# Bottlenecks Everywhere: NVIDIA's Call, Hot Chips, and a Ten-Year Packaging Bet > Notes after listening to the 2026-08-27 Caibaogou markets episode: NVIDIA underwriting compute with its own balance sheet, Hot Chips turning into a memory conference, and Powerchip's decade-old panel-level packaging finally landing orders. Distilled into one way to read 'capacity-constrained' versus 'demand-starved' growth. Educational only, not investment advice; no tickers recommended and no price targets. Published: 2026-08-28 Locale: en Tags: NVIDIA, panel-level packaging, HBM, AI chips, Caibaogou, industry notes ![Deep inside a data center hall at night, a freshly finished square packaging substrate glows under inspection lamps in the foreground while the same aisle recedes into rows of half-installed racks and hanging cable, darkening with distance](/covers/caibaogou-2026-08-27-548-nvidia-x-hotchips-x-foplp-cover.png) > Ten years I ground this single blade; > its frost-bright edge has yet to be tried. > Today I hold it out to you — > who here has been wronged? > > —— Jia Dao, "The Swordsman" (Tang dynasty, c. 9th century; translated by the author) ## What This Episode Covers The 2026-08-27 episode of Caibaogou's markets segment puts three seemingly separate stories side by side: NVIDIA's just-concluded earnings call, a wave of memory announcements at the Hot Chips conference, and Powerchip loudly staking its claim in fan-out panel-level packaging. What ties them together is **supply**. NVIDIA says demand will double but only commits to seventy percent revenue growth, because it cannot build fast enough. The star of Hot Chips shifted from logic to memory, because memory is the tightest square on the board. And a packaging approach Powerchip has worked on for ten years suddenly has customers — because the incumbent route got expensive enough to make switching worthwhile. My overall takeaway: every story this week is answering the same question. **When demand is obviously there, which layer does the money flow to?** What follows is my own synthesis, not a transcript, and it does not substitute for the show's own judgment. ## The Key Points **1. NVIDIA answered the "circular financing" critique head-on — by expanding the practice, not walking it back.** The worry had been that NVIDIA invests in AI companies and then sells hardware to them. This call didn't dodge it. The company laid out that it will backstop minimum compute revenue for neocloud startups and participate in securing land and power for data centers. The show's read is the right one: this **moves the customer's credit risk onto NVIDIA's own balance sheet**. Those startups were never short on demand; they were short on financing. NVIDIA fills that hole with its own credit and gets one thing in return — certainty that the hardware gets bought. Revenue splits into two lines: selling boxes, and renting compute. Whether this is brilliant or reckless rests entirely on one premise: demand has to keep climbing. As long as the capacity rents out and the investment gets recovered, it's a money machine. The moment demand turns, the same guarantee flips from asset to liability. This isn't magic — it's turning a conviction about the future into a signed contract. **2. "Which link is the bottleneck?" — "Haven't you noticed I'm out having dinner with everyone?"** The funniest and most informative moment. Asked to rank power, land, memory, and wafer capacity, the answer was that all of them are constrained, with the dinner schedule offered as evidence. That joke carries more information than a ranked list would: **no ranking means there is no single bottleneck you can attack first.** The numbers reinforce it. Demand is described as more than doubling; revenue confidence is set around seventy percent. That gap isn't conservatism, it's capacity. One easy trap to avoid: NVIDIA's fiscal year runs ahead of the calendar, so the "2028" in its commentary maps roughly to calendar 2027. **3. What analysts actually pressed on wasn't demand — it was competitive intensity in inference.** Nobody doubts that agentic AI generates enormous inference load anymore. The repeated question was how durable NVIDIA's edge is specifically in inference, given the parade of custom ASICs arriving. The show's example: OpenAI presented its in-house silicon at Hot Chips claiming it beats GB300, and got teased because GB300 is already last generation. Fair rebuttal from the hosts — you can only benchmark against what exists when you design it, and being one generation behind isn't embarrassing. **4. The hard part of an ASIC isn't the chip. It's the chain from chip to full rack.** This is the most valuable sentence in the episode. Getting silicon back is the starting line. You still have to go from die to board to rack to a working system, and each stage is a separate discipline with its own sub-problems. Hence the conclusion: nobody will run entirely on one vendor, and nobody will build entirely in-house either. The one player that looks genuinely capable end-to-end is Google — largely because much of the foundational research came out of there, and it has been quietly building for years. For scale: per-rack revenue moves roughly from $18M for GB200 to $25M for GB300 to $40M for Vera Rubin. That escalation is exactly why everyone is forced into custom silicon. **5. On the same TPU shipment forecast, the closer you sit to the supply chain, the higher you guess.** The detail I most wanted to write down. Next year's TPU unit forecasts: US brokers around six to seven million, Hong Kong around eight million, Taiwan-based analysts around nine million. One future event, three geographies, three numbers — ordered exactly by physical distance from the fabs. The show's posture toward the sensational "TPU overtakes NVIDIA by 2028" headline is worth copying: nobody knows who's right, the gap next year is still very large, and the program was strategically delayed by a year anyway. But one framing landed well — for supply chain companies, a second large customer isn't a split of the pie. It's a bigger pie. **6. Hot Chips became a memory conference.** Last year's theme was optics; this year the hall was full of memory. The reason is blunt: AI compute scales roughly with HBM capacity times bandwidth, so whoever pushes bandwidth up one notch controls the line. The competitive frontier has moved down a layer, to the logic base die underneath the stack — it needs a different process from the memory itself, and it directly sets the speed of the whole stack. Custom HBM grows out of exactly that. The more interesting branch is 3D DRAM. Startup D-Matrix stacks compute logic on top; Samsung puts logic underneath — two opposite extremes chasing the same goal: a middle point between SRAM's speed and HBM's capacity, with claimed bandwidth above current HBM and better power, at the cost of capacity. The hosts' stance is right: it sounds lovely, don't expect it tomorrow. Drilling through DRAM is genuinely hard, and 2028–2029 is the optimistic window. The big three memory makers will follow this path eventually. **7. Powerchip's panel-level packaging started in 2016 — ten years, exactly.** Its message this week amounts to: this field shouldn't be handed entirely to the display makers, we've been at it longest. Of its four approaches, the pivotal one is chip-middle, conceptually the counterpart to TSMC's CoWoS-L, embedding a silicon bridge for fine interconnect. The difference lies in where the bridge goes: one route builds it into an organic layer, while Powerchip cuts into the ABF substrate and drops the bridge in. **That shifts the entire difficulty onto the substrate rather than the layer above it.** Why now, after ten years? The show's explanation is mechanical and convincing: wafer-level packaging prices are already up thirty to forty percent, and panels are square, so the cut efficiency is better. The technology didn't get better — the alternative got expensive. Reports indicate the first line has orders and is full, with large customers in discussions. ## Going Further ### 1. The news is this good — should I be chasing it? This is what surfaces in most people's heads after a big earnings call. Great quarter, better guidance, stock pops after hours, and you start suspecting you're about to buy the emotional top. My suggestion is to shift attention from *how good it sounded* to **what the company chose to address**. The most meaningful move on this call wasn't a growth number; it was voluntarily taking on the circular-financing critique. Whatever a management team spends time dismantling is usually the variable they know the market is pricing on. What they say is one thing; what they feel compelled to say is where the actual argument lives. Treat it as a free checklist of what to watch next. The second layer: split the guidance into **demand rhetoric** and **falsifiable commitments**. "AI demand is exceptionally strong" is the former — unverifiable a year out. "Demand doubles, revenue up seventy percent" is the latter, and the thirty-point gap even tells you where to look: is power connected, is memory available, has packaging capacity come online. The first kind sets a mood; only the second can be graded. If you finish a call without a single falsifiable sentence written down, you effectively didn't listen. The third layer is the uncomfortable one: better visibility and a cheaper valuation are not the same thing. As the show put it, everyone's confidence got topped up. But confidence sits in the numerator; price sits in the denominator. The same guidance at two prices is two entirely different trades. So "should I chase" is the wrong question. The right one is: **if everything he said is true, how much of it does today's price already assume?** ### 2. A headline says X will overtake NVIDIA. Should I switch positions? This episode hands you a near-perfect teaching case, and I think it's worth more than the forecast itself: US brokers at six to seven million TPUs, Hong Kong at eight, Taiwan at nine. The lazy read is "Taiwan is too optimistic." The better read is that **each analyst is forecasting the slice they can see.** The supply chain sits in Taiwan, so Taiwanese analysts can check real lines and real pull-in behavior. Those signals are genuine, warm, and concrete. What they can't see — whether another link jams, whether the customer eventually cuts — simply isn't in frame. Visible things get amplified; invisible things get treated as zero. Nobody is being stupid; this is a systematic bias set by information position, and its direction is predictable: closer to the line, higher the number. So next time a dramatic forecast lands, don't start with "is this right." Start with **"who produced it, and where on the chain do they stand."** The same figure sourced from fab checks, from customer surveys, or from a macro model deserves entirely different haircuts. On switching positions, the show's angle is more fundamental: for a supplier that used to have one giant customer, a second and a third means a larger addressable market, not a divided one. **A fight over architecture and a fight over total volume are different fights.** If your position is a bet that one architecture wins, this headline matters enormously. If it's a bet that everything has to pass through your layer regardless, the headline is a point in your favor. Knowing which one you own is far easier — and far more reliable — than predicting the winner. ### 3. This technology has been "coming" for ten years. How do I know this time is real? Panel-level packaging is a term anyone following semis has heard for years. Every so often a wave of coverage says it's about to take off, and then nothing. This "wolf" pattern hurts retail investors both ways: too early and time grinds you down; wait for certainty and it's already priced. The episode supplies a usable set of verification signals, all mechanical facts you can check yourself: **First, the price of the alternative.** Wafer-level packaging is up thirty to forty percent. Technologies usually get adopted not because they improved, but because the incumbent got expensive. What sits in a lab is often perfectly workable and merely uneconomic — so the number to watch isn't news about the technology, it's the price increase of what it replaces. That number typically lives in someone else's earnings call. **Second, the state of the first production line.** Patents, conference papers, and letters of intent can all be generated without a single order. "The first line has orders and is full" cannot — that's a fact somebody paid for. Orders are the only signal PR can't manufacture. **Third, whether the difficulty has moved.** This is the subtle one. Embedding the bridge into the ABF substrate relocates the hardest part from the upper layer down to the substrate. The path from impossible to viable usually looks exactly like this: **the hard part keeps migrating**. Solve one, and the difficulty relocates to the next station. So when the conversation about a technology shifts from "can it be built" to "how do we cut the cavity in the substrate," that shift *is* the progress. **A problem becoming specific is the most reliable maturity signal there is.** Conversely, if a decade of debate is still about the same question of principle, it hasn't moved. **Fourth, remember that "not yet" is information too.** 3D DRAM in the same episode is the control group: lovely simulations, interesting architectures, everyone interested — and a 2028–2029 window with the through-silicon-via problem unsolved. Both are new technologies worth following, but only one currently has cash flow you can check. Giving them the same level of excitement is the most common misjudgment. ## Worth Reading - Caibaogou Podcast, episode 548, markets segment, 2026-08-27. The show notes that full call transcripts with Chinese translation live on their site — go to the source if you want the exact wording - NVIDIA's investor relations page: the deck and, more importantly, the Q&A around compute guarantees and data center deployment - The public Hot Chips agenda and presentations, with an unusually heavy memory track this year - Earnings materials from Taiwanese packaging and substrate suppliers, to cross-check the wafer-level price increases - Multiple brokers' accelerator shipment forecasts — the point isn't picking one to believe, it's lining up three or four and finding which assumption they differ on ## The One Thing to Take Away Of everything in this episode, one idea is worth keeping for life: **"Able but unwanted" and "wanted but unable" look identical from outside — growth came in below full marks — and they mean opposite things.** NVIDIA said demand doubles and committed to seventy percent. Read only the second half and it's a shortfall. Read the first half and it's growth pinned by capacity. Same numbers, opposite conclusions, and the only difference is whether you asked one more question: **is it can't, or is it won't?** This distinction is even more useful outside markets. A team slipping schedule, a child's grades dropping, a relationship going cold — all you ever observe is that a number got smaller, and the reflex is to jump to "not trying hard enough" or "doesn't care anymore." The truth is often on the other side: they want to, and one square is jammed. And if you keep prescribing for "won't," that square never loosens. **A practice you can do today**: take a sheet of paper and write down three things you didn't finish this week. Next to each, write exactly one word — "can't" or "won't." No third option, and you may not write both. One of them will make you hesitate for a long time. That one is the point. Force yourself to pick a side and you'll notice your past remedies were all aimed at the wrong square: the things genuinely stuck on "can't" got pep talks, and the things stuck on "won't" got new tools, new methods, and new schedules. Then act on only the square you picked. If it's "can't," go find the bottleneck — go have dinner with everyone, the way NVIDIA does. If it's "won't," stop optimizing the workflow and start by admitting you don't want to do it.