Nvidia’s Jensen Huang just claimed 6 million GPUs shipped this year alone, fueling a $500 billion order backlog for Blackwell and Rubin chips through 2026. Sounds like the AI gold rush is hitting escape velocity, right? Wall Street high-fives, stock pops 5%, and suddenly everyone’s toasting to infinite scaling laws till they start digging into the numbers, as I did in “NVIDIA EARNINGS: MORE QUESTIONS THAN ANSWERS ON THE STATE OF THE AI BUBBLE“, and something does not add up. As someone who’s spent way too many late nights poring over earnings transcripts and supply chain tea leaves, I can’t shake this nagging itch: Is this 6 million GPUs shipped number for real, or is it the kind of “pro forma” magic that makes Enron blush?
In this deep dive, I will do my best to reverse-engineer those 6 million units against the cold, hard reality of TSMC’s wafer counts, Taiwan’s export tallies, and Nvidia’s own revenue math. Spoiler: The pieces fit… mostly. But the seams? They’re starting to fray in ways that scream “bubble alert.” Buckle up: we’re talking implied ASPs that’d make a luxury yacht salesman jealous, growth rates that lap the field like it’s standing still, and a value chain where Taiwan’s getting crumbs while Nvidia feasts. By the end, you’ll be asking the same question I am: How much longer can this house of cards defy gravity?
“We’ve shipped 6 million GPUs over the last four quarters,” Jensen Huang claimed to make Wall Street analysts rejoice (“Nvidia ships 6M Blackwell GPUs, expects $500B in sales“). That’s Q4 FY25 through Q3 FY26—November ’24 to October ’25—for those unfamiliar with Nvidia’s accounting calendar. Primarily Blackwell beasts, with some Hopper holdovers to pad the stats. And yeah, it’s tied to that monster $500 billion backlog, which sounds epic until you remember: Backlogs are promises, and by now investors should have learned promises don’t bode very well when they aren’t backed up by guarantees like in the Oracle case (“Oracle is already underwater on its ‘astonishing’ $300bn OpenAI deal“).
But let’s play along. Nvidia’s Data Center segment—the AI cash cow that’s 88% of their total revenue—racked up $167 billion over those four quarters. Strip out the networking fluff and Grace CPU sideshow, and you’re left with ~$137 billion in compute sales. Now, divvy that by 6 million units, and boom: an effective ASP of just $22,833 per GPU. Wait, what? That’s below street estimates for bare H100s ($25-35k) and laughably low for bundled Blackwell systems (think $40-50k effective when they’re slinging 8-GPU HGX racks at $400k a pop). At a more realistic blended $42k ASP (70% bundled, per the analyst herd), those revenues imply only ~3.26 million units shipped. So where’s the other 2.74 million hiding? Deferred revenue from software licenses? OEM inventory dumps not yet booked? Or creative counting of “shipped” as “promised”?
Here’s the quarterly breakdown, because numbers don’t lie:
Jensen Huang’s 6M GPUs shipped claim feels like it’s counting every die that left TSMC’s fab door, not every finished, revenue-generating box that hit a customer’s dock. Coincidence? Or convenient?
Here is TSMC’s side of the story: wafer wars and the missing millions. TSMC fabs ~90% of Nvidia’s advanced GPUs—those 4/5nm wonders that power the Blackwell hype train. Their High Performance Computing (HPC) segment? That’s code for “Nvidia’s playground,” with the chip giant sucking up 70-80% of it. Over the same four quarters, TSMC’s HPC hauled in ~$64.5 billion, implying $45-52 billion straight to Nvidia’s coffers if you buy the attribution.
At $8-12k per GPU in TSMC costs (wafer pricing + yields; think $20k per advanced wafer yielding 100-150 dies at 80% good stuff), 6 million units should’ve burned through $48-72 billion in foundry fees. It fits—barely—within that $64 billion bucket, leaving scraps for AMD’s MI300 or Broadcom’s products. QoQ growth? Nvidia’s “shipments” are up ~14% on average; TSMC’s HPC clocks in at +10-12%. But here’s the rub: TSMC’s whispering about 90%+ utilization and U.S. fab ramps that won’t goose output till ’26. If Nvidia’s really cranking 1.5M GPUs per quarter (their 6M average), that’s ~225,000 wafers quarterly just for Nvidia. TSMC’s total advanced-node output? Peaking at 1.5-2M wafers/year across all customers. Do the math: Nvidia alone would eat 40-50% of it, hence it kind of makes sense. Nvidia claims they’re “sold out”, or is it that the numbers are getting… stretched?
Quarterly HPC teardown for the skeptics:
See that total? 4.84M implied GPUs—cozy with revenue math, but still short of 6M if you factor in yields and testing overhead.
Question for Jensen: If demand’s apparently infinite, why isn’t TSMC’s top line exploding harder?
Furthermore, suppose such an astounding demand for GPUs is real. Why is TSMC not growing CAPEX exponentially to fulfill it like everyone else is doing, frustrating the whole AI bandwagon in the process?
Let’s zoom out to the TSMC mothership: Taiwan’s IC exports. Over four quarters, ~$207 billion flowed out, up 28% YoY—respectable, sure, but yawn-inducing next to Nvidia’s 84% revenue sprint. AI’s the culprit, driving 30%+ surges in Q3 alone. At $10-15k export value per GPU die, that’s capacity for 13-20M units shipped globally. Plenty of room for Nvidia’s 6M… on paper. But peel the onion: Exports capture ~20-25% of Nvidia’s end-game ASP—the raw silicon slice. The rest? Assembly in Malaysia, software voodoo in Santa Clara, and 75% gross margins that turn $10k wafers into $40k+ systems. Taiwan’s GDP gets a 5% AI jolt; Nvidia’s market cap balloons to $5 trillion. Fair? Maybe. Sustainable? Exports grew 28%; Nvidia’s revenues grew 84%. That’s not alignment—that’s a value-chain heist. Backlog meets reality: fits like a glove… until it doesn’t.
Bundling, bare-chip dumps, and software deferrals paper over the ASP gaps; supply constraints might explain the “sold out” sob stories. But let’s not kid ourselves: This is AI bubble fuel, pure and simple. Growth lapping suppliers by 2-3x? Backlogs bigger than some countries’ GDPs? It’s the dot-com playbook, 2.0—except with more electrons and fewer pets.com. So, what’s next? Blackwell ramps in ’26 could validate the fairy tale, or delays (cough, production hiccups) could pop the myth. Rubin? Vaporware until proven. And that $500 billion backlog? Show me the cash, not the IOUs. To be fair, investors are sleeping with one eye open.
SYNNAX PROMO
Get exclusive intelligence on stocks you can’t find anywhere else – insights that could give you a serious edge. Claim your free Synnax trial month now:https://synnax.app/PROMO/JUSTDARIO