Nvidia CEO Reveals GPU and Software Moat in AI Chips

Nvidia is banking on its software expertise and broad GPU ecosystem to stay ahead in the fiercely competitive AI chip market, CEO Jensen Huang said in an interview with Goldman Sachs Wednesday. Huang pointed to NVIDIA's large base of installed GPUs and their software compatibility as key strengths.

Huang highlighted three key elements of Nvidia's competitive moat: a large installed base of GPUs across multiple platforms, the ability to enhance hardware with software like domain-specific libraries, and expertise in building rack-level systems. The CEO said Nvidia's chip design prowess, noting the company has developed seven different chips for its upcoming Blackwell platform.

These comments come as Nvidia faces increasing competition from rivals. Addressing supply chain concerns, Huang said NVIDIA has sufficient in-house intellectual property to shift manufacturing if necessary without significant disruption. The company plans to begin shipping Blackwell-based products in the fourth quarter of fiscal 2025, with volume production ramping up in fiscal 2026, according to Huang.

From the note that Goldman Sachs sent to its clients: 1) Accelerated Computing: Mr. Huang highlighted his long-held view that Moore's Law was no longer delivering the rate of innovation it had in the past and, as such, was driving computation inflation in Data Centers. Further, he noted that the densification and acceleration of the $1 trillion data center infrastructure installed base alone would drive growth over the next 10 years, as it would deliver material performance improvement and/or cost savings.

2) Customer ROI: Mr. Huang noted that we have hit the end of transistor scaling that enabled better utilization rates and cost reductions in the previous virtualization and cloud computing cycles. He explained that, while using a GPU to augment a CPU will drive an increase in cost in absolute terms (~2x) in the case of Spark (distributed processing system and analytics engine for big data), the net cost benefit could be as large as ~10x for an application like Spark given the speed up of ~20x. From a revenue generation perspective, Mr. Huang shared that hyperscale customers can generate $5 in rental revenue for every $1 spent on Nvidia's infrastructure, given sustained strength in the demand for accelerated computing.

Read more of this story at Slashdot.