Malcolm Reeves: Cole — before we get into anything, I need to ask: have you been following the Jensen Huang Bloomberg interview closely, or are you coming in fresh?
Cole Brennan: I've been following it, yeah, and honestly it kind of broke my brain a little — in a good way. What's your read going in?
Malcolm Reeves: My read is that what Huang said at Dell Technologies World 2026 is one of the more honest things a CEO has said publicly in years. He told Bloomberg the whole industry supply chain — wafers, packaging, silicon photonics — is in short supply. And that these constraints will persist for several years. That is not a hedged answer.
Cole Brennan: No, it's not. And the thing that hit me — like, the specific thing — is that the binding constraint he keeps naming isn't a competitor, it's not regulation, it's... hard hats. Skilled construction labor. Bloomberg reported in July 2026 that data center builds are stalling because there aren't enough specialized construction workers. That's wild to me.
Malcolm Reeves: But here is the part that anchors it structurally — SK Hynix, the major HBM manufacturer, its CEO has forecast that 2027 will be the worst-ever memory shortage in the industry's history. Beyond 2030, demand still outstrips supply. You cannot solve that by hiring more construction workers.
Cole Brennan: Huh. So even if you built every data center you'd ever want — you still can't fill it with the memory you need.
Malcolm Reeves: That's the shape of it. And NVIDIA's Vera Rubin NVL72 is where it gets concrete — they are trimming memory capacity on that rack because HBM4 prices, per GF Securities, could consume twenty-nine percent of the whole rack's cost. Design decisions are bending to physics.
Cole Brennan: So the question driving today — is Huang's 'doubling each year' ceiling an honest physics report, or is it something else? Because I'm not fully sure those are the same thing yet.
Malcolm Reeves: But here's where I'd pump the brakes on the whole framing — think of it like a gas station owner who filled his own tanks the night before, then went on the news the next morning to warn everyone about a regional fuel shortage. The warning is completely true. And he is not hurting the same way his customers are.
Cole Brennan: Wait — you're saying Huang's warning is real AND self-serving at the same time?
Malcolm Reeves: NVIDIA signed long-term memory supply agreements ahead of the shortage. So when Huang goes on Bloomberg at Dell Technologies World and says the whole industry is constrained — that is accurate. But NVIDIA already hedged itself. The companies scrambling are everyone else.
Cole Brennan: Huh. So the 'doubling each year' ceiling — I mean, is that actually NVIDIA's ceiling, or is it everyone else's ceiling that NVIDIA is describing from a pretty comfortable seat?
Malcolm Reeves: That's exactly the question to hold. And then Meta — Meta announced its own in-house AI chip going into production September 2026, targeting doubled computing capacity. That's not a contingency plan. That's vertical integration explicitly designed to route around the shortage.
Cole Brennan: Wait, September 2026 — that's not future tense anymore.
Malcolm Reeves: No, it's now. And the pattern — NVIDIA locks in supply, Meta vertically integrates — those are the moves of companies that heard the Huang warning early and acted. The 'doubling each year' framing is both a genuine structural ceiling and, you see, a justification for why NVIDIA's supply agreements command premium pricing. Both things are true.
Cole Brennan: So we should hold the Huang warning with — I mean, one eyebrow up. Not dismissing it, because SK Hynix's CEO said the same thing independently. But the constraint hits Micron's customers, hits smaller hyperscalers, differently than it hits NVIDIA.
Malcolm Reeves: The shortage is real. Who suffers it equally — that's the part the Bloomberg headline doesn't carry.
Cole Brennan: But that asymmetry — okay, that's actually what makes the 29% number land so hard for me. GF Securities ran the math: HBM4 alone could eat twenty-nine percent of every Vera Rubin NVL72 rack's total cost. And NVIDIA's response wasn't 'we'll absorb it' — they trimmed the memory capacity on the Vera Rubin itself. The flagship product shipped with less. That's not a pricing footnote. That's the constraint bending the hardware.
Malcolm Reeves: And that's where the shortage stops being rhetorical.
Cole Brennan: Right — because picture a Google infrastructure team. It's September 2026. Land's approved, the power grid connection is signed off, construction crews are actually on-site. Everything is green. And then they find out HBM4 orders are sitting on an 18-month backlog. The whole facility just... waits. A billion dollars of concrete doing nothing.
Malcolm Reeves: And SK Hynix's CEO is saying 2027 is the floor of that problem — worst-ever shortage — and demand doesn't catch supply until past 2030. That's not a quarterly guidance hedge.
Cole Brennan: Micron's leadership flagged the same thing independently — so this isn't one manufacturer talking their own book. Two major memory producers saying the same thing at roughly the same moment. That's... I mean, that's hard to wave away.
Malcolm Reeves: And Michael Dell said it too — co-spoke with Huang at Dell Technologies World 2026, identified memory as the primary bottleneck. Three voices, same constraint.
Cole Brennan: Wait — Dell said memory specifically? Not the construction labor thing?
Malcolm Reeves: Memory. Primary. His word, essentially. And you see, that matters — because Dell is on the customer side of this. He's not a memory manufacturer with an interest in scarcity pricing. That's a hardware integrator saying the component he needs most is the one he cannot get.
Cole Brennan: So HBM is — actually, I want to be precise here, because I don't think we've said what it actually is. High Bandwidth Memory is stacked directly on or near the AI accelerator chip itself. It's not regular server RAM sitting across a bus. It's integrated, it's specialized, and you cannot substitute another memory type for it. Which means the shortage has no workaround at the component level.
Malcolm Reeves: No substitution. And that's the kernel Huang is actually right about — though whether his ceiling is *everyone's* ceiling or just a very convenient frame, that's the part we still need to crack open.
Cole Brennan: And that's the part Huang actually said to Mike Allen — told Axios he doesn't see this as a hype cycle. He said the investment wave reflects a fundamental shift in computing infrastructure. Which, I mean — if you believe that, the 'doubling each year' framing isn't a ceiling he's worried about. It's a floor he's frustrated by.
Malcolm Reeves: Now that's the distinction worth landing. A ceiling is a warning. A floor is a complaint.
Cole Brennan: Right — and he projected the semiconductor industry may need to expand ten times its current size over the next decade. Ten times. That's not supply-chain hedging. That's someone who genuinely believes demand is sitting there waiting.
Malcolm Reeves: Which is where the permitting problem becomes almost darkly funny. You've got this projected tenfold expansion, and data center projects are sitting in multi-year regulatory waits across multiple jurisdictions. The ambition and the approval process are not on speaking terms.
Cole Brennan: Construction workers AND permits. Yeah.
Malcolm Reeves: So here's the calibrated claim — and I think this is what actually holds once the hype is stripped. The supply constraint is real and it's genuinely multi-front. HBM, construction labor, permits, power — the Vera Rubin NVL72 stress-tests every single one simultaneously, in one rack. That is not a rhetorical convenience. But the ceiling Huang is naming is not equally distributed. NVIDIA is the most credible voice on this constraint and, you see, the most insulated from it.
Cole Brennan: Most credible and most conflicted. Both.
Malcolm Reeves: Both. The shortage is documented — SK Hynix independently, Micron independently, Michael Dell from the customer side. That's the real signal. Huang's framing of the ceiling, though — that's where you keep one eyebrow up.
Cole Brennan: So the verdict: believe the constraint, interrogate who's delivering it. That's — yeah, I can stand on that.
Malcolm Reeves: The guy who already bought all the gas is telling us there's a fuel shortage. He's right. And he's also fine.
Cole Brennan: That's — yeah. That's actually where we started, isn't it. And we've spent the whole time confirming both halves of it. The constraint is real — SK Hynix, Micron, Michael Dell all saying the same thing from different seats. And NVIDIA, I mean, they trimmed the Vera Rubin's memory not because the shortage wasn't real but because... it was. So the warning and the insulation coexist. That's not a contradiction. It's just — it's a complicated position to be in.
Malcolm Reeves: The open question now is whether new fabs, power infrastructure, construction capacity — whether all of that scales fast enough before sustained premium pricing convinces hyperscalers the capex returns don't justify it. That's the bet the whole industry is making, and nobody has a clean answer.
Cole Brennan: Good talk. Genuinely.
