HBM alone is no longer enough for the next phase of AI.
At FMS 2026, major memory and storage companies outlined hierarchical memory architectures designed for inference and agentic workloads.
As AI moves beyond training into multi-step reasoning and repeated data access, both capacity and bandwidth requirements are rising. Systems need more than just stacking higher HBM.
SK Hynix and SanDisk presented High Bandwidth Flash, or HBF. The technology stacks NAND in 8-high or 16-high configurations, targeting capacities up to 512 GB and bandwidth ranging from 0.4 TB/s to 3.0 TB/s. It uses the open UCIe chiplet interface and is positioned between HBM and conventional SSDs.
SK Hynix also described a tiered-memory approach in which HBM handles the highest-bandwidth data while HBF covers larger-capacity needs.
The two companies have been working on HBF standardization under the Open Compute Project, with participation from Google and others.
Micron showed a hierarchical structure linking HBM, DRAM and SSD so that data can be placed according to its access characteristics.
Samsung highlighted next-generation V-NAND above 400 layers along with zHBM, which stacks memory directly on the accelerator, and zNAND-O for higher-performance NAND use cases.
The common theme is that inference and agentic AI are expanding the memory working set.
Pure HBM scaling is necessary but not sufficient. A broader hierarchy that mixes bandwidth, capacity and cost is becoming the practical path.
#semiconductors
How important do you expect intermediate layers such as HBF to become in AI system designs over the next few years?
Morgan Stanley: "NAND market is stronger than feared"
On Sandisk:
"LTAs now cover about 50% of next year and 2⁄3 of the following year, at gross margin floors of about 80%. That's below market, a little (current gms are mid 80s) but there are pricing escalators and there is protection both ways. We aren't buying the stock for the LTAs, but to us the LTAs provide evidence of just how durable the customers perceive this tight supply situation to be, which is key."
$SNDK
SNDK's gross margin plateau is a welcome move. It means they are not squeezing customers to the extreme, and that they are looking at a long term vision instead. The disciplined, balanced capex execution I have been arguing for all along, the kind that prevents a company from investing to extremes at the top of the cycle and taking severe damage in the downcycle, requires SanDisk's NBM model. There is no way around it.
So is now the right time to buy SanDisk? The bears argue that the gross margin plateau marks the peak of this cycle. That is also why the stock is falling right now.
As I wrote in my SK Hynix earnings review(You can find it in the Citrini Substack chat), this is the point where we should be looking at Q, not P. As long as you believe in the AI supercycle, as long as big tech keeps spending, as long as compute companies like Nvidia remain in short supply, there is no scenario where the shortage eases for memory alone. That is the corresponding effect of LTAs (NBM) and disciplined capex. Even if gross margins stall, as long as Q expands and the floor set by those LTAs prevents an excessive collapse, the total profit pool will grow enormously.
Layer massive shareholder returns such as buybacks on top of that, and memory companies should be rerated, and they will be rerated. From cyclical stocks to essential infrastructure of the AI world.
$SNDK
Wistron Chairman Simon Lin said AI servers currently account for 95% of its server business and that the AI trend is so strong that even high memory chip prices haven’t slowed customer spending, media report, adding he doesn’t fear even higher memory prices because “customers will still buy.” He said Wistron’s business has been mainly from one major enterprise client, but new enterprise clients will start taking shipments in the 2nd half. $NVDA $DELL $HPE https://t.co/eCaK7wvqal
#education
Before you get your panties in a bunch over an earnings move, zoom out and look at the higher TF
Some of the best opportunities show up after earnings. You won't see them if you're reacting like you just found blood in your underwear.
This earnings season has been full of the same pattern.
Sell algorithms slam the stock on the initial reaction, then buyers step in over the next day or two.
If you're glued to the one minute chart and letting every candle dictate your emotions, you'll miss the bigger trade.
The market rewards patience more often than panic.
- Zoom TF out
- Find support
Let the emotional traders create the opportunity for you.
Absolutely insane $SNDK guidance comments from earnings call.
- "NAND market will exceed $300B in revenue in calendar year 2026, up 3x year over year."
- "NAND market will approach $500B in revenue in calendar year 2027."
- "Data center share of total revenue to expand from ~30% in 2025 to ~50% in 2026, and to continue outpacing the market in 2027."
- "Demand from our customers is growing faster than our supply."
$SNDK: "We estimate the NAND will exceed $300 billion in revenue in calendar year 2026, up 3x year over year. Looking further ahead, we estimate that the NAND will approach $500 billion in revenue in calendar year 2027."
SOUTH KOREAN MEDIA: APPLE RECENTLY NEGOTIATED WITH CXMT OVER SUPPLY PRICES FOR MOBILE DRAM, INCLUDING LPDDR5X, IN AN EFFORT TO EASE MANUFACTURING COST PRESSURES ON ITS NEXT-GENERATION IPHONE AND SMART DEVICE LINEUP, BUT CXMT REJECTED ITS REQUEST FOR PRICE CUTS
SOUTH KOREAN MEDIA: CXMT INSTEAD INSISTED ON PRICES THAT WERE HIGHER THAN OR COMPARABLE TO THOSE OFFERED BY SAMSUNG ELECTRONICS AND SK HYNIX
$AAPL