on which its corporate, retail and other businesses can develop their own services.

“We spend a substantial amount of time building the controls and the governance and the guardrails to make sure that we deploy the AI in a safe way,” Garrido said.

Standard Chartered operates across jurisdictions with different regulatory requirements. Garrido said it needed common technology foundations that could support reliable services and let businesses develop products for their own customers and markets.

Engineers and business process owners are rethinking how they work, with employees proposing applications and building agents they can share between groups. Garrido said the bank was investing in training and retraining while keeping that experimentation within its governance framework.

Connecting data and moving controls into development

Shebani Baweja, Group Chief Data Officer, said Standard Chartered was bringing together information from about 150 disconnected systems through what it calls a global data supply chain. Some components of its global data platform already exist. Baweja expects the broader approach to develop over the next two years.

The platform is being built first around regulatory, financial and non-financial reporting. It is designed to bring data from the different systems into a common environment and check it as it enters. Baweja said teams could then use the same verified information for reporting, analysis and AI applications.

Baweja said AI could help identify weaknesses in data quality, but would not resolve them. As AI is deployed more widely, it can amplify both the strengths and shortcomings of the underlying data. Its quality, consistency and trustworthiness will therefore shape how effectively the bank can use AI across markets, products and client journeys.

Discussing data governance, Baweja said 80% of the bank’s key data controls have been codified and 90% are continuously monitored through automated AI guardrails. She also reported 95% data accuracy in AI-powered document processing, although the scope for that figure was not specified.

The bank is changing when data controls are applied. Baweja said retrospective reviews could uncover data problems close to launch, creating backlogs when an application was otherwise ready. Standard Chartered is seeking to “shift left” by incorporating data standards and controls into design and development. “Both speed and control can coexist,” she said.

Baweja said lower-risk applications need room for rapid experimentation, while higher-risk ones require more rigorous review. Teams therefore need to know which data and controls apply to each application they build.

The bank is developing data-management learning modules, from foundational training for a broad workforce to more advanced material for engineers, architects and data specialists. Baweja said data literacy was everyone’s responsibility, although what an employee needs to understand varies by role.

In practice, Baweja said relationship managers were spending less time finding information on multinational clients across separate systems and more time deciding how to serve them.

Applying common data standards across markets

Business teams can develop their own analytics and AI applications using data supplied through the platform, within common controls. For more transactional uses, Baweja said they may connect directly through application programming interfaces to the relevant systems of record. The choice depends partly on whether a team needs current operational data or a view that includes historical information.

Baweja said Standard Chartered sets common policies and minimum controls centrally, including requirements for data protection, privacy and record-keeping. Its market teams then work with compliance colleagues to incorporate additional local requirements on issues such as data sovereignty.

Baweja described this as a combination of centralisation and federation. The central platform checks and prepares data, while businesses determine how to use it within the controls that apply to them. Local requirements add to the global baseline where necessary. Garrido said security standards needed a high degree of consistency because threats do not follow national boundaries.

Choosing models and building AI capacity

John Sharratt, Global Head of Technology and Infrastructure, said growing AI use was forcing the bank to plan its computing capacity and choose models more carefully. “As you start to scale, things start to creak and break,” he said.

For coding work, engineers face restrictions on which models they can use and work within budgets. Those who need to exceed their budgets must request an extension. For inference applications going into production, the bank selects a model suited to each use case at the design stage. It has produced an internal paper to guide those decisions across teams.

Sharratt said Standard Chartered had already purchased or placed committed orders for capacity sufficient for its projected needs this year and next, including hundreds of graphics processing units (GPUs) for inference run by the bank.

External frontier models can serve tasks that need their capabilities, while models the bank operates itself offer another option where data requirements, availability or cost favour them. Sharratt calls inference that can run on a single GPU “commodity AI”. He described the mix as a hedge, giving the bank more than one source of AI capacity as model capabilities and costs change.

He said the bank’s OCR-based document processing uses a mid-sized model and is among the AI Factory’s larger token consumers. Customers use the AI Factory indirectly through processes such as this.

Sharratt expects capabilities now associated with frontier models to move into smaller models as they are distilled. An architecture designed for mid-sized models today could, in six months, run “frontier-grade inference on mid-tier cost base”, he said. This expectation informs the bank’s investment in capacity to run smaller models, although the pace of improvement and the cost advantage remain to be tested in practice.

The cost case for owned GPUs depends on utilisation and the costs of power, cooling, software and operation. Workload allocation also affects the comparison with external models. Sharratt estimated a payback period of six to 36 months for a GPU running AI applications that generate business value. Standard Chartered is targeting six months, although he did not specify how the bank calculates payback.

Garrido expects the economics to change again. He suggested that in three to five years token costs may cease to be the main point of competition, with greater weight placed on how banks connect models to their wider technology and business processes. For now, he said, hardware scarcity and rising prices remain pressing concerns.

Making room for AI in the private cloud

Running more inference on the bank’s own infrastructure also depends on the capacity of its private cloud. Sharratt said roughly 70% of the bank’s global technology estate had moved to a virtualised architecture. The new environment has handled around 300,000 virtual-machine deployments cumulatively, with approximately 50,000 running at present. He attributed the difference to workloads being deployed and replaced repeatedly.

Sharratt said the bank’s data centres in Hong Kong and Singapore were around 70% virtualised. Equipment was being put in place in locations in the United Kingdom over 200 kilometres apart, with application migrations there planned for next year. The bank is also extending its private-cloud capacity to more of its markets alongside the public-cloud services it uses.

Sharratt said power consumption is 30% to 40% lower on the new private-cloud architecture, freeing about five kilowatts per rack. He said that headroom would allow the bank to add GPUs to existing racks.

As more AI applications move into production, Standard Chartered will have to keep updating the data, controls and computing infrastructure beneath them. The measure of this work will be whether it can make those changes more frequently without increasing operational failures or disrupting customers.