Quick View: NVIDIA still more boom than bust
Portfolio Manager Richard Clode joined NVIDIA’s most recent earnings call. He highlights the main considerations for investors on the company, as well as the generative AI tech wave more broadly.
3 minute read
Key takeaways:
- The near-term initial teething problems of ramping an incredibly complex supply chain for rack level Blackwell (AI chips) systems now appear to be in the rear-view mirror.
- Longer term, NVIDIA made the case for a new exponential AI scaling law in test time scaling reasoning models as well as Jevon’s Paradox to drive future AI infrastructure growth.
- The team looks towards NVIDIA’s GTC event in March where the company will unveil its next generation suite of AI chips and other innovations.
As ever, NVIDIA founder and CEO Jensen Huang is very aware of the market concerns centred around companies overspending on the AI infrastructure build out and sought to address them head on during the earnings call. The company also plans to provide more detail on its latest Blackwell Ultra, Rubin and Vera chips, as well as a host of other new innovations at its next generation roadmap at GTC, NVIDIA’s AI and tech conference in March, as the company looks to break the stock out of an eight-month digestion period.
Here we summarise 4 key highlights from the earnings call:
1 What Blackwell supply issues?
After significant angst around a problematic initial ramp of a very complex and inaugural rack level system, the company delivered US$11 billion of Blackwell sales last quarter (ending 26 January 2025) well ahead of guidance of ‘several billion’. For context, 350 plants are producing the 1.5 million components in each single Blackwell rack. A more aggressive supply ramp through 2025 to meet ‘extraordinary’ demand for Blackwell can be expected.
2 A new AI scaling law will further accelerate LLM development
Test time scaling and the advent of reasoning models is a major new exponential vector for AI infrastructure spending. Long think reasoning can be 100 times more compute intensive versus standard one-shot inferencing. That is the first generation, with the possibility of subsequently more thoughtful reasoning models that could be hundreds of thousands or millions of times more compute intensive. The more the model thinks, the more compute processing it does, the more accurate and intelligent the answer.
3 Jevon’s Paradox is kicking in
NVIDIA has delivered a 200 times reduction in inferencing costs in just two years. DeepSeek was just the latest iteration of AI innovation that is driving down the cost of this new technology. Far from reducing demand for AI infrastructure, Jevon’s Paradox means we are experiencing the opposite, as we have witnessed during prior technology cost down curves in the PC and internet eras.
4 NVIDIA ahead on the ASIC versus GPU debate
Broadcom’s ASIC chips are designed for single purpose tasks and are also cheaper versus NVIDIA’s GPUs which are general purpose, being able to handle multiple tasks like AI training, image and video rendering, and gaming. As such, ASIC has emerged as a challenger to NVIDIA GPUs.
NVIDIA’s CEO made the point that the main data centre bottleneck is power supply, hence customers want to maximise the revenue potential of each data centre. The way to do that in an AI world is to maximise the number of tokens generated per watt. NVIDIA currently leads the way by a factor of 2-8 times versus the competition. Its general-purpose processors can multi-task, including handling the significantly greater complexity of AI software, which can be a challenge when operating multiple underlying hardware.
We believe generative AI is not just a theme. The complexity of a major new technology wave and a ‘winner takes most’ industry requires bottom-up fundamental research and active management to identify the key beneficiaries who can exhibit underappreciated earnings growth in this highly innovative and disruptive investment landscape.
Note: source of all NVIDIA earnings information as at 26 February 2025: https://nvidianews.nvidia.com/news/nvidia-announces-financial-results-for-fourth-quarter-and-fiscal-2025.
Blackwell chip: NVIDIA’s next-generation Blackwell graphics processing units (GPUs), have significantly better energy consumption at lower cost to complete tasks for AI processing.
Inferencing: refers to artificial intelligence processing. Whereas machine learning and deep learning refer to training neural networks, AI inference applies knowledge from a trained neural network model and uses it to infer a result.
Jevon’s Paradox: a theory suggesting that greater efficiency in the use of any given resource can result in increased demand for that resource. Applying this to technology/AI chips, as technological and efficiency improvements in chips increase, the overall demand for chips actually increases rather than decreases.
LLM (large language model): a specialised type of artificial intelligence that has been trained on vast amounts of text to understand existing content and generate original content.
Long think reasoning: a deliberate and extended process of considering information and potential outcomes, by analysing multiple perspectives, considering long-term implications, carefully weighing various factors before reaching a conclusion.
One shot inferencing/prompting: refers to the method where a model is provided with a single example or prompt to perform a task. It relies on a single, well-crafted prompt to achieve the desired output.
Test time scaling: a language modelling approach that uses extra test-time compute to improve performance.
Token: AI tokens are the fundamental building blocks of input and output that Large Language Models (LLMs) use. They are the smallest units of data used by a LLM to process and generate text/output that is useful.
3 minute read
Key takeaways:
- The near-term initial teething problems of ramping an incredibly complex supply chain for rack level Blackwell (AI chips) systems now appear to be in the rear-view mirror.
- Longer term, NVIDIA made the case for a new exponential AI scaling law in test time scaling reasoning models as well as Jevon’s Paradox to drive future AI infrastructure growth.
- The team looks towards NVIDIA’s GTC event in March where the company will unveil its next generation suite of AI chips and other innovations.
