Eliza Ward: Brian, hey — I've been staring at this one number all morning and I genuinely cannot tell if it's the most important data point in tech right now or complete vibes.
Brian Reed: The Ives number.
Eliza Ward: The Ives number. Dan Ives, Wedbush Securities, goes on CNBC on July 27th and says Nvidia's AI chip demand is outpacing supply at twelve-to-one. Twelve customers for every chip Nvidia can actually ship.
Brian Reed: And that's — hang on, that's his estimate, right? Nvidia didn't put that in a filing.
Eliza Ward: Correct. That is Ives's own analytical estimate — not a metric Nvidia has disclosed or verified. Which is exactly why what happened next is so interesting. By August 2nd — six days later — he revised it. Upward. To fifteen-to-one.
Brian Reed: Wait — fifteen? In six days?
Eliza Ward: Twenty-five percent increase in a week. And that's — that's the episode. Not whether the ratio is real. What does it mean that the ratio can move that fast? If it jumps twenty-five percent in six days, what exactly is it measuring?
Brian Reed: So the question isn't twelve or fifteen — it's whether this number is tracking actual chip orders or tracking the temperature of the narrative.
Eliza Ward: That's the thing — and the backlog number has the same problem. Jensen Huang said in March and again in June 2026 that Nvidia is sitting on roughly $500 billion in what he called 'very high confidence demand and purchase orders' for Blackwell and Rubin chips. Five hundred billion dollars.
Brian Reed: Okay, but — 'Very high confidence demand' is doing a lot of work in that sentence. That's not the same as a signed contract.
Eliza Ward: It is not. Those are different things.
Brian Reed: It's like — okay, think about a restaurant reservation list. Lots of people say they're coming. Some of them show up, some cancel, some just don't come. The restaurant didn't actually sell those meals. The reservation is real, the revenue isn't — not yet. And we don't know how many of Huang's $500 billion are reservations versus, like, someone's already paid and there's a penalty for walking out.
Eliza Ward: Right — and Nvidia's own financials give you some grounding here. Q1 2026 revenue came in at $82 billion. Q2 guidance is $91 billion. That's real, audited money. But $500 billion? That came from Huang's public statements, not from anything in the financials.
Brian Reed: So think about the procurement lead at a mid-size AI lab who placed a Blackwell order six months ago. Her commitment is counted in that $500 billion figure — but she still doesn't have a firm ship date. Her order is a reservation. It's in the pile. It moves the number. It hasn't moved a chip.
Eliza Ward: And we don't know — actually, wait — we don't know how many of those orders would evaporate if she could get comparable silicon from AMD six months sooner. That's the durability question the $500 billion figure just doesn't answer.
Brian Reed: Which is why the stock matters. Nvidia is up four percent year-to-date. Four. With $82 billion in quarterly revenue. The market is looking at the same backlog number and pricing in something the headlines aren't saying out loud.
Eliza Ward: The market's read might be: we believe some of it converts, not all of it. 'High-confidence demand' is a vision of revenue, not a guarantee of it.
Brian Reed: But that's — hang on, that's still assuming the constraint is on Nvidia's end. Like they're not building fast enough, not designing fast enough. And I don't think that's where the ceiling actually is.
Eliza Ward: It's not. That's the part that's actually been driving me crazy. The bottleneck moved — and almost nobody's talking about it. It's not the chip design. It's the packaging.
Brian Reed: The packaging.
Eliza Ward: CoWoS — Chip-on-Wafer-on-Substrate. It's the assembly process TSMC uses to actually put Nvidia's high-end accelerators together. And you cannot — I mean, this is the critical piece — you cannot rapidly scale it. It's not like spinning up a new fab line in six months. Nvidia has reportedly locked up around 60% of TSMC's entire 2026 CoWoS capacity. Sixty percent. Which sounds dominant until you realize that means the other 40% is gone and Nvidia still might not have enough.
Brian Reed: Wait — so they won the capacity race and they're still constrained?
Eliza Ward: Because 60% of a fixed ceiling is still a ceiling. And then layer the memory on top of that. SK Hynix, Samsung, Micron — the HBM suppliers — all booked solid. Twelve to fifteen months out, no meaningful relief before late 2026. So picture an engineer at a robotics startup who finally gets a confirmed Blackwell slot — she still can't ship her system because the HBM that goes inside that chip isn't coming.
Brian Reed: So Ives is actually right that supply is severely constrained — that part of the hot take holds. He's just pointing at Nvidia when the actual wall is TSMC and three memory companies.
Eliza Ward: That's the partial win, yeah. Unmet demand is real. The 12-to-1 direction is probably real. But the constraint doesn't live inside Nvidia's design decisions — it lives upstream with suppliers Nvidia does not own and cannot accelerate on its own timeline.
Brian Reed: And that gap — between what Huang's $500 billion assumes and what SK Hynix can actually deliver by end of 2026 — that's probably what the stock's already trying to tell us. Which, the four percent thing, that's — we should actually get into that, because I think it rewrites this whole story.
Eliza Ward: Yeah, and that four percent is the verdict — like, the Street read the same Jensen Huang disclosures, the same $82 billion Q1, the $91 billion Q2 guide, and landed on: we already knew. The boom was priced in before the boom printed.
Brian Reed: Which means the question isn't whether Nvidia is dominant — it's whether there's anything left in the story the market hasn't already absorbed. And I think the answer is actually yes, but it's coming from somewhere weird.
Eliza Ward: The consumer side.
Brian Reed: The consumer side. Because the memory squeeze — the SK Hynix, Samsung, Micron situation — that's not just an AI infrastructure problem. Advanced memory component prices are already surging, and that ripples into gadgets. Like, your next phone or laptop gets more expensive because the HBM supply chain is locked up serving Blackwell orders.
Eliza Ward: Wait — that one actually surprised me. The AI chip shortage showing up in consumer electronics pricing. That's a second-order effect I don't think Ives was pointing at on CNBC.
Brian Reed: He wasn't. And then there's the Physical AI piece — humanoid robots, autonomous vehicles. Ives said on July 27th that Physical AI hasn't even started to play out yet. That's either the bull case getting dramatically larger, or — and I think this is the honest read — it's demand that's still theoretical. No firm orders from those sectors are in that $500 billion figure in any meaningful way.
Eliza Ward: So the calibrated version is — actually, let me just say it plainly. The constraint is structural and real. Ives's direction is right. But the 15-to-1 ratio is sentiment, not data. The $500 billion is a vision of revenue that depends on TSMC packaging and three memory companies hitting timelines they haven't hit yet. And the market's four percent gain isn't confusion — it's the Street saying we priced the good news already, now show us the deliveries.
Brian Reed: And Goldman Sachs flagged that Nvidia's GTC event gave clearer demand visibility — but clearer isn't the same as certain. Visibility into a constraint isn't relief from it.
Eliza Ward: The thing to watch is late 2026 — when the memory relief either shows up or doesn't. That's when the backlog either converts or the ratio starts moving the other direction.
Brian Reed: The date. That's actually — when I strip everything else away, that's the only number that settles this. Not the ratio, not the backlog figure. When does TSMC unlock more CoWoS capacity. That date either validates the $500 billion or starts chipping away at it.
Eliza Ward: Yeah. And fine — maybe 15-to-1 is a vibe, not a metric. I'll half-concede that. But the vibe is pointing at something structurally real. The packaging constraint, the HBM timeline — Ives didn't invent those. He just dressed them up in a ratio that revised 25% in six days, which, I mean, that's a sentiment thermometer, not a balance sheet.
Brian Reed: So we're landing on: demand real, ratio approximate, supply ceiling not Nvidia's to fix. That's — honestly, that's an uneasy place to sit with $500 billion on the table.
Eliza Ward: Uneasy is accurate. Late 2026 either arrives or it doesn't.
Brian Reed: Good one to dig into. Thanks for walking through all of it.
