AI infrastructure · a read, not yet a living one
Is memory now what limits AI?
The current read
Memory is the binding constraint right now, and the argument is about whether that is a fact about memory or a fact about how bottlenecks behave. Gavin Baker ranks DRAM above power and everything else, expects it to hold, and puts a number on it: 30 to 40 percent of hyperscaler capex next year. Jensen Huang, speaking ten weeks earlier, described bottlenecks as things the whole industry swarms, and used HBM as his example of one that has already moved from specialty to mainstream. Dylan Patel holds both halves at once: the DRAM cell is about forty years old with no major breakthrough since, and asked what marks somebody as repeating a meme, the first thing he reaches for is “memory is the bottleneck”, which he grants is true and then moves past. Companies are already trying to build the way out.
What changed · between 15 April and 27 June 2026
Memory was already named as a constraint in April. What was missing was a ranking.
The April framing is that instantaneous demand runs ahead of supply everywhere at once, that whichever component falls too far behind gets swarmed by the whole industry, and that the next shortage is bought ahead of rather than reacted to.
“But as, as you know, the bottleneck often in training and doing inference on these models is the amount of bandwidth.”
Dwarkesh Patel · asking Jensen Huang · 15 Apr 2026 · 1:08:44 Named in April, and put to Huang directly. This is why 27 June is a change in ranking, not a first appearance.
On 27 June, after Micron's quarter, Baker ranked it and sized it: DRAM ahead of lasers, capacitors, power supplies, NAND flash and disk, and 30 to 40 percent of hyperscaler capex. The sharpest move is inside a single person. Sixteen days before that, the same investor explained why customers stay with Nvidia by pointing at a watt-constrained world.
“…that is an extremely unlikely outcome, and especially as long as we are in a watt-constrained world.”
Gavin Baker · Bg2 Pod · 11 Jun 2026 · 1:01:45 Sixteen days later he puts power below DRAM, in the list below.
The moment that moved it
Gavin Baker
Atreides Management, on All-In Podcast,
“Socialists Sweep NYC, China Catches Up in Coding, AI Memory Crunch, Micron's Blowout
Quarter”
No, well, one, DRAM is the most important bottleneck. … The bottleneck that matters is DRAM, and DRAM and HBM DRAM, this is the most important bottleneck, simply because memory capacity and bandwidth are foundational to the performance of every AI model…
Word-timestamped · speaker-attributed · verbatim
The strongest case that memory is the constraint
“…Elon is focusing the TeraFAB on memory 'cause he sees it as the most important bottleneck. You know, not lasers, not capacitors, not power, power supplies, semiconductors, not NAND flash, not HDDs, DRAM. Uh, and I think this bottleneck is gonna be with us for a while…”
Gavin Baker · All-In Podcast · 27 Jun 2026 · 1:04:03 The ranking, given as a list of what it beats.
“But for the DRAM you need in these AI servers, there are three companies that can make it. It's really hard to do.”
Gavin Baker · All-In Podcast · 27 Jun 2026 · 1:05:43 Why he expects the constraint to hold rather than clear.
“…memory is, DRAM is probably gonna be 30 to 40% of all hyperscaler CapEx next year. Every do- hundreds of billions of dollars that are spent, you know, it's going straight to DRAM.”
Gavin Baker · All-In Podcast · 27 Jun 2026 · 1:06:07 Sized, not only named.
“…Making the HBM DRAM, making what NVIDIA calls SoCAM, making LPDDR, these are the co- types of DRAM that are really hard to make, not consumer-grade DRAM, and they are increasingly what you need in these AI data centers.”
Gavin Baker · All-In Podcast · 27 Jun 2026 · 1:07:12 The DRAM in an AI server is a different product from the DRAM in a phone, which is what makes three suppliers a real number.
The strongest counterpoint
Made ten weeks earlier, by the person whose company buys that memory rather than making it. The three suppliers he depends on are named in the quotation above.
“…it could be, at any instant, a- a- any instance, we could be limited by the number of plumbers.”
Jensen Huang · Dwarkesh Patel · 15 Apr 2026 · 9:37 His account of which constraint binds, at its bluntest.
“If we're too far apart, uh, if one particular item, one particular component is too far, too far away, um, obviously, obviously the industry swarms it. So for example, I notice people aren't talking very much about co-was anymore.”
Jensen Huang · Dwarkesh Patel · 15 Apr 2026 · 10:05 The mechanism, with the packaging step he says people stopped talking about as the worked example.
“…for a long time, co-was was rather a specialty, and, um, uh, HBM memory was rather a specialty. But they're not specialties anymore. People now realize they're mainstream computing technology.”
Jensen Huang · Dwarkesh Patel · 15 Apr 2026 · 10:47 HBM is his own example, and it is the thing Baker ranks first ten weeks later.
“…each one of these generation, each one of these bottlenecks gets a- a great deal of attention, um- And now we're, we're prefetching the bottlenecks, uh, years in advance.”
Jensen Huang · Dwarkesh Patel · 15 Apr 2026 · 11:55 Buying ahead of the shortage rather than reacting to it.
Where the two readings actually meet
“…memory is an easy one that everyone's talked about, but I'm not gonna talk about it from a supply chain angle, I'm talking about it from a technology angle, right? Memory, um, capacity and bandwidth have been improving very slowly. The NAND cell was invented, like, 25 years ago, the DRAM cell was invented, like, 40 years ago, and there's been no major breakthrough in, in cell…”
Dylan Patel · Sequoia Capital · 30 Jun 2026 · 29:38 The physics, separated from the trade.
“…are there, like, trigger topics in SEMYs for you? You know, like, if someone's, like, which is such a meme, you think this person must be a moron.”
Shaun Maguire · Sequoia Capital · 30 Jun 2026 · 42:12 The question. Patel's first answer, ten seconds later, is the memory take.
“…if it's like, oh you, like, memory is the bottleneck. I mean, it's true, but like…”
Dylan Patel · Sequoia Capital · 30 Jun 2026 · 42:22 Granted as true, and left there. The take he says really gets him, moments later, is that AI has no ROI.
What would change this answer
Watching: whether memory moves onto the chip. Patel describes companies working on stacking the memory directly on the chip instead of beside it, a few years out. If that lands, the question stops being who can supply HBM and becomes an ordinary design problem, which is close to Huang's account of how a bottleneck ends.
“…there's, like, new innovations coming in the next few years where instead of, you know, stacking the HBM separately from the chip, you stack the memory directly on the chip and that makes your bandwidth explode. Um, and so there's interesting companies in that space and interesting POCs that companies are trying to do there.”
Dylan Patel · Sequoia Capital · 30 Jun 2026 · 30:11 Companies trying it, not shipping it. The read moves when it ships or when it visibly does not.
Hear it yourself, in order
- 9:37 How a shortage behaves once the whole industry notices it. Jensen Huang · Dwarkesh Patel · 15 Apr 2026 1 min 30 s
- 11:55 Buying ahead of the next one. Jensen Huang · Dwarkesh Patel · 15 Apr 2026 40 s
- 29:38 Why memory improves slowly, and what companies are trying instead. Dylan Patel · Sequoia Capital · 30 Jun 2026 1 min 30 s
- 1:03:35 DRAM ranked first, then sized. Gavin Baker · All-In Podcast · 27 Jun 2026 2 min 50 s
- 1:07:12 What AI-grade DRAM actually is, and why the supplier list is short. Gavin Baker · All-In Podcast · 27 Jun 2026 1 min 20 s
- 42:22 Granted as true, then left behind. Dylan Patel · Sequoia Capital · 30 Jun 2026 20 s
Six moments, 8 minutes 10 seconds end to end. Each one opens the original conversation at the second it was said.
Method & sources
Quotations are verbatim from this engine's word-level transcripts, with omissions marked and nothing reordered, so the hesitations are left in. Speaker names are resolved against a voice database and corrected by hand where it has no match. Every timestamp on this page links to the second it was said, and a check runs before publication that refuses any quotation whose words, speaker or second do not match the transcript.
How this page was made, plainly. A language model read the four transcripts, chose these moments and wrote the sentences between them. The machine check binds the quotations; it cannot bind a sentence, so every claim outside a quotation is written to point at a quotation on this page rather than to stand on its own. Three reviews of this page found four such claims that did not, and they were cut or cited before it went up.
- Dwarkesh Patel, Jensen Huang: Will Nvidia's moat persist? · 15 Apr 2026
- Bg2 Pod, The SpaceX IPO, Fable 5, AI Capex Update & Market Check w/ Gavin Baker, Andrew Fox & Clark Tang · 11 Jun 2026
- All-In Podcast, Socialists Sweep NYC, China Catches Up in Coding, AI Memory Crunch, Micron's Blowout Quarter · 27 Jun 2026
- Sequoia Capital, Why Hardware-Software Co-Design Is AI's Real 100x, Dylan Patel · 30 Jun 2026
What this is not yet. Nothing updates this read when new conversations are processed, and no email has been sent from Superlisten to anyone. The struck line above is a comparison between April and June sources, not an edit history: this is the first published read and there is nothing to correct yet. The companion question, is CUDA still what protects Nvidia, is on the front page.