Hot Chips 2026：DRAM 与闪存现货价暴涨约七倍，Nvidia 警告 NVL72 系统明年涨价 15%

STORE

The Hot Chips 2026 conference kicked off with half-day tutorial sessions on Sunday, and Jim Handy, the general director at semiconductor market research firm Objective Analysis, kicked off the memory tutorial in the first part of the day by giving us a rundown on the state of the DRAM, HBM, and flash markets to set the stage for the presentations about high bandwidth memory and high bandwidth flash that followed.

We will be getting into all of those technical presentations from Micron Technology, Samsung, SK Hynix, d-Matrix and Meta Platforms, and Oxmiq Labs separately, but like Handy, I think it is important to level set on the absolutely stratospheric pricing that is happening with DRAM and flash as we have a massive demand shock in a head-on collision with a dearth of supply.

The main culprit is the hyperscalers and the cloud builders and their aspirations to be the platforms upon which ever more powerful GenAI models will be trained and which will also be where a lot of the inference will be done when these models are run in production. This is driving the HBM and GPU/XPU booms, as well as the ancillary DRAM and flash booms. And the rise of agentic AI, which needs tons of CPUs as well as some DRAM and flash to make them useful to create sandboxes in which scads of possibly crappy Python code generated to perform tasks we give agents runs as well as code for other platforms is generated as applications are modernized or created whole cloth from the cosmic Ether(net).

Here is the spending as Handy has cased from the beginning of 2021 through the second quarter of this year:

What I find interesting in this chart is how Alibaba, the biggest of the Chinese hyperscalers that also has a cloud business, barely registers against the vast spending by the big four tech titans in the United States: In order, that is Amazon Web Services, Google (Alphabet is not really doing the investing, the search engine, advertising, and cloud division is), Microsoft Azure, and Meta Platforms. And Apple, which uses AI but does not create it, is barely a line, and in fact, its capital investments in AI are smaller in 2026 than they were for the prior five years.

I mean, if you can sell AI at a profit and not have to do the capital investing, then go, be, do.

Because of this demand from the hyperscalers and cloud builders and a new round of investments from enterprises, governments, and academia as they seek to get more efficient back office, Web infrastructure, and data analytics computing to make way for GenAI systems, prices for all kinds of storage, all the way back to hard disk drives, which are largely consumed by the tech titans as nearly cold storage, are skyrocketing. And so the big five memory and flash makers are making out like bandits, as you can see from Handy’s aggregation of their revenues:

Look at that memory and flash crash back in late 2022 and in early 2023, when GenAI ironically was just getting started. If the GenAI boom keeps going – and all the prognostications I have seen see it growing through 2030 at the least and shortages continuing until that time at least – then there is no reason to believe that revenues will keep going up and to the right at a linear rate.

“It's all made on the same manufacturing line,” Handy explained to the assembled Hot Chips multitude, referring to flash, DRAM, and HBM. “And so because of that, it's causing DDR prices to go crazy, and all of NAND flash prices are going crazy, and you have the huge revenues. The huge revenues are actually caused by the fact that they are commodity products. Basically, you can buy NAND flash from supplier ABC, and the transition is very easy. That's what a commodity is. You don't really care about the manufacturer; you mostly care about what kind of price you can get. But it also means that the manufacturers, when there's a shortage, then they don't care about you. And they can say, okay, we dealt with this company for a while, except they don't pay their bills on time, and anyway, we can get a higher price from that guy over there, and so they go for the higher price, and so you've got these prices that have gone up about seven times on the spot market, which is not a very good indicator, but it's an easy one to measure and so I put it up here.”

Spot pricing is a kind of average, and everyone making systems or chips with integrated memory on the package (whether it is HBM or LPDDR5X DRAM) are making their own deals with their own pricing and volume commitments. Some do better than spot, and some do worse, depending on these factors.

For instance, as the highest volume buyer of HBM memory for many years in the era of deep learning and the beginning of the AI boom, Nvidia has largely controlled the price of HBM. (See He Who Can Pay Top Dollar For HBM Memory Controls AI Training from February 2024 for more on that.) But now, thanks to its custom XPU business, Broadcom is now the third largest buyer of HBM memory, behind Nvidia and AMD, and it has the financial firepower to compete for HBM capacity (based largely on price) on behalf of the myriad AI hardware startups could not hope to get chip etching and packaging without the help of companies like Broadcom and Marvell.

Hence, with all of the GPU components being long since sold out for 2026 and component allocations almost certainly done through the end of 2027, Nvidia is warning customers who are buying “Grace-Blackwell” and “Vera-Rubin” NVL72 systems next year that there is a 15 percent price increase coming, according to reports in Bloomberg. I assume Nvidia has put flexible pricing provisions into its contracts so it can pass through rising HBM, DRAM, and flash costs on to its ODM and OEM customers, who will in turn pass it on to the actual system buyers. I am surprised the price increase is not higher.

Nvidia’s business can only grow as fast as the HBM memory it can get, the same as every other GPU and XPU maker. And if the memory makers and Nvidia don’t want to get hit with antitrust lawsuits, Nvidia has to make sure its competitors can get capacity. Generally speaking, 85 percent revenue or shipment market share and higher has historically been the threshold for being called a monopoly by the US Department of Justice, but these days, what really matters is what President Trump thinks about your business at any given moment.

What the industry needs to do is figure out ways to accelerate the accelerators, and we are seeing that with externalized flash-based, DPU-driven storage tiers for KV caches as well as adding compute to HBM base dies and also adding compute into memory subsystems – so-called Processing In Memory, or PIM, which we have written about many times here at The Next Platform. Handy agrees, saying that the bandwidth requirement of a communications channel is inversely proportional to the intelligencer at either end of the channel linking two compute devices.

“In the way of getting you know past the memory wall, we are probably headed towards having smarter memory chips that will end up reducing the amount of traffic it needs to go to the processor,” Handy said in his opening keynote. “And that means that you might have the processor telling the memory do a sort, and the memory does the sort on its own, and then it comes back and it says done, and then the processor says what's on top, and one byte comes back, and that's the end of the job, and that's something that would have involved an awful lot of communication if there weren't the smarts inside the memory chip. So with that, I'm expecting to see AI permeate everywhere. AI can be done on microcontrollers, and I think that there's a new way of thinking that people are working on, that they are saying, oh, you know, it seems to be working over here. Let's try it over there, and that isn't something that people were doing before. There wasn't this readiness to accept it, and so I'm expecting to see AI used at all of the endpoints. I'm expecting to see it in the communication channel, and it will get to the point where it's invisible.”

SPONSORED FEATURE: HPE and NVIDIA help organizations apply accelerated computing, software, enterprise infrastructure, networking, control plane, and services to the workloads and business outcomes that matter most.

October 1, 2026

September 30, 2026

AMD Buys World Labs For $8.2 Billion To Get AI Model Cred