Image generated by Deeptech Times using Google Gemini
As enterprises race to operationalise AI, AI success is becoming less about models and more about the data foundations beneath them.
In an interview with Deeptech Times on the sidelines of AIMX Singapore 2025, Yuanhao Sun, founder and CEO of Transwarp, offered a candid and technically grounded view of why many organisations struggle to become truly AI‑ready, and what it will take to move from fragmented data environments to intelligent, scalable infrastructure.
According to Sun, one of the most persistent misunderstandings among enterprises is the assumption that adopting AI tools or LLMs automatically translates into AI readiness. In practice, many organisations are still operating in what he describes as a state of data chaos: siloed systems, inconsistent data quality, fragmented governance frameworks and architectures built for yesterday’s analytics rather than tomorrow’s AI workloads.
“AI magnifies existing weaknesses in data infrastructure,” Sun explained. “If your data is fragmented or poorly governed, AI does not fix that; it amplifies the problem.” The consequence is a proliferation of proof‑of‑concept projects that fail to scale and deliver limited business value despite significant investment.
From silos to a unified intelligence layer
Transforming data chaos into a unified intelligence layer, Sun argued, requires rethinking infrastructure from the ground up. Traditional enterprise data platforms were designed primarily for structured data and batch analytics. Today’s AI systems, particularly those driven by generative and agentic AI, demand far more.
Enterprises must consolidate structured and unstructured data: documents, images, video, audio and sensor data into a single, multimodal data platform. This shift is not trivial. Much of this information is scattered across enterprise applications, employee devices and legacy systems, and historically lacked the tools needed for meaningful analysis.
GenAI changes that equation by making it possible to process and extract value from unstructured data at scale. But Sun cautioned that consolidation alone is insufficient. Data must also be made “AI‑ready” through processes such as data engineering, transformation and contextualisation, which enable models and agents to act on it reliably.
Scalability without runaway costs
In fast‑growing markets such as APAC, scalability and cost efficiency are inseparable concerns. Sun noted that many organisations are deterred from deploying AI by the high upfront cost of GPU infrastructure, which can run into hundreds of thousands of dollars per server. The return on that investment is often uncertain, particularly in early deployment stages.
Transwarp’s approach, shaped by years of experience supporting data‑intensive workloads in China, emphasises distributed systems and incremental scaling. Enterprises can start with a single server, workstation or even a laptop‑class AI PC, and expand capacity by adding nodes as demand grows, without interrupting applications.
Crucially, Sun highlighted recent advances in optimising AI software to run efficiently on lower‑end GPUs. By fine‑tuning smaller language models and improving inference efficiency, Transwarp enables organisations to deploy AI locally, reduce hardware costs and maintain acceptable performance. This approach, he argued, lowers the barrier to entry and allows enterprises to validate business value before committing to large‑scale infrastructure investments.
Data sovereignty and the limits of the cloud
As AI adoption accelerates, concerns around data privacy and sovereignty are intensifying. Sun observed that many enterprises and governments remain uneasy about cloud‑based AI deployments, particularly when sensitive or strategic data is involved.
While data residency requirements (keeping data within national borders) are often seen as a solution, Sun described this as only a partial safeguard.
If infrastructure and software are controlled externally, the risk of data leakage persists. For organisations with the necessary capabilities, on‑premises deployment remains the preferred model, which offers full control over data, infrastructure and access policies.
This preference is driven by a broader realisation: as AI models become increasingly commoditised through open source and commercial offerings, proprietary data becomes the primary source of competitive advantage. Protecting and governing that data is therefore mission critical.
Fine‑grained governance in the age of AI
Effective AI governance, Sun argued, requires far more than high‑level policies. Enterprises must implement fine‑grained access controls, robust data classification and mechanisms to prevent sensitive information from being inadvertently exposed through AI training or inference.
To address this, Transwarp is introducing what Sun described as a “cognitive database” – a system that embeds AI directly into the data layer. Such databases can understand content semantically, support natural‑language queries across multiple languages, and enforce granular access controls based on data sensitivity. Confidential information can be isolated in external knowledge bases, while non‑sensitive data is safely leveraged for AI workflows.
How GenAI and agentic AI are reshaping data infrastructure
The rise of generative and agentic AI represents a fundamental departure from earlier machine learning paradigms. Previously, enterprise data platforms focused on labelled, structured datasets designed for predictive models. Today’s AI systems must ingest and reason over vast volumes of unstructured, multimodal information.
Sun also emphasised the growing importance of knowledge graphs and domain‑specific knowledge capture.
Much of an organisation’s most valuable expertise remains undocumented, residing in the experience of employees rather than formal systems. By structuring this knowledge and integrating it into AI‑driven workflows, enterprises can automate complex processes while reducing dependence on individual expertise.
Rethinking hardware and storage
Looking ahead, Sun believes some of the most transformative changes will occur at the intersection of hardware and software. Today’s AI deployments are constrained by memory limitations and the high cost of high‑bandwidth memory, forcing vendors to restrict data context sizes and concurrency.
Emerging architectures that enable GPUs to access high‑speed storage directly could change this equation. By offloading large context windows, indexes and knowledge graphs from memory to storage, enterprises could support far more users and much larger datasets on lower‑cost hardware.
Transwarp is developing software optimised for this new paradigm, alongside so‑called “cognitive storage” systems that decouple data scaling from compute scaling.
For Sun, sustainable AI adoption is not about chasing the latest model, but about building resilient, intelligent data infrastructure. Organisations that invest in scalable architectures, rigorous governance and data‑centric innovation will be best positioned to unlock long‑term value from AI.
