MediaTek Executive: By 2030, One Million H100s Will Be Considered Only Mid-Scale Computing Power
Nashnova编辑部
MediaTek senior director Bor-Sung Liang projects that by 2030, more than 10 data centers worldwide will exceed one-million-H100 scale, each drawing 3–4 gigawatts — AI's next bottleneck is shifting from chips to the physical limits of power and infrastructure.
How big are AI data centers getting?
Liang benchmarks against H100-equivalent compute: today's largest AI data center equals roughly 1 million H100 GPUs, consuming close to 1 gigawatt.
By 2030, that scale may rank only as "mid-tier" — he expects more than 10 facilities to exceed it, each drawing 3–4 gigawatts.
This means → within five years, today's ceiling becomes the starting line; compute demand is outpacing most people's intuition.
How did GPU demand explode step by step?
Before the Transformer architecture, training a model required fewer than 10 GPUs.
By the ChatGPT era, that number climbed to thousands, approaching ten thousand.
Today's hyperscale data centers run 400,000–500,000 GPUs — and are still expanding.
In plain terms = in under a decade, the GPU count per model grew more than 50,000-fold — the curve is nearly vertical.
Where does the power come from — and why is anyone talking about space data centers?
Liang cited SIA data: global compute-power consumption is growing faster than terrestrial energy systems can comfortably support.
That partly explains why Elon Musk floated the idea of space-based data centers — ground-level power may simply not be enough.
If the industry's end goal is 1 terawatt of compute capacity, that equals roughly twice the total U.S. electricity consumption and could require deploying millions of satellites.
This reflects a fundamental pivot: AI's next bottleneck is no longer the chip itself, but the physical availability of power and infrastructure.
Is chip design alone still enough — what does "systems engineering" mean here?
Liang stressed that building AI infrastructure at this scale can no longer rely on chip design alone.
It must span TSMC's advanced process nodes, advanced packaging, and system-level engineering covering scale-up (expanding interconnects within a single machine) and scale-out (networking many machines together).
Over $1 trillion has already been committed to AI infrastructure, but Liang says that only covers current demand — the future capacity gap remains enormous.
This means → Taiwan's semiconductor supply chain is set to grow even more strategically critical, because process and packaging capabilities are concentrated there.
How is AI architecture changing?
AI is shifting from relying on a single large language model (LLM) to collaborative teams of AI agents — programs with distinct roles and capabilities.
Liang proposed a seven-layer AI architecture. Compared to Jensen Huang's five-layer "cake," it adds a dedicated context layer and an agent layer, arguing that coordination and application are the real keys to unlocking productivity.
In plain terms = future AI is not one "generalist" but a squad of "specialists" working together, linked by a coordination layer.
Could humans become AI's bottleneck?
Liang outlined a software evolution path: prompt engineering → context engineering → workflow engineering — an "inverted pyramid" progression.
If humans keep verifying every AI output one by one, humans themselves become the bottleneck.
The future model: AI agents cross-verify each other, while humans supervise the overall process through metrics and anomaly monitoring.
This means → this shift could give rise to an entirely new discipline — AI management — where the team being managed is not human, but artificial.
Content is for reference only, not financial advice.