Carlo Ruiz, Vice President for Enterprise Solutions at NVIDIA, used his keynote interview at Datacloud Global Congress 2026 on June 3rd to advance a reframing of data centre value that has significant implications for how operators, investors, and customers measure and price infrastructure. The unit of value, Ruiz argued, is no longer the GPU or the megawatt-it is the token.

NVIDIA’s position in the data centre market needs no introduction. Its GPUs underpin the majority of large-scale AI training globally, and its roadmap-from Hopper to Blackwell to the GB300-has effectively set the pace of infrastructure investment across the sector.

What Ruiz brought to DCGC 2026 was not a product announcement but a framework shift: a way of thinking about what data centre infrastructure is actually for, and how its value should be measured. The argument is both technically precise and commercially provocative, and it arrives at a moment when operators, investors, and hyperscaler customers are all renegotiating the metrics by which data centre performance is assessed and priced.

The tokenomics argument

Ruiz’s central proposition is that the data centre industry has been measuring the wrong thing. Historically, infrastructure has been measured in compute units-chips, cores, GPUs-or in physical capacity-megawatts, square metres, rack units. Ruiz argued that in the AI era, the correct output metric is the token: the discrete unit of intelligence produced by a large language model or other generative AI system.

Measuring tokens per watt or tokens per dollar invested, he suggested, provides a more commercially meaningful basis for evaluating infrastructure performance than any hardware-centric metric. This is not merely a vocabulary change. It reorients the entire investment thesis for AI infrastructure away from asset acquisition and toward output optimisation-a shift with significant implications for how capacity is designed, procured, and priced.

The DSX reference architecture

To support this framework, Ruiz outlined NVIDIA’s DSX reference architecture, which the company developed to help organisations achieve 30% to 40% greater data centre performance within the same physical footprint. The architecture is designed to address a specific problem: as AI workloads scale, the naive approach of adding more hardware produces diminishing returns on performance per unit of space and power.

The DSX framework optimises the relationship between hardware, networking, and software to maximise token output per physical resource unit. Ruiz framed this as an invitation to operators to map their own value within a five-layer AI infrastructure model encompassing energy, specialised hardware, networking, models (including sovereign and enterprise fine-tuned models), and applications. The practical implication for operators is that differentiation in the AI infrastructure market will increasingly be won at the systems level, not the component level.

The open ecosystem argument

Ruiz also used the keynote to position NVIDIA’s partner ecosystem as a strategic asset. Rather than presenting NVIDIA as a vertically integrated provider seeking to capture the full value chain, he emphasised the company’s commitment to an open, partner-driven model that allows operators and system integrators to build differentiated offerings on top of NVIDIA infrastructure.

This is a commercially significant positioning choice. It signals that NVIDIA sees its role as defining the performance standard and measurement framework-tokenomics-while leaving the delivery and integration layer to partners. For data centre operators attending DCGC 2026, the message was that aligning with NVIDIA’s architecture is a prerequisite for competing in the AI infrastructure market, but that the opportunity to add value above the chip layer remains open.

Key Takeaways

• Ruiz’s tokenomics framework reorients AI infrastructure value from hardware metrics (GPUs, megawatts) to output metrics (tokens per watt, tokens per dollar), a shift with direct implications for how operators price and differentiate their services.
• The DSX reference architecture promises 30% to 40% performance improvement within the same physical footprint-a compelling proposition for operators facing land and power constraints.
• NVIDIA’s five-layer AI infrastructure model provides a map for where value can be added above the chip layer, including in sovereign models, enterprise fine-tuning, and application integration.
• The open ecosystem positioning signals that NVIDIA intends to set the performance standard while enabling a partner-driven delivery market-an approach that should reassure operators concerned about vendor lock-in.
• The shift from measuring data centres as facilities housing hardware to measuring them as factories producing intelligence is not a marketing metaphor-it is a structural change in how AI infrastructure investment will be evaluated.

The tokenomics framework will not immediately restructure how data centre contracts are written or how capacity is priced. But it establishes a direction of travel that sophisticated operators and investors should take seriously.

As AI inference scales and the question of “how much intelligence does this facility produce per dollar invested” becomes a standard due diligence question, the operators with the clearest answer will have a meaningful commercial advantage.