Jensen Huang used Nvidia Corp (NASDAQ:NVDA, XETRA:NVD)’s keynote at the GTC global AI conference on Monday to lay out a much larger demand picture for AI infrastructure — and a clear shift in where that demand is heading next.
The headline figure was a sharp upgrade to expected orders. Nvidia now sees as much as US$1 trillion in combined demand for its Blackwell and next-generation Vera Rubin platforms through to 2027, roughly double the US$500 billion estimate disclosed last October. The figure excludes several product configurations, suggesting the total pipeline could be larger again.
That scale underscores how quickly AI spending continues to build, even as markets have become more selective in how they respond to strong earnings and guidance from major technology names.
Inference takes centre stage
A key theme from the keynote was the growing weight of inference — running AI models in real-world applications — as usage expands beyond training large language models.
Huang pointed to a surge in token generation and the rise of “agentic” AI systems, where models can carry out multi-step tasks with limited human input. These applications demand persistent compute, lower latency and far greater efficiency than earlier training-focused workloads.
Nvidia’s roadmap is increasingly geared towards that shift. The upcoming Vera Rubin architecture, expected to begin shipping later this year, is designed to deliver around 10 times the performance per watt of current Grace Blackwell systems — a step change as operators look to scale deployments without a matching jump in energy costs.
The company also unveiled new rack-scale systems tailored for inference-heavy environments. One configuration, combining Vera Rubin with next-generation infrastructure, is targeting up to 35 times higher inference throughput per megawatt, alongside a lift in potential revenue generation for large models.
Power and efficiency move to the forefront
Energy is quickly becoming one of the defining constraints in the next phase of AI expansion. Data centre operators are already running into capacity limits in parts of the US and Europe, with access to power shaping how quickly new infrastructure can come online.
Nvidia’s emphasis on performance per watt reflects that reality. Efficiency gains of this scale can ease operating costs and allow higher utilisation of existing infrastructure, particularly as inference workloads run continuously rather than in bursts.
These pressures are flowing through the broader data centre ecosystem, influencing everything from chip design and system architecture to cooling technologies and power supply planning.
Expanding the software layer
Alongside hardware, Nvidia continues to push deeper into software, reinforcing its role across the full AI stack.
The company introduced new enterprise-ready frameworks and reference architectures designed to support agent-based applications, including tools built around emerging open-source ecosystems. The aim is to simplify deployment while keeping customers within Nvidia’s platform.
Bringing hardware, software and developer tools together has long been a strength for Nvidia. As enterprise adoption accelerates, that integrated approach is becoming more valuable — particularly for organisations looking to move from experimentation to production.
Market expectations and valuation context
Heading into GTC, expectations were already high following a strong earnings season. Nvidia’s valuation has eased from earlier peaks, with shares trading at a lower multiple than several semiconductor peers despite its dominant position in AI infrastructure.
That leaves less room for surprise. The keynote added to the visibility around Nvidia’s roadmap and demand outlook, but near-term share price moves are still likely to hinge on delivery and broader market sentiment.
Broader implications for the AI buildout
The direction set at GTC points to a longer, more layered phase of AI investment.
As workloads shift towards inference and real-time applications, infrastructure requirements are becoming more continuous, more energy-intensive and more tightly integrated across hardware and software. That evolution is expected to support ongoing demand across semiconductors, data centres and supporting technologies over the next several years.
Nvidia’s updated outlook suggests the buildout still has momentum — with the next phase defined less by how fast models can be trained, and more by how efficiently they can be deployed at scale.