New Tricks in Memory, an Old Problem in Robotics: Everyone Is Routing Around the Same Bottleneck

Contents

1. What the episode covers
2. The main points
3. 1. A new architecture gets a paper written against it two weeks after launch
4. 2. The doubts aren’t new — this time they came with numbers
5. 3. Intel’s XBM: stop building it together, build it apart and bond it back
6. 4. The number 2029 tells you more than XBM does
7. 5. A fat gross margin is a public invitation to bid
8. 6. The Cathie Wood example runs backwards
9. 7. Unitree’s prospectus is more interesting than its dancing
10. Going further
11. ”Every new-technology headline sounds impressive. How do I tell which ones are real?”
12. ”Is it too late to chase an industry earning eighty percent margins?”
13. ”Unitree went up ninefold in a day. Did I miss it again?”
14. Worth a look
15. The one thing to take away

Standing by the river envying the fish is worth less than going home to weave a net.

—— Huainanzi, Shuolin (Western Han dynasty)

Two thousand years ago that line was already making the point: what decides whether you catch anything isn’t how badly you want it, it’s the net. This episode kept bringing the line back to me. The whole AI industry is standing at the same river right now, staring at the same fish, and everyone has run home to weave a differently shaped net. Some weave fast. Some get halfway before a stranger points out the holes.

What the episode covers

Statementdog’s August 20, 2026 news roundup covers two things. First, the cluster of memory news that broke in the week right after the Flash Memory Summit: HBF challenged by an academic paper, Intel returning to the memory market with XBM and ZAM, and a well-known investor explaining why she won’t buy the memory makers. Second, where robotics actually stands, now that Unitree has gone public and spiked.

The two topics look unrelated and turn out to share a structure. On one side, everybody is jammed against the same number, so everybody is improvising. On the other, the hardware is already respectable but nobody can say clearly what it’s for. Demand too certain versus demand not yet born.

The main points

1. A new architecture gets a paper written against it two weeks after launch

HBF’s formal spec had been out for a week or two when researchers published a paper arguing its real-world performance falls far short of the claims — the gap large enough that the hosts laughed at how brutal it was. What’s interesting isn’t only the conclusion, it’s the tempo. A new spec used to take a year or two before it drew systematic criticism. This one took two weeks. The paper title was also written with unusual flair; one host joked he wished he’d known you were allowed to write them that way.

The scene itself is the signal. A lane has to be extremely crowded before people race to dismantle a rival architecture in public inside a fortnight.

2. The doubts aren’t new — this time they came with numbers

The hosts note the industry had been uneasy about HBF for a while, on physical grounds. Flash is more heat-sensitive than DRAM to begin with, and today’s systems run far hotter than yesterday’s. Enterprise SSDs have hot swap: a drive dies, you pull it and slot a new one without powering down. But if the thing is soldered onto the board, what then? As one host put it bluntly — you can’t exactly keep the power on and yank it out. Downtime is money.

The line worth copying down isn’t “HBF is bad.” It’s the question shape: when evaluating a new component, don’t only ask how fast it runs, ask what happens when it breaks. Speed is in the spec sheet. What breaking costs you never is.

3. Intel’s XBM: stop building it together, build it apart and bond it back

XBM stands for Crossbar Memory — a name one host teased for freeloading off HBM. Technically it’s stacked DRAM: stack the memory cell arrays first, fabricate the controller and base die separately on a different process, then bond them, with the bottom layer talking over the UCIe protocol.

The hosts compare it to Yangtze Memory’s Xtacking, which separated the memory cells from the peripheral circuitry precisely because flash can’t take a high-temperature process, then joined them at the end. Same idea, moved to DRAM.

They also flag the hard part: the published patent material calls for double-sided hybrid bonding interconnect, and nobody has that yield looking good yet — everyone is still trying. So it’s a clean concept sitting on brutal engineering, and it’s still at the freshly-published-patent stage.

4. The number 2029 tells you more than XBM does

Intel also has ZAM, with samples next year and commercialization in 2029; XBM comes later still. One host’s reading is the sharpest cut in the episode: the timeline alone tells you Intel does not believe the memory cycle ends before 2029, or it would never schedule a product to launch then. It plainly intends to earn the money that comes after 2030.

A historical aside the episode digs up: Intel was a memory company first, driven out in 1985 by Japanese competitors, and only then pivoted to processors and struck gold. That’s forty years ago. Intel doing memory isn’t a foray into someone else’s business — it’s going home.

5. A fat gross margin is a public invitation to bid

Why are new architectures sprouting everywhere at once? The hosts’ answer isn’t technical, it’s arithmetic: memory gross margins are running north of eighty percent. If it’s that lucrative, why on earth would you not be building something.

This is the part of the episode that travels furthest outside semiconductors. Any layer with an absurd gross margin is simultaneously evidence of two things: that layer is genuinely hard (so there’s a moat), and everyone alive is working to take it from you (so the moat is under permanent attack). Both are true at once. It isn’t a choice between them.

6. The Cathie Wood example runs backwards

The episode covers an interview in which a well-known investor says she won’t buy memory makers because HBM has gotten absurdly expensive, and she’d rather own companies whose technology reduces dependence on it — citing Groq’s SRAM approach as her example.

Both hosts found this odd. SRAM costs vastly more per unit of capacity than HBM, not marginally more. SRAM is chosen for speed, never for being cheap or dense. So “HBM is too expensive, therefore I’ll buy the SRAM route” has a premise fighting its own conclusion. And Groq’s team was subsequently acquired outright by Nvidia, which makes it a strange banner for the alternative camp.

The hosts also explain where SRAM genuinely sits in the stack: it’s fast enough to spit out several candidate answers at once, which the larger memory behind it then evaluates for which is most promising. Like guessing three or four exam questions before the test — guessing is quick, checking is what takes time. It’s a speed tier, not a capacity tier. It doesn’t replace HBM.

7. Unitree’s prospectus is more interesting than its dancing

Unitree spiked close to ninefold on debut. Reading the prospectus, the hosts found two things genuinely impressive: it is actually profitable, still profitable with subsidies stripped out; and the transition from quadrupeds to humanoids was startlingly fast, going from hundreds to thousands of humanoid units within a year. The method is very Chinese — ASP cut from around 260,000 to a bit over 100,000, and volume follows.

The same prospectus exposes the problem: not one vertical application shipping steadily. As a host asked plainly — if you put a robot dog in your house, what is it supposed to do? Play with you? Playing with you is not an industrial use case.

Going further

”Every new-technology headline sounds impressive. How do I tell which ones are real?”

Three new memory acronyms surfaced in a single week, each claiming to relieve the AI memory bottleneck, each sounding perfectly reasonable, and you don’t have a lab.

The episode demonstrates a test that needs no lab. I’d write it as three questions.

First: does it move the denominator? The structural line of the whole episode is that the ceiling for AI is still HBM capacity multiplied by bandwidth, and that equation hasn’t changed. PCIe generations, CXL, KV cache offload, new SSD controllers — all of it exists to serve the same thing, getting data into the compute fast enough. So when a new solution appears, ask whether it enlarges that denominator or merely routes around using it. Routing buys time. If the equation is untouched, the pressure comes back.

Second: what is the failure mode? The HBF dispute isn’t about benchmarks, it’s about whether you can replace the part after it dies. That’s the column the spec sheet never has and the production floor pays for first. The same question generalizes: when this breaks, do I reboot, halt the line, or scrap the batch?

Third: where’s the timeline? Something commercializing in 2029 barely touches the competitive picture for the next three years — but it does tell you what that company believes about the years after. A roadmap is a public statement of conviction, not just a schedule.

Run three questions and the equally-impressive-sounding headlines sort themselves into order.

”Is it too late to chase an industry earning eighty percent margins?”

One passage deserves to be thought through word by word: stock prices reflect expectations of the future, and expectations here have already arrived somewhere specific — not the expectation of eighty percent margins, but expectations beyond eighty, because eighty-something is now simply a reported fact. Once a number is that good and that public, it becomes easier to imagine it moving down. Which is why the volatility is enormous.

Two layers, and readers usually see only one.

The first is valuation discipline: you aren’t buying how profitable the company is, you’re buying the gap between that and what the market already believes. When the gap is zero, an excellent company still leaves nothing for you.

The second gets less airtime: high margins cannibalize themselves. Eighty percent is a tender notice nailed to the door reading “this layer prints money, come take it.” Intel returning, academics dismantling new specs, investors hunting for routes around it — all responses to the same notice. Moat and magnet are two faces of one object; you can’t count only the first.

So what do you do? The stance the episode takes is honest: these companies really did earn a lot this round, and that gives them resources to fund the next wave, but where the stock goes short term is genuinely hard to say, and volatility will be violent. Admitting there’s no answer to the second half is more useful than manufacturing one. What you can actually do is split “is the industry still growing” from “is this price worth paying” and answer them separately. The episode separates them cleanly — the host finishes the industry read, explicitly says that was the industry picture and people will now be wondering about the stock, then changes glasses. That move is worth imitating.

”Unitree went up ninefold in a day. Did I miss it again?”

The good parts are genuinely good: real profits, real volume, ruthlessly executed price cuts. But the hosts spend more time on what’s missing, and the gap is instructive.

Robotics isn’t short of hardware. China’s EV and drone industries produced an entire cheap supply chain of motors, gearing and controllers that transplants straight into robots. Unitree’s balance control is strong — in the hosts’ framing, the cerebellum works well.

What’s missing is the cerebrum, and both routes to it are unfinished.

One route is learning from vision: the robot watches camera footage and learns the motion, which is what those humanoid training clips show. That route costs time and data.

The other is simulation to reality: practice folding laundry inside a virtual factory, then load the model into a real machine. More seductive, because the training data is synthetic and you don’t fold anything ten thousand times. But the episode lands a great counterexample. You might imagine training under triple gravity produces a superhuman fighter, anime-style. In reality every step now applies triple the force and the knees fail shortly after. The host adds a more personal version: train boxing with a ten-kilogram dumbbell and your form goes strange once you put it down, because throwing a punch was never about fighting gravity. You trained the wrong thing.

Both metaphors say the same thing: the moment your simulated physics diverges from reality, the harder you train the more spectacularly the transfer fails. That’s the real reason everyone is chasing world models. Not showmanship — if the simulation isn’t faithful, the synthetic data is worthless.

And the episode names a gap almost nobody mentions: even once you can generate simulated data, who verifies it’s correct? In software we write a second program to test the first. Robot simulation needs the equivalent, and that layer is essentially empty today.

So, back to the anxiety. A ninefold move prices in expectation of an enormous market, and the expectation isn’t unreasonable — research houses put humanoids at 600,000 units by 2030, and Musk’s number is higher. But equating “this is a big market” with “this company captures it” skips exactly the part the episode dwells on: the use case isn’t settled, the cerebrum isn’t there, the verification layer isn’t built. The way to tell is simple — look at whether shipments concentrate in one steady vertical or scatter across demos and trials. The first is a business. The second is a proof of capability. Both have value; they aren’t worth the same money.

Worth a look

• Statementdog podcast, episode 545, August 20, 2026 — the source of this discussion, both the memory and robotics segments.
• Post-event coverage from the Flash Memory Summit — where this week’s memory news originated.
• WPG Holdings’ earnings call — the hosts note first-half pull-forward driven by a July price increase, yet guidance still has the second half beating the first. A good exercise in separating data from interpretation: the same call reads as slowdown to some and capacity migration to others.
• Unitree’s prospectus — revenue, margins and the quadruped/humanoid shipment mix are all in there, and it carries more information than any dancing video.
• Intel’s published XBM patent material — for anyone wanting to check engineering constraints like double-sided hybrid bonding, primary documents beat secondhand reporting.

The one thing to take away

What stuck with me most from this episode: the thing that keeps you stuck is usually a denominator nobody says out loud. HBF, XBM and SRAM were all shouting at once, and one line from the hosts — “the ceiling is still capacity and bandwidth” — lined them all up. That line is what made me notice that most of the things I complain about have a denominator like that too, and I’ve been circling it.

Here’s a small thing I tried; you’re welcome to try it tonight. Go back to the thing you complained about most this week, write down the three things you’ve done about it, and next to each put a single word — “root” or “around.” The first time I did it, all three said “around.” A little embarrassing, but it was also the first time that denominator existed anywhere on paper.

A month from now, look at the sheet again and see whether a first “root” has shown up among the new entries.

Raw Markdown

This article is an educational discussion of investment method. It is not advice to buy or sell any individual security, offers no target prices, and does not analyze any current holding. Investing carries risk; make your own decisions or consult a qualified professional.