Interview with P Equity, the "influencer chip analyst": Computing power gaps, memory cycles, hidden bottlenecks, the shift from copper to optical interconnects, and the "grand lie of 10-year demand visibility"

Amidst the market's intoxication with expectations of endless capital expenditure by hyperscale cloud providers for AI, veteran analyst Mr. P offered a sharp critique: the claim of "ten-year demand visibility" is merely a grand deception. What truly constrains AI development is no longer just the quantity of GPUs, but rather expensive memory bills, outdated power grids, and the overlooked segment of ABF substrates.

On September 26, Mr. P, the anonymous analyst behind P Equity Research who has long tracked AI infrastructure construction, appeared as a guest on the latest episode of the Logan Jastremski podcast. Currently, the AI competition centered on large language models is extending from a race in model parameters to an ultimate struggle over underlying infrastructure. The allocation of trillions of dollars in annual capital expenditure by hyperscale cloud providers (Hyperscalers) has become the primary focus for global investors. In this in-depth conversation lasting over an hour, Mr. P provided sobering and demystifying insights on core issues such as the AI computing power gap, the major memory cycle, the competition between optical and electrical interconnects, and geopolitical maneuvering, based on detailed data and supply chain research.

"Ten-year demand visibility is a lie": Memory remains bound by the cyclical fate of boom and bust

Recently, the market has firmly believed that hardware investment driven by AI would usher in an inexhaustible supercycle, with some voices even claiming that cloud providers have order visibility extending ten years into the future. In response, Mr. P offered a highly restrained assessment.

"If a hyperscale cloud provider tells you they have ten-year demand visibility, they are simply misleading you. They cannot possibly predict what will happen two years from now," Mr. P stated bluntly. "I believe memory is cyclical and will remain so, primarily due to the nature of spending. People assume that hyperscaler spending is sustainable and can grow indefinitely, but this is merely an assumption. I believe spending will eventually slow down."

Despite skepticism about "infinite spending," AI's ability to absorb capital in the memory sector remains staggering in the short term. As data centers shift their primary focus from training to the more memory-intensive inference phase, memory (including HBM, DRAM, and NAND) is consuming massive amounts of capital.

"Next year, capital expenditure by hyperscale cloud providers is expected to range between $1.1 trillion and $1.2 trillion, with memory alone potentially accounting for 50% to 60% of that, amounting to approximately $500 billion to $700 billion. UBS Group even estimates that spending on memory alone could reach $900 billion next year, which would exceed the total capital expenditure of all hyperscale cloud providers combined this year."

To mitigate cyclical volatility, memory manufacturers are aggressively pushing for Long-Term Agreements (LTAs). Mr. P revealed that Samsung has begun receiving requests for contracts spanning up to ten years. However, an industry expert who previously worked at Samsung and AMD warned him that LTAs might be overly hyped: "If demand plummets, these long-term contracts may be silently canceled behind closed doors by both parties without any penalty clauses, simply to avoid completely alienating customers. If you force customers to absorb $50 billion worth of products they no longer need, they will merely stack the memory in warehouses, leading to a genuine cliff-like drop in your revenue over the following years."

The computing power gap persists, but "invisible bottlenecks" are shifting toward electricity and substrates

Currently, there is significant divergence in the market regarding whether computing power is constrained. Microsoft CEO Satya Nadella has hinted that chips are sufficient, with the bottleneck being a lack of "warm racks" to house them; whereas Google and Amazon have explicitly stated in their financial reports that they are constrained by shortages in computing power.

Mr. P corroborated the current shortage of computing power using market price data: "Rental rates for older chips from four or five years ago are now higher than before. For instance, the second-hand price for the H100 remains as high as $25,000, and the hourly rental rate for B-series GPUs has risen from $5 to $7–8. The service life of chips is being significantly extended, with some reaching up to 10 years."

However, the sheer number of bare GPU dies may not be the core bottleneck. An AMD expert confirmed to Mr. P that while supply of raw chips is adequate, the true constraints lie in data center power, advanced packaging, and memory storage.

An analysis of the Bill of Materials (BOM) reveals that the genuine shortage areas are coming into focus. Mr. P pointed out: "In the BOM for Vera Rubin (NVIDIA's next-generation architecture), components seeing triple-digit price increases include ABF substrates, PCBs, and memory. Capacity for ABF substrates is extremely tight, and supply shortages are likely to persist until 2028 or even beyond 2030. In recent financial reports, Dell, HP Inc, NVIDIA, and Broadcom have consistently listed DRAM, NAND, ABF substrates, and wafers as capacity constraints."

On the energy front, gas turbines required for building new computing centers face extremely long delivery lead times. "Backlogs for Mitsubishi, Siemens, and GE Vernova extend beyond 2030. If you order a gas turbine today, you may not receive it until 2030."

Copper Recedes, Optics Advance: Optics Is the Endgame, but Copper Will Be 'Forced to Survive'

In the realm of data center network connectivity, the debate between copper cabling and optical communication routes is intensifying. Although the market generally anticipates the rapid adoption of Co-Packaged Optics (CPO), Mr. P believes this will not happen overnight.

"Nothing travels faster than the speed of light, so optics will ultimately prevail; speed is king." However, he emphasized that copper and optics will coexist for an extended period until 2027, with cloud providers striving to extend the lifecycle of copper infrastructure.

"Because memory storage is prohibitively expensive, hyperscale cloud providers must cut costs elsewhere. CPO is not only very expensive but also currently faces issues with yield rates and heat dissipation. You may see the adoption of Near-Packaged Optics (NPO) first in 2027, followed by a ramp-up in CPO capacity in 2028 or 2029, but its penetration rate will remain low. CPO is unlikely to become dominant until after 2030."

Regarding calls in the U.S. to slow the development of frontier AI models due to security concerns, Mr. P dismissed such views as lacking "any commercial logic," stating, "If you slow down for security reasons, you will simply cede market share to competitors."

Below is the full transcript of the conversation, translated by AI:

Logan Jastremski: Excellent. P Equity Research, thank you very much for joining my show. It’s a pleasure to have you here. I’ve always enjoyed your content on Twitter and Substack. As we discussed just before recording, congratulations on your continued rise in the Substack rankings. Your current output is impressive, and I look forward to our discussion.

Mr. P: And everything you’re seeing in the market. Yes, thank you. I’m excited to discuss compute and AI with you. It feels great to gain traction on Substack and X (formerly Twitter), and I’m happy to connect with you. You are always active. It seems like you’re online 24/7, which I truly appreciate.

Logan Jastremski: Yes, I am often online. However, I must admit that I frequently use the “schedule tweet” button to plan my posts. Perhaps that is the secret to generating alpha.

Mr. P: Haha, yes, that is alpha. Using scheduled tweets.

Logan Jastremski: Fantastic. I’d love to learn how you began diving deep into the compute sector. You are clearly very active in publishing your research, and I believe you have gained significant influence within the investor community, particularly regarding the construction of various data centers. So, if you’re willing to share, I’d be interested to hear about your background and how you entered the field of equity research.

Mr. P: Sure. I started tweeting around early May this year, when I had only about 30 followers. So, I was quite unknown at the time. I had registered the account for a while but hadn’t used it much. Then I decided to start using it actively, posting tweets and publishing numerous articles on Substack. Initially, I often wrote about niche companies in Japan, Taiwan, and South Korea. The response was fairly positive. As I gained more recognition on X, I increased my posting frequency.

Mr. P: Around July, I genuinely changed my approach to posting. I aimed to become someone who shares abundant information, maintains a neutral stance, and allows readers to interpret the data themselves. I have consistently adhered to this strategy, and my account has truly taken off since then. However, I have absolutely no technical background. All my knowledge is self-taught. I do not hold an engineering degree. I have bachelor’s and master’s degrees in accounting and previously worked in public accounting. That is my basic background.

Logan Jastremski: I am not an engineer either. So if someone asks me something like “what is the difference between this and that,” I might not be able to answer; my knowledge isn’t that extensive. But that’s fine. I’ve generally found that individuals who excel in their fields are often self-taught. In any case, before delving deeply into AI, I spent considerable time in the cryptocurrency market. Prior to that, I worked at Tesla. A good friend once told me that you are not obligated to remain who you were yesterday; you can teach yourself new skills. Therefore, as long as you keep learning, I believe you will always stay at the forefront.

Mr. P: Yes, I completely agree. I think this is also one of the reasons I enjoy posting so much: to learn as much as possible for free. Because there is no limit to what you can learn. So, sharing information is always a good thing.

Logan Jastremski: I agree 100%. Perhaps we can start the discussion here. Clearly, a great deal is happening in the AI data center sector. To a large extent, I believe this area began to take off following the release of Opus 4.5 (around late 2025). Since then, for the general public, we have shifted more from OAuth and subscription models toward API-based models. This has truly triggered an explosive growth in revenue, although this takeoff may have begun even earlier. As a byproduct, you see many so-called hyperscale cloud providers investing hundreds of billions of dollars annually in capital expenditure (CapEx). I know you have been delving into bill of materials (BOM) analysis and tracking where these capital expenditures are actually going. So, I would like you to discuss from a macro perspective what you have been monitoring and your views on the current state of the market.

Mr. P: Regarding hyperscale cloud providers, I believe there are four key areas: Logic, Memory, Power, and Networking. A significant portion of funding will now be spent on logic chips, namely GPUs and ASICs. Meanwhile, memory is expected to account for 50% to 60% of cloud providers' capital expenditures next year. For reference, if the projected capital expenditure for cloud providers is $1.1 to $1.2 trillion, spending on memory alone will reach $500 to $700 billion. The actual figures could be higher, but this includes a mix of various products: HBM, DRAM, commodity DRAM, and NAND flash.

Mr. P: However, looking at computing power, they are spending heavily on procuring ASICs and GPUs. We all know that Amazon and Google are striving to increase production of their custom chips (ASICs) to reduce reliance on NVIDIA. This raises a question: Is computing power still constrained? Many argue that we have sufficient computing power but insufficient electricity. Others claim we have enough power—well, I have yet to hear anyone say that power supply is sufficient. But people do say we are not lacking in other areas, only in computing power; that computing power is the bottleneck. This is the narrative coming from various AI labs.

Mr. P: So, I would like to address two viewpoints. I previously conducted an expert interview with someone from AMD. But before mentioning him, here are the two prevailing perspectives. The first viewpoint holds that we have sufficient computing power, but lack enough electricity to plug in and operate these chips. This discussion stems from remarks made by Satya Nadella, CEO of Microsoft, in a podcast. I cannot recall the specific host—it might have been SemiAnalysis or another party—but he essentially stated that we do not have ready-made 'hot racks' (plug-and-play cabinets) available to deploy these chips.

Mr. P: The second viewpoint, expressed by AI labs and vendors, is that demand for computing power is immense, and we are currently in a state of compute constraint. If you listen to the earnings calls of Google and Amazon, they mention being compute-constrained; if not for this limitation, the proportion of revenue from their cloud businesses would be even higher than currently reported. Meanwhile, a research report released today by CLSA states that the supply-demand gap for ASICs and GPUs is as high as 73%, and they expect this shortage to persist until 2030. Therefore, if you look at comments from computing power users, they report a lack of capacity. Rental prices also confirm this, as leasing rates for older chips aged four to five years are now higher than before, even though these chips should be fully depreciated or obsolete. (Note: As you know, Michael Burry sparked a major debate about depreciation a year ago, and now you are beginning to see these chips lasting beyond five years, even approaching ten years.)

Logan Jastremski: Yes, absolutely correct. You are starting to see these chips lasting beyond five years, even approaching ten. I believe NEBIUS recently released some data—please correct me if I am wrong or if you have better data—showing they are using chips that are 8 or 9 years old, which essentially extends the lifespan of chips like the H100.

Mr. P: I think used H100s are still selling for around $25,000, which is quite astonishing. Yes, I believe NVIDIA's Volta architecture chips are still in use, which are nearly 10-year-old products. But indeed, these chips are being used beyond their natural or expected service life. This is also interesting because a comment made a month or two ago mentioned that a small startup tried to acquire a small batch of chips but simply could not obtain them. Once they finally secured them, someone on their team bought a cake to celebrate, illustrating just how critical it is to secure computing power nowadays.

Logan Jastremski: Yes, we have heard similar situations from several portfolio companies we have invested in. They are trying to do many interesting things, but without access to computing resources, you are truly out of luck (SOL). So, as you said, everyone seems to be desperately crying out for computing power. Even Elon Musk, while building the Tera Fab, stated that there is not enough computing power to simultaneously meet the needs of robots, the cars they want to build, and ultimately, space data centers. Therefore, from a macro perspective, we do indeed appear to be constrained by computing power.

Mr. P: However, regarding your point, is the constraint really due to energy, or is it the physical computing chips themselves? I believe energy is a major issue. I would say that the current problem is that too many data center construction projects are being delayed. We do not know exactly how many data centers have been postponed. SemiAnalysis suggests this is not a major issue, but based on all the news I have read, it appears that data center delays are quite severe. There is also significant social and political resistance to data center construction. So, I do believe power is a major problem. I think power is absolutely one of the worst bottlenecks at present. I will elaborate on this later. But returning to the comments from the AMD expert I mentioned earlier, he stated that in terms of raw chips and pure computing power, he believes we have sufficient supply; the bottlenecks lie in other areas, such as data center power, advanced packaging, and most importantly, memory.

Mr. P: Therefore, I believe that regarding GPUs, you will see much debate over whether we have sufficient quantities. However, rental prices indicate that our current computing power is insufficient.

Logan Jastremski: Yes, I think I saw it this morning, or perhaps it was a post you shared. It mentioned that the hourly rental price for B200 or B300 GPUs has risen from $5 to $7 or $8.

Mr. P: Yes, it’s crazy. Prices have been climbing continuously.

Logan Jastremski: Perhaps we could delve deeper into the memory aspect. This is an area I have spent some time researching and discussed in detail with Bubble Boy. Specifically, as these data centers largely shift their focus from “training” to “inference,” which is more memory-bandwidth constrained. On one hand, model weights are becoming increasingly large; on the other, context windows are generally lengthening. As you pointed out, demand for memory is surging dramatically among next-generation cloud providers (“Neo clouds”) and across the entire industry. Do you have a more detailed breakdown of the striking figure that “memory accounts for 60% of capital expenditure”? Given that trillions of dollars are invested annually in infrastructure construction, this is indeed astounding.

Mr. P: Indeed. In fact, it is difficult to determine precisely how much capital expenditure is allocated to memory. Let me clarify this for you. First, forecasts vary depending on whom you ask. I do not believe anyone knows exactly how much is being spent on memory. A few months ago, a CLSA research report stated that memory would account for 48% of capital expenditure by 2027. SemiAnalysis cited 36%. Citrini projected 60% by 2028 (I am unsure of their 2027 forecast). JPMorgan forecasted 49% for 2027 and 60% for 2028. UBS Group initially projected 63% for 2027, later revising it to 73%. You see my point? We cannot determine who is correct. Moreover, I believe we may never obtain an accurate figure because the composition is complex, involving significant amounts of HBM, NAND flash, and commodity DRAM. Consequently, estimates vary widely depending on the source. However, generally speaking, no one disputes that “memory is very expensive,” representing a substantial cost. UBS Group estimates that spending on memory alone will reach $900 billion next year. To put this in perspective, this basically exceeds the total capital expenditure of all hyperscale cloud providers this year combined. It is quite insane, isn’t it?

Logan Jastremski: Yes, that is indeed insane.

Mr. P: Yes, it represents a massive sum for memory. Currently, in terms of supply tightness, HBM is likely the most constrained, or possibly commodity DRAM, depending on the specific context.

Logan Jastremski: I think you might have missed... sorry to interrupt. Have you seen any detailed breakdowns of specific memory proportions in the research you have reviewed or in other analyses you track? So far, my macro-level discussions with Bubble Boy suggest that the industry is not truly focused on “maximizing throughput.” This is understandable, as cloud providers charge based on token volume. In terms of throughput, companies like Groq and Cerebras are largely leading the effort. Then you have components like High Bandwidth Memory (HBM), which are ultimately used to store model weights and context. However, as agents operate over longer durations and execute tasks spanning extended timeframes, an increasing amount of data is being offloaded to NAND flash. Have your studies revealed any indications of divergent trends within memory segments?

Mr. P: I have not really seen detailed information on this aspect. However, I am aware that memory suppliers are attempting to make HBM a larger component of their revenue. From a purely economic perspective, they aim for a higher proportion of revenue from HBM, as they believe this is sustainable in the long term. Commodity DRAM is just a commodity. The majority of their DRAM sales still exclude HBM, and they likely consider these products more price-sensitive in the long run. HBM is more customized and can be regarded as non-commodity. Therefore, they believe they can sustain higher profit margins for HBM than for standard DRAM in the long term. Thus, I think the current trend is that memory manufacturers wish to reduce their reliance on commodity DRAM sales and sell more HBM.

Logan Jastremski: So they are shifting away from standard commodity DRAM used in personal computers and memory modules, moving more towards data centers that consume high-bandwidth memory, as the latter may be slightly less price-sensitive?

Mr. P: Yes, exactly. For instance, SK Hynix expects that by 2030, 35% of its DRAM revenue will be linked to HBM. This is nearly double the figure for this year. While this may not sound like a large amount, I believe every penny counts for these memory manufacturers, as they recognize that they operate in a typical “boom-bust” cyclical industry.

Logan Jastremski: What is your personal view? Clearly, the price-to-earnings (P/E) ratios of these storage companies are relatively low because they have historically been more cyclical. There is that famous cautionary phrase: "This time is different." But based on your personal perspective or the research you have read, do you believe the sector will remain cyclical? Or do you think it will become less cyclical due to AI bringing more gigawatts of power online and these models requiring more general-purpose storage?

Mr. P: No, I still believe this is a boom-and-bust cycle, although the downturn may be slightly softened by the existence of long-term agreements (LTAs). This is actually a very good question. I have been asked this before, and I have publicly stated that I believe storage is cyclical and will remain so. This is primarily determined by the nature of spending. There is an assumption that cloud providers' spending is sustainable and can continue to grow. However, for the sector to cease being cyclical, I believe such spending would need to continue indefinitely, which I do not think is possible. I believe there will always be a point where spending slows down.

Mr. P: Regarding the boom-and-bust cycle, I would note that the long-term agreements (LTAs) signed by these storage manufacturers actually comprise four aspects. The CEO of SanDisk (now part of Western Digital) simply broke this down. They introduced something called "New Business Models" (NBMs). In reality, they should just call them LTAs. I am not sure why they chose the term NBMs, perhaps to make it sound more sophisticated.

Logan Jastremski: Yes, to sound more upscale.

Mr. P: They want to appear distinctive. But it consists of four parts: time commitment, volume commitment, pricing, and financial guarantees. I think most people are familiar with the pricing component, but regarding time commitments, they are typically four to five years. I believe current LTAs are generally five years. I do not know if many people are aware of this, but Daishin Securities reported two days ago that Samsung has begun receiving requests for contracts lasting up to 10 years. So, this is basically double the initial LTA duration. Moreover, they have already started contract extensions with some customers. They previously mentioned that if customers require higher demand or other adjustments at some point, they can extend the LTA and modify it. So, a five-year term might become seven years, for instance. Ultimately, the contract length is determined by the supplier, so it would not be surprising if it reaches 10 years.

Mr. P: Furthermore, I believe even the CEO of SanDisk mentioned at an event on September 9 that there is a possibility of achieving demand visibility for up to 10 years, but they are "not yet" at that stage. He emphasized the word "yet." However, I do not think they are likely to reach that stage, as a 10-year contract seems disadvantageous for both parties. You do not want to lock in a customer for 10 years because the risk during a downturn could be very high. No one has 10-year visibility. If a hyperscale cloud provider tells you they have 10-year demand visibility, they are simply misleading you. They do not have it; they cannot predict outcomes beyond two years, correct? Therefore, there is significant volatility in a 10-year contract. Another negative aspect of signing a 10-year contract is that you lose the ability to capitalize on higher prices. That covers the time commitment aspect.

Logan Jastremski 21:02: Regarding the capital expenditure (Capex) of cloud providers, initially, they mainly funded it using funds from their balance sheets and excess cash accumulated from previously established profitable businesses. I think if you look at the free cash flow of all of them now, it is trending toward zero, or they are borrowing money. And I believe many of them must continue to fund it. Is this your base case assumption? That is, because revenue may lag behind their earlier capital expenditures, this spending will slow down?

Mr. P: Well, I believe every major technology goes through a period of heavy investment, after which it begins to slow down. Because the premise of "unlimited spending" assumes that the AI industry will never mature, right? But I believe every industry matures at some point, whether it is PCs or smartphones. You have spending, but then you also reach a stage of industry maturity, so you do not need to spend as much. You only need to maintain a certain level of capital expenditure, either by cutting Capex or keeping it at a certain level in the coming years. This is what I believe will happen in the future. I do not think you will see capital expenditure rising indefinitely. Of course, the situation would change if all cloud providers could generate substantial free cash flow in the coming years. This is expected to possibly happen in 2028, but we shall see.

Logan Jastremski: It seems you are not very convinced by the so-called "AI supercycle"—that is, we achieve recursive self-improvement, and then everything is set, with AI agents taking over everything.

Mr. P: Well, I do believe in AI, but I think we need to see how high a return on investment (ROI) they can achieve. This is a major question because currently, they are merely consuming free cash flow. So, until free cash flow turns positive, people will remain skeptical.

Logan Jastremski: Returning to the memory segment, I previously discussed volume, pricing, and financial guarantees. Regarding volume commitments, Long-Term Agreements (LTAs) stipulate the contracted quantities that cloud providers must purchase. I believe 50% to 70% of SK Hynix’s supply is tied to long-term agreements. Then there is pricing, where participants have essentially established a price floor. For instance, SanDisk recently stated its intention to maintain an 80% profit margin through the end of 2030, reflecting its price floor mechanism. However, there is a caveat: they can remove the price ceiling. I spoke with a Micron employee who indicated that, if desired, the company could raise DRAM profit margins to 94% or 96%.

Logan Jastremski: That is insane.

Mr. P: Yes, it is quite extraordinary, isn't it? One rarely hears of corporate profit margins reaching such levels. It effectively equates to selling DRAM at near-zero cost. Naturally, neither he nor I believe this will actually occur. The key point, however, is that while a price floor exists, a price ceiling may not. Consequently, memory manufacturers can continue to raise prices if they choose to do so.

Logan Jastremski: Indeed, throughout the industry, long-term agreements mandate fixed increments of new capacity coming online, with specific volumes varying by company. On the other hand, there is uncommitted 'floating capacity.' Potentially, because this portion is not bound by long-term agreements, it can command even higher profit margins. It operates somewhat on a first-come, first-served or highest-bidder basis. Therefore, they can potentially achieve higher margins than those secured under long-term agreements.

Mr. P: Yes, they can charge higher margins. However, the obvious reason they seek to secure more long-term agreements is their awareness that this is a cyclical industry. They aim to stabilize demand. Regardless of fluctuations in hyperscale cloud providers' capital expenditures, they can lock in fixed revenue streams for the coming years. I will elaborate on this shortly.

Mr. P: The final component is financial guarantees, correct? Micron has indicated that it holds numerous strategic customer agreements, which are also LTAs. They mentioned receiving $22 billion in advance payments. These arrangements include certain penalty clauses, whether in the form of prepayments or cancellation fees. I believe this distinguishes the current cycle. This is why I argue that the downturn will be mitigated. We will not see the severely depressed margins characteristic of previous downcycles, which previously drove memory manufacturers into bankruptcy due to unprecedented spending scales and LTA obligations.

Mr. P: Most importantly, one can make an argument regarding what I call 'maintenance capital expenditure.' Allow me to explain. Imagine if cloud providers announced tomorrow that they had overbuilt capacity and capital expenditure in recent quarters and would begin cutting capex. Personally, I do not believe AI spending will reach a point where cloud providers cut capital expenditure by 50% or even 30%. In fiscal year 2023, the total spending by Oracle, Google, Microsoft, Meta, and Amazon was approximately $150 billion. Even if next year's projected capital expenditure of $1 trillion were cut by 50%, it would still amount to $500 billion. This remains three to four times the capital expenditure level of 2023. So, do you really think they will cut spending by 50%? Even if they do, it is still a substantial sum, correct? Therefore, I believe that even if cloud providers reduce spending, memory manufacturers will still maintain respectable profit margins, establishing a new average baseline.

Logan Jastremski 27:13: Personally, I consider such spending cuts unlikely. When I examine OpenRouter and observe the volume of tokens generated by open-source or closed-source models, the trend appears to be consistently upward. Of course, token usage is distinct from revenue. But merely looking at token consumption, including for applications like Agents, Muse, Hermes, and Grok Bot, seems to represent the next frontier, especially with the emergence of 'Computer Use' capabilities. As laypersons, perhaps not as deeply versed as those in computer science or software engineering, we are beginning to build agents for basic tasks such as accounting or operations without writing code. This field appears to be in a very early stage, which I find exciting.

Mr. P: Yes, memory demand will be immense, correct? The only caveat is ensuring that cloud providers do not become dissatisfied with you. I anticipate seeing comments from companies—perhaps Micron, which I expect to release its earnings report next week (on the 30th)—stating that profit margins will stabilize from this point forward. Margins may not increase further as they secure more LTAs. There will be commentary similar to SanDisk's. Thus, yes, I believe the memory sector remains cyclical, but the downturns will not be as severe as in the past.

Mr. P: However, I must mention that the individual I spoke with, who previously worked at AMD, also has experience at Samsung. He is quite versatile. He has worked at Samsung, AMD, and Renesas—he is currently at Renesas but previously worked in their memory business. He essentially argued that long-term agreements (LTAs) are overhyped. He believes these contracts can be quietly canceled bilaterally without any penalty clauses. He pointed out that during the last downturn, we signed LTAs with major clients. When demand fell, we were obligated to continue delivering the guaranteed quantities. However, if we forced them to accept the goods, these companies would simply stockpile our products in warehouses for years. Cloud providers might take the DRAM and store it. Once the contract expired, they would no longer need our memory because they had already accumulated inventory. Consequently, your revenue would decline further. His argument is that if I force customers to accept $50 billion worth of products they no longer want, my revenue will eventually appear to drop off a cliff to zero. Right? So, not only does the stock price halve and demand evaporate, but you also damage client relationships. That is his perspective. Basically, if AI spending deteriorates, this scenario could occur. They might quietly cancel LTAs privately.

Logan Jastremski: That makes sense. I hope that scenario does not materialize. I hope the momentum continues, with consumers continuing to pay for AI services, or even through advertising models—I believe we have largely yet to unlock this market, which is currently dominated by the freemium model. I recall seeing a statistic on X, though I have not verified it, so take it with a grain of salt: only 2% to 3% of the global population is actually paying for AI. This directionally seems accurate, as relatively few people have truly engaged with these tools. Even among those who have, their usage has historically been quite basic. In reality, adoption is driven by a small cohort of extreme super-users.

Mr. P: Yes, that is absolutely true.

Logan Jastremski: Based on my observation of your work, you have conducted extensive and in-depth research into broad bills of materials (BOMs). I find this particularly interesting given the trillions of dollars in capital expenditure being directed toward infrastructure. Understanding exactly which components are absorbing this capital is crucial. You have clearly highlighted computing power and storage. What other areas have caught your attention, or where are you seeing significant increases in overall costs and expenditures?

Mr. P: Indeed, if you examine the bill of materials for Vera Rubin (NVIDIA’s next-generation architecture), components showing triple-digit growth include ABF substrates, PCBs, and memory. Therefore, I believe ABF substrates represent a critical area to watch. Substrate capacity is tightly constrained. The shortage of ABF substrates is likely to persist beyond 2028; some even suggest it could last until 2030. If you review the earnings call transcripts from companies such as Dell, HP Inc., NVIDIA, and Broadcom released just a few weeks ago, they cited DRAM, NAND, ABF substrates, and wafers when discussing supply chain constraints. There is a consensus that substrates are a bottleneck. I also see supply shortages in indium phosphide, evidenced by companies like AXT, Sumitomo Electric, and JX Advanced Metals expanding their indium phosphide supply for laser and optical applications. So yes, PCBs and ABF substrates are definitely two areas facing tight supply, as reflected in the bill of materials.

Logan Jastremski: Another area where I would value your perspective is photonics/optics. It appears efforts are underway to disaggregate different racks to perform specialized tasks. Potentially, if throughput is sufficiently high, very interesting opportunities arise in this domain. I know many investors have been focusing on optical communication stocks, such as Lumentum. What is your macro view on network connectivity, including components that connect racks or interface with technologies like NVLink?

Mr. P: While I do not have granular details on specific technologies, regarding optical communications, there is an ongoing debate about 'copper vs. optics,' correct? Which will prevail? Which will become obsolete? There are concerns that copper has limited longevity, yet companies are still seeking ways to extend its useful life. I believe copper will coexist with optical solutions for a considerable period; the transition will not be as rapid as some anticipate. This is because Co-packaged Optics (CPO) still faces significant challenges regarding yield and thermal management. Consequently, in 2027, you may see widespread adoption of Near-package Optics first, followed by CPO production ramp-ups in 2028 and 2029. This aligns with narratives from Ayar Labs, Lumentum, and other industry players. However, penetration rates will remain low. While production volumes will increase, CPO is unlikely to dominate until after 2030. Much of this delay stems from it being a new technology that requires further testing. Moreover, CPO is expensive. As we discuss the rising cost of memory, cloud providers may become more cost-conscious in how they network their racks. I am aware that some cloud providers continue to emphasize the importance of sticking with copper for as long as possible. They are reluctant to switch to optics prematurely. They aim to maximize copper utilization because they anticipate substantial memory expenditures in the coming years, necessitating cost savings during the transition to optics, thereby extending the lifespan of copper infrastructure. In my view, the landscape for copper and optics will remain hybrid over the next few years.

Mr. P: That said, optical communication is the ultimate goal. Co-packaged Optics will eventually prevail; the question is merely when. Perhaps between 2030 and 2040, because speed is critical, correct? As you push higher speeds over copper, you encounter increasing thermal issues. When pursuing higher speeds with cables bundled closely together, you face crosstalk. This is a major disadvantage of copper. One does not need to be an expert to know that nothing travels faster than light, correct? For this reason alone, optical communication will ultimately win, as speed is paramount.

Logan Jastremski: Yes, I am very interested in the optics side and the various trade-offs, such as thermal management versus copper. This is fascinating. I believe that for optics, repeaters are often required, which can consume more energy. Given the power constraints you mentioned, there is a strong industry-wide push to reduce total power consumption across data centers. I think this is another reason why people generally prefer to stay on copper for as long as possible before switching to optics. But as you pointed out, optics are significantly faster. Light transmission through glass is much quicker. Indeed, we saw this during the transition from dial-up to broadband, and then to fiber optics. Obviously, the fiber optics used in most homes today are vastly superior to dial-up.

Mr. P: Oh, yes. My home uses AT&T's fiber network. It is much faster than the modems I used previously.

Logan Jastremski: Very interesting. For me, one of the key macro questions—which you touched on earlier—is how many new gigawatts (GW) of power are coming online and our actual ability to energize them. I understand SpaceX has approximately 2 to 2.5 GW of power in Memphis, which I believe is the world's largest contiguous data center campus. They have several data centers there and continue to expand. I recall them mentioning in a recent earnings call that their target is approaching a total of 8 to 10 GW, which should be very interesting. However, looking at the big picture, how many new gigawatts of power can actually come online? It is not just about constructing data centers and erecting the structures, but truly operationalizing them from a power perspective. Many factors hinge on this, as it is the primary revenue driver for generating tokens from a hardware standpoint.

Mr. P: Personally, I have posted extensively about gigawatts, but I have not even been tracking the construction progress of newly added gigawatt capacity. However, I am aware that SemiAnalysis mentioned that frontier labs alone are set to add 14 gigawatts of power next year. That said, I have not been closely monitoring this, so I cannot provide an exact figure. This is also a highly volatile area, as news emerges daily regarding delays or moratoriums on certain data center projects. Consequently, tracking gigawatt additions has become increasingly difficult.

Logan Jastremski 39:25 Or, to diverge slightly, what is your view on the notion of “slowing down frontier model development”? I largely consider this to be pure nonsense. It seems to be merely a form of psychological coping mechanism to make the situation appear more palatable. However, their alarmist rhetoric is somewhat irritating because, to a large extent, these data centers have already been planned, and capital expenditures have been committed. If they were genuinely concerned about AI alignment or AI safety, I do not believe they would slow down; alternatively, they might decelerate data center construction. But these involve contracts spanning decades, and data centers typically take many years to build. Therefore, I find it hard to envision them actually reducing data center output. If anything, as you pointed out, we are observing an acceleration in these figures. Whether they can truly achieve these targets remains to be seen, but based on their actions, there is no sign of slowing down.

Mr. P: Indeed, I also believe that discussing a slowdown is meaningless. If there were a two- to three-year generational gap between your technology or models and Chinese models, perhaps it would be feasible. However, Chinese open-source models lag by only about six months, correct? They are not significantly behind, and many enterprises have already begun shifting towards using Chinese models. Therefore, if you talk about slowing down model development, you are merely allowing China to catch up. And they will not slow down, right? Moreover, I believe China may have already implemented many of the “safety guardrails” discussed in the U.S. If you wish to slow down, fine, no problem. But this could pose risks to your business model. I do not think you can afford such a cost, as you would simply invite more competition. You would allow China to catch up, and they would capture greater market share.

Mr. P: As Jensen Huang stated today at the All-In Summit (I believe there is a relevant video on YouTube), he said that if there are AI safety concerns, then halt the model, correct? Do not release it. You do not need government intervention. You do not need government regulation. There are already numerous existing laws in cybersecurity and data breach contexts that can be used to hold these labs accountable. You do not need to formulate some new government policy. This is no different from human hackers infiltrating and damaging a system. You hold them accountable. You can use the same laws to hold these labs liable.

Logan Jastremski: Yes, I believe the entire narrative about human extinction before 2030 is sheer absurdity. It is quite crazy. I recall seeing similar comments from Mark Zuckerberg, who stated that they had Muse ready months ago but chose not to release it because they wanted to address certain safety issues. They did not go around proclaiming how dangerous Muse was; they simply fixed the problems from an engineering perspective and released it only when they felt comfortable doing so.

Mr. P: Yes, Muse is excellent. I have not used it yet, but I have heard many positive reviews. I suppose I need to start using Muse.

Mr. P: Are you currently using any AI agents?

Logan Jastremski: I am not using any agents. It seems like you might be, given how timely your posts always are.

Mr. P: You are the second person to ask me if I use agents, but no, I do not use agents. I handle everything manually. I do utilize certain models.

Logan Jastremski: Personally, I actually prefer using Gemini. I am unsure why people dislike it, but I suppose it depends on the use case. I switch between several models. Regarding agents, I initially used Hermes, which allows you to switch underlying models, but it is quite technical. I would say it is more suitable for power users. If you desire more customization, they certainly allow for that. A notable advantage is the ability to switch models—for instance, when Anthropic (Claude) surpassed ChatGPT, you could switch to their model; or when ChatGPT launched Astra, you could switch back. This aspect is beneficial. However, when Elon Musk and xAI announced the launch of Grok Bot, I switched to it. It is much simpler, and I can execute many of the workflows I have already established. Now you are locked into that specific model. I am not sure if Muse discloses which underlying model it runs on; I assume it is Facebook's internal model.

Logan Jastremski: I find all of this fascinating, and it brings us back to the topic of storage. When you require these models to execute increasingly long-horizon tasks, where in the memory hierarchy do you store this information? It is unlikely to be SRAM due to capacity constraints, or HBM as you network more chips together. However, I believe SanDisk and the broader NAND flash ecosystem present an interesting opportunity, as they are offloading more long-horizon tasks to flash memory.

Mr. P: Yes, I believe NAND is a sector where we could see a significant surge in storage usage driven by the adoption of AI agents. Many banking reports, such as those from KB Securities, suggest that widespread agent deployment could lead to increased CPU and GPU utilization. This, in turn, would drive higher demand for HBM, DDR5, and flash memory. Overall, I view this as highly constructive for storage demand. I anticipate increased utilization across the board; I do not foresee one component completely displacing another. Perhaps flash usage might outpace others, but I am less familiar with that specific domain, so I cannot elaborate further.

Logan Jastremski: Indeed, this is intriguing because AI agents truly represent a broad frontier for everyday applications. Hopefully, the trend moves in this direction rather than requiring individuals to write code, but we shall see. It is certainly fascinating.

Logan Jastremski: When considering a timeframe not just for this year, but leaning towards 2027 or 2028, what research have you conducted in that area? Or, as you look ahead several years, what has come into your view?

Mr. P: I assume you have discussed HBF (High Bandwidth Flash) with others, correct? As HBF is poised to become a prominent topic within the next year or two, with increasing commercialization and adoption. I found Bubble Boy’s comments on HBF interesting; he mentioned hearing that Chinese customers are interested in HBF. Personally, I am unsure of their rationale, and I do not doubt his assertion given his greater experience and network. However, the issue with HBF is its low endurance, correct? Thus, it does not appear to be a substitute for HBM. Moreover, CXMT is still striving to develop HBM3. So, why pursue HBF? To my knowledge, HBF is placed in close proximity to heat-generating GPUs. How can endurance be guaranteed under such conditions? It seems you would achieve virtually no endurance. This presents a double whammy—a worst-case scenario. How would you mitigate the thermal challenges of HBF? I prefer not to delve too deeply into engineering specifics, as I lack the requisite expertise.

Mr. P: However, there are two credible sources highlighting the potential role of HBF. First, during a Twitter Space discussion with Nick Doyle of SemiAnalysis, he suggested that HBF might be suited for low-batch inference scenarios. This would not apply to hyperscale cloud providers, but rather to smaller deployments, such as on-premise setups in private enterprises with limited GPU resources. Another source is the Technical President of Nintendo, who, in an expert interview conducted by Bank of America, agreed that HBM and HBF might coexist. This is due to the aforementioned challenges, and uncertainty remains regarding whether HBF is sufficiently mature, given its issues with thermal management, endurance, bandwidth, and cost structure. He mentioned they utilize a different substrate, though I cannot recall the specifics. Nevertheless, we will observe how this technology evolves. I find HBF exciting because, in the semiconductor industry, unforeseen changes can occur overnight.

Logan Jastremski: Yes, I have always been fascinated by storage architectures across different designs, as you pointed out, significant capital expenditure is being directed toward this area, primarily in the private markets. If you are building any storage accelerators, please contact Frictionless Capital (note: Logan's VC fund). We would be delighted to connect. A shameless plug.

Logan Jastremski: My perspective on High Bandwidth Flash (HBF) is that, potentially, as models grow larger—starting from 1 trillion parameters—I believe Groq's latest model is around 4.7, roughly 2.1 or 2.3 trillion parameters, while 4.6 is 1.5 trillion parameters. Elon Musk has also hinted that over time, Grok and total parameter counts could reach 4 trillion, 6 trillion, 10 trillion, or even 100 trillion. Where will the weights for these models be stored?

Logan Jastremski: Perhaps we will discover some effective algorithmic compression techniques—I have discussed this macro perspective with Bubble Boy as well—so that you do not need to store the entire model's weights using Mixture of Experts (MoE) or other technologies. But if that is not feasible, and we must load the entire weights into High Bandwidth Memory (HBM) or even High Bandwidth Flash (HBF), I believe HBF could be highly suitable. As you noted regarding endurance, I think it is well-suited for 'reads' rather than random-access 'writes' like those in KV caches. I believe model weights themselves could be retained in high-bandwidth flash. Alternatively, if you perform pre-computation—for instance, as Logan has discussed, which is a common operational practice—you have already invested prefill compute time. Perhaps you simply store the result in high-bandwidth flash to avoid recalculating it each time, rather than discarding it. Thus, I do see unique opportunities. Of course, this remains to be seen. Given that it is not yet in mass production, it is still in the early stages. However, I find it exciting that people are beginning to experiment with different custom storage accelerators, similar to our approach in training. Therefore, I consider this an interesting area to watch. To a large extent, we are shifting substantial macro-level compute or data center construction from pre-training (which will certainly continue) more towards inference, decoding, and how this transforms the underlying hardware.

Mr. P: Yes, you are more knowledgeable in this area than I am. I wish I could elaborate further. However, I have not delved deeply into those specifics, so my apologies.

Logan Jastremski: It is fascinating to delve into these topics.

Mr. P: Yes, it is interesting to see what people are attempting. HBF is quite intriguing.

Logan Jastremski: What other areas broadly interest you, or specifically excite you? Clearly, you devote all your time to this; you are at the forefront and truly have your finger on the pulse of the industry. What personally excites you?

Mr. P: Indeed, I am excited to see how optical communications will evolve. I find optical communications still very captivating. There will be massive network demand in the coming years. I believe networking is currently the fastest-growing component of capital expenditure, or something to that effect. For instance, the compound annual growth rate (CAGR) over the next five years is approximately 64%. Therefore, I think network build-out will be interesting because speed is ultimately critical for all these models; they aim to achieve maximum throughput. Consequently, there will be intense competition regarding who can provide the best lasers.

Mr. P: We have Lumentum, Broadcom, and Coherent. I consider Lumentum a great company with excellent lasers. Based on conversations I have had, their technology appears superior to that of Coherent. Even firms like Rational Analysis state that products such as Coherent’s lasers are inferior.

Mr. P: Optical communications are exciting. I also find power infrastructure exciting. However, the issue with power is that it may not currently be a suitable area for investment, primarily due to excessively long lead times. For example, consider gas turbines. Investing in gas turbines might have been exciting two years ago, but it may no longer be advisable. All major manufacturers—Mitsubishi, Siemens, and especially GE Vernova—have order backlogs extending beyond 2030. Their capacity is already tightly constrained. Even if they wanted to, they cannot accept additional orders. The lead times are too long. If you order a gas turbine today, you might not receive it until 2030. That is how lengthy the timeline is.

Logan Jastremski: Yes, I saw Elon discuss turbine blades and vanes on a podcast. Furthermore, a significant portion of power generation is shifting toward “behind-the-meter” solutions, which is also quite interesting. However, when conducting research, I questioned exactly how many new gigawatts of power are coming online. Recalling the SemiAnalysis report on projected gigawatt additions, many different players can only contribute at the megawatt scale. This highlights a substantial disconnect between the total capacity discussed by these participants from an energy perspective and the gigawatt-scale reality we actually need to achieve. Personally, I do not focus heavily on the energy sector. That said, power is obviously required to operate these chips; otherwise, nothing else matters.

Mr. P: Yes, energy is currently the primary bottleneck. Referencing comments made by Michael Dell—I believe this was at an event hosted by Goldman Sachs or Citi a few weeks ago—he discussed how he views next-generation cloud providers (“Neo-clouds”) as a metric for gauging real demand for computing power. These providers are the ones securing limited resources such as land, power, and shell space. Therefore, he believes that whoever has the ability to secure all these resources has the best outlook on demand, as one can infer the required computing power and server volume through them. So yes, energy is likely the biggest bottleneck right now.

Mr. P: And then there is storage, correct? But in my view, nothing currently compares to gas turbines, given the excessively long lead times and highly concentrated market structure.

Logan Jastremski: Yes, that is interesting. There is certainly a need to continue aggressively building out these data centers. I recall seeing statements from either CoreWeave or NEBIUS indicating they would bring several gigawatts of power online. Hopefully, everyone contributes incrementally—one gigawatt here, another there—and we will reach our targets.

Mr. P: Yes, everyone needs to do their part. There can be no further delays.

Logan Jastremski: Yes, absolutely.

Mr. P: Yes, thank you for inviting me. It was an excellent discussion.