P Equity says the triple-digit cost increases on Vera Rubin are ABF substrate, PCB, and memory, and the substrate shortage lasts past 2028 "If you look at the bill of materials for Vera Rubin, the components that have seen triple-digit increases have been ABF substrate, PCB, memory. So I think ABF substrate is something that is important to look at." "Substrate capacity is tight. It's likely that ABF substrate supply shortage lasts beyond 2028. I think some people have said 2030." "And even if you read some of the earnings calls from companies like Dell, HP, Nvidia, Broadcom that reported just like a few weeks ago, in their constraints, they mentioned DRAM, NAND, ABF substrate, wafers." "So all of them are in line with substrate being an issue." "I think there's also a case that indium phosphide is short in supply, because you see a lot of companies like AXT Incorporated, Sumitomo Electric, JX Advanced Metals expand the supply of indium phosphide that is needed for lasers and optics." "But yeah, I think PCBs, ABF substrate are two areas where there's definitely a supply crunch, and you can see that in the bill of materials."

Elon Musk reveals Grok Bot is growing ~100% a month, and says xAI ships a new model every month or two and expects to catch the frontier next year "Well, we release a new model roughly every month or two. And it's called Grok 4.7, a solid workhorse of a model. It's not as good as, say, Opus 5.5 that just got released, but also we've only been around for about three years doing AI, whereas Anthropic's been around for six years." "So I would expect that we probably catch up to Frontier sometime next year, most likely." "In the meantime, I think Grok is very useful for many things. And Grok Bot is kind of like a personal digital assistant, has been very popular and is growing about 100% a month. So it's growing very rapidly."

$NBIS CRO Marc Boroditsky reveals Nebius has four customers ready to sign for every GPU and is turning away 75% of demand, and says there's another 18 months of reasonable clarity "If you look at our pipeline today, for every GPU, we have a customer ready to sign. We have four customers ready to sign for every GPU. You can just think about that for a moment. We're saying no 75% of the time. That's today." "Now, does that continue into the future? If you look at the plans that customers are sharing with us and are looking for commitments with us, it actually moves well into next year. And their willingness right now to sign up for tens of thousands of GPUs in the middle of next year is very high." "And there's a sort of growing level of fear that if you don't actually secure your future, you're not gonna have the compute that you need in order to be able to succeed." "So I would say that there's probably reasonable clarity for another 18 months at the current rates. Beyond that, we're trying to figure it out by actually utilizing new pricing mechanisms. We introduced an auction about a month ago. This week, we also introduced spot price."

VAST Data's Renen Hallak says one AI cloud asked for 500 petabytes, then came back weeks later wanting 2 exabytes more "I think the demand is there today. I'm not sure it gives me comfort and sometimes it scares me as to how much demand is there and the acceleration in demand, because every quarter our customers are coming back to us and telling us they need a lot more than they thought they did." "We had a customer, one of these AI clouds, one of the smaller ones, come to us a quarter ago, and we told them we need to plan ahead for the next 3 years because there's supply chain implications and you need to be aware of this." "And they said we're probably going to need about 500 petabytes over the next 3 years. And last week they came back to us and said we're going to need an extra two exabytes on top of that 500 petabytes." "And my expectation is that, from what I'm seeing in other parts of the landscape, that they'll come back again in a quarter or two and say we need a double-digit number of exabytes. We undershot." _______ For a look at a similar dynamic from an incumbent's perspective, $P Everpure's Analyst Day notes are here: firesidealpha.substack.com/p…

Satya Nadella reveals his coding agent pulls every hyperscaler and neoclouds' SEC filings into a dashboard that refreshes every day for real-time ROIC "And this is the other aspect of it, which is the enterprise context combined with the world's context. In fact, I go to the SEC filings of every cloud provider, hyperscaler, each of these neoclouds. It's in real time." "I have a data runner in Fabric that brings all that data, puts it into a semantic model that then gets read by my coding agent and then surfaces it as a dashboard. And every day it's fresh." "So I have the entirety of every SEC filing that goes out there, plus all of my internal analysis constantly coming together, giving me real-time ROIC by layer."

Two $MRVL EVPs were interviewed to talk about the shift from copper to optics, the memory wall, and their custom silicon strategy. Notes below. Clip is of Dave Lazovsky (EVP, GM of Data Center Networking) saying scale-up optics on an analog SerDes can cut total data-center power more than 25%. On with Dave is Will Chu, EVP and GM of the Custom Cloud Solutions Business. 1. The pitch, Marvell is not just an XPU company The market honed in on the XPU as the one custom element, but Marvell sells "XPU attach," everything around it across networking, memory, storage and security. For every XPU there are three, four or more attach opportunities in the next-generation design. 2. A concentrated market and a custom business model This is "the largest infrastructure investment in the history of humanity," and it is very concentrated. Four companies represent north of 75% of the total addressable market in data-center infrastructure. 3. The forcing function, copper to optics and the memory wall "AI infrastructure is the forcing function," and the shift from copper to optical scale-up networks "comes down to the models." Foundation models first drove raw memory capacity, then reasoning models and inference-time compute piled on KV cache. KV caches grew about 10X in nine months, part of why Micron and Hynix now carry more than $1 trillion in market value. Housing that HBM capacity forces scaling beyond the rack, past 144 interconnected XPUs toward 576 in a pod. 4. The interconnect portfolio, SerDes plus CPO and NPO Marvell claims "the world's strongest team" in optical interconnect, backed by 224 gig SerDes moving to 448 gig, integrated onto the XPU. It deploys CPO directly on its own switches for both Ethernet and UAL scale-up, and offers XPU-side links with both CPO and multiple NPO engagements. NPO, or onboard optics, eases customers into the first wave of optics before committing to CPO. Will's addition is that the portfolio walks a customer from copper to NPO to CPO end to end, and putting the same IO on both sides of the scale-up link avoids "link flops." The optics come as "fast pipe" at 200 gig going to 400 gig and "flat pipe" at 56 gig going to about 112 gig, with the analog SerDes far more energy efficient. 5. Getting over the memory wall, the in-box techniques Dense SRAM, specialized in-chip IP tightly coupled to compute, since off-the-shelf memory compilers no longer meet customer needs. Custom HBM for more bandwidth and capacity, with 3D stacking of memory on the logic die coming next. Memory expansion outside the chip through CXL or proprietary methods to add raw capacity. Disaggregated memory like the Photonic Fabric Memory Appliance, which puts 32 terabytes in a separate rack. Near-memory compute, placing CPUs right next to memory to offload the CPU or GPU, since every technique still drives more IO and connectivity. 6. Memory's flip to best-in-industry margins and the CXL moment Memory went from commoditized, at "multi-digit negative gross profit margins," to "the highest margins in the industry," led by HBM. High memory prices are pushing hyperscale architects to rethink whole infrastructures. CXL demand is rising to recycle old memory onto new servers, with compression giving about 2X the memory at 1X the chips, and pooling still to come. The next bleed-over is into flash, both HBF and traditional flash, to tier memory below DRAM cost. 7. The Photonic Fabric Memory Appliance The appliance holds 32 terabytes and is built to remove the trade-off between DDR capacity and cost and HBM bandwidth. Multiple pseudo-channels per HBM stack plus parallel memory transactions hide latency. For the first time it lets customers scale memory capacity and bandwidth independent of compute. 8. Customization in scale-up networking No two hyperscalers have standardized, so each builds its own customized scale-up network at both the physical and logical layers. Some lean toward Ethernet scale-up, some toward UAL (memory-semantic load-store, more like NVLink and more efficient than moving Ethernet packets), and one runs a proprietary topology. Marvell customizes down to the protocol layer and tunes the FEC per customer, since "this is not a merchant silicon world any longer" and it co-architects switches years ahead. Will frames the multi-trillion-dollar build as the economic rationale, with the XPU and its scale-up network "very, very tightly coupled." 9. Scale-across, coherent optics, and the power problem Will's three elements are connectivity, memory and compute, with memory the key to the memory wall and compute the better-understood piece. Data-center campuses are exploding, and Zuckerberg said Meta's newest is about the size of Manhattan. A new class of optics for sub-10 kilometer reach, coherent light, has taken off, and Marvell claims it is first to market with 1.6T optical. Scale-up is roughly 85% of data-center traffic, processor to processor, versus 15% scale-out, so efficiency there drives total power. Power turns into a burning problem over the next three to five years, with the US "nowhere near enough" to build out, so Marvell's optical scale-up on analog SerDes cuts link power about 4X versus conventional 224 gig optics. Fully deployed, Marvell believes that lands a 25% plus cut in total data-center power.

$INTC Lip-Bu Tan reveals Intel has about 70 people working on quantum computing and is making it a strategic direction for the next five to 10 years "Good news, right now Intel Capital is also under my watch, so Intel Capital also invests into some of these quantum computing startups" "And by the way, I just asked my CTO, Pushkar, quantum computing we have like 70 people working on this, the spin quantum, and we also have a fab, we're also working on some of this testing of whether it works." "So I think we're going to put that into work, and this becomes part of the very important strategic direction that Intel want to have in terms of five year, 10 years from now." _______ More from Lip-Bu Tan and his AI playbook: firesidealpha.substack.com/p…

Peter Thiel says everyone calls the Antichrist a crazy idea but secretly agrees with him "When I started giving these speeches about the Antichrist three years ago, I thought this was an extremely unconventional view that was very far outside the zone. And so I would have said that that's an answer." "And now I would say, and then most people would say, yeah, that's a really crazy thing to talk about. I would now say that it's actually not contrarian, that most people think something like the Antichrist is right around the corner, that we have incipient totalitarianism throughout the world, and so it was less contrarian than I thought." "But then the meta-level truth is, I think everybody thinks, everybody says that the Antichrist is a crazy idea, and the reality is they actually secretly agree with me."

SemiAnalysis's Jordan Nanos says CoreWeave is so backed up a customer with $100 million to spend was told to wait until May, so Nebius has been capitalizing "They're serving this middle tier of the market. CoreWeave has set the standard. I mean, they are the best in terms of technology, this reliability stuff I'm talking about. They just have so much data for monitoring." "I mean, these guys deploy 10,000 GPUs a week at the peak when they're building data centers. And so that's incredible. But like their balance sheet is full. There's only so many people that you can marshal and resources to build out another gigawatt when they're going to 1.5 and beyond." "So Nebius has really capitalized on this middle tier of the market, which is not small." "Like I said, this is customers who are coming to us and saying, hey, I saw you write such good things about CoreWeave. I just have $100 million. I'm trying to give it to them. They're telling me they can't accept it until May of next year because they're backed up. And therefore, Nebius has been capitalizing." ______ For more commentary on CoreWeave and Nebius: firesidealpha.substack.com/p…