By Roger Cummings, CEO of PEAK:AIO
GTC 2026 was, by any measure, a exceptional occasion. Jensen Huang’s announcement of $1 trillion in projected orders by way of 2027, double final 12 months’s $500 billion projection, set a brand new benchmark for AI infrastructure ambition. The Vera Rubin structure, the Groq LPU integration, and the gigawatt-scale AI manufacturing facility imaginative and prescient – all of it factors to speedy growth of the market.
As spectacular because the GTC keynote was, it solely captured a small portion of what we’re seeing.
The AI manufacturing facility narrative NVIDIA offered at GTC is correct for hyperscalers. It displays how the biggest cloud suppliers and know-how firms are fascinated with infrastructure at excessive scale. Nevertheless, it doesn’t describe the vast majority of organizations constructing and deploying AI infrastructure techniques as we speak.
87% of PNY’s prospects – PNY being considered one of NVIDIA’s main distributors – run fewer than ten DGX techniques. Essentially the most impactful medical AI applications within the UK are operating on six DGX techniques. Conservation AI at a worldwide scale is operating on two GPU servers.
This isn’t the perimeter of the market. It’s the mainstream.
This sample is constant throughout earlier infrastructure waves. The headline numbers have a tendency to explain the highest finish, the place scale and capital expenditure are highest. The broader market usually develops within the center – organizations with critical necessities and budgets, however no urge for food for hyperscale complexity. That’s the place a good portion of long-term adoption takes place.
One of many extra notable elements of this 12 months’s keynote was Jensen explicitly naming storage as one of many 5 pillars of the AI manufacturing facility, alongside compute, reminiscence, networking, and safety. That framing displays a rising recognition of storage as a first-order concern in AI system design.
Nevertheless, the dialogue largely stopped at identification. The sensible query – what purpose-built AI storage appears to be like like for organizations working outdoors hyperscaler environments – didn’t come up within the keynote, regardless of being a key subject in each infrastructure dialog.
In lots of deployments, GPU utilization falls in need of {hardware} capability, not as a result of the GPUs are fallacious, however as a result of the storage techniques feeding them weren’t designed for AI workload profiles. For organizations operating 10, 15, or 20 GPUs, this could change into a persistent bottleneck. It’s not often seen on a specification sheet however exhibits up day by day in efficiency that falls in need of what was promised.
These challenges are usually not new, and in lots of circumstances, they’re or have already been solved. The difficulty is much less concerning the existence of options and extra about their adoption throughout the broader market.
Ongoing reminiscence constraints
One other vital assertion from GTC got here from the sidelines, moderately than the keynote stage itself. SK Group Chairman Chey Tae-won, whose firm SK Hynix is NVIDIA’s main HBM provider, mentioned that the industry-wide reminiscence provide shortfall will persist at over 20% by way of 2030, that means 4 to 5 years of elevated costs and constrained provide.
For a lot of organizations, this modifications the infrastructure equation solely. When {hardware} refresh cycles change into considerably dearer and provide is constrained, the crucial shifts towards extracting extra efficiency and effectivity from current infrastructure. On this setting, software-defined storage that delivers AI-grade efficiency from commodity {hardware} isn’t a workaround. It’s the suitable architectural reply.
What this implies for the broader market
Nevertheless, the story that issues extra to the vast majority of enterprise IT leaders, analysis establishments, and domain-specific AI groups is the one GTC quietly confirmed by way of its session catalogue and present ground: AI infrastructure at a smaller scale is maturing quickly. DGX Spark was on sale on the present; NemoClaw runs on a laptop computer.
The potential is transferring down the stack. Techniques have gotten extra accessible, extra modular, and simpler to deploy outdoors of hyperscale environments. Edge and near-edge use circumstances are clear examples, as constraints on energy, house, and latency require a special strategy to infrastructure design.
The fact is that the AI infrastructure market just isn’t outlined by the biggest deployments – a degree GTC 2026 each highlighted and, at instances, neglected. Whereas Jensen Huang’s keynote targeted on hyperscale techniques, GTC as a complete mirrored a a lot wider vary of real-world adoption.
This being mentioned, a very powerful developments in many of the market won’t be the biggest techniques described on stage. As a substitute, they’re the continued progress in making AI infrastructure usable, environment friendly, and efficient throughout a broader vary of real-world environments.
Concerning the writer
Roger Cummings is the CEO of PEAK:AIO, an organization on the forefront of enabling enterprise organizations to scale, govern, and safe their AI and HPC purposes. Below Roger’s management, PEAK:AIO has elevated its traction and market presence in delivering cutting-edge software-defined information options that remodel commodity {hardware} into high-performance storage techniques for AI and HPC workloads.
Associated
Article Subjects
AI infrastructure | GPUs | semiconductors
