Alex A.C.

Alex A.C. @AnalysisOp

I'm seeing a new kind of Memory Panicans, but much worse: the former Memory Bulls that have sold their Memory names recently and now are saying: "ey, the cycle! downturn is near!"🤦‍♀️ Preaching against the ones we're holding. You're pathetic and know nothing🤷 $MU SK hynix $SNDK

Damnang @damnang2

🚨🚨 There will be no memory oversupply, even beyond 2030 As you all know, HBM demand is enormous, and meeting that demand is constraining the production capacity available for conventional DRAM. HBM already requires far more wafers than conventional DDR5 to produce the same number of bits. So what happens when new 3D stacked DRAM architectures, integrating DRAM directly with logic, begin to scale around 2029 or 2030? Could that put even more pressure on conventional DRAM supply? As far as I know, the memory companies have not even finalized which DRAM process nodes they will use for these architectures. And if my understanding of the architecture is correct, 3D stacked DRAM could impose even greater constraints on bit output than the DRAM used in today’s HBM. I think many analysts still aren’t even considering the supply constraints from 3D stacked DRAM. We’ll probably only start hearing about this next year. Many companies in the industry are already collaborating with memory vendors on 3D stacked DRAM. These are things I already discussed in my articles two months ago, but I’m bringing them up again. With SK hynix posting strong results, this seems like a good time to revisit the topic 😊

dylan ツ @demian_ai

most of those trillions of tokens on OpenRouter are not fresh work for GPUs, they’re agents repeatedly rereading the same convo OpenRouter’s daily numbers for the main Deepseek models sit around 10 trillion tokens. On the biggest one (V4.1 Flash), roughly 99% of that is input, and about 95% of the input is a cache hit. so that leaves something like: - 114 billion tokens the model actually generates - 460 billion tokens it has to process from scratch - 8.9 trillion tokens that are just the model being shown context it already saw earlier Agents do this constantly. Every turn they resend the entire session so far. The Openrouter counter counts all of it. -> but the GPU part is surprisingly small The fresh work (the 460B + 114B) is what actually needs compute. Spread over a full day that works out to a few million tokens per second. A single GB300 rack (so around 72 GPUs, roughly $7 million of hardware) can handle that load. So you could serve nearly all of Openrouter’s Deepseek traffic with one of these racks. where does the real cost shows up? Those 8.9 trillion cached tokens don’t rerun the model. They come back from Deepseek’s longer lived cache (something they started talking about in 2024). The model has already squeezed the stored state down to about 7 KB per token, yet the traffic still writes roughly 4 petabytes of new cache data every day. Keep just the last hour and you’re looking at around 165 TB (several times more than the high speed memory sitting inside the entire rack) In other words, the Openrouter meter is mostly measuring agents rereading their own history. It tells you less about GPU demand than it first appears, and more about the memory and storage sitting next to those GPUs. If you’re serving agents, the number that is the most important is cache footprint, not tokens per day. GPU compute for the fresh work is real but modest. What scales with agent traffic is the cache tier, the high speed memory plus the storage that feeds it. What would make this wrong: routing and caching get good enough that almost everything stays inside the GPUs’ own memory and the storage layer becomes a rounding error. But right now, that isn’t the case.

SemiAnalysis @SemiAnalysis_

Anthropic’s S-1 shows inference revenue per megawatt nearly quadrupling in a year. “Inference revenue per megawatt went from sixteen million last year to sixty plus currently, across the entire fleet. AI model businesses scale extremely well.” “Revenue grew 1,368% year over year for the three months ended June 30, compared to compute and infrastructure expenses that grew 300% during the same period.” “How much of that is Anthropic being geniuses versus market-wide efficiency? We’re seeing GB be 50x better than Hopper.” “They have TPU, Trainium, NVIDIA GPU and AMD GPU all online. New hardware is always more performant, but you need the software to use it. Who’s got the best software? Probably the company with the top engineers and the best model.”

Illiquid @illyquid

AI model revenue from cybersecurity work could grow 10x as the cadence of response and remediation work increases. The incremental opportunity for security vendors from model revenue alone could be equal to the combined 2026 revenue of $net $panw and $crwd.

Carrio @carrioresearch

$LITE 🚨 LUMENTUM'S CEO SAYS IT IS BOOKED OUT THROUGH 2029 AND IS STILL ADDING CAPACITY: FIVE WAFER FABS NOW, AND CAPEX PLANNED FOR FOUR TO FIVE MORE YEARS OF AI DEMAND SANTA CLARA, Calif. - Lumentum ($LITE) is "essentially booked out through 2029" and is investing "as if this will continue for another four or five years," chief executive Michael Hurlston told SiliconANGLE's theCUBE at CASPA's 35th annual conference, in an interview published Oct. 27. He said the bigger constraint in AI infrastructure now is laser chips: "The indium phosphide problem is becoming more acute as we see co-packaged and near-packaged optics technologies come to bear." - Five wafer fabs now "under Lumentum's jurisdiction," and the company is weighing whether to add more - Supply sits about 70% below the demand signal - Order book essentially full through 2029 - Capex sized for four to five more years of demand, plus a packaging investment he called "underappreciated in the optics industry" - Copper reach: about 10 meters at 800G, 1 to 2 meters at 1.6T, "far less than a meter" at 3.2T Hurlston described Lumentum as "sort of the TSMC and the NVIDIA at the same time." Laser chips mix process and design, so the company has to run its own fabs and add capacity "way upstream of when the demand signal hits." He called NVIDIA, which agreed in March to invest $2 billion in Lumentum, "a great teaching customer" that started on co-packaged optics before anyone else. The rest of the industry is now taking its own route, he said, "which is generally NPO rather than the co-packaged optics." I set this interview next to his Sept. 24 keynote in Malaga. There he put the 2027 laser shortfall at 70% and said he hoped for "some level of balance" by 2029 or 2030. On Oct. 9 he told Bloomberg Television in Tokyo that its components were effectively sold out through early 2029, according to coverage of that interview. Lumentum's revenue in the quarter ended June 27 was $1.006 billion, up from $480.7 million a year earlier. The order book runs to 2029, and the balance he expects arrives in 2029 or 2030. His capex plan runs to 2030 or 2031, past both.

Jukan @jukan05

SK HYNIX: DRAM AND HBM VOLUMES ARE VIRTUALLY SOLD OUT FOR THE NEXT THREE YEARS THROUGH 2029. APPROXIMATELY 50% OF 2027–2028 CAPACITY IS COVERED BY LONG-TERM AGREEMENTS (LTAS), AND THE COMPANY EXPECTS TO MAINTAIN A SIMILAR SHARE GOING FORWARD.

Fireside Alpha @firesidealpha

Gavin Baker says Elon Musk giving every child in America a share of SpaceX is way better for America than a foundation, and more wealthy people should do what Elon, Michael Dell, and Brad Gerstner have done “And maybe some of them will do what Elon did. You know, every child in America is going to have a share of SpaceX. That’s awesome. I think that’s way better for America than setting up some foundation that, you know, a board or a bunch of bureaucrats are going to decide how to allocate the money.” “You know, just give people the money. Just let people make individual decisions. And I would encourage more companies and more wealthy people in America to do what Elon and Michael Dell and, of course, my friend Brad Gerstner have done.”

Fireside Alpha @firesidealpha

Gavin Baker says the compute market is now in contango per Silicon Data and his portfolio companies are shocked at how much more they have to pay for clusters in 2027 “You know, Silicon Data, which I do think is, my firm is an investor in them, but I do think they’re the most authoritative data source. Carmen Li, the founder, stood up at GTC and said that the compute market is now in contango. It’s not just that the spot market is tight. If you want a cluster a year from now, you have to pay a lot of money." "And then this is maybe another discontinuity between public and private markets. But just the number of companies we have in our portfolio who are like, okay, we got a cluster of 4,000 Blackwells, and we got it in May or June ‘26, and we wanna bring on another cluster of 4,000 in September 2027, and they’re shocked by how much more they have to pay in the future." "And it’s not like everybody is unaware of this. Everybody’s aware of the supply that’s coming on. But even with all of that supply coming on, it is hard for the companies in our portfolio to find compute even well into ‘27. So for, you know, all the commodity experts who are focused on backwardation and what that implied, I got news for you. We’re not in backwardation anymore.”

Fireside Alpha @firesidealpha

Elon Musk explains how to build AI satellites on the moon from lunar silicon and aluminum, and says 500 to 1,000 terawatts a year cannot be launched from Earth “Well, the lunar soil is 20% silicon or something like that. So you can mine the silicon on the moon, refine it, and create the solar cells and the radiators on the moon. So get the radiators out of aluminum. So there’s plenty of silicon and aluminum on the moon to make the cells and the radiators. The chips you could send from Earth because they’re pretty light. But maybe at some point you make them on the moon, too." "Like I said, it does seem like a sort of a video game situation where it’s difficult but not impossible to get to the next level. I don’t see any way that you could do 500 to 1,000 terawatts per year launched from Earth.”