Nvidia’s H100 shipped with 80GB of high-bandwidth memory. Its rack-scale GB300 NVL72 now carries 20TB across 72 GPUs, about 278GB each (our arithmetic), per Nvidia’s specifications.
Micron’s chief executive says AI “just needs more memory and storage,” and that 75% of its 2027 supply is already sold. Here is how AI memory works, in six numbers. Nvidia’s market story is on our Nvidia stock page. Micron’s filed numbers are tracked on our Micron stock page.
Important: not investment advice
This article is for information and education only. It is not investment, financial, legal or tax advice, and it is not a recommendation or an offer to buy or sell any security, ETF, option or other financial product. FourWeekMBA is not a broker-dealer or a registered investment adviser. IPO shares, leveraged and inverse ETFs, and options carry significant risk, including the loss of your entire investment; leveraged and inverse ETFs reset daily and can lose value quickly. Prices, availability, fees and eligibility change and can differ by country and broker, so check the prospectus and your broker’s terms. Figures marked “our arithmetic” are our own calculations from the cited filings and may contain errors. Do your own research and consider speaking to a licensed financial professional before making any investment decision.
Business Pill · Free Cash Flow
An evergreen Business Engineer explainer. It teaches the general idea only and does not discuss this report.
The key insight: Memory per Nvidia GPU rose from 80GB on the H100 to about 278GB on GB300 systems (our arithmetic), and Micron says 75% of its 2027 supply is sold, with $13.0 billion paid by customers before delivery.
What HBM Is
High-bandwidth memory, or HBM, is DRAM stacked in layers and placed right next to the GPU. Micron sells it in stacks of 8 or 12 dies, its product page shows.
Each Micron HBM3E stack has 1,024 input-output pins and moves more than 1.2 terabytes a second, the company says. That width is what lets a GPU feed its compute cores fast enough.

Number 1: 80GB to About 278GB Per GPU
Nvidia lists 80GB for the H100 SXM and 141GB of HBM3e for the H200, “nearly double the capacity” of the H100. Its DGX B200 system holds 1,440GB across eight GPUs, or 180GB each (our arithmetic).
The GB300 NVL72 lists 20TB of GPU memory for 72 Blackwell Ultra GPUs, about 278GB per GPU (our arithmetic). That is about 3.5 times the H100 (our arithmetic). Nvidia says Blackwell Ultra offers “1.5x larger HBM3E memory” than its predecessor.
Number 2: 12 Dies in One Stack
Micron’s 24GB HBM3E uses 8 stacked dies and ships in Nvidia’s H200; its 36GB version uses 12, per Micron. More layers per stack is how capacity per GPU keeps rising.

Micron’s chief executive told an audience on Wednesday that the companies have been “pushing it toward higher die stack and lower power,” as reported by Seeking Alpha.
Number 3: 75% of 2027 Already Sold
“Seventy-five percent of our 2027 is sold out. We are in discussions with customers now moving to 2028,” Micron CEO Sanjay Mehrotra said, as reported by Seeking Alpha.
He said AI “just needs more memory and storage” as frontier models get larger and context windows longer. In October he said HBM prices for 2027 “are much higher than 2026 prices,” PCMag reported.
Number 4: $13 Billion Paid Before Delivery
Micron’s contract liabilities, money customers have paid before receiving products, were $13.0 billion at 3 September 2026, up from $169 million a year earlier, its 10-K data show. That is about 77 times higher (our arithmetic).
The 10-K data do not say which products the payments cover. Mehrotra has said many of Micron’s Strategic Customer Agreements now run five years, which “give you that visibility into the demand.”
Number 5: $30.71 Billion of Capital Spending
Micron spent $30.71 billion on plants and equipment in fiscal 2026, about 23.1% of its $133.19 billion revenue (our arithmetic), per its 10-K. Revenue was about 3.56 times the prior year (our arithmetic).
New capacity is slow to arrive. “Clean rooms take a long time to build,” Mehrotra said in October, and he expected tight supply to continue even with new space coming up in 2028, PCMag reported.
Number 6: Three Makers
HBM comes from three companies: SK Hynix, Samsung and Micron. This week Samsung handed its memory business to its HBM chief and SK Hynix took a 14.1% stake in Kioxia.
Earlier we reported SK Hynix and Micron margins above Nvidia’s. Our earlier explainer covers why memory, not only GPUs, sets the pace, and our NAND piece covers storage.
The Structural Read
As we read it, each new GPU generation raises memory demand faster than chip counts alone, which is why memory makers can sell output years ahead.
As we read it, customer prepayments turn part of a cyclical memory business into contracted supply, at least while AI demand holds.
Sanjay Mehrotra, Micron CEO, TomorrowX Summit (per Seeking Alpha)
“Seventy-five percent of our 2027 is sold out. We are in discussions with customers now moving to 2028.”
Three Implications
FOR AI BUILDERS Each Nvidia generation carries more HBM per GPU, per Nvidia’s specifications.
FOR MEMORY MAKERS Micron reports $13.0 billion of contract liabilities and $30.71 billion of capital spending, per its 10-K.
FOR SUPPLY Mehrotra says new clean rooms take a long time to build, PCMag reported.
What Is Not Established
Micron did not split the 75% figure by DRAM, NAND or HBM, or give prices. The per-GPU figures for B200 and GB300 are our division of Nvidia’s system totals; individual configurations can differ.
The Bottom Line
Memory per AI GPU has risen from 80GB on the H100 to about 278GB on GB300 systems (our arithmetic), and Micron says 75% of its 2027 supply is already sold.
Important: not investment advice
This article is for information and education only. It is not investment, financial, legal or tax advice, and it is not a recommendation or an offer to buy or sell any security, ETF, option or other financial product. FourWeekMBA is not a broker-dealer or a registered investment adviser. IPO shares, leveraged and inverse ETFs, and options carry significant risk, including the loss of your entire investment; leveraged and inverse ETFs reset daily and can lose value quickly. Prices, availability, fees and eligibility change and can differ by country and broker, so check the prospectus and your broker’s terms. Figures marked “our arithmetic” are our own calculations from the cited filings and may contain errors. Do your own research and consider speaking to a licensed financial professional before making any investment decision.
97,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. GPU memory figures are from Nvidia’s product specification pages and HBM stack figures from Micron’s HBM3E page, read on 20 November 2026. Revenue, capital spending and contract liabilities are from Micron’s 10-K data via SEC company facts. Mehrotra’s remarks are as reported by Seeking Alpha and PCMag. Per-GPU figures for B200 and GB300, the 3.5 and 77 times ratios and percentages are our arithmetic.
Sources: NVIDIA H100 Tensor Core GPU specifications · NVIDIA H200 Tensor Core GPU specifications · NVIDIA DGX B200 specifications · NVIDIA GB300 NVL72 specifications · Micron HBM3E product page · Micron Technology 10-K filings (SEC EDGAR, XBRL company facts) · Seeking Alpha via TradingView, Micron at the TomorrowX Summit, 18 Nov 2026 · PCMag via Yahoo Finance, Micron CEO: No End in Sight for Memory Shortage, 1 Oct 2026









