MediaTek Executive Projects Million-GPU Clusters as Mid-Tier by 2030

The New Baseline for AI Infrastructure

The artificial intelligence industry is racing toward compute scales that would have seemed absurd just two years ago. Bor-Sung Liang, senior director at MediaTek, told an industry forum that deployments of one million Nvidia H100 GPUs will be considered mid-tier infrastructure by 2030, a projection that underscores how rapidly the baseline for competitive AI development is rising.

Speaking at a semiconductor outlook event, Liang described an industry trajectory that extends beyond terrestrial data centers. The compute demands of next-generation AI systems are pushing providers to consider space-based infrastructure as a viable, perhaps necessary, expansion path within this decade.

From Single Models to Agent Teams

The architectural shift driving this compute explosion is the move away from monolithic large language models toward what Liang characterized as collaborative teams of AI agents. Rather than scaling a single model to handle all tasks, the emerging paradigm involves multiple specialized agents with distinct roles and capabilities working in concert.

This approach multiplies infrastructure requirements. Where a single LLM might require tens of thousands of GPUs for training and inference, a team of coordinated agents demands compute resources for each specialist model plus the orchestration layer that enables them to communicate and collaborate effectively.

The economics are forcing a rethink of what constitutes adequate scale. Companies that once aimed for clusters in the tens of thousands of accelerators are now planning deployments an order of magnitude larger. Liang's projection suggests that by 2030, clusters below the million-GPU threshold will lack the capacity to compete in frontier AI development.

Competing Architectural Visions

Liang also revisited his seven-layer AI architecture framework, positioning it against the five-layer model previously outlined by Nvidia CEO Jensen Huang. While both frameworks attempt to map the stack from silicon through systems to applications, the divergence in layer count reflects different perspectives on where meaningful abstraction boundaries lie.

MediaTek's seven-layer approach breaks out components that Nvidia consolidates, potentially offering finer-grained optimization opportunities for chip designers and system integrators. For a company whose business model depends on differentiation in mobile and edge AI, a more granular architectural view provides additional surfaces for innovation and value capture.

The debate over architectural layers is more than academic. How the industry converges on a reference model will shape tooling, interoperability standards, and the division of value across the supply chain. A five-layer consensus favors vertically integrated players with strength across multiple levels; a seven-layer model creates more niches for specialists.

Space as the Next Frontier

The trajectory Liang described points toward infrastructure challenges that terrestrial data centers may struggle to solve. Power availability, cooling capacity, and physical space constraints already limit the growth of hyperscale facilities in established markets. Moving compute infrastructure to orbit offers theoretical advantages in power access via solar arrays and thermal management via radiative cooling, though the engineering and economic hurdles remain formidable.

Several startups and established aerospace firms are exploring space-based data centers, but none have deployed commercial-scale AI training infrastructure beyond Earth's atmosphere. If Liang's timeline holds, the industry has less than five years to prove the model at scale.

The Implications for Chipmakers

For semiconductor companies, the shift to million-GPU clusters and potential space deployment creates both opportunity and risk. Demand will remain robust, but the technical requirements for chips operating in space environments differ substantially from terrestrial parts. Radiation hardening, thermal cycling tolerance, and long-term reliability without physical maintenance become paramount.

MediaTek's public positioning on these trends suggests the company sees an opening to expand beyond its traditional mobile stronghold into data center AI, a market Nvidia currently dominates. Whether that ambition translates into meaningful market share will depend on execution in chip design, software ecosystem development, and the ability to convince hyperscalers to diversify their accelerator suppliers.

The million-GPU baseline Liang projects also implies a concentration of AI capability among a shrinking number of well-capitalized players. If mid-tier infrastructure costs run into the tens of billions of dollars, the number of organizations capable of competing in frontier AI development contracts sharply, with consequences for innovation diversity and competitive dynamics across the technology sector.