GTC 2026: Moving from Chip Launches to Infrastructure Architecture
NVIDIA’s GTC 2026 emphasized software and infrastructure integration for “AI factories,” with DSX platforms, heterogeneous rack designs, 800 VDC power discussions, and grid-connected power orchestration becoming central themes for enterprise DCPI planning.
Platform push for designing and operating AI factories
The report highlights NVIDIA DSX as a full-stack approach to designing, building, and operating Artificial Intelligence (AI) factories, and notes it counted “over 200 partners” in its ecosystem. In the same framing, NVIDIA CEO Jensen Huang said the conference covers “every single layer of the five-layer cake” from facility infrastructure to applications.
Among DSX elements, the Omniverse DSX blueprint is described as a generally available platform for modeling data center layouts, power topologies, and thermal behavior with simulation-ready 3D models in OpenUSD format. The note argues that many data center designs still rely on traditional Cohort Analysis Dashboard (CAD) and Boot Integrity Measurement (BIM) tools and that digital twin adoption remains early, making the Omniverse Blueprint a forward-oriented approach.
Rack designs shift away from a single workhorse
The report states that GTC 2026 moved beyond planning around one rack architecture by introducing multiple rack configurations under the Vera Rubin umbrella. It describes the NVL72 as a liquid-cooled, fanless, cableless enclosure exceeding 200 kW per rack, and it outlines additional options including a CPX rack, a Vera CPU-only rack, and an LPX rack with third-party silicon in NVIDIA’s reference design.
It connects the broader implication to operational planning, citing “heterogeneous clusters” as a driver for managing mixed rack densities, uneven heat loads, and different liquid cooling needs within the same row. The note also frames the change as aligned with workload diversification beyond large-scale training.
800 VDC and liquid cooling move to the vendor foreground
While the report says 800 VDC received limited attention in NVIDIA’s official channels, it describes vendor engagement as extensive on the show floor. It adds that vendors including Delta Electronics, Texas Instruments, and STMicroelectronics focused announcements on 800 VDC developments.
The note also points to open architecture details, stating that Eaton’s J.P. Buzzell referenced “an OCP white paper expected in the coming weeks” during a session titled “A Safe, Efficient, and Scalable Approach to 800 VDC Architecture.” It further says supply chains and safety guideline codification remain prerequisites before broad deployment.
On thermal delivery, the report says liquid cooling discussions centered on scaling Coolant Distribution Unit (CDU) capacity with “a new class of multi-megawatt CDUs.” It also describes a direct-current CDU capable of connecting to an 800 VDC bus as an exhibition finding, and it links the trend to tighter coupling of power and cooling design decisions.
Standardization pressure under the MGX ecosystem
The note describes NVIDIA’s reference architectures as increasingly specific, raising the question of how vendors differentiate as NVIDIA defines more of the design. It describes the “MGX wall” as showing components from multiple vendors within a standardized MGX ecosystem.
It characterizes the MGX approach as standardizing interfaces, form factors, and performance specifications across the infrastructure stack, which the report says supports mixing and matching components. The report also states that such standardization compresses vendor space for differentiation toward areas beyond physical product characteristics.
Software for grid access and dynamic power adjustment
The report identifies DSX Flex as a launch tied to connecting AI factories to grid services and orchestrating dynamic power adjustment. It frames the problem as a gap between power needed for forecast chip shipments and the pace of grid updates.
It reports that in a direct exchange during the event, Jensen Huang said data centers must change their relationship with the grid and “be willing to accept less stringent Service Level Agreements (SLAs) in exchange for faster access to capacity.” The note further states that the software orchestration and alignment of utilities, grid operators, and regulators factor into scaling access to power.
DCPI market update cited during GTC week
The report cites Dell’Oro Group’s DCPI market update stating the market reached $10.9 billion in 4Q 2025, up 20% year-over-year, and it attributes the movement to synchronized backlog surges across vendors in power and cooling. It also references the update as released during GTC week.
It then describes GTC 2026 as reinforcing expectations around record investment driven by AI deployments, while also noting that the overall tone involved confidence in the trajectory of compute buildout. The report also includes a set of vendor press release items, but does not analyze individual announcements beyond the DSX, rack, power, and cooling themes.
This Analyst Signals brief reflects a neutral, fact-based summary of the original research note.