Tess Hollis: Felix, hey — I want to ask you something before we even start. If I told you the single most-cited number in AI right now came from one analyst's conversations with buyers, would that change how you've been thinking about the chip supply story?
Felix Ortiz: I mean — yeah, probably? Which number are we talking about?
Tess Hollis: Twelve-to-one. Nvidia demand over supply. Dan Ives, Wedbush, CNBC, July 27th. Not an Nvidia disclosure. Not from TSMC. Channel checks.
Felix Ortiz: Okay, that's — huh. I've been repeating that number. I think everyone has.
Tess Hollis: And he then revised it to fifteen-to-one, which means the number is not fixed. It's a range someone is estimating from conversations. And the market treated the revision like a data update.
Felix Ortiz: Right, but — and I'm genuinely asking, not defending it — is it wrong? Because I read that Tae Kim checked thirteen different providers on August 3rd for a single B200 or B200s node. Nothing. Zero availability across thirteen providers.
Tess Hollis: Huh — no, that's exactly the right thing to bring in. Tae Kim's ground check is real data. The scarcity is real. Jensen Huang called Blackwell demand 'insane,' said customers are 'emotional' and 'tense.' That's not marketing copy, that's a CEO describing procurement pressure in unusually raw language.
Felix Ortiz: So the shortage exists — we're just not sure twelve-to-one is the right way to measure it.
Tess Hollis: What I want to know is whether twelve-to-one is measuring anything, or whether it's the story the moment needed badly enough that the market just... kept it.
Felix Ortiz: Okay, but here's the part that actually breaks my brain — even if the ratio is wrong, it might be wrong *because* of itself. Like, think about a concert where only a hundred tickets exist. Every fan buys ten hoping to get two. The box office suddenly sees demand for a thousand tickets and announces a ten-to-one shortage. That's not a measurement of how many people want to go. That's a measurement of how scared people are that they won't.
Tess Hollis: That's the whole thing. That's the feedback loop.
Felix Ortiz: Right — so Microsoft, Amazon, Google, Meta, collectively they're committing over two hundred billion dollars to AI infrastructure in 2026. And when chips are rationed, rational procurement means you order five times what you need hoping to receive two. So Dan Ives does his channel checks, he's hearing enormous order volumes, and — wait, no — he's hearing *order volumes*, not deployment volumes. Those are completely different things.
Tess Hollis: And Jensen Huang projected a trillion dollars in purchase orders for Blackwell and Vera Rubin through 2027. A trillion. That number is orders — not chips in racks running workloads.
Felix Ortiz: Which — okay, that number is staggering, and I want to sit with it for a second. Nvidia's own Q1 2026 revenue was eighty-two billion. Q2 guidance is ninety-one billion. So the financials are real. The money is moving. But the ratio —
Tess Hollis: The ratio is measuring fear of allocation, not end-use pull.
Felix Ortiz: Exactly — and that's the thing that nobody wants to say out loud because it sounds like you're calling the shortage fake. The shortage is real. Tae Kim found zero B200 availability across thirteen providers. But the twelve-to-one number? That might be real scarcity amplified by rational over-ordering, which then becomes the evidence for the scarcity, which justifies more over-ordering.
Tess Hollis: So the question isn't whether the shortage is real. It's whether the ratio is measuring the shortage or measuring the response to the shortage.
Felix Ortiz: And if you're measuring the response, then the ratio doesn't collapse even if hyperscalers get rational — because the underlying constraint isn't behavior. It's physical.
Tess Hollis: That's exactly where the hot take survives. Because the bottleneck isn't Nvidia's chip design. It's not wafer yield at TSMC. It's CoWoS — the packaging step where HBM3E memory stacks get bonded directly to the GPU die. TSMC said on its July 17th earnings call that CoWoS demand outstrips capacity through at least mid-2027. Their newest facility, AP8, fully booked through Q1 2027.
Felix Ortiz: Wait — so it's not the chip, it's how you wrap the chip.
Tess Hollis: And the memory attached to it. SK Hynix, Samsung, Micron — HBM3E booked solid twelve to fifteen months out. Each B200 needs eight stacks. Each GB200 superchip needs sixteen. So imagine a CoreWeave purchasing manager, call it a Saturday, pulling the latest allocation spreadsheet — she's not waiting on Nvidia's designers. She's waiting on a lamination process at a factory in Hsinchu.
Felix Ortiz: Sixteen stacks per GB200. That's — okay, I hadn't actually broken that number down before. That's not a chip shortage, that's a memory math problem.
Tess Hollis: And Nvidia has locked up roughly sixty percent of TSMC's entire 2026 CoWoS allocation. So even if AMD has a competing design, they're fighting over forty percent of a constrained line. The scarcity is structural.
Felix Ortiz: Okay — and TSMC announced they're outsourcing some CoWoS to ASE and SPIL to relieve it, right? I saw that. So that's the fix?
Tess Hollis: TSMC shares rose three percent pre-market on August 4th on that news. Three percent. The market is pricing that as 'eventually, maybe' — not relief. ASE and SPIL are third-party partners with unknown timelines on a process TSMC has spent years perfecting. You're introducing fresh quality risk at the exact layer that determines whether your billion-dollar GPU cluster works.
Felix Ortiz: Three percent is not a 'problem solved' reaction. That's a 'we noted this information' reaction.
Tess Hollis: And here's what nobody's saying yet — Vera Rubin hit full production July 21st, shipping to OpenAI, CoreWeave, Google Cloud, Azure, Meta, Dell. While Blackwell is still at zero availability across thirteen checked providers. Nvidia just stacked a second generation into the same choked packaging line. That part — two simultaneous ramps, one constraint — is where this gets worse, and we'll get there.
Felix Ortiz: And that's actually the part that breaks the hot take open — because Vera Rubin shipping on schedule sounds like good news, right? GTC Taipei, May 31st, announced ramping with Taiwan's top server makers, then July 21st confirmed full production. The roadmap executed. But wait — it executed into the same CoWoS line. Blackwell didn't clear before Vera Rubin showed up. They're now both in the queue.
Tess Hollis: That's the surgical correction. The schedule working proves the chip design and yield are fine. What it doesn't do is add a single square centimeter of CoWoS capacity.
Felix Ortiz: Two generations. One ceiling.
Tess Hollis: And HBM3E is the same wall — booked through the same 12-to-15-month window. Vera Rubin doesn't use less of it.
Felix Ortiz: Okay, so here's my concrete version of this — imagine a Google procurement team gets an allocation notice. Five hundred Blackwell units available next quarter. Twenty percent above list price. They need five thousand. Do they walk away? No — they take the five hundred. And then, quietly, they accelerate the TPU roadmap. Same week. Because Google's building their own silicon anyway, Amazon has Trainium, Microsoft has Maia — they're all signing long-term Nvidia agreements with one hand and writing internal chip team checks with the other.
Tess Hollis: Wait — that's the real hedge hiding inside the two-hundred-billion capex number. The public commitment to Nvidia is real. The private TPU acceleration is also real. Those aren't contradictory.
Felix Ortiz: They're actually the same rational move. You lock in allocation *and* you build the exit. Because if CoWoS is the ceiling through mid-2027 and Vera Rubin inherits it — actually, no — the next chip doesn't solve the constraint. It adds to it. So the calibrated take isn't that Nvidia's dominance breaks. It's that the packaging wall doesn't move just because the chip got faster.
Tess Hollis: New product generations are not ladders over it. They're more people in the same elevator. That's the thing that holds up once the hype is stripped.
Felix Ortiz: And that's — yeah, that's my take too. But fine, I'll give you this: the twelve-to-one number might be more sermon than science. Maybe it's fifteen-to-one, maybe it's eight, maybe it's genuinely unmeasurable because the orders are inflated. But Nvidia posted eighty-two billion in Q1, guided ninety-one billion for Q2, and the CoWoS line is physically booked through mid-2027 regardless of what ratio you believe. When the sermon generates those numbers, maybe that's not a bug in the story.
Tess Hollis: Yeah. Fair. The sermon is doing real work. I'll half-concede that.
Felix Ortiz: The part I keep staring at, though — and I think this is actually where it resolves — it's not when supply catches up to demand. It's when some CFO at Microsoft or Google or Meta is sitting with two hundred billion in committed capex through 2027, deployment economics that still haven't landed cleanly, and a CoWoS line that won't clear until mid-2027 at best. And they look at the gap between their conviction and their calendar and they decide that's no longer someone else's problem to solve.
Tess Hollis: That's the inflection point. Not a supply event — a patience event. Good conversation.
