NVIDIA continues to believe that it is still in the relatively early innings of a major refresh of around US$1 trillion worth of data centre infrastructure that needs to be accelerated by GPUs for modern data processing, notably focused on generative AI. The company is about to embark on their first major new product cycle of the post-ChatGPT era with the new Blackwell chip that is scheduled for mass production in the fourth quarter of 2024 – new product cycles have historically been a key driver of NVIDIA’s growth and earnings potential.

Confirmation of only a short delay to Blackwell was therefore a key takeaway for investors. NVIDIA’s competitive moat in AI training so far remains untouched. But as gen AI applications are scaled out, infrastructure spending will pivot to inferencing, where barriers to entry are typically lower. This will create opportunities for other companies to emerge as challengers to NVIDIA.

Meanwhile, the law of big numbers makes it more challenging for NVIDIA to continue to maintain its stellar revenue growth rates (and exceed market expectations), given the quantum of additional absolute dollars needed for capital spend each year. We believe the next generation infrastructure theme remains alive and well, as evidenced by NVIDIA’s Big Tech key customers during the recent results season. Looking ahead, investment opportunities are likely to broaden out beyond NVIDIA from 2025.

Key insights from NVIDIA’s 2Q fiscal year 2025 results:  

• Blackwell delivery delay fears allayed

The key concern heading into the latest results announcement was the impact of some initial teething issues with the next generation Blackwell platform. Given the accelerated roadmap of an annual cadence of new product delivery, these initial issues were to be expected. Management confirmed a mask redesign to improve yields but no functional issues with product sampling to customers today. NVIDIA appears to be on track to deliver several billion dollars’ worth of Blackwell chips in the end January quarter. Blackwell is key to driving growth for NVIDIA into the next two years. Hence the confirmation of a minor delay was welcome. As the first new architecture to launch in the post ChatGPT era, Blackwell should meet the challenges of next generation large language models requiring 10-20x the compute power. It should also enable NVIDIA to capture more content in the data centre by offering a full-rack solution encompassing GPU, CPU, memory, networking and software. Having all those pieces of the puzzle under one roof remains NVIDIA’s competitive moat, particularly in AI training, facilitating 3-5x the AI throughput in the same data centre floorspace, an increasingly scarce commodity – Blackwell is an AI infrastructure platform, not just a GPU.

• The enterprise AI wave has begun

Major deployments to date have been to the hyperscalers and internet companies. But NVIDIA now sees enterprises across verticals and geographies deploying AI infrastructure, which remains an early innings opportunity. Automotive companies training up next generation autonomous driving models, as well as healthcare companies were cited as key verticals spending several billion dollars each with NVIDIA – the company now has most of the Fortune 100 as customers.

• Increasing sovereign demand

NVIDIA raised its revenue forecast from sovereigns (countries as customers) and related entities to low double-digit billions from  high single-digits previously. This is a new customer base that did not exist in the internet era. Countries are now realising that their data is their natural and national resource – they have to use AI, build their own AI infrastructure so that they could have their own digital intelligence.

• China comeback

Historically, China comprised 20-25% of NVIDIA’s data centre sales, but that collapsed post US semiconductor export restrictions. The recent launch of a compliant new chip, the H20, led to a significant rebound in China sales, albeit still only back to around 12% of sales. In a recent discussion we had with a major Chinese internet company, we learned that while the Blackwell’s predecessor the H20 is not necessarily a better chip to local alternatives, when using thousands of them to train an AI model, NVIDIA’s proprietary networking really comes into its own, driving far superior system performance.

• Mild normalisation of margins

NVIDIA gross margins peaked at almost 80% in the end April quarter; in contrast, management had guided for a normalisation. Even with a current new product, the H200, that is just reselling extra high bandwidth memory and heading into the ramp up of a low-yielding major new complex product like Blackwell, the company has still been able to maintain a mid-70% gross margin and a mid-60% operating margin. We think this shows that NVIDIA’s competitive moat and pricing power is likely to remain strong.

• New US$50bn share buyback

NVIDIA generated US$13.5bn in free cashflow last quarter. Given that backdrop, the company’s board authorised a new US$50bn share buyback authority. This is reminiscent of other highly cash- generative big tech companies that have delivered sizeable shareholder returns, while also investing heavily in innovation and preserving their virtuous growth flywheel.

*Source of NVIDIA earnings information and announcements: https://nvidianews.nvidia.com/news/nvidia-announces-financial-results-for-second-quarter-fiscal-2025

JHI

There is no guarantee that past trends will continue, or forecasts will be realised.