The enterprise technology world is hitting a structural turning point of its own design. Over the past few years, the enthusiasm surrounding generative artificial intelligence has driven a frenzy of exploratory deployments across the Asia-Pacific region.
Yet, as organisations attempt to scale these systems into live environments, they are discovering that their underlying architectures require fundamental updates.
This requires a shift toward low-latency interconnection, sovereign compute nodes, and deeply integrated hybrid clouds. A recent FutureCIO roundtable in Hong Kong highlights how industry leaders are navigating these critical architectural choices.
The framework of control
For many enterprises, the sheer scale of modern large language models presents an immediate infrastructure constraint. Thomas Shum, head of Digital and Data Architecture at Sun Hung Kai Real Estate Agency Ltd, highlights this operational bottleneck.
“The most challenging factor when deploying into full production is managing the capacity of the model, particularly because organisations hit the strict rate limits imposed by cloud providers when user volumes spike unexpectedly,” says Shum.
This constraint forces a deeper re-evaluation of how algorithms are managed, prompting organisations to look closely at code security and agentic workflows. As Shum explains, companies must build comprehensive workflows to govern how natural language code is protected and how model outputs are validated before they reach end-users.
In heavily regulated sectors like financial services, this tension between raw scale and strict control becomes even more pronounced. An executive from the banking sector emphasises that the primary obstacle is not the capability of the AI itself, but the systemic risk surrounding its deployment.
He posits that: “When you provide autonomous agents to thousands of users across an organisation, the overarching challenge is risk management.”
To combat the threat of fragmented, unmonitored model creation, his institution has implemented rigid internal frameworks to govern agent deployment, testing capabilities with 10,000 users to ensure structural alignment before a wider rollout.
This mismatch between business expectations and operational realities frequently manifests as financial friction. Tom Lloyd, executive director and head of Infrastructure and Security, highlights the increasing focus among senior corporate leadership on achieving clear returns from investment in cognitive tools.
“We see strong momentum and interest in AI, and as adoption grows, senior executives are placing greater emphasis on ensuring these investments deliver clear and measurable business outcomes,” Lloyd says.
He also notes that as technology adoption accelerates, governance and data protection capabilities must continue to evolve at the same pace. To address this, we must take a proactive approach by strengthening controls and guiding users toward trusted and secure platforms, supported by advanced edge security solutions such as Cloudflare, to protect corporate information and maintain strong data governance.
The regulatory landscape itself adds a layer of complexity for product developers. Shweta Jain, vice president of Product Strategy, AI and Fintech Ecosystem at Finastra Hong Kong Limited, believes that the lack of clear mandates from regional oversight bodies complicates production timelines.
Jain adds that because regulators in different jurisdictions operate with fluid or poorly defined parameters regarding what constitutes a validated AI model, “bringing automated financial software to market safely remains a moving target.”
Data architecture and jurisdictional sovereignty in regional operations
As data sovereignty laws tighten across the Asia-Pacific region, the physical location of information dictates where compute infrastructure must reside. This geographical anchoring, known as data gravity, presents a profound challenge for multinational organisations that must balance localized legal mandates with the global nature of foundational models.
In the industrial and infrastructure sectors, Horace Chu, CIO of Gammon Construction, shared that data residency requirements are often defined upfront, particularly in public sector projects, where sensitive data must remain within specific jurisdictions and be securely encrypted.
Rather than adopting a single deployment model, Chu takes a workload-based approach distributing systems across on-premises environments, localized data centres, and cloud platforms depending on regulatory, latency, and risk considerations.
The priority is not just compliance but maintaining architectural flexibility and control as requirements evolve across different markets.
For established financial infrastructure operators like Octopus Cards Limited, data protection and operational resilience remain core priorities. Nigel Ng, section head, ITS at Octopus’ Technical Department, noted that the organisation’s longstanding, highly stable and scalable network environment requires careful consideration when evolving its architecture.
He highlighted that transitioning toward more flexible infrastructure models involves both technical and organisational adjustments, given the need to maintain service continuity, security, and reliability that customers expect.
Ng added that Octopus’ well-established payment processing network, which processes over 15 million transactions every day, accepted at over 190,000 transit, retail, and dining locations across the city of Hong Kong, —while robust—require ongoing optimisation to support emerging digital and future AI-related workloads. As such, any architectural evolution is approached in a measured and pragmatic manner, ensuring alignment with internal governance requirements and applicable regulatory expectations.
One delegate emphasises the difficulty of sourcing single technological platforms that satisfy highly disparate regulatory environments simultaneously.
“Deploying global models is rarely a uniform process, particularly when your footprint spans complex regulatory boundaries like mainland China, India, and Taiwan,” he observes.
Local compliance teams in each jurisdiction possess distinct definitions of acceptable risk, slowing massive adoption and keeping many projects constrained to highly controlled pilot stages.
This architectural fragmentation requires localized engineering teams to spend significant resources assessing regulatory gaps. One delegate from the banking sector notes that a simple copy-and-paste methodology across borders is fundamentally flawed.
“When launching solutions developed for one market into another, we discover that different central bank guidelines alter structural controls entirely,” he added.
This compliance friction is exacerbated by the soaring capital costs of specialised hardware, meaning that data science teams must constantly re-justify their infrastructure choices to corporate boards amidst shifting economic environments.
Security concerns must also address the direct conflict between localized mandates and international corporate policies. Another delegate comments on the delicate task of maintaining data privacy within mainland China while simultaneously adhering to strict United Kingdom corporate security policies.
He points out that the fundamental architectural question is “how to leverage advanced models within restricted markets without exposing confidential corporate intellectual property to external frameworks.”
The data engineering foundation
Legacy technical debt, accumulated over decades of mergers and siloed software acquisitions, acts as a primary drag on cognitive computing. To unlock the value of distributed AI, enterprises must engage in deep systemic cleaning.
Barry Sears, group director of Global IT Audit for a multinational insurance firm, targets the structural accountability of information management. He notes that many organisations are failing to achieve true strategic transformations because their data programmes remain overly focused on tools rather than systemic governance.
“Data ownership belongs explicitly to the business, yet there is a chronic misconception that the IT department is responsible for its quality,” Sears states.
He adds that moving forward with AI requires senior leaders to fund the necessary investment of data normalisation and data cleansing. Without this foundational work, any automation initiative is bound to underperform. Within his own domain, Sears is using agentic AI to transform the auditing lifecycle, aiming to reduce audit cycles from 120 days to under 30 days while boosting risk coverage from a minor 7% to a comprehensive 70% by using predictive models.
In the consumer retail world, data engineering takes the form of stitching fragmented customer journeys. A delegate from a leading jewellery retailer cites a critical challenge: capturing clean data from anonymous web and application traffic.
He explains that the enterprise must design deliberate digital incentives to encourage consumer authentication, “enabling the system to stitch anonymous digital footprints into a unified Customer Data Platform.”
This gathered data can then feed AI models to generate actionable commercial insights.
Nextgen DC: Power, interconnection, and advanced infrastructure
High-performance computing requires specialised facilities designed to handle unprecedented power densities and thermal loads.
Franco Lan, senior director at Equinix, points out that the operational landscape has moved past speculative exploration.
“We are observing a massive, region-wide race for physical GPU capacity, with service providers rapidly consuming power, space, and high-density space across Tokyo, Seoul, Singapore, and Malaysia,” he continues.
This physical infrastructure expansion is critical as enterprises adapt to rapid changes in financial markets, such as the transition toward 24/5 global trading desks and shortened T+1 settlement cycles, which demand real-time data processing and robust hybrid architectures.
Ultimately, the transition from legacy setups to AI-ready environments requires a fundamentally new networking philosophy. Anthony Ho, director of regional product and solutions marketing at Equinix, warns that moving successful pilot projects into live production fails if the network architecture remains isolated.
“A significant pitfall for organisations is executing a proof-of-concept in an isolated, low-cost environment, only to realise that full scale production requires a completely different ecosystem of connectivity,” Ho explains.
He emphasises that standard, hardware-dependent connections are insufficient for distributed workloads. Instead, the industry is shifting toward specialised AI networking designed to support complex business workflows.
These next-generation data centres, as Ho describes, must evolve beyond basic colocation, offering advanced liquid cooling, high-speed upgrades from 100Gbps to 400Gbps, and dense ecosystems that bring enterprise data, multi-cloud access, and specialised GPU-as-a-service providers into a unified physical environment.
