From Data Centres to AI Factories: What NVIDIA GTC 2026...

Walking into NVIDIA's GTC 2026 in San Jose felt like watching an industry reckon with what comes next. More than 30,000 developers, engineers, and infrastructure leaders filled the SAP Center, and the energy reflected something much bigger than a product launch.

I attended on behalf of my consulting practice and as an advisor to IES to understand the infrastructure implications of NVIDIA's roadmap. What emerged has fundamentally reshaped how I think about the near-term challenges facing data centre owners, operators, and design engineers.

**The Year of Inference is Here**

Jensen Huang framed the moment decisively. His GTC keynote declared 2026 the "year of inference"—a tipping point where AI workloads shift from training to real-time token generation. "AI need to think now," he said. "Thinking requires inference, and inference requires generating a large number of tokens."

The scale behind this is staggering. NVIDIA now projects at least $1 trillion in cumulative demand for Blackwell and Vera Rubin systems through 2027—double last year's forecast. Computing demand has jumped 10,000 times in two years alone. The direction of travel is unmistakable.

The core metric for this era is tokens per second per megawatt—a system-level measure, not a chip metric. That distinction matters enormously for anyone designing or operating the facilities housing these systems. GPU, CPU, LPU, networking, storage, memory, liquid cooling, air cooling, and power delivery architecture from substation to chip must all work in coordination.

**The AI Factory: A Fundamental Architectural Shift**

The "AI factory" is the most important concept to emerge from GTC. These are not server rooms or traditional HPC facilities. They are tightly coupled, co-designed industrial systems engineered for one purpose: producing tokens at scale.

The Vera Rubin platform sits at the heart of this vision—a full-stack architecture of seven chips and five rack-scale systems designed to operate together as a single AI supercomputer. The NVL72 configuration pairs 72 Rubin GPUs with 36 Vera CPUs via next-generation NVLink interconnects, achieving up to 35x improvement in token throughput compared to Hopper-generation systems at equivalent power.

Alongside this is the Groq 3 LPX rack, born from NVIDIA's multi-billion-dollar licensing of Groq's Language Processing Unit technology. Purpose-built for ultra-low latency inference, the LPX rack holds 256 LPUs and handles the serving side of token generation while Rubin processes heavy workloads. Together, they can theoretically scale token output for a 1-trillion-parameter model by an extraordinary margin over previous systems.

Also unveiled was Kyber, NVIDIA's next major rack architecture featuring 144 GPUs in vertically oriented compute trays for greater density and lower latency. Kyber forms the basis for Vera Rubin Ultra, expected in 2027, with the Feynman generation arriving in 2028. Step-change hardware evolution is occurring every 12 to 18 months and accelerating. Each leap resets design assumptions for power density, cooling strategy, and physical architecture.

**The Infrastructure Consequences No One Has Fully Solved**

Liquid cooling is no longer optional. Vera Rubin and Kyber are designed for predominantly liquid cooling. Highest-density platforms cannot be adequately served by air. Yet liquid cooling does not eliminate air-side loads entirely. A 200–300kW cabinet can still generate meaningful residual air heat load—a design reality often underappreciated. While OEMs define acceptable coolant temperature ranges, no accepted facility-level standard yet exists for specifying and operating cooling infrastructure.

800VDC is becoming a reality in the white space. Legacy data centres run 208VAC at the rack level, sufficient for CPU-era loads but impractical for 150–300kW racks. 415VAC has emerged over 12–18 months as a response to rising densities, but it too reaches practical limits at AI factory densities. 800VDC and OCP's +/−400VDC—with higher voltage, fewer conversion steps, and lower distribution losses—is emerging as the standard for megawatt-scale compute blocks. Yet no established standard exists for connectors, protection schemes, rack-level interfaces, or safety codes. Vendors are creating noise around "the right answer" while operators face major capital commitments in the absence of clarity.

No stable hardware target exists beyond 18 months. This may be the most consequential challenge. Owners of AI compute want to deploy capacity in enormous blocks at unprecedented speed. But the equipment is novel to them and to the data centre host. OEMs each have their own implementations of the NVIDIA reference architecture, with distinct thermal signatures and power requirements. No single standard defines how heat is rejected or power distributed at facility level. By the time a facility is designed, built, and commissioned, the hardware target may shift again.

**Simulating Reality Before Committing Capital**

The question facing operators has shifted. It is no longer simply "will this design meet spec?" but rather "will this system perform to expected levels as an integrated whole, not as aggregated components?" That requires scenario testing before capital commitment, validating vendor projections in a modelled environment, and understanding power, cooling, compute, and networking behaviour as an integrated system.

NVIDIA signalled this importance at GTC by announcing the Vera Rubin DSX AI Factory Reference Design and the Omniverse DSX Blueprint—a physically accurate digital twin for planning, building, and operating large-scale AI factories. The DSX Air platform extends this further with full logical simulation of NVIDIA hardware before installation, allowing customers to reduce time to first token from weeks to hours.

Industry partners—Switch, Trane Technologies, Vertiv, and Eaton (which recently acquired Boyd Cooling and Flexnode)—are all contributing to this reference ecosystem, reflecting broad recognition that pre-deployment simulation is now a competitive necessity, not a luxury.

This is precisely where IES brings distinctive capability. While platforms like Omniverse DSX focus on network topology and ROI, IES's tools operate at the thermal, HVAC, and mechanical level—providing the facility-level physics beneath the IT layer. For operators determining whether existing buildings can handle next-generation loads, whether cooling strategies will perform under real conditions, or how mixed OEM environments behave across multi-tenant spaces, that simulation fidelity is critical.

The feedback from hyperscalers at GTC—teams managing some of the world's largest deployments—was consistent: they are making infrastructure decisions under supply chain pressure, accepting whatever transformers and cooling equipment they can source, and analysing performance after.

[Note: The source text appears to be truncated at the final sentence.]