NVDA: Vera Rubin is closing in on prime time
Vera Rubin, the AI world’s first inference-optimized platform, is closing in on prime time. The platform has been discussed endlessly on the Street. Investors should not allow a false sense of familiarity with the platform to miss a potential return of NVDA to alpha generation once Vera Rubin (VR) starts shipping in volume.
We called out delay in VR two quarters ago, before it was noted on the Street (link). Ahead of Computex this year, we turned modestly positive as we felt VR delay was priced in (link). VR has had a long gestation period as NVDA and its partners worked to iron out dauting challenges. It is still a work in progress. Going into the event today, we think NVDA management may be close to taking the covers off this secretive project. We think Nvidia has largely ironed out the challenges arising from its choices in the physical design. Changes to the software stack and the operating system are getting close to completion, we believe.
The earnings call today is less about the typical +$2bn/+$3bn upside to print/guide expectations. We are not looking for management to raise the $1trillion outlook for Blackwell + Rubin during 2025-2027. The call today is about management giving a firm timeline for meaningful revenue recognition from VR. Also, investors would be happy to hear of a status update to the $20billion revenue expectation of Vera CPU racks. Our checks show CRWV is scheduled to start taking delivery of Vera racks in the next few months.
Has VR already started shipping? It has, but only as test systems. CRWV took delivery of a unit a few months ago. MSFT has had a few test units to help NVDA co-develop the software stack. We believe that at least one other hyperscaler, Google, too may have taken delivery of a Vera Rubin test unit. We are gaining confidence that Google is likely to deploy NVDA’s Vera Rubin platform ahead of its internally developed inference-optimized TPU v8i. That would be an important endorsement of NVDA’s VR platform.
We like the stock at current level. We maintain our modest $250 PT based on 19x to consensus 2027 estimates and will be looking to raise the target if the VR narrative appears on firm ground at the earnings call. We think investors have little expectation of NVDA returning to the alpha generator it once was. Expectations therefore are modest. If the CEO could explain VR’s journey so far and lay out a narrative for 2027, we think investors could be nudged into giving the stock another look.
–xx–
.
Believability: Street consensus models NVDA datacenter revenue in 2025-2027 slightly ahead of management’s $1trillion outlook. The question is whether the Street believes in the estimate, for if it did the stock would not be trading at sub-20x to 2027 EPS. We think Google is in no particular hurry to install its internally developed TPU v8i, no sooner than 4Q27. NVDA may have the inference-optimized AI market at all hyperscalers, including GOOG, all to itself for most of next year. If so NVDA’s $1trillion bogey sounds more believable.
Nvidia has ironed out the challenges arising from the physical design. We were among the first to point out heating issues in VR related to HBM and interposer. We think those concerns have been addressed to within manageable levels via re-definition of memory content, not just at the HBM but also at the main memory. What remains to be completed are the final pieces of the software stack. For instance, Dynamo, Nvidia’s inference-serving framework, is still being worked on.
Resolving issues with the physical hardware has been no small task. The platform had been beset with heating-related issues. Heating issues had dogged both the HBM as well as the lpDDR5x main memory. In our view, heating issues get worse with higher DRAM content. HBM and lpDDR5X have different paths to heat-related challenges, the details of which we will not get into in this note. We think the solution has been to manage the DRAM content in the HBM and the main memory vs the original design.
We are aware noisy headlines have caused investor confusion. In our view, NVDA merely came up with an engineering compromise by optimizing hardware yield, workload performance and in-field reliability. It is our view that the reduction in DRAM content per board vs the original intent has little to do with DRAM cost or DRAM supply. Content management is just part of a design optimization process.
One thing we are reasonably certain of – HBM stacks 12-hi and above are unfeasible due to heat-induced high bit error rate. We think it took NVDA some time to come to this conclusion, by trial and error. This was part of the VR gestation period. Vera Rubin is NVDA’s first step in reducing dependency on scaling up HBM stack height.
Investors in memory stocks need not panic. Reduced memory content per board simply means more boards are needed to load up an instance of a given LLM model. The total DRAM bits needed to load up a given model remains largely unchanged.
The Vera Rubin platform makes a clean break from the traditional GPU-based designs of the past. VR breaks free off the constraints that came with the original Microsoft-created game-centric nature of GPUs. The main memory in VR no longer runs on the traditional PCIe bus. The main memory is no longer isolated from the GPU memory (aka the HBM) as it is on traditional Windows machines. One the most important changes in VR is the creation of a unified memory where the main memory and the graphics memory (i.e. HBM) are on the same NVLink network.
This offers the VR platform the enormous advantage of being able to distribute the model parameters across the HBM and the main memory, a feature not available on Blackwell or its predecessors. This allows VR and its follow-ons the advantage of being able to tolerate smaller HBM memory vs. Blackwell, for the simple reason the model parameters can be offloaded to the main memory. Smaller HBM memory allows less heat-related issues.
The main memory can be as large as 1.5TB, which is ~5x the HBM density in Blackwell or Rubin. As the model parameters can be stored across the main memory and the HBM, the system architect has the design choice of being able to reduce the size of the HBM and offload the model to the much larger main memory, though at some cost to latency.
Such design choices were not available to the system architects of Blackwell and Hopper where the model parameters can be loaded only on the HBM.
Is NVDA suffering from DRAM supply issues? It is not, in our view. We think NVDA may be the only large AI player who anticipated its DRAM demand way ahead to time and made supply arrangements with SK Hynix. This allows NVDA to adjust the DRAM content on HBM and lpDDR5x modules without being at the mercy of the ever-changing spot market. With the final configuration of the main memory and HBM now settled, far from facing shortages, our checks show NVDA has adequate DRAM supply at SK Hynix which it has allocated to Vera CPU racks and to DGX client devices.
Vera CPU stand-alone racks to ship soon: Due to ready supply of DRAM bits NVDA enjoys, we believe Vera CPU racks have an advantage over competitors in the ARM and x86 space. We will not be surprised if the $20bn outlook for Vera CPU racks is raised. We believe CRWV to be one of the early recipients of Vera CPU racks.
Net/net: Vera Rubin has had a long gestation period as NVDA ironed out dauting challenges with its new architecture. Going into the event today, we think NVDA management may be close to taking the covers off this secretive project. Meanwhile, investor expectations are muted; most investors believe the stock could continue to underperform hyperscaler peers and the broad market. We like the stock at current level as Vera Rubin gets close to primetime. Not just the traditional partners of NVDA, such as MSFT and CRWV, our checks show even the likes of GOOG seem ready to take delivery of VR. We maintain our modest $250 PT based on 19x to consensus 2027 estimates and will be looking to raise the target if the Vera Rubin narrative appears on firm ground at the earnings call.
KC Rajkumar
Disclosures and Disclaimers
Lynx Equity Strategies, LLC is an independent equity research provider. The Company is not a registered broker dealer or investment adviser. No employee or member of the Company, or immediate family member thereof, exercises investment discretion over securities of any issuer analyzed in this report two days prior and/or two days after this report is issued. It participates in “Alpha Capture Systems” that seeks investment or trading ideas from the sell-side and may pay for the participation based on relative performance.
Limitations of Information
This report has been prepared for distribution to only qualified institutional or professional clients of Lynx Equity Strategies, LLC (the “Company”). The contents of this report represent the views, opinions, and analyses of its authors. The information contained herein does not constitute financial, legal, tax or any other advice. All third-party data presented herein were obtained from publicly available sources which are believed to be reliable; however, the Company makes no warranty, express or implied, concerning the accuracy or completeness of such information. In no event shall the Company be responsible or liable for the correctness of, or update to, any such material or for any damage or lost opportunities resulting from use of this data. Nothing contained in this report or any distribution by the Company should be construed as any offer to sell, or any solicitation of an offer to buy, any security or investment. Any material received should not be construed as individualized investment advice. Investment decisions should be made as part of an overall portfolio strategy and you should consult with a professional financial advisor, legal and tax advisor prior to making any investment decision. Lynx Equity Strategies, LLC shall not be liable for any direct or indirect, incidental or consequential loss or damage (including loss of profits, revenue or goodwill) arising from any investment decisions based on information obtained from Lynx Equity Strategies, LLC. Reproduction and Distribution Strictly Prohibited. No user of this report may reproduce, copy, distribute, sell, resell, transmit, transfer, license, assign or publish the report itself or any information contained therein. This report is not intended to be available or distributed for any purpose that would be deemed unlawful or otherwise prohibited by any local, state, national or international laws or regulations or would otherwise subject the Company to registration or regulation of any kind within such jurisdiction.
Copyright, Trademarks, Intellectual Property
Unless otherwise indicated, all copyrights, trademarks, service marks, logos and other intellectual property included in this report are proprietary materials of Lynx Equity Strategies, LLC and the unauthorized use of such terms, marks, and logos is strictly prohibited. The Company reserves all rights with respect to the intellectual property ownership of all materials in this report and will enforce such rights to the full extent permissible by law.