No, not that kind of bar. I’m referring to the hurdles we attempt to leap over. The walls we try to run through. There are dozens of updates to digest from GTC ‘26 and they go beyond just NVIDIA and GPUs, to include every facet of IT infrastructure. The impacts are captured in Jensen’s 5 Layer Cake: Energy, Chips, Infrastructure, Models and Applications.
On the surface, many attendees felt GTC ’26 was a let down when compared to last year’s conference. There was sentiment that announcing VeraRubin at CES rather than waiting for GTC was perhaps the reason for the deflated reactions in San Jose. But, if we take a step back and remind ourselves that each decision across every pillar of IT is pivoting on the extreme co-design concept, then we begin to understand that at any given moment one IT pillar can appear to have priority or seemingly advancing faster than the others. And there’s no way to ignore the seismic financial announcement during the keynote when Jensen stated that NVIDIA has $1T in Revenue/Orders already lined up for their 2027 Fiscal year (Feb 2026 - Jan 2027).
For the network engineers, some of the most impactful announcements were around DOCA 4.0 and Bluefield4 (BF4). BF4 takes a more prominent role in the overall networking configurations of the VeraRubin NVL72. With VeraRubin and DOCA 4.0 the BF4 moves into a pivotal role for configuration of networking settings not only relative to the BF4s but also the ConnectX-9s. The BF4 aims to consolidate North-South (FrontEnd) and East-West (BackEnd) management under one configuration point. Additional emphasis on support of multi-tenancy with Astra, Weave, EVPN VXLAN VTEPs and IPSec capabilities.
And now Storage too. Also included in DOCA 4.0 is a Storage Controller framework with STX and CMX (Context Memory Storage). CMX addresses Inference use cases where the KV Cache needs have expanded exponentially as conversations lengthening and armies of agents leverage tools and accomplish tasks on our behalf. This is the first time we’re seeing a reference to storage placement in the East-West network rather than solely living in the North-South portion. And of course all of the security implications that come with that storage including encryption and data integrity.
The highly anticipated updates regarding the assets gained from the $20BN purchase of Groq, licensing and their LPU chip were expected and yet also surprising. The initial emphasis seems to be the inclusion of the LPU, LP30, as an adjacent part of the VeraRubin NVL72 rather than its own dedicated and separate Inference node. Similar to the GB and VR GPU/CPU combos, the LPUs are installed on dedicated compute trays that go into the LPX racks. But the desire to repatriate Inference back to on-premises builds will require a focus of driving down overall power consumption as the GB NVL72 expects ~130kW and the VR NVL72 pushes that up to ~285kW with the 256 LPUs from Groq possibly needing up to 160kW.
In other news, OpenClaw rose like Icarus, flying directly into the NVIDIA sun and featured predominantly on stage during Jensen’s keynote as well as all over the conference grounds. NVIDIA even offered OpenClaw building stations to allow attendees a chance to test drive some of the latest advancements in AgenticAI.
There’s plenty of published blogs and debriefs after the event so I’ll leave you to read those interpretations as you deep dive into DSX, CUDA and other elements of the 5 Layer Cake. With each major event throughout the calendar year we should probably be realistic about where the bar is set and calibrate our expectations of as well as our ability to digest these updates. Perhaps not setting the bar so high for each element of these complex, building-sized and campus wide factories of tokens. So while GTC ‘26 might not have been perceived as the same tsunami as past GTC events, there’s still plenty to unpack across all IT silos and even more on the horizon. No one expects this to slow down. The next couple of generations of advancements are road-mapped well beyond 2030. We all have our eyes focused down the road to Hot Chips, Hot Interconnects and of course, the Supercomputing Conference (SC ’26) in November, hosted in Chicago this year. Hope to see some of you there. And if you’re not at SC ’26 then you better be at NAF’s AutoCon6.