Architecting Inference: The OCP Path from Token Demand to Data Center Design
A one-day deep dive on the infrastructure that will serve Southeast Asia’s AI inference workloads through 2030 – built from the top down, grounded in open compute principles.
Day 1 – September 29, 2026 · 10:00 – 17:00
Overall Theme for the Day
Southeast Asia’s AI build-out has moved from speculation to construction. Malaysia has formally restricted non-AI data centres to protect grid capacity. Singapore has reopened its allocation process under stricter green-power criteria. Indonesia, Vietnam and the Philippines have revived nuclear programmes citing AI demand explicitly. Johor is one of the fastest-growing data center markets in the world. None of this was true two years ago and the operational decisions being made now – on what clusters get built, what packaging gets specified, and what facilities get permitted – will define regional AI capacity for the rest of the decade. Architecting Inference is a day to think through those decisions from the top down, through the lens of the OCP Community focusing on token demand, cluster architecture and facility design.
The conversation about AI infrastructure too often starts with hardware. This track starts with demand. Over the course of one day in Singapore, we trace the full top-down chain: from forecasting how many tokens users will actually consume, to designing the GPU clusters that will serve them, to building the facilities (power, grid, and cooling) that make those clusters physically possible. The format is OCP-flavored throughout: open hardware, open specs and open architectural patterns, applied to the operational reality of six SEA markets.
01 · Tokenomics : Demand Forecasting for AI Inference
Demand sets the floor for everything else. The first part of the day builds a bottom-up token-consumption model for the region: input/output token ratios by use case, country-level sizing across Indonesia, Singapore, Vietnam, Thailand, Malaysia and the Philippines, and the diverging trajectories of hyperscaler, sovereign, and neocloud deployments. Attendees leave with a shared vocabulary for regional inference demand at the granularity infrastructure decisions actually require.
02 · Cluster Architectures & Advanced Packaging
Once the demand picture is clear, we translate tokens into GPU hours and GPU hours into cluster designs. This section covers the architectural choices that matter for inference: network fabric design for east-west traffic, accelerator modules and DC-MHS reference designs from OCP, and the advanced-packaging supply chain (HBM, CoWoS, chiplets, 3D stacking).
03 · Power, Grid Readiness & Cooling Strategy
Megawatts, not GPUs, are the binding constraint in most of SEA. We close the day with the physical layer: grid capacity, power and cooling structures needed to serve the Southeast Asian region. Attendees leave with a defensible point of view on where regional capacity actually lands and why.
Special Sessions : Photonics & Quantum
Two interstitial panels between modules look beyond the GPU paradigm: silicon photonics and optical interconnect as cluster-scale fabric, and the state of quantum deployments in Singapore and the country’s emerging hybrid-quantum pilots along with the OCP’s efforts to ensure data center readiness for quantum deployments at scale.
Who this is for
Inference platform engineers, hardware engineers, cluster architects, data centre developers and operators, sovereign-cloud strategists, GPU neocloud operators, and regional policy leads working on AI infrastructure in or for Southeast Asia.