Wolfe Research Reiterates NVIDIA at Outperform, $275.00 Price Target
Wolfe Research reiterated its Outperform rating and $275.00 price target on NVIDIA following NVDA’s financial Q&A at GTC Taipei.
NVDA held a Q&A session with financial analysts at GTC Taipei at Computex. Our key takeaways follow.
Buybacks. NVDA expects to return 50% or more of its FCF going forward – building on its prior statements that it will return 50% of FCF this year, and institute a $0.25/shr dividend.
We expect NVDA to generate FCF of $216bn / $274bn in FY27 / FY28, respectively, driving cash returns of $245bn through FY28.
We expect ~$40-45bn of that to be distributed as dividends, leaving ~$200bn in share buybacks over the next 2 years.
We believe this is a modest tailwind to EPS over the next 2 years, driving a ~$0.25-$0.30 (+2.0%-2.5%) benefit to our ~$12.19 FY28 EPS.
Server CPU market sizing inline with our expectations. NVDA’s Vera CPU is purpose-built for high single-threaded performance – each CPU is needed for its own agentic workload.
Each Vera is a monolithic CPU, which contains 96 Olympus Arm-based cores, and can run 10 instructions per cycle (IPC).
Jensen indicated that the use of CPUs for workload orchestration, storage servers, data processing, etc. could effectively double the CPU overall number of CPUs in the datacenter.
Given head node CPUs typically attach with GPUs at a 1 : 2 ratio, this suggests that NVDA expects the number of CPUs to rise to about 1:1 – consistent with its earlier NVDA Pod commentary (link) as well as our prior CPU analysis (link).
Notably, Vera is a fully custom NVDA core (based on the ARM instruction set), whereas Grace is a customized ARM Neoverse core – so Vera is more substantially differentiated than Grace (which shares its Neoverse roots with GOOGL Axion and AMZN).
In addition, NVDA noted their decision to build Vera as a monolithic die to avoid the “chiplet tax,” which trades off performance/latency for yield (and therefore cost).
The absence of the chiplet tax is one of the features that drives single threaded performance (which NVDA feels is more important than multi threaded performance in agentic AI) – but it will also drive higher cost.
Architectural details of the RTX Spark. The RTX Spark Superchip has a Blackwell RTX GPU connected to a 20-core CPU based on Arm cores, and 128 GB LPDDR5X memory.
With 1 PF of FP4 performance, NVDA expects agents can run locally and support a wide variety of tasks.
We have modest expectations for the RTX Spark; NVDA is trying to “redefine the category.” The company believes its AI PC will “redefine the category,” supporting higher long-term pricing for consumer PCs and expanding the overall TAM.
While the ASP for the RTX Spark is as yet unpublished, the DGX Spark (announced at March GTX in 2025) sold for ~$5k – given the specifications for the RTX Spark, we expect a similar ASP for the new PC product.
We expect the addressable market for the RTX Spark to be limited by the high ASP, and in any case, we expect the size of NVDA’s PC business relative to its datacenter business to remain limited.
This research note is reproduced verbatim from the issuing firm. Price Target never edits, paraphrases or alters analysts’ words — we only republish them in one place.
