NVIDIA Vera Rubin Bill of Materials and the Investment Opportunities
Leonardo Boquillón Briefs
SubscribeSign in
NVIDIA Vera Rubin Bill of Materials and the Investment Opportunities<br>Vera Rubin BOM: A $7.8M Rack. Where Could the Margins Hide?
Leonardo Boquillón<br>May 22, 2026
Share
Since yesterday, May 21, a Morgan Stanley report has been circulating online showing the Bill of Materials price increases on racks, from Blackwell (the previous NVIDIA architecture) to the newest Vera Rubin. A GB300 rack costs almost $4 million, while a VR200 NVL78 comes in at $7.8 million, and yes, you are reading that correctly, one single rack consuming around 150 kW of power is almost $8 million.
Thanks for reading Leonardo Boquillón Briefs! Subscribe for free to receive new posts
Subscribe
But what I want to write about is the Bill of Materials of a Vera Rubin rack and the price spikes versus the previous generation, plus what can be extracted from this as an investment opportunity, and as for those incumbents that have been raising prices relentlessly, can they be challenged?<br>Let's talk first about the Morgan Stanley report. I'd rather not paste the screenshot they published, but the numbers tell a clear enough story on their own. Memory moved the most: HBM went up 435% generation over generation, from about USD 375k per rack to USD 2 million. Basically, memory today is half of the total cost difference between GB300 and VR200. PCB (printed circuit board) came in second at 233%, climbing from roughly USD 35k to USD 117k. MLCC (capacitors) also posted a 182% increase. MLCC is not a big part of the BOM in dollar terms, but it has increased a lot in unit counts. NVLink switch silicon roughly doubled, plus a bit. 122% increase, from USD 65k to USD 144k. ABF substrate followed at 82% and the GPU itself was up "only" 57%, from USD 2.52 million to USD 3.96 million. Others that were not up "that much" (especially compared to memory): power supply at 32%, rack assembly at 29%. Cooling barely moved at 12%. CPU stayed flat at USD 180k, the only line in the entire bill of materials that did not change.<br>Now a few thoughts about this and how it can be used as an investment opportunity, if there is one.<br>First of all, look at memory. This is a two-player game, SK Hynix and Samsung, with a third trying to enter, Micron, but still far behind. How long can the memory makers sustain this? It is not a secret that NVIDIA’s CEO is very “unhappy” about memory prices, but memory makers are their own world. They own their own fabs, they are vertically integrated, so they have a kind of freedom no one else has. Have you seen the recent stories about SK Hynix and Samsung employees? They are single-handedly pumping the Korean economy. There is even a Korean term, 성과급 잔치 (seonggwa-geup janchi), which translates roughly as “bonus feast”. That is the headline phrase Korean media uses every time SK Hynix or Samsung’s semiconductor division pays out a big PS (초과이익분배금, “profit sharing”) bonus and the surrounding economy goes into overdrive. Macquarie Securities projected that SK Hynix employees would receive an average bonus of 1.29 billion won (~USD 910k) per person next year. One SK Hynix employee told a reporter not so long ago:<br>It used to be mostly German cars like Mercedes-Benz or BMW around here, but this year you see noticeably more people driving Porsche
SK Hynix last year performance<br>But is this sustainable? Aren’t high prices the biggest driver of innovation? How long will your customers happily keep paying those prices? Believe me, people are already working around the memory wall.<br>A few hours ago, this headline came out on Tom’s Hardware:<br>Intel is going with one massive GPU for its next-generation Xe3P AI GPU, and LPDDR5X memory rather than HBM.
So Intel is actively trying to find a way around HBM prices, because eventually there will be an answer to this, that is what capitalism does. Just as ASICs have been the answer, slowly biting into NVIDIA accelerators. The reason TPU and Trainium were born is because hyperscalers are tired of paying such prices. How long until there is a solution to break the HBM pricing power of the Korean giants? Two years? Five maybe? We will see, but solutions are being worked on right now. Nobody just sits and watches their provider milk them tirelessly while doing nothing.<br>But let’s move on to what I find more interesting. If you look at the numbers, PCB and ABF substrate, even though they went up a lot, represent a smaller part of the BOM, which is now basically memory. So any substantial price increase in the future can be absorbed faster, there is less motivation to find alternatives and enter new qualification rounds, because if it is still affordable and reliable, it creates less pressure than memory, for instance.<br>PCB is a less explored area than memory. I wrote a piece not long ago called "When It Rains, It Pours" talking about how the Middle East crisis and naphtha shortage will drive up the prices...