Given the GPU gold rush compatible with AI, there is hardly a surprise that the NVIDIA GPU Technology (GTC) conference has become a magnet magnet with Wall Street, with serious note analysts after note on what to expect during the Nvidia highlight event.
JPMorgan’s prospect on NVIDIA GTC
Blackwell Ultra: a performance beast on the horizon
JPMorgan expects the Blackwell Ultra (GB300) to be the highest point of this year’s GTC conference. This chip will include a logical structure similar to the B200 but with significant improvements …
– Jukanlosreve (@jukanlosreve) March 14, 2025
Now JP Morgan is expecting what to expect during the NVIDIA GTC 2025.
Of course, the ultra GPU Blackwell (GB300) should be the proverbial star star. The Titan of Wall Street is expecting the new chip presents a logical structure similar to that found in the B200 chip, but with a high bandwidth (HBM) capacity of 288 GB and by taking advantage of the HBM3E stacking technology at 12 tall, thermal design power (TDP) up to 1.4 kW, and an increase in performance of 50% in the FP4 compromise. In addition, JP Morgan expects Nvidia to start sending the GB300 chips in the third quarter of 2025.
In addition, even if the Rubin architecture of Nvidia de Nvidia will enter the production phase in 2026, JP Morgan expects that the GPU giant reveals incremental details. As such, the Wall Street Titan expects the GPU Rubin to appear:
- A logical structure similar to Blackwell but equipped with 2 TSMC N3 process chips in what is essentially a double logical chip structure.
- Eight HBM4 batteries with a total capacity of 384 GB.
- A TDP ~ 1.8 kW
- Vera Arm CPU upgraded to the N3 process of TSMC and taking advantage of a 2.5D packaging structure.
- Network 1.6T comprising two NIC Connectx-9.
- Possible introduction of Rack NVL144 and NVL288 Rack structures.
- The initial production begins at the end of 2025 or at the beginning of 2026, the mass expeditions which should begin the T2 2026.
In addition, JP Morgan believes that Nvidia will reveal additional details on its roadmap for Co-Emballée Optics (CPO), which aims to increase bandwidth, reduce latency and reduce energy consumption. However, the CPO in the GPU continues to deal with important technical challenges, including heat dissipation problems (due to the generation of optical optical heat heat), reliability problems and the risk of substrate deformation. Consequently, the Titan of Wall Street only expects a broad adoption of the CPO in 2027, the Rubin 2026 iterations likely to see this technology incorporated as an optional characteristic in switches such as quantum (infiniband) and the spectrum (Ethernet).
In the meantime, like us note Recently, Nvidia asked for a new patent – bearing the publication number US20250078199A1 – March 06, 2025. The patent envisages discrete sections of a GPU working within the local limits to store and access the data, and carry out calculations, thus reducing the delays inherent in access to distant calculation resources. It is possible that Nvidia reveals new details compared to its efforts to locate the current subgpu in the GTC.