Just back from NVIDIA GTC 2026: The era of the single-chip AI data centre is over

Just back from NVIDIA GTC 2026 in San Jose — and one thing is crystal clear: the era of the single-chip AI data centre is over.

Jensen Huang's keynote laid out what we've been saying for years: the AI factory of the future isn't built on GPUs alone. It's a heterogeneous stack — CPUs, GPUs, LPUs, and specialised ASICs working together. NVIDIA announced seven new chips in the Vera Rubin platform. Seven. That's not a product line. It's an admission that one chip can't do it all.

This wasn't always the narrative. For years, the industry chased bigger GPUs, more HBM memory, more interconnects. The assumption was simple: more compute solves everything. But GTC 2026 marked a distinct pivot. Jensen himself framed the future around the Pareto curve of AI workloads — where GPUs serve some tasks, but ASICs are purpose-built for others. The age of the one-size-fits-all AI chip is dead. The market has spoken.

Two themes dominated every conversation:

1. Sustainable Performance

First, performance. Tokens per second matters — but now it's about sustainable performance. The new metric isn't just speed. It's tokens per watt. The industry is waking up to what we've been shouting about for over a year now: raw performance means nothing if your electricity bill makes the model economics unsustainable. At SCX.ai, we've been measuring Intelligence per Watt as our No 1 goal since foundation. It's validating to see NVIDIA adopt the same framing.

2. The Energy Bottleneck

Second, energy. Power is the bottleneck. Every major player I spoke with is grappling with the same problem: how do you scale AI infrastructure without melting the grid? Data centres are hitting power limits across the globe. NVIDIA's answer is liquid cooling, more efficient architectures, and specialised chips that do more with less. Their Vera Rubin platform targets 10x inference efficiency gains over Blackwell. But here's the catch — liquid cooling isn't optional for everyone. Not every data centre can — or should — be redesigned from the ground up.

The SCX.ai Approach

This is where SCX.ai comes in.

While the industry debates how to cool bigger GPUs, we're taking a different path: ASICs purpose-built for inference. Not general-purpose chips that try to do everything — but specialised silicon designed for one job: delivering the highest performance per watt without water cooling.

Our latest benchmarks speak for themselves: less power consumption than equivalent GPU systems, up to 10 times the tokens per second per watt, and we achieve all of this with standard air cooling. No exotic infrastructure. No reconfiguring your data centre. Just higher throughput, lower total cost of ownership, and a smaller environmental footprint.

Why ASICs for Inference?

Because inference workloads are predictable, repetitive, and fundamentally different from training. You don't need the flexibility of a GPU when your workload is well-defined. You need a chip optimised for exactly what you're running — nothing more, nothing less. That's what an ASIC delivers.

And here's the market reality no one can ignore: inference is where the money is. Jensen called it the "inference inflection point" at GTC. The bulk of AI spend in 2026 and beyond will be inference, not training. Every company building AI infrastructure needs to ask themselves a simple question: am I optimising for training, or for the 90% of my spend that will be inference?

GTC 2026 confirmed what the market is starting to realise:

The future of AI infrastructure is specialised, energy-efficient, and heterogeneous. The era of the GPU-monoculture is over. The multi-chip AI factory is here.

We've been building for that future — longer than most.

Scaling Australian Innovation

But honestly? The people made the week.

Two days after the keynote, we co-hosted "SYD > SJC — Scaling Australian Innovation at GTC" with Investment NSW and the Australian American Chamber of Commerce at Hapa's Brewing Company in San Jose. More than 150 people showed up — founders, investors, government reps, all hungry to talk about sovereign AI. Great conversations, cold beers, and a whole lot of optimism about what Australia is building.

And throughout the show #Australia shows up as a leader in AI - the whole community is definitely showing what Australia can create and do.