GP-You missed it? Nvidia’s wild ride is not over

Update 1 (Nov 2024): The only pushback to this narrative that I can see currently, is that while training compute is highly skewed in favour of Nvidia, because of its CUDA moat, inference compute is ‘currently’ not that dependent on CUDA, and NVDA doesn’t have much of an advantage there, with some benchmarks showing Groq and Cerebras outperforming Nvidia in inference. Some things to consider here.

1. There is a brand value of new infrastructure project being deployed with Nvidia vs risky/unproven newcomers (although this could mean nothing if builders can’t get their access on Nvidia GPUs)
2. The results were based on smaller context windows. As the context window increases of the models, it skews the need for larger and larger inference clusters. This is where Nvidia shines. The context windows for agents will continue to grow, these systems want to only consume more data. As you deploy these solutions for real work, we’re looking at always-on always listening agents. So I see this headed in the pseudo infinite context window.
3. Nvidia will still have the supply chain and production edge over the competitors for atleast 1–2yrs.

Alright, let’s skip the history lesson — everyone knows Nvidia’s origins by now. And unless you’ve been living under a rock, you know that the AI wave is here, leading to the craziness we’ve seen.

But let’s really hit you on the head with the magnitude of what’s ahead.

The AI Revolution: The True Driver of Nvidia’s Value

Let’s start with this: we’ve already achieved AGI. ChatGPT, Claude, and Gemini are outperforming average human performance in various tasks, such as undergraduate knowledge assessments and advanced reading comprehension.

But the rate at which this is accelerating is mind-bending, and there are ZERO signs of this stopping.

“It is possible that we will have superintelligence in a few thousand days (!); it may take longer, but I’m confident we’ll get there”

-Sam Altman

While he is biased, I do think that here he is sandbagging it, and ASI will happen within 2–3 years.

In a recent Lex Fridman podcast, Dario Amodei, CEO of Anthropic, discussed how we are standing at the doorstep of Artificial Superintelligence (ASI), predicting 2026–2027.

Listen the first 45 seconds.

Although the definition of ASI and when it happens doesn’t matter, but we are on this race to build it.

People are still not fully grasping the sheer scale of what’s coming next. And what this means for compute requirements?

Well, Ilya Sutskever, co-founder of OpenAI, emphasized the immense computational demands of advanced AI systems, stating, “To reach its full potential, AI would require a computer the size of a city”. This isn’t just hyperbole — the largest supercomputer, XAI, has 100,000 Nvidia H100 GPUs, and OpenAI’s GPT-3 was trained using about 10,000 GPUs. We are entering the age of 500k and 1M GPU superclusters.

We are likely 10x to 100x in training demand away, with every major company in this arms race to build the foundation models.

But of course, you're thinking, this is all priced in! I mean, Nvidia is already sold out through 2025. Nvidia’s stock has delivered a remarkable 2,600% return over the past five years.

I’ve already missed the boat, right? It’s already in the top 2 largest companies in the world by market cap!

But while the training demand is priced in, what people are missing is the inference compute demand.

Right now, you have an AGI on your fingertips, but how much do you use it? The usage practically rounds down to zero. By Jan 2025, OpenAI is releasing its Operator Agents, and Claude already just released Computer-Use, its form of live agent. Imagine a sales agent, or HR agent, given a job, sitting on LinkedIn, messaging leads, setting up meetings. CSMs checking in on customers, Finance analyst agents, personal assistant agents, the possibilities are endless. The value this will create for each company is so massive, that they will wants hundreds, if not thousands of agents, running 24/7. The reason ChatGPTs o1 model is limited in tokens, is because of a lack of compute! The compute demand is about to skyrocket, by at-least 10000000000x (ten orders of magnitude). And what all does this run on? Nvidia GPUs.

Listen to Jensen, highlighting this same fact, which many analysts are missing.

But what about the valuation?

P/E Ratios and Growth Potential: Still Room to Grow?

Nvidia currently trades at a Price-to-Earnings (P/E) ratio around 67. The valuation is sky-high by traditional standards. But this isn’t a traditional situation. It’s growing at rates more similar to startups, and a better metric is the forward PE, which is currently at 33. For perspective, many high-growth startups, especially in the tech and AI sectors, often trade at forward P/E ratios well above 50, reflecting investor expectations of rapid expansion. Given the unprecedented demand, this is going to be supply constrained for many years to come.

Risks on the Horizon

Further, competition from Groq, and other emerging AI chip startups like Cerebras and Graphcore could eventually challenge Nvidia’s market dominance. However, the real limitation for these competitors is Nvidia’s moat — established through its CUDA software ecosystem and even more significantly through its supply chain and manufacturing capabilities. Ramping up manufacturing for a chip company is NOT EASY, and will take many years. Nvidia itself can’t ramp up fast enough; even its own suppliers are sold out. Bernstein analysts have noted that Nvidia’s ecosystem lock-in gives it a near-unassailable competitive advantage in the AI market. The main risk in my mind if the supply can capture this demand, or will we remain in a compute drought for many years…

The AI wave is the largest wave in our lifetime. If we follow the trends, I think Nvidia is still undervalued as of today.