SK Securities: Memory Revaluation Stemming from a Shift in the Perception of Compute Perception of compute: from consumable to asset AI compute is not a consumable. Rather than a technology good whose value erodes quickly, it could become an infrastructure asset that is redeployed across generations and generates recurring cash flows. From this perspective, AI CapEx is not simply a recurring cost but a process of accumulating a compute stock that can be monetized multiple times going forward. CoreWeave's recent earnings release shows this possibility. Even A100 based resources, six years past launch, are being signed into long term contracts running through 2029, and prices for older GPUs are also holding at meaningful levels. This does not have to be read purely as a supply shortage. Customers choose the cost relative to the performance their workload requires, not the fastest GPU. The newest GPUs handle frontier training and high performance inference, while existing H100/A100 can shift to general inference, fine tuning, batch workloads and the like. This is the concept of compute tiering, analogous to memory tiering. Once the perception of compute changes, the lens through which the expansion of AI infrastructure finance is viewed also changes. If AI infrastructure such as GPUs can generate revenue repeatedly through recontracting after the initial contract ends, it becomes a financeable infrastructure asset rather than simply a piece of electronic equipment. BEP within three years, and the period after that may be the real profit zone The same logic can apply to hyperscaler FCF. Amazon stated that its investments in servers and networking equipment reach BEP on average in less than three years. This means that if the economic life of compute assets including GPUs extends beyond three years, the subsequent period of use may be not simple principal recovery but a zone that generates additional profit. Revenue from new GPUs and second life revenue from the existing compute stock would occur at the same time. If, as with the A100, assets remain applicable into their ninth year, it means a considerable, highly profitable excess return zone opens up between the payback period the market worries about and the actual economic life. The change in memory's status is a structural phenomenon. A change that has to be accepted Once the perception of AI CapEx changes, it is entirely natural that memory's status changes. AI is an asset heavy industry, and physical supply capacity constraints in the real world are not something that can be changed. The reason five year LTAs with high binding force have been signed also becomes clear. Memory has a material impact on the customer's "true" profit zone. If memory does not make money, meeting demand will remain out of reach, volumes will not be secured while memory prices alone rise, and the future revenue curve will be revised downward. Memory making a lot of money is structural, and it is a change that has to be accepted. Inflection point in the change of memory valuation metrics: shareholder returns strengthen in earnest through greater earnings durability and visibility SK Securities maintains its Overweight view on memory. We focus on memory's structurally enhanced earnings power and on the shareholder returns that will begin in earnest, driven by greater durability based on a long cycle. Combined estimated operating profit for Samsung Electronics and SK Hynix is KRW 641tn in 2026 and KRW 977tn in 2027, among the highest globally, and aggressive shareholder returns are expected to be the key foundation for building confidence in the long cycle and for memory's revaluation. We are wary of peak out logic based on the inevitable slowdown in the rate of memory price increases. This is now an era in which the strength of the cycle is absorbed not only through "price" but also through "durability and visibility."

$AMD CEO Lisa Su is heading to Korea to secure HBM4 supply from $SKHY and Samsung as MI450 ramps with 432GB of memory per GPU. That makes HBM supply a big piece of AMD’s 2027 ramp while Samsung’s push for a major AI customer gives Su more leverage on allocation.

🎶 On to the next one… predicting the next token From when I first started studying them in the 80’s, I have been fascinated by the emergent homologies between artificial and biological neural networks. From the emergence of similar hierarchical layers in vision systems to critical periods of curriculum learning, we now have evidence that our brains and our LLMs predict the next noun in a spoken sentence before hearing it. TLDR; from the Nature study below: when listening to an audiobook wired to EEG and MEG scanners, the brain activates structures for the anticipated next noun or adjective before hearing it. The brain does a form of next-token prediction. They fed the same audiobook into a Llama LLM and found the artificial neural net was doing the same thing, and like the humans, it was much better with nouns than verbs. This would come to no surprise to Jeff Hawkins, pictured here at his neuroscience institute. From his books, On Intelligence followed by A Thousand Brains, Jeff bravely presents a framework for how the brain works to produce intelligence from cells organized into ~150 thousand cortical columns. His decades of self-funded dedication to studying how the brain works affords a possibly unique and unifying perspective. In his first book, Hawkins presents a memory-prediction framework for intelligence. The neurons in the neocortex provide a vast amount of memory that learns a model of the world. These models continuously make low-level predictions in parallel across all of our senses. We only notice them when a prediction is incorrect. Higher in the hierarchy, we make predictions at higher levels of abstraction (the crux of intelligence, creativity and all that we consider being human), but the structures are fundamentally the same. If that is not mind-bending enough, in his latest book, Jeff extends the memory framework to the construct of “reference frames”. Everything we perceive is a constructed reality, a cortical consensus from competing internal models resident in many cortical columns, the amalgam of 1000 brains. Those models are updated by data streaming from the senses. But our reality resides in the models. And in particular, sequence memory predicts the next note in a melody and “is also used for language. Recognizing a spoken work is like recognizing a short melody.” With that in mind, let’s jump to summary passages from the recent study: “The brain continuously generates predictions about incoming sensory inputs, including the next word during language comprehension” Large language “models rely almost entirely on predictive processes to generate language enabling them to approach - or even surpass - the Turing-Test with unprecedented proficiency. The human brain, too, is not a passive receiver, but an active ‘prediction machine’, constantly anticipating upcoming words and events” “If language indeed represents the fundamental ability required for the development of general intelligence, chain-of-thought reasoning and abstract cognition, and if grammar naturally emerges through language usage – thereby aligning brain mechanisms with patterns observed in deep neural networks – this raises the critical question: can LLMs trained solely on next-word prediction evolve into artificial general intelligence (AGI)?” From a prior study, “even a relatively simple neural network trained only on next-word prediction can spontaneously internalize basic grammatical structures. Hence, it seems plausible that the human brain, with its approximately 100 billion neurons, could accomplish this feat through continuous language-based prediction alone.” “We observed significant pre-onset activity for nouns… in a complementary analysis, we trained a linear probe neural network on the hidden states of Llama 3.2, revealing that nouns and adjectives are more easily predicted than verbs. We discuss our findings in the context of anticipatory frameworks in artificial neural networks, highlighting potential parallels between biological and computational approaches to language anticipation.” “In the EEG, nouns, adjectives, and proper nouns elicited a significant negative peak beginning before 0 seconds, whereas verbs showed no evidence of early anticipatory activity. The MEG data also revealed a significant peak around 0 seconds in left frontal channels for nouns, but not for the other word types, suggesting early predictive mechanisms specifically associated with this word class. [And with the LLM Llama] we found that the prediction probability for the next word is highest for the word class ‘noun’, and considerably lower for the other three word classes.” “We acknowledge that this represents only a preliminary step toward understanding syntactic and semantic processing by comparing LLMs and human brain activities. We present initial evidence suggesting that different transformer blocks may functionally correspond to distinct cortical regions, though further exploration is needed.” “Our analyses show that the predictive signal in the brain seems to consist of two components: a syntactic readiness, which can be observed in temporal regions, and a semantic readiness potential, mainly located in frontal areas. These convergent findings are consistent with the Bayesian framework of the brain, suggesting that both neural and computational systems continually update their internal models by integrating prior expectations with incoming information. Taken together, this evidence highlights the presence of distinct but complementary predictive mechanisms underlying semantic and syntactic processing in language comprehension.” “The question remains whether LLMs can serve as a valid model for understanding the human brain. Large language models are constructed as layered stacks of transformer blocks that operate via self-attention rather than explicit recurrent connections, yet their repetitive structure - characterized by self-similarity and fractal organization being a universal principle in biological structures - may serve as an analogue to the recurrent transmission of signals through the arcuate fasciculus. In the human brain, the arcuate fasciculus facilitates dynamic bidirectional communication between Broca’s and Wernicke’s areas, a pathway long proposed as the neural substrate for a universal innate grammar. Although LLMs do not replicate the full complexity of biological recurrence, the iterative processing achieved by stacking transformer blocks appears to approximate the brain’s integration of syntactic and semantic cues. These observations suggest that the self-similar structure of the LLM may provide valuable insights into the neural strategies underlying anticipatory activity and integrative processing during language comprehension.” “Our study demonstrates that prediction-related readiness in language processing operates through both syntactic and semantic anticipation, as reflected in distinct pre-word onset activity captured by combined MEG and EEG recordings.” “In addition, our results suggest that LLMs provide a computational framework that approximates human prediction-related readiness, with their stacked transformer blocks potentially mirroring the recurrent interactions between Broca’s and Wernicke’s areas via the arcuate fasciculus. The self-similar organization of these transformer architectures may reflect universal hierarchical principles of cognitive processing” “The fusion of generative AI and neural data has the potential to refine cognitive computational neuroscience (CCN) and provide deeper insights into the hierarchical organization of language processing in biological and artificial systems. Ultimately, bridging neuroscience, AI and linguistic theory may not only reveal the cognitive mechanisms that govern human language, but also drive the development of artificial intelligence - bringing it closer to the way the human brain anticipates and processes language” 🧠 -------- And I have transcribed my favorite passages from Hawkins most recent book, A Thousand Brains. I revisit them to learn. Travelling without moving, as we’ll see… “The cells in your head are reading these words. Think how remarkable that is.” “If you ignore folds and creases, then the neocortex looks like one large sheet of cells, with no obvious divisions. The neocortex looks similar everywhere. Every part of the neocortex generates movement. In every region we have examined, scientists have found cells that project to some part of the old brain related to movement. The complex circuitry seen everywhere in the neocortex performs a sensory-motor task. There are no pure motor regions and no pure sensory regions.” The cortex is relatively new development by evolutionary time scales. After a long period of simple reflexes and reptilian instincts, only mammals evolved a neocortex. “At some point millions of years ago, a new piece of the brain appears that we now call the neocortex. It starts small, but then grows larger, not by creating anything new, but by copying a basic circuit over and over. As the neocortex grows, it gets larger in area but not in thickness.” Given the recency, it’s “probably not enough time for multiple new complex capabilities to be discovered by evolution, but it’s plenty of time for evolution to make more copies of the same thing.” • Vernon Mountcastle’s proposition from 1978: “All the things we associate with intelligence, which on the surface appear to be different, are, in reality, manifestations of the same underlying cortical algorithm. Darwin proposed that the diversity of life is due to one basic algorithm (evolution). Mountcastle proposed that the diversity of intelligence is due to one basic algorithm.” Beyond the evolutionary time-scale argument, the brains’ vast flexibility to accept different, even prosthetic, sensory input changes and its ability to learn many different things point to a universal framework for learning. • Cortical Columns are “the largest and most important piece of the puzzle.” They are roughly one square millimeter in size with 100K neurons. A mouse has one column per whisker. “Every cortical column is making predictions. We are not aware of the vast majority of these predictions unless the input to the brain does not match.” • Learning through movement: “The brain learns its model of the world by observing how its inputs change over time. There isn’t another way to learn. Every time we take a step, move a limb, move our eyes, tilt our head, or utter a sound, the input from our sensors change. For example, our eyes make rapid movements, called saccades, about three times a second. With each saccade, our eyes fixate on a new point in the world and the information from the eyes to the brain changes completely.” We don’t perceive any of this because we are living in the model, which is predicting the next input to come, across all the senses. “Vision is an interactive process, dependent on movement. Only by moving can we learn a model of the object.” “To avoid hallucinating, the brain needs to keep its predictions separate from reality. We are not aware of most of the predictions made by the brain unless an error occurs.” “Thoughts and experiences are always the result of a set of neurons that are active at the same time (about 2% of the total). Individual neurons can participate in many different thoughts or experiences. Everything we know is stored in the connections between neurons. Every day, many of the synapses on an individual neuron will disappear and new ones will replace them. Thus, much of learning occurs by forming new connections between neurons that were not previously connected.” • Locus of Predictions: “Oddly, less than 10% of the pyramidal cell’s synapses are in the proximal area. The other 90% are too far away to trigger a spike. For many years, no one knew what 90% of the synapses in the neocortex did. The big insight I had was that dendrite spikes are predictions. A dendrite spike occurs when a set of synapses close to each other on a distal dendrite get input at the same time, and it means the neuron had recognized a pattern of activity in some other neurons. When the pattern of activity is detected, it raises the voltage at the cell body, putting the cell into what we call a predictive state. The cell is primed to spike… and the cell spikes a little bit sooner than if it would have if the neuron was not in a predictive state.” And this inhibits other neurons from ever firing, the ones who were behind in that race. “When an input arrives that is unexpected, then neurons fire at once. If the input is predicted, then only the predictive-state neurons become active. This is a common observation about the neocortex: unexpected inputs cause a lot more activity than expected ones.” Predictions prime the pump, sub-threshold. “Predictions are not sent along a cell’s axon to other neurons, which explains why we are unaware of most of them.” “Most predictions occur inside neurons. With thousands of distal synapses, each neuron can recognize hundreds of patterns that predict when the neuron should become active. Prediction is built into the fabric of the neocortex. As few as 20,000 neurons can learn thousands of complete sequences. The sequence memory continued to work even if 30% of the neurons died or the input was noisy.” • Reference Frames: “The secret of the cortical column is reference frames. A reference frame is like an invisible, 3D-grid surrounding and attached to something” (like a map) “Predicting the next input in a sequence and predicting the next input when we move are similar problems. Our sequence-memory circuit could make both types of predictions if the neurons were given an additional input that represented how the sensor was moving.” “Most of the circuitry is there to create reference frames and track locations. The brain builds models of the world by associating sensory input with locations in reference frames. You need a reference frame to specify the relative position and structure of objects. Roboticists rely on them to plan the movements of a robot’s arm or body. Reference frames were the missing ingredient, the key to unraveling the mystery of the neocortex and to understanding intelligence. We showed that a single cortical column could learn the 3D shape of objects by sensing and moving and sensing and moving. Each cortical column must know the location of its input relative to the object being sensed. To do that, a cortical column requires a reference frame that is fixed to the object. The brain must have neurons whose activity represents the location of every object that we perceive.” “Mammals have a powerful internal navigation system. There are neurons in the old part of our brain that are known to learn maps of the places we have visited” — the hippocampus and enthorhinal cortex, organs roughly the size of a finger. “Place cells tell a rat where it is based on sensory inputs, but planning movement requires grid cells. Grid cells form a grid pattern. The two types of cells work together to create a complete model of the rat’s environment. Every time a rat enters an environment, the grid cells create a new reference frame to specify locations and plan movements.” In the new brain, these same cells and structures create models of objects instead of environments. “Every cortical column learns models of complete objects. The columns do this using the same basic method that the old brain uses to learn models of environments. It is as if nature stripped down the hippocampus to a minimal form, made tens of thousands of copies, and arranged them side by side in cortical columns. That became the neocortex. Each patch of your skin and each patch of your retina has its own reference frame in the neocortex. Your five fingertips touching a cup are like five rats exploring a box.” “Not all cortical columns are modeling objects. Language and other high-level cognitive abilities are, at some fundamental level, the same as seeing, touching, and hearing. The reference frames that are most useful for certain concepts have more than three dimensions.” • Thinking is a form of movement: “The brain arranges all knowledge using reference frames, and thinking is a form of moving. Thinking occurs when we activate successive locations in reference frames.” “A cortical column is just a mechanism that tries to discover and model the structure of whatever is causing its inputs to change” whether the structure of environments, physical objects or conceptual objects. “Reference frames are not an optional component of intelligence; they are the structure in which all information is stored in the brain. Every fact you know is paired with a location in a reference frame. Organizing knowledge this way makes the facts actionable” to “determine what actions are needed to achieve a goal.” “To recall stored knowledge, we have to activate the appropriate locations in the appropriate reference frames. Thinking occurs when the neurons invoke location after location in a reference frame, bringing to mind what was stored in each location. The succession of thoughts we experience when thinking is analogous to the succession of sensations we experience when touching an object with a finger, or the succession of things we see when we walk about a town.” • What and Where Pathways. “Your brain has two vision systems. If you follow the optic nerve as it travels from the eye to the neocortex, you will see that it leads to two parallel vision systems, called the ‘what’ visual pathway and the ‘where’ visual pathway.” If you disable one, you can identify what something is but not where, or vice versa. “Similar pathways also exist for other senses. There are what and where regions for seeing, touching, and hearing.” “Cortical grid cells in What columns attach reference frames to objects. Cortical grid cells in Where columns attach reference frames to you body.” The distinction depends on where the inputs come from. “If a cortical column gets input from the body, such as the neurons that detect the joint angles of the limbs, it will automatically create a reference frame anchored to the body.” “Your body is just another object in the world. However, unlike external objects, your body is always present. A significant portion of the neocortex — the Where regions — is dedicated to modeling your body and the space around your body.” For abstract concepts like mathematics, there are difference reference frames one could use to learn. “Part of learning is discovering what is a good reference frame, including the number of dimensions.” History can be learned on a timeline, or geographically. “They lead to different ways of thinking about history. They might lead to different conclusions and different predictions. Becoming an expert in a field of study requires discovering a good framework to represent the associated data and facts. Discovering a useful reference frame is most difficult part of learning, even though most of the time we are not consciously aware of it.” It's no surprise that the memory trick called the memory palace, is a good method for remembering a large sequential list of nouns. From fMRI studies, “the process of storing items in a reference frame and recalling them via ‘movement’ is the same.” “Nested structure and recursion are key attributes of language. Each cortical column has to be able to learn nested and recursive structure. Cortical columns create reference frames for every object they know. Reference frames are then populated with links to other reference frames. The brain models the world using reference frames that are populated with reference frames; it’s reference frames all the way down.” • The Thousand Brains Theory of Intelligence: The prevailing view of the neocortex was a hierarchy of feature detectors, from edge detectors up to face detectors. Jeff argues that each and every column is a sensory-motor system. “When the eyes saccade from one fixation point to another, some of the neurons in the V1 and V2 visual regions do something remarkable. They seem to know what they will be seeing before the eyes have stopped moving. These neurons become active as if they can see new input, but the input hasn’t yet arrived. There are connections between low-level visual regions and low-level touch regions.” Mouse vision occurs in the V1 region; it does not depend on a hierarchy of vision abstractions. “All cortical columns, even in low-level sensory regions, are capable of learning and recognizing complete objects. A column that senses only a small part of an object (e.g., from a patch of retina) can learn a model of the entire object by integrating its inputs over time.” “Learning is not a separate process from sensing and acting. We learn continuously. When a neuron learns a new pattern, it forms new synapses on one dendrite branch. The new synapses don’t affect previously learned ones on other branches. Thus, learning doesn’t force the neuron to forget or modify something it learned earlier.” It’s additive. “What a column learns is limited by its inputs. Columns in V1 can recognize letters and words in the smallest font. V1 and V2 learn models of objects, such as letters and words, but the models differ by scale.” “Knowledge of something is distributed in thousands of columns, but these are a small subset of all the columns. This is why we call it the Thousand Brains Theory: knowledge of any particular item is distributed among thousands of complimentary models. The columns are not redundant, and each is a complete sensory-motor system.” • The Solution to Sensor Fusion and the Binding Problem: “Columns vote. Your perception is the consensus the columns reach by voting.” “If you touch something with only one finger, then you have to move it to recognize the object. But if you grasp the object with your entire hand, then you can usually recognize the object at once. In almost all cases, using five fingers will require less movement than using one.” (made me think of reading Braille with multiple fingers). “Voting works across sensory modalities (sight, touch, etc.)” How? “Cells in some layers send axons long distances within the neocortex” between left and right-hand brain regions or between V1 and A1, the primary vision and auditory regions. “These cells with long-distance connections are voting. Cells that represent what object is being sensed can vote and will project broadly. Often a column will be uncertain, in which case its neurons will send multiple possibilities at the same time. Simultaneously, the column receives projections from other columns representing their guesses. The most common guesses suppress the least common ones until the entire network settles on one answer. The voting mechanism works well even if the long-distance axons connect to a small, randomly chosen subset of other columns” • The Stability of Perception with ever-changing inputs: “What we perceive is based on the stable voting neurons. We are not consciously aware of the changing activity in each column.” Roughly 98% are silent at any given time and 2% are continuously firing. Consider the experience of an optical illusion duality (like the drawing of a pair of faces or vase); you can only see one at a time, and there is a delay if you force yourself to switch. “Recognizing an object in one sensory modality leads to predictions in other sensory modalities.” • Attention: We have the perception of multiple objects in our visual field even though we can only attend to one at a time. “Attention plays an essential role in how the brain learns models. The brain can attend to smaller or larger parts of the visual field. Exactly how the brain does this is not well understood, but it involves a part of the brain called the thalamus, which is tightly connected to all areas of the neocortex. It is so intimately connected to the neocortex that I consider it an extension of the neocortex.” • Consciousness: “Neurons form a continuous memory of both our thoughts and actions. It is this accessibility of the past — the ability to jump back in time and slide forward again to the present — that gives us our sense of presence and awareness. This is the core of what it means to be conscious. If we couldn’t replay our recent thoughts and experiences, then we would be unaware we are alive.” “The neocortex does not directly control any muscles. The neocortex has to be attached to something that already has sensors and already has behaviors (the primitive brain). It does not create completely new behaviors; it learns how to string together existing ones in new and useful ways.” “Reverse engineering the brain and understanding intelligence is the most important scientific quest humans will ever undertake.”

"NO ONE KNOWS HOW LONG THE SHORTAGE WILL LAST" - FULL $SKHY CEO INTERVIEW SK Hynix CEO Kwak No-jung predicted that the memory semiconductor supply shortage will continue until the end of 2030, and that there will not be a sharp decline in demand thereafter. Regarding plans for additional investment in the U.S., he reaffirmed the existing stance that investment is possible anywhere sufficient support is provided. CEO Kwak gave this response to a question regarding how long he expects the memory supply shortage to continue during a press conference following the groundbreaking ceremony for an advanced packaging production base for AI memory at Purdue University’s Holloway Gymnasium in West Lafayette, Indiana, on the 27th (local time). "No one knows how long the shortage will last, but there are no clear signs of a downturn," said CEO Kwak. "We expect the shortage to continue until the end of 2030." He further explained, "I believe the next downturn will be different from the ones we have experienced over the past few decades. While adjusting production volumes appropriately is not easy, our business model for AI is transitioning from 100% standard products to partially customized or fully customized products." In the case of AI memory, since it is custom-made for client systems, demand forecasting is easier and ties with manufacturers are stronger compared to the past. Consequently, the explanation is that even if a downturn occurs, the damage will not be as severe as in the past." Even if a downturn hits, I believe demand will slow down slightly or remain at a stable level rather than plummeting," CEO Kwak said. "Therefore, I am not concerned about the period after 2030. I believe supply and demand will balance out, and the entire AI industry will grow together." When asked about plans to invest in additional semiconductor production facilities in the U.S., he replied, "We are open to any location where sufficient resources, such as water, electricity, talent, and subsidies, are provided," adding, "Since nothing has been finalized yet, I cannot discuss specific details or directions." Last month, SK Group Chairman Chey Tae-won stated, "If there is a suitable place, it does not matter whether it is the United States or anywhere in the world," expressing the same position. However, CEO Kwak specifically added, "I hope to be able to deliver good news to the U.S. AI industry next time." When asked if he felt pressured by the Donald Trump administration's demand for investment in memory semiconductor front-end process facilities, he replied, "Our current situation is very well known through media reports." He reiterated, "We are reviewing any location in the world as long as it offers customers and new business opportunities," adding, "If the conditions presented by the other party are met, we would like to deliver another piece of good news to the AI industry in the U.S. or other countries, just like at today's event." Regarding whether to continue maintaining its stake in the Japanese NAND flash company Kioxia (formerly Toshiba Memory), he answered, "There are no concrete or confirmed plans yet." He further explained, "However, just as we contribute to the U.S. AI industry, we also intend to contribute to the Japanese semiconductor industry. Since we are investing in Japan just as we have in the U.S., we view it as our responsibility in some respects to contribute to that local community and the nation. Although there are no confirmed plans, we are carefully examining how we can grow the market together." Source: newsis $MU $SNDK $KXIAY

$SKHY is bringing a critical piece of the AI memory supply chain to the U.S. with a new $4B HBM packaging facility in Indiana. Wafers will still come from South Korea but packaging and testing will move stateside with ~$1B of CHIPS Act funding supporting the buildout.

15 stocks Anchoring the AI Infrastructure Buildout: $MU — Micron's HBM ramp has been one of the clearest beneficiaries of the AI memory supercycle alongside SNDK, and pricing power across DRAM and NAND has held up far better than prior cycles suggested it would. $NVDA — Still the anchor of the entire buildout. Every other name on this list exists in some relationship to Nvidia's roadmap, whether as a supplier, a customer, or a competitor trying to chip away at share. No large-cap AI thesis is complete without acknowledging where the center of gravity still sits. $AMD — Two separate catalysts stacking here. Helios and the next-gen CPU lineup represent a real inflection point on their own. But the more interesting development is AMD effectively borrowing Nvidia's NeoCloud flywheel playbook — helping partners raise capital that gets recycled into buying AMD GPUs. $COHR — Coherent's optics and photonics exposure puts it right in the middle of the bandwidth bottleneck story — as data center interconnect speeds keep climbing, this is one of the picks-and-shovels names benefiting from every upgrade cycle. $AMZN — The Trainium ramp is hitting its stride right now. Custom silicon is becoming a bigger piece of Amazon’s AI infrastructure story, reducing reliance on third-party GPUs while giving AWS more control over cost and supply — a dynamic that should keep showing up in margin trends. $JBL — The story here is the 1.6T LRO optical transceiver ramp. What's underappreciated is the margin profile — as these ramps mature and volumes scale, blended margins tend to surprise to the upside relative to what the Street models in. $SNDK — The recent investor conference did real work reinforcing the bull case: this isn't a cyclical memory blip, it's durable, higher-margin revenue that should hold up over the next several years. Memory has been one of the more volatile corners of the thesis, so a credible durability argument matters. $RDDT — Now sitting in the S&P 500, and honestly this one has just quietly compounded, quarter after quarter, without much drama. Index inclusion tends to bring a fresh wave of forced buying and attention — this name has earned it on fundamentals first. $TSM — Still the center of gravity for the entire AI buildout. Every layer of this thesis — GPUs, custom silicon, networking chips — ultimately routes through TSMC's fabs. As close to a toll-booth position on the whole AI capex cycle as exists. $AMKR — Backstopped growth via Nvidia and TSMC contracts gives this one a floor most packaging/assembly names don't have. Advanced packaging is a bottleneck layer in this whole buildout, and Amkor's contract visibility reflects that. $SMCI — The ~$70B revenue ramp targeted for 2027 is a genuinely staggering number for a server integrator. If that materializes anywhere close to plan, the scale of the AI server buildout embedded in that estimate says a lot about where hyperscaler and enterprise capex is headed. $SKHY — SK Hynix has been the clearest HBM leader through this memory cycle, posting record profits on the back of that positioning, and still seems underdiscussed relative to how central it is to the AI memory stack. $GFS — strategic domestic semiconductor capacity is clearly a policy priority now. On top of that, I think CPO exposure gives GlobalFoundries another leg of revenue acceleration going forward. $AVGO — Broadcom sits at the center of the custom ASIC and networking layer, designing the XPUs that hyperscalers use to reduce Nvidia dependence while also owning critical switching/networking silicon. $CEG — Constellation Energy is one of the cleanest ways to play the power-for-AI-datacenters theme. Nuclear capacity is increasingly the preferred baseload source hyperscalers are contracting for directly, and Constellation's fleet gives it real leverage in those negotiations. Not financial advice.

CEREBRAS $CBRS IS ABOUT TO GO PUBLIC ... HERE IS A DEEP DIVE ON WHAT THEY ACTUALLY DO Most people have no idea what this company actually does. Here is the plain-English version: THE BIG IDEA Every AI model you have ever used (ChatGPT, Claude, Gemini, Llama) is trained on chips. The dominant company is NVIDIA $NVDA, which sells lots of small chips that get wired together into clusters by the thousands. Cerebras went the opposite direction. They build ONE giant chip the size of a dinner plate. Not exaggerating. Their flagship "Wafer-Scale Engine" (WSE-3) is 46,225 square millimeters of silicon. A normal AI chip is smaller than a postage stamp. THE NUMBERS On that one giant chip: - 4 trillion transistors (a top-end NVIDIA GPU has roughly 80 billion) - 900,000 AI cores - 44 GB of on-chip memory - 125 petaflops of compute power - Built on TSMC's 5nm process WHY THAT MATTERS When you wire thousands of small chips together, the wires become the bottleneck. Data has to travel between chips constantly. That eats time, power, and money. Cerebras keeps everything on one piece of silicon. No cables between chips. No slow networking. Just one giant brain. The pitch: faster training, faster inference, fewer engineers needed to manage cluster bottlenecks. WHAT THEY SELL Three ways to use Cerebras: - Buy the system outright (the "CS-3" is the box that holds the chip) - Rent compute via Cerebras Cloud - Dedicated capacity contracts for big customers They have 6 new AI inference data centers coming online across North America and Europe. WHO ACTUALLY USES IT The customer list is the validation: - OpenAI: $20B+ committed over three years - Meta $META: powers the Llama API for developers - Perplexity: runs its Sonar search model on Cerebras - Mistral: the French AI lab runs Le Chat on Cerebras - Mayo Clinic: trains genomic AI models on Cerebras infrastructure - GSK $GSK: trains biological language models - Argonne National Lab: has used Cerebras hardware since 2019 - AWS: hosts Cerebras chips inside Amazon data centers, accessed through Bedrock - US Department of Energy: signed an MOU for the Genesis Mission THE TRADE-OFF Cerebras is small compared to NVIDIA. 2025 revenue: $510M. 2025 operating loss: $146M. Concentration is the risk most coverage will not flag: - G42 (the UAE conglomerate) was 85% of 2024 revenue per Reuters - G42 plus MBZUAI (the Abu Dhabi AI university) were 86% of 2025 revenue per FT - The OpenAI deal is the big bet to diversify away from that concentration THE STORY IN ONE LINE NVIDIA bet that the future of AI is millions of small chips working together. Cerebras bet on one giant chip doing the work in one place. The market just decided their bet is worth nearly twice what they priced it at.

Under the hood of dots3-note Preview: BigMac 🚀 One piece of the training infra behind dots3-note Preview is BigMac, our pipeline scheduling system purpose-built for VLM training. BigMac co-schedules heterogeneous vision and language workloads through a dependency-safe nested pipeline — preserving the efficiency of optimized LLM schedules while reducing pipeline bubbles and activation memory. Across our evaluated workloads, it delivers 1.08–1.9× training speedups over existing baselines. We’ve now open-sourced it.

# Useful but Little-Known Features of ADK 2.0 🌍 How do you keep track of user preferences, intermediate results, and shared data across agent conversations? ADK 2.0's State system makes it straightforward. ADK 2.0's State feature is a key-value scratchpad for persisting and sharing data within and across sessions. Different prefixes control the scope of each piece of state. 📌 Title: State Management 🔗 URL: 🧩 Overview State is a key-value data store with four scopes determined by prefixes. No prefix means session scope (valid only within the current session). The user: prefix shares state across multiple sessions for the same user. The app: prefix makes state global across all users and sessions. The temp: prefix creates temporary state that is discarded after the invocation ends. You can reference state values in agent instructions using the {key} syntax, enabling dynamic prompt construction. 🛠 How to use it There are several ways to write state values. ```python # 1. Auto-save agent output with output_key agent = Agent( name="summarizer", output_key="last_summary", ... ) # 2. Explicitly set via EventActions.state_delta from import EventActions actions = EventActions(state_delta={"user:preference": "dark_mode"}) # 3. Set from within tools via ToolContext def my_tool(query: str, tool_context: ToolContext) -> str: tool_context.state["app:global_counter"] = 42 tool_context.state["temp:intermediate"] = "temporary_value" return "done" ``` Reference state in instructions like this: ```python agent = Agent( instruction="User preference is {user:preference}. Previous summary: {last_summary}", ... ) ``` 🏗 Building it into production ・Choose the right scope for each piece of data — use temp: for throwaway intermediate results ・Design user: and app: scoped state carefully, as changes affect multiple sessions ・Always read and write state through CallbackContext or ToolContext to ensure event tracking ・Establish consistent naming conventions for state keys across your team 💡 Use cases 👤 Persist user preferences across sessions with the user: prefix 📊 Track application-wide statistics and counters with app: prefix 🔄 Automatically reference the previous agent output in the next step via output_key 🧹 Store intermediate computation results temporarily with temp: to keep sessions clean ⚠️ Watch out Do not directly modify session.state outside of a context. Bypassing CallbackContext or ToolContext skips event tracking, meaning state changes won't be recorded in the event history. This can break rewind functionality and make debugging significantly harder. ✨ By leveraging State's four scopes, you can flexibly manage agent memory and context. Proper state management is the foundation of high-quality agent experiences. #ADK# #AIAgent#

TL;DR of @leerob's 45-min Stanford lecture on always-on agents, the last 10 minutes especially: Build as little UI as possible, because you'll rebuild it every six months. Every app either goes agentic or becomes "a headless piece of data for another agent." My takeaway: the UI becomes the least important part of a personal agent. What matters is the stuff you can't see: memory, agents learning on the job, knowing when to speak up and when to stay quiet. You don't feel any of it on day one. You feel it on day ninety.