Not long ago, I did a deep dive into Micron’s fundamentals and it became apparent how important High Bandwidth Memory (HBM) has become for its revenue. Some even argue that this new type of memory is not a traditional cyclical, which led me down the rabbit hole of “Who’s buying this memory?”.
One of the names that stands out is Nvidia, and the numbers indicate that this is something that is not receiving as much attention as it deserves. Using Micron’s $100B TAM forecast for HBM, multiple sources claim around 60% of this memory ends up in GPUs and accelerators. Which seems realistics after their $40B revenue for Q3 2026.
This leaves us with $60B of memory going into GPUs. Using Nvidia’s numbers, around 10% of GPU cost is HBM, so… if there’s demand for $60B of HBM just for GPUs and accelerators, that’s around $600B of reasonable TAM for the same GPUs containing this memory. If we compare this to Morgan Stanley’s forecast, who puts GPU TAM at $400B out of their $1.1T total AI CapEx forecast, and we adjust it to these numbers results a total over $1.3T.
Nvidia currently captures around 35% to 40% of AI CapEx, so simply sustaining that pace would put their Data Center revenue at \~$465B between compute and networking. Already over the $390B Wall Street total consensus. This is before adding $15B–$30B for their edge/retail GPUs and compute, which would put that number on the high end at \~$495B in revenue.
As of today, Nvidia can convert 66% of that revenue into EBITDA, with estimates going as high as 70% for 2027. Even on the low end of those numbers, it would put the company at \~$325B EBITDA, which at today’s enterprise value of \~$4.5T gives us a 14x forward EV/EBITDA multiple, or $192 per share, for the company currently leading the compute market. That’s half of today’s multiple and around three times less than AMD’s forward multiple of x46.
[](https://substackcdn.com/image/fetch/$s_!AFkE!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fda3f52a0-141c-4731-b5c5-55b999daabb5_1556x467.png)
With that in mind after capturing 35% of total hyperscaler CapEx, with current enterprise value, the multiple of 14x would leave us with the cheapest large-cap AI name in the market. Considering their large footprint in the sector, I would consider that to be grossly undervalued. Even if their multiple were to contract somewhere between 14x and today’s \~30x, and they traded at a discount to peers at 25x, we land at $336 a share. Putting it in an odd situation where it seems like the market has almost forgotten it’s even there. With AMD (\~110x) and Intel (\~50x) trading at multiples well above the leader of the segment, and even its own supplier Micron trading at \~30x.
Let me know what you think, and do not take anything in this piece as investment advice. Just my observation while doing research.
