GM ☕️
Think outside the frame. Create beyond the ordinary.
Leave your mark. The world is your canvas.
Rethinking Robot Safety: Moving Beyond the Binary
If robot safety was ever a yes-or-no question it can't be any longer. When people ask whether a humanoid robot is safe, they usually want a clean answer. The difficulty is that no complex autonomous system can really offer one. Some level of risk is always present.
When safety is treated as a single gate that a system either passes or fails, the conversation tends to stall. We end up with debates about whether a robot is "ready" in the abstract, when the more useful questions are about the specific context in which it operates, the kinds of harm that are possible there, and how confident we are that we've measured and managed them.
In 2023,
@Waymo published a detailed framework organized around the idea of "absence of unreasonable risk." The phrasing accepts that some residual risk will always remain, and it shifts the work toward establishing what counts as an acceptable level of risk in a given context, and then demonstrating that the system meets it.
It guides how Waymo assesses risk along several dimensions:
Severity potential — the scale and extent of the harm that could occur
Conflict role — whether the system initiated a hazardous situation or was responding to one created by others
Behavioral capability — distinguishing among regulatory compliance, conflict avoidance, and collision avoidance
Functionality status — performance under normal conditions versus performance when something is degraded
Level of aggregation — reasoning about individual events alongside statistical, fleet-wide rates
Each dimension can be measured, and acceptance criteria can be set for each.
Waymo describes this as methodology-agnostic, meaning the broad approach could in principle be adopted by others working in different parts of the autonomous systems world. Look, I just got here, but it strikes me as correct to treat safety as something you build a credible, evidence-backed case for, across many dimensions, rather than a single threshold you clear once.
@Figure_robot has established a Center for the Advancement of Humanoid Robot Safety and has committed to publishing periodic reports covering its testing procedures, results, and challenges. The company's safety lead has been candid that the standards humanoid manufacturers are currently required to meet remain ambiguous, given that there is no dedicated regulatory regime for this class of machine yet.Some of the companies building humanoid robots appear to be moving in a similar direction.
Publishing testing methods and acknowledging open challenges is consistent with the spirit of the Waymo approach: building trust through transparency and a documented safety case rather than a one-time declaration. It reflects an understanding that safety in this space is an ongoing, evidence-driven undertaking.
Other developers seem to be working from related premises. Another company building a humanoid for industrial settings has described leaning on the established safety standards that already exist for mobile manipulators and autonomous mobile robots, while anticipating that humanoid-specific standards will emerge over time. Its robots incorporate perception-based behaviors that slow or stop the machine when a person enters a defined zone. That’s more of a concrete way of managing risk according to context rather than asserting blanket safety.
The common thread, as I read it, is an effort to ground safety claims in defined criteria and operating conditions, and to treat international standards as a foundation to build on as they develop.
From here, the industry could wait for regulators to produce a clear, binary certification gate. Think of some official stamp that a humanoid is approved for a given environment. That has obvious appeal in its simplicity but clearly that solution has many problems. It may take years to arrive, could freeze design around whatever assumptions were current when it was written, and might prove ill-suited to a fast-moving field.
On the other hand, the industry could work now to build out spectrum-style frameworks, grounded in existing manufacturing and functional safety standards and in honest measurement of residual risk. This is harder and less tidy, asks more of the companies doing it, and asks observers to get comfortable with safety claims that are conditional and contextual.
The conceptual approach to safety deserves as much attention as the engineering itself. The framework we reach for shapes which questions get asked, what evidence gets collected, and ultimately what gets deployed.
For humanoid robots that will operate in warehouses, factories, stores, and eventually homes, the credibility of the safety case is what earns the trust everything else depends on.
Treating safety as table stakes doesn't require pretending a robot is perfectly safe. It requires a credible, transparent, multi-dimensional account of how risk is being measured and managed in the contexts where the machine actually works. The companies that seem to take this seriously tend to define their criteria explicitly, measure across several dimensions, share evidence of how they're performing, and stay engaged with the standards bodies shaping the field. As operating data accumulates, their safety case can grow more confident over time.
Again, the binary question of whether a robot is "safe" is too blunt for the systems now being built, and the spectrum-based frameworks emerging from autonomous vehicles seem to offer a more honest and more workable way to reason about it.
One of the more important conversations the humanoid robotics field can have right now is how we conceive of safety in the first place.
Show more
Journey through Hell made with seedance 2.5
prompt :One continuous 30-second chaotic amateur first-person smartphone video filmed by a standing passenger inside a completely packed magnetic-levitation commuter train. Single unbroken take with no cuts, jump cuts, dissolves, crossfades, double exposures, portals, morphing or artificial scene transitions. The train begins as an ordinary weekday commute on Earth and then physically travels downward on an impossible journey — through subway tunnels into bedrock, into a colossal void beneath the crust, past a soot-blackened waiting platform, through an immense eroded gate, out above a burning plain, down the terraced wall of a vast pit, and into the inhabited depths of Hell. Everything must feel physically connected, as though the same train is genuinely travelling down through each environment.
The tone is bureaucratic dread, not horror-movie shock. This is a scheduled service. The route is old, the infrastructure is worn, and the train runs it the way it runs any other line. The horror comes from how ordinary the journey is and how enormous the destination turns out to be.
The interior is a completely packed standing-room-only maglev commuter car. Passengers are pressed shoulder-to-shoulder, gripping overhead straps and vertical poles. Backpacks are squeezed between bodies, coats and loose clothing shift with acceleration, straps swing on their inertia, and the entire carriage constantly vibrates and rattles. The camera is a cheap smartphone held at chest height by one standing passenger who grips a pole with the other hand. The phone itself is NEVER visible because the phone is the camera. The framing is crooked, slightly off-centre, partially blocked by shoulders and arms, and imperfect like genuine accidental footage. The camera constantly shakes, rolls, yaws and gets thrown around by acceleration. Use realistic rolling-shutter distortion, autofocus hunting, exposure pumping, blown highlights, crushed noisy shadows, low-bitrate compression, macroblocking and smeared motion blur. It must look like genuine spontaneous smartphone footage, not professional cinematic footage.
The camera always looks through the LEFT-SIDE WINDOWS at approximately 90 degrees to the train's direction of travel. The train always travels forward and the outside world always streams past the windows from front-to-back. Never switch to a forward-facing train-nose view. Never show the front of the train. The same carriage, same passengers, same poles, same straps and same windows remain visually consistent throughout the entire journey. The interior is the constant realistic anchor while the outside world becomes increasingly impossible.
0 to 3 seconds. Begin with an ordinary overcast weekday commute on an elevated urban line. Grey apartment blocks, rooftop water tanks, a scrapyard, overhead wires, a canal and traffic on a road below streak past the left windows at different distances with realistic parallax. The passengers are tired and mostly uninterested — some on phones, some staring out, some talking quietly. A calm public-address chime sounds and an announcer quietly says, "Next stop: Hell." Nobody reacts. One passenger glances up briefly and goes back to their phone. The train accelerates and everyone instinctively tightens their grip as the carriage gives a hard lateral jolt.
3 to 6 seconds. The line drops into a cutting and then into a tunnel. Tiled subway walls, cable runs, service lights and a passing platform strobe across the windows in hard bands of light and dark, throwing the carriage into stuttering illumination. The tunnel ages as the train descends: modern concrete becomes older brickwork, then rough-cut stone. The fittings become scorched and soot-caked — blackened signal lights, corroded brackets, cabling burnt down to bare metal. This route has been running a long time. The gradient steepens noticeably; passengers lean back against the pitch and the straps hang visibly off-vertical.
6 to 9 seconds. The tunnel wall becomes raw geology and the train is now clearly descending through the crust at impossible speed. Through the left windows the rock face streaks past in visible strata — pale limestone, dark shale, red iron-stained bands, seams catching the carriage light. The air begins to heat and passengers loosen collars. Then, under the mechanical roar, a sound arrives before anything is visible: a vast, distant, continuous mass of human voices, far away and heavily reverberant, never resolving into individual words. It is not loud. It is simply there, and it does not stop. Two passengers look at each other. Nobody says anything.
9 to 11.5 seconds. The rock wall falls away entirely and the train emerges into a void larger than a city, with no visible far side. It runs along a ledge on the wall; below the windows the space drops away into darkness. Stone columns kilometres tall stand in the void with strong parallax — near ones sweeping past, distant ones barely shifting. Far below and far ahead, a faint orange glow is already visible. It must already exist in the frame from this moment and grow continuously from here forward without ever popping or suddenly enlarging. The voices are louder here, spread across an enormous space. One passenger near the glass leans forward slightly. Another quietly says, "Ohh."
11.5 to 14 seconds. The train passes a station. A soot-blackened platform is cut into the rock wall — worn edge, dead lamps, faded markings unreadable under the grime — and figures are standing on it in an orderly queue, waiting, facing the track, completely motionless. They are seen for less than a second at speed, as dark shapes against the platform light, never close enough to read faces. The train does not slow. Nobody boards. The platform is gone behind the carriage. This is the moment the passengers understand. Several stop looking at their phones. Embers begin drifting upward past the windows from below.
14 to 16 seconds. Ahead, a wall crosses the entire void — a single continuous barrier extending beyond sight in both directions and upward past the ceiling. The train passes through an immense gateway cut into it, an arch hundreds of metres tall, its stone eroded smooth by an unimaginable volume of traffic, its carved markings worn past legibility. For half a second the carriage is in the shadow of the arch and everything goes dark. Then it is through, and the light on the far side is completely different: warm, hard, and coming from below.
16 to 19 seconds. The train races above a burning plain. Outside the left windows: a vast crusted expanse of dark solidified ground cracked into slow-moving plates with brilliant orange fissures running between them, distant fountains of molten material rising and falling in slow motion because of their true scale, and smoke columns standing kilometres high. Running across it are raised stone causeways, and on the causeways are crowds — dense, continuous, moving slowly, all in the same direction, extending to the limit of visibility. They are seen only as masses at distance, never in detail. The light entering the carriage is now dominant and hard, throwing sharp orange edges on faces, poles and straps with deep black shadows behind them. The interior air is visibly hazy. The voices are constant.
19 to 21 seconds. The plain ends at the rim of a colossal pit — a shaft so wide the far wall is only a suggestion in the haze. The train races over the edge and begins descending along the interior wall. The wall is terraced: enormous concentric ledges receding downward, each deeper and darker than the last, disappearing into smoke layers. The descent acceleration pushes the passengers down and forward against the poles. Ash begins striking the windows and streaking backward.
21 to 23 seconds. The terraces are populated. Endless slow processions move along every ledge, strings of figures following the curve of the wall down into the smoke and out of sight, lit from below by the fires beneath them. There is no chaos in it — it is orderly, patient and entirely without end, which is worse. Furnace mouths open in the rock face, glowing white at their throats, and long queues stand before them. Everything is at distance. Nothing is close enough to resolve. The light entering the carriage takes on a deeper red and the fluorescents are overwhelmed. Passengers nearest the window press slightly closer to the glass.
23 to 25 seconds. Enormous shapes move among the terraces — figures many times the height of the crowds around them, walking slowly along the ledges, their scale established only by comparison. They do not look at the train. Distant winged forms cross beneath the carriage at low altitude with slow heavy wingbeats appropriate to their size and vanish into the smoke. Stone pens and barred openings are cut into the wall in their thousands, receding into the haze like housing. Ash accumulates in the corners of the windows. The carriage is hot enough that the glass fogs and clears in waves.
25 to 27 seconds. The train reaches the deep layer and the full extent of Hell reveals itself. It is not a city on a human plan and it is not a cavern. It is a continuous inhabited geology extending in every direction — towers of black stone fused into cliff faces, vaults hollowed out of the rock at cathedral scale, rivers of molten material running in cut channels between districts and falling in slow luminous cataracts to levels further down, bridges spanning gaps kilometres across, and everywhere on all of it, crowds. Smoke columns rise for kilometres and flatten against unseen ceilings. Every level is lit by its own fires. The train races along a ridge line through the middle of it; arches and aqueduct spans pass overhead and terraces stream past below the window with violent parallax. It is impossible to see where any of it ends.
27 to 29 seconds. The depths continue past the windows. The camera struggles badly with the contrast, blowing out the fires and crushing everything else into noise. Ash cakes the corners of the glass. Passengers are pressed against the windows in total silence, faces lit from below in hard orange, expressions stunned and completely still. One person quietly whispers, "What is that?" Nobody answers. Autofocus hunts between the ash on the glass and the world beyond it.
29 to 30 seconds. The fires end. The train passes out of the smoke and the lowest region opens — and it is not burning. It is a vast frozen plain stretching beyond the visible horizon, an impossibly wide sheet of dark grey ice, absolutely still, lit by nothing but a faint pale glow from within itself. Shapes are visible held motionless within it, spaced far apart, receding to the horizon, never shown in detail. Every exterior sound stops at once — the voices, the fires, everything — leaving only the carriage. The heat drops out and frost blooms instantly across the outside of the windows. The scale feels planetary. The passengers stare in complete silence. At approximately 29.7 seconds the calm PA voice says quietly, "Welcome to Hell." The train continues moving. The camera keeps shaking naturally as the frozen expanse extends endlessly beyond the left window.
Lighting
The lighting must evolve naturally throughout the journey. Begin with flat overcast daylight around 6500K. In the tunnels use harsh strobing bands from passing service lights against near-total dark, then let the sources become sparse until the carriage's own weak fluorescent tubes are the only illumination. In the void the interior lights fall off into nothing. Then introduce a growing warm orange from below — first a faint wash on the lower half of faces, then mixing with the cold interior white, then completely dominating it: hard, high-contrast, sharp-edged, with deep black shadows. Above the burning plain it should be strong enough to blow out the phone's sensor at the window. In the deep layer it becomes red-orange and omnidirectional from countless fires at every distance. In the final second all warm light vanishes and is replaced by a faint pale luminance from the ice, cold and almost sourceless. All exterior light must enter naturally through the LEFT-SIDE WINDOWS and fall across passengers' faces and clothing at every stage. Do not add artificial interior lighting to create the colours.
Passengers
Keep all passengers completely ordinary throughout the journey. Realistic skin texture, subtle capillary variation, natural blinking, breathing, eye movement and imperfect facial symmetry. Clothing has realistic folds and responds to acceleration. Passengers must continuously perform small independent movements: shifting weight, adjusting grip, loosening a collar in the heat, wiping fog from the glass, turning their heads, tightening their hands around poles. Do not make them freeze. Do not make everyone react at the same moment. Do not let them perform dramatic acting.
Their emotional progression is subtle: indifference at the announcement → mild confusion in the tunnels → unease when the voices arrive → the exact moment of understanding as the waiting platform passes → dread on the descent → complete stunned silence in the depths. No screaming. Only one soft "Ohh" at the first glow and one whispered "What is that?" near the end. The most powerful reaction is no reaction — just faces pressed against the glass, lit from below by something that should not exist.
Physics
The train always travels forward and always downward after the tunnel. Passengers sway according to acceleration and lean against the gradient; straps hang off-vertical on the descents. Loose clothing reacts to movement. Exterior objects have different velocities and distances with strong realistic parallax at every stage — near stone columns sweep past, mid-distance terraces move steadily, distant districts barely shift. Ash and embers must move independently of the train, drifting upward on thermal currents rather than streaming with the carriage. Heat must be expressed physically: window fogging and clearing, haze inside the carriage, frost at the very end. Enormous figures and flying forms must move slowly, because they are enormous. Nothing teleports, nothing freezes, nothing suddenly appears, and nothing changes position without physical cause.
No transitions
There are absolutely no visual transitions. No dissolves, crossfades, portals, morphing or ghosted overlays. Every environment change happens because the train physically travels into the next environment. The city becomes a cutting, the cutting becomes a tunnel, the tunnel becomes bedrock, the bedrock opens into the void, the void contains the waiting platform, the void ends at the wall, the gateway gives onto the burning plain, the plain ends at the pit rim, the rim becomes a terraced descent, the terraces fill with processions, the processions become the full inhabited depths, and the depths open onto the frozen plain. It must feel like one physically continuous impossible journey where every step is the inevitable consequence of the step before it.
Audio
Audio must be entirely diegetic with no music. Use cheap compressed smartphone microphone quality. The train produces a constant deep maglev roar, mechanical clatter, rattling poles, vibrating windows, swinging straps and low-frequency carriage rumble, plus passenger breathing and clothing movement. In the tunnels add hard reverberant slap-back and pressure changes as the train passes openings.
The defining sound of this video is the voices. They must arrive at 6 to 9 seconds, before anything is visible, as a distant continuous mass of human voices with enormous reverberation, never resolving into words and never rising to screaming. From that point they never stop. They grow with depth and spread across a wider stereo field as the spaces open up. Beneath them, a very low frequency resonance builds steadily, felt more than heard. Above the burning plain add a broad distant roar and irregular deep concussions at long intervals. Ash strikes the hull as fine irregular ticking. In the depths the voices are vast, layered at many distances, and mixed with immense indistinct industrial-scale sound that never resolves into detail.
In the final second every exterior sound stops at once — total silence outside, leaving only the carriage, the passengers' breathing, and ice creaking against the hull. At approximately 29.7 seconds the calm PA voice says quietly, "Welcome to Hell."
Style and content limits
The overall style must remain dirty photorealistic amateur smartphone footage despite the spectacular environments. Vertical 9:16, real-time speed, no slow motion, no stabilisation, no cinematic camera, no perfect composition, no clean VFX presentation. Use heavy shadow noise, blown highlights, low-bitrate compression, rolling-shutter skew, autofocus hunting, auto-exposure pumping, smeared motion blur and crooked framing. The camera should occasionally be partially blocked by a shoulder, arm or nearby passenger, and should struggle badly with the extreme contrast between fire and darkness.
Hell is conveyed through scale, architecture, crowds at distance, fire, ash and sound. Do not show suffering, injury, bodies, blood, restraints or any graphic content. Every figure outside the train stays a distant silhouette or part of a mass — none is ever close enough for its face or condition to be read. Do not use recognisable religious iconography or legible text of any kind; all markings are eroded past legibility. Do not make it look like a game cinematic or a clean VFX render. The world outside should be overwhelming and physically impossible, but the recording itself must look raw, accidental and believable.
The most important rule
ONE SINGLE CONTINUOUS TAKE. No cuts, no jump cuts, no dissolves, no crossfades, no scene resets and no artificial transitions. The same train, same passengers and same camera remain present from beginning to end. The orange glow must be visible as a faint distant smudge from the first cavern and grow continuously and inevitably until it fills the window. The voices must arrive before the fire and never stop until the final second. The heat must build physically across the whole descent so that its total disappearance at the end lands as a shock. The population must emerge gradually — the waiting platform, then crowds on the causeways, then processions on the terraces, then the full inhabited depths — never appearing all at once. End while the train is still moving, the frozen plain extending endlessly beyond the left window, passengers silent, carriage still shaking, the PA announcement fading into the sound of ice.
Show more
The Three Tailwinds Nobody's Pricing Into Nvidia
🔊
On Monday's edition of The Macro Show, Director of Research Daryl Jones
@HedgeyeDJ and $HGRO Portfolio Manager Sam Rahman
@SamofAmerica explained how a PM actually positions into a print like Nvidia's, $NVDA.
As Rahman noted, he owns both Nvidia and Marvell, $MRVL — the only two semis currently on the signal strength list — but keeps his active Nvidia position small, with far bigger active weights in software names like Amazon, $AMZN and Microsoft, $MSFT.
His approach into big benchmark events is not to overthink the positioning, because get on the wrong side of the news flow and you can get hurt badly.
The number that frames the quarter, Rahman explained, is that Nvidia's data-center business has the potential to nearly double year-over-year, an extraordinary feat for a company its size: "Think about the size of this company doubling its core sales line year over year."
Beyond the print, he flagged three tailwinds:
1. The transition from Blackwell to the faster, more efficient Vera Rubin chip, with hyperscaler orders already landing.
2. SpaceX emerging as a large, genuinely incremental new buyer of data-center infrastructure post-IPO.
3. A roughly 15% price increase on chips and servers next year, driven by memory-chip inflation, that flows straight to revenue.
Subscribe to The Macro Show for daily market analysis you won't find anywhere else:
Show more
🎶 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.”
Show more
MiniMax H3 (Hailuo 3) video on 2× DGX Spark 🎬
• 5 s 1344×768 clip with audio: 643 s (one Spark: 1399 s)
• 10 s: 1690 s, 15 s: 3080 s (one Spark: out of memory)
• 20 s: 6740 s, beyond the model's 15 s spec and still coherent to the last frame (clip below, one continuous generation, seed 1101, 20 steps)
• Text-encoder TP2 + Ulysses SP2 over one ConnectX-7 cable (RoCE, 19.5 GB/s)
• Online FP8 DiT, cuDNN attention, 24 fps, native audio
• Frames leave the GPU as uint8, VAE tiles decoded on both GPUs: same bytes, 15 s fits with 25 GB to spare on each host
• H3-Base open weights (768p), vLLM-Omni, no Ray and no changes to the vLLM-Omni tree
vllm serve cannot span two hosts, so the recipe adds a cross-host executor: rank 1 runs on the second Spark and is driven over the link through vLLM's own message queues, behind the unchanged single-node API. Quality checks, memory traces and
the open levers are in the repo.
Full recipe, measurements and credits:
Show more
Nice expert call on $CBRS with a Director at Figma who's sending 40% of AI traffic to Cerebras' cloud
There are clear use cases for hyperfast inference and the customer is willing to pay a premium for these fast tokens
Figma spent roughly a year building an in-house language model for its assistant product, a chat interface intended to make Figma AI-native. The model is in the 50 to 100 billion parameter range, which he characterizes as roughly five to six times smaller than the frontier models it replaces, and better in quality for the specific task. Figma ships the weights to Cerebras and Cerebras handles all serving and optimization. He explicitly frames the relationship as a cloud provider relationship, not a hardware purchase.
The AI traffic split:
- 20% of assistant traffic routes to external frontier models (Anthropic and others) for general question answering and RAG-type requests. This share is structurally permanent. Figma has no intention of training its own model for it.
- 80% routes to the in-house model.
- Of that 80%, half goes to Cerebras, so roughly 40% of total assistant traffic.
- The remaining 40% goes to Fireworks, Modal and Baseten.
Figma Make, the vibe-coding product, sits in the same org but runs almost entirely on external models from Anthropic and Google. Critically, the AI assistant is currently exposed to only around 30% of the Figma user base.
Current Cerebras spend is described as low to mid seven figures annually. His forward view: "I would expect at least in the next one year to easily double... maybe 2X-3X growth in the next one to two years."
Beyond that he expects the use case to plateau, since the core objective is penetrating the existing user base rather than acquiring new users. This is a useful anchor for how a mid-size, non-AI-native enterprise customer scales on this platform: fast doubling off a small base, then aiming to flatten once penetration completes, with the next leg dependent on new products rather than the same product growing.
# Why Cerebras Does Not Get the Other 40% of Traffic
1. Capacity is reserved, not metered. Figma has weekday peaks and effectively zero weekend and overnight traffic. Reserving for peak means paying for idle silicon most of the week, so Figma deliberately sizes its Cerebras reservation at only 40% to 50% of average traffic and sends the peaks to on-demand GPU providers. This single design choice is what creates the opening for Fireworks, Modal and Baseten.
2. Reliability - Error rates are still above what he wants from a production system, which forces Figma to maintain fallback paths.
3. Observability - He wants first-try success rates, retry counts and internal failure data, which he is not currently getting, and without which he cannot engineer around the reliability gap.
"If Cerebras did not have any limitation, we would've used the whole 80% of traffic through Cerebras."
# The Deployment Friction, Scored 7 out of 10
Model updates are not self-serve. Figma has to notify the Cerebras team roughly a week ahead. Hand a model over Monday or Tuesday, and it is deployed by Friday. Two to three days per iteration, versus effectively immediate self-serve deployment at every GPU-based competitor.
He rated the pain at seven out of ten on a scale where ten is severely damaging.
"It's of course something we can live with, but it definitely slows our execution a lot... we build a model, we do some testing, we do this three-day wait, do some testing, find that there is a small bug, and then we have to retrain the model."
This is a developer infrastructure maturity gap, and it is the kind of thing that does not show up in benchmark comparisons but does show up in renewal conversations and in how much of a customer's roadmap a vendor can capture.
# The Price of Speed
Against the frontier model, running Figma's own smaller model on Cerebras costs roughly half to one third as much for equivalent traffic. But the model is five to six times smaller, so the like-for-like inference premium is real.
Against a GPU-based provider running the same model, his estimate is that Cerebras is around 30-50% more expensive. Figma has not run a formal side-by-side, which he attributes to the fact that reserved pricing versus per-token pricing makes the answer entirely dependent on the traffic profile.
In a bake-off against hyperscalers, Modal, Baseten, Fireworks and Groq, Cerebras came out roughly 10 to 15 times faster than most of the market. Figma found Cerebras through the Artificial Analysis public benchmark, then spent months on POCs specifically to de-risk infrastructure maturity before scaling to production.
The latency threshold he describes is a genuine product constraint rather than a nice-to-have. For assistant tasks a user could perform manually in about 30 seconds, an AI response taking longer than that adds no value. Delivering in three to five seconds changes the product.
"Speed is extremely valuable... even that 50% to 2X increase in price, I think is totally something we are willing to pay for the speed."
Cost optimization at scale would likely not come from switching hardware. It would come from shrinking the model further or routing simple requests to a smaller model.
"Maybe a workflow user makes a request, it takes three hours to run. If that's the case, then speed is not a big criteria... For those use cases, we may switch to the cheaper inference provider."
Figma has not built an in-house model for Figma Make because the output format is React and HTML, which frontier models have seen extensively, leaving little quality headroom. But he notes cost and latency wins are still available, and if cost becomes a concern as Make scales, Figma would likely train a smaller model and host it on Cerebras for the base load. Additional projects beyond the assistant are early stage, with some expected to scale in the second half of the year.
He also notes Figma experimented with Groq prior to its acquisition by NVIDIA.
$CBRS $NVDA
Show more
$HUBB KEY READ-THROUGHS FROM HUBBELL’S Q3 2026 EARNINGS CALL
Hubbell’s results reinforce a durable North American electrical-infrastructure investment cycle, but the most actionable implications concern the divergence between demand, shipment capacity and earnings conversion. Utility backlog increased by more than $500 million year to date, data-center sales grew more than 80% organically, and substation components grew more than 20%. Conversely, control-house capacity expanded more slowly than expected, inflation required additional pricing actions, and tariff refunds materially supported reported margins. The strongest positive read-throughs favor qualified grid-equipment suppliers, electrical balance-of-system vendors and transmission contractors. The clearest negative read-through concerns expectations for rapid margin recovery despite persistent cost inflation. Country designations below identify primary listing markets, not corporate domicile. Impact assessments describe fundamental exposure and incremental information value, not predicted share-price returns.
ELECTRICAL EQUIPMENT
PRICE-COST RECOVERY IS A MORE IMMEDIATE EARNINGS RISK THAN DEMAND WEAKNESS (READ-THROUGH 1)
Affected company: Eaton (ETN: United States). Direction: Negative for near-term margin-recovery confidence, despite a positive demand backdrop. Magnitude: Moderate potential consolidated earnings sensitivity; high relevance to Electrical Americas margin expectations.
The critical commentary was that Hubbell’s Q3 price-cost-productivity equation was approximately neutral only “including the net impact of IEEPA tariff refunds.” Approximately $30 million of refunds, less approximately $5 million of customer considerations, produced a $25 million net benefit. Removing that benefit analytically reduces Hubbell’s adjusted operating margin from 25.5% to approximately 24.1%, compared with approximately 23.9% a year earlier. Thus, most of the headline margin expansion was not evidence that ordinary pricing and productivity had decisively overtaken inflation.
This matters particularly for Eaton because its previously disclosed Electrical Americas margin recovery depends heavily on precisely that equation. On its Q2 call, Eaton attributed approximately 300 basis points of expected H2-versus-H1 margin improvement to price-cost recovery, supplemented by output and productivity. Eaton also expected negative price-cost to become approximately neutral during H2. Hubbell’s October disclosure that inflation had intensified and that Electrical pricing was being accelerated into November is therefore incrementally adverse evidence for the ease of achieving that recovery, rather than evidence of weaker electrical demand.
The transmission mechanism is a lag between higher material, purchased-component and labor costs and their recovery through realized selling prices. Backlog can extend that lag when older orders retain earlier pricing. Even complete dollar recovery can dilute margin percentages because revenue and costs rise without a corresponding increase in operating profit. Higher volume therefore does not guarantee the anticipated margin progression.
The immediate trading catalyst is Eaton’s 11/03/26 earnings release, particularly Electrical Americas price-cost realization, factory productivity and the Q4 margin bridge. Hubbell does not establish that Eaton will miss its forecast: Eaton implemented pricing earlier, has different contracts and retains company-specific productivity opportunities. The high-conviction conclusion is narrower but important: a demand-led revenue beat should not automatically be interpreted as validation of the margin-recovery thesis. Over 2027, successful repricing could reverse this pressure, making the issue one of earnings timing and execution rather than a broken electrical-investment cycle.
SUBSTATION STRENGTH SUPPORTS GRID-EQUIPMENT DEMAND MORE DIRECTLY THAN GENERAL AI SPENDING (READ-THROUGH 2)
Affected company: GE Vernova (GEV: United States). Direction: Positive. Magnitude: High fundamental relevance to Electrification; moderate incremental relevance to consolidated earnings because Power and Wind have separate drivers.
Hubbell reported more than $500 million of year-to-date utility backlog growth, approximately 30% above the beginning-of-year level, with particular strength in transmission and substation projects. Substation components, including connectors, bushings, arresters and switches, grew more than 20%. Management also described a growing 765-kV project pipeline, with an initial project beginning shipments around year-end 2026 and into 2027.
These are physical grid-development signals rather than indirect inferences from hyperscaler capital budgets. Substations require coordinated procurement of transformers, switchgear, protection equipment, connections and control systems. Hubbell’s order and component-shipment strength therefore supports the duration of the addressable market for GE Vernova’s Power Transmission and Grid Systems Integration businesses. The economic channel is stronger equipment demand, improved capacity utilization and a longer period over which manufacturing investment can earn returns.
GE Vernova had already reported $6.3 billion of Q2 Electrification orders, approximately 1.7x segment revenue, and $40.6 billion of equipment backlog. Hubbell’s disclosure corroborates that established demand picture; it is not a newly identified GE Vernova award. The incremental information is that demand remained strong into Q3 and that constraints extended beyond the most widely discussed large equipment categories into smaller components and integrated substation systems.
The immediate catalyst is GE Vernova’s 10/28/26 earnings release. The relevant tests are Electrification orders, shipment capacity, delivery schedules and margins, rather than gas-turbine reservations alone. The longer-duration implication extends through 2027 and subsequent transmission investment cycles, including higher-voltage infrastructure.
The important limitation is project sequencing. Strong component orders may accompany transformer or switchgear orders already placed, so the same project should not be counted again as entirely new future equipment demand. Hubbell’s control-house delays also warn that excellent equipment bookings can coexist with uneven project completion. The strongest positive conclusion concerns demand duration and backlog quality, not an automatic acceleration of near-term revenue recognition.
ENGINEERED ELECTRICAL HOUSES ARE A CAPACITY BOTTLENECK, NOT AN UNDIFFERENTIATED FABRICATION MARKET (READ-THROUGH 3)
Affected company: Powell Industries (POWL: United States). Direction: Positive for the value of qualified engineering and manufacturing capacity. Magnitude: Moderate incremental information value, with potentially high longer-duration relevance to Powell’s relatively concentrated business.
Hubbell disclosed that its substation portfolio represents approximately 20% of Utility Solutions, divided roughly equally between components and modular control houses. Components grew more than 20%, but control-house revenue was flat sequentially from Q2 to Q3 because capacity proved harder to ramp than expected. Management stated that expansion had “just proved a little harder” and would take longer to translate into growth.
The implication is that component availability alone does not resolve the substation bottleneck. Engineering, systems integration, fabrication space, skilled labor, testing and delivery coordination can constrain the output of complete electrical assemblies. Those capabilities create economic value for established suppliers able to deliver qualified, project-specific systems.
Powell’s engineered electrical distribution equipment and power control rooms provide relevant exposure. Its fiscal Q3 disclosures already showed approximately $2.4 billion of backlog and $934 million of quarterly orders, including substantial utility and data-center activity. Hubbell therefore provides independent confirmation of the demanding industry backdrop in which Powell is expanding capacity.
The transmission mechanism is sustained demand for engineering-intensive assemblies, utilization of qualified manufacturing space and the ability to compete on delivery reliability rather than component price alone. However, Hubbell’s utility control houses and Powell’s various power control rooms are not universally interchangeable products. No transfer of a Hubbell customer order to Powell was disclosed, and market-share gains should not be assumed.
The longer-duration catalyst is Powell’s manufacturing expansion, including a planned approximately 300,000-square-foot facility expected to become available late in fiscal Q2 or early in fiscal Q3 2027. Its Jacintoport expansion initially targets LNG-related power control rooms, so that capacity should not be modeled immediately as dedicated utility or data-center capacity.
Near-term results remain dependent on Powell’s own project milestones and staffing. The actionable distinction is between a favorable market for qualified capacity and the ability to commission that capacity successfully. Hubbell strengthens the former thesis while simultaneously highlighting the execution risk embedded in the latter.
DATA-CENTER ELECTRICAL INFRASTRUCTURE
COMPONENT CONTENT AND CAPACITY ARE EXTENDING GROWTH BEYOND THE LARGEST POWER SYSTEMS (READ-THROUGH 4)
Affected companies: Legrand (LR: France), nVent Electric (NVT: United States), and Eaton (ETN: United States). Direction: Positive for relevant data-center electrical products. Magnitude: High relevance to Legrand’s and nVent’s growth profiles; moderate-to-high relevance to Eaton’s electrical businesses, with less direct applicability to their unrelated operations.
Hubbell’s data-center sales increased more than 80% organically. Management explicitly attributed the performance to “content gains, new product introductions and capacity additions,” highlighting wiring devices and Pin & Sleeve products. This is important because the growth was not described merely as more data centers being constructed. Hubbell was supplying more of the available electrical component opportunity while removing production constraints.
The precise transmission mechanism is greater revenue opportunity within each installation, combined with rising facility deployment. Power-distribution connections, cable management, busway, enclosures and other supporting electrical products can expand alongside larger UPS, switchgear and transformer systems. Manufacturers with broad product portfolios, specification relationships and sufficient capacity can participate in more categories within the same project.
Legrand offers particularly direct exposure through critical power, busway, electrical distribution and related data-center infrastructure. Its H1 disclosure placed data centers at 32% of group revenue, with organic growth above 30%. nVent had already reported strong growth in cable management and engineered buildings, alongside liquid cooling, and expected more than $2 billion of 2026 data-center sales. Eaton’s electrical businesses also address the broader power-distribution requirements of these facilities.
Hubbell strengthens the electrical-component demand case, but its 80% growth rate cannot be transferred to these peers. Some of that growth reflects Hubbell-specific penetration and capacity additions, potentially including competitive gains. Its wiring-device results do not directly validate nVent’s liquid-cooling growth rate, and the call does not establish the eventual winners from changes in data-center power architecture.
Near-term catalysts are nVent’s 10/30/26 results, Eaton’s 11/03/26 results and Legrand’s 11/05/26 results. The critical disclosures are electrical-product orders, capacity conversion and customer deployment schedules. These peers had already disclosed strong data-center activity, so the new signal is continued strength into Q3, not discovery of an unrecognized end market.
The longer-duration opportunity extends into 2027 as new capacity and broader product offerings support additional content capture. Geographic interpretation is important: Hubbell most directly corroborates North American deployment. It does not establish a comparable recovery in European construction or Chinese industrial demand.
TRANSMISSION ENGINEERING AND CONSTRUCTION
BACKLOG AND 765-KV ACTIVITY SUPPORT A MULTI-YEAR CONSTRUCTION PIPELINE, NOT NECESSARILY AN IMMEDIATE REVENUE SURGE (READ-THROUGH 5)
Affected company: Quanta Services (PWR: United States). Direction: Positive for transmission and substation activity. Magnitude: High relevance to the electric-infrastructure business and material consolidated exposure; moderate incremental information value because Quanta already has substantial committed work.
Hubbell’s utility orders continued to exceed expectations, its backlog increased by more than $500 million year to date, and management described transmission and substation investment as being in the early stages of a multi-year cycle. It also indicated that capacity additions would need to continue beyond 2027 to serve subsequent years. The developing 765-kV pipeline provides a more specific signal of large, technically demanding transmission projects.
The transmission mechanism runs from approved utility investment and equipment procurement into engineering, site preparation, line construction, substation installation, testing and commissioning. Quanta participates directly in those execution activities. A longer and broader pipeline can support engineering utilization, specialized labor deployment and multi-year customer programs rather than reliance on isolated construction awards.
Quanta had already reported $53.4 billion of total backlog in Q2. Hubbell’s evidence strengthens the durability of that opportunity set but does not establish additional Quanta bookings. Some of the component orders may support projects already included in Quanta’s backlog or work already underway. Treating every upstream order increase as a new contractor award would double-count portions of the same investment cycle.
The immediate catalyst is Quanta’s 10/29/26 results, with emphasis on electric-infrastructure bookings, project starts, workforce utilization and 2027 visibility. The more substantial fundamental effect lies in 2027-2029, as higher-voltage networks and associated substations progress through construction.
The negative qualification is delivery sequencing. Hubbell’s control-house constraints can leave individual projects waiting for essential assemblies even when engineering and other equipment are ready. That creates schedule risk, but not a sufficient basis to infer a broad Quanta cash-flow problem. Quanta’s own Q2 release reported approximately $0.9 billion of free cash flow, providing clear counterevidence to any automatic transfer of Hubbell’s working-capital pressure. The preferred read-through is stronger demand duration with project-level timing risk, not an indiscriminate acceleration of quarterly revenue estimates.
ELECTRICAL AND UTILITY DISTRIBUTION
SUPPLY COMPLEXITY INCREASES THE VALUE OF PROCUREMENT AND LOGISTICS, BUT ALSO THE CAPITAL REQUIRED TO DELIVER GROWTH (READ-THROUGH 6)
Affected company: Wesco International (WCC: United States). Direction: Positive for demand and service relevance, with a negative working-capital qualification. Magnitude: Moderate-to-high revenue relevance because Wesco participates in both data-center infrastructure and utility supply chains; cash-flow implications depend on inventory and customer-payment execution.
Hubbell reported strong data-center, transmission and substation demand, while describing component shortages, uneven capacity availability and continuing inflation. It also cited constructive distributor discussions around large projects and ongoing success in cross-selling and channel conversions.
For a distributor and supply-chain services provider, these conditions create value beyond simply reselling equipment. Coordinating specifications, sourcing multiple product categories, staging inventory and sequencing delivery can reduce the risk that a small missing component delays a much larger project. The economic channel is higher addressable project spending and greater demand for procurement, logistics and inventory-management services.
Wesco’s own Q2 results already showed approximately $1.5 billion of data-center sales, growth of approximately 45%, and total backlog growth of approximately 60%. The company also disclosed a significant multi-year Grid Services award from a hyperscale customer. Hubbell therefore corroborates both sides of Wesco’s exposure: electrical infrastructure inside the data center and the utility infrastructure required to serve it.
The positive read-through is not that Hubbell’s share gains automatically accrue to Wesco. Distributor allocation, project awards and commercial terms were not identified on the Hubbell call. Rather, strong demand across multiple constrained product categories increases the relevance of Wesco’s established distribution and service capabilities.
The negative channel is cash intensity. Higher unit costs, larger projects and uneven equipment delivery can require more inventory and receivables before cash is collected. Wesco’s Q2 free cash flow was positive at $32 million but modest relative to its $6.7 billion revenue base, illustrating why revenue growth and cash realization must be evaluated separately. Hubbell’s experience reinforces this monitoring requirement without proving deterioration at Wesco.
The near-term catalyst is Wesco’s Q3 reporting cycle in late October, particularly backlog conversion, inventory, receivables and supplier-price recovery. Hubbell’s November pricing action also creates a possible distortion from purchasing ahead of price increases, although the call did not establish that such behavior had occurred. The longer-duration benefit is a larger role for integrated procurement and project logistics; the investment condition is that service value and earnings growth translate into acceptable returns on the additional working capital.
REGULATED UTILITIES
GRID INVESTMENT IS WELL SUPPORTED, BUT CONTRACTED LOAD SHOULD NOT BE EQUATED WITH IMMEDIATE CASH EARNINGS (READ-THROUGH 7)
Affected company: American Electric Power (AEP: United States). Direction: Positive for the duration of transmission and distribution investment; negative for assuming rapid, frictionless monetization of large-load commitments. Magnitude: Material long-term capital-allocation relevance, but limited evidence for an immediate EPS revision.
Hubbell’s backlog growth and strong transmission and substation orders support the need for substantial utility investment. However, the same call showed that modular control-house output failed to ramp as expected, despite strong demand. Management also indicated that capacity additions would need to continue across several years.
The positive transmission mechanism is additional investment in infrastructure needed to connect and reliably serve load, supporting regulated asset growth where projects receive appropriate approval and cost recovery. AEP has particularly relevant exposure through its transmission footprint and large-load development. Its Q2 disclosures outlined a $78 billion 2026-2030 capital plan and 69 GW of contracted load additions through 2030.
The negative mechanism is the gap between customer commitments, construction spending and operational readiness. A substation cannot necessarily be commissioned merely because the line, transformer or other major components have arrived. Missing control and protection assemblies can defer energization. Depending on the regulatory framework, accounting earnings may receive support during construction, while cash recovery and customer electricity consumption occur later. Strong project demand can therefore coexist with elevated funding requirements.
This is an industry execution read-through, not evidence that a named AEP project has been delayed by Hubbell. No specific supplier allocation was disclosed. AEP also has important protections that limit the bearish interpretation: its own commentary described long-term customer commitments, large-load tariffs and take-or-pay arrangements intended to mitigate development-timing risk. Those protections vary by project and should not be assumed to cover every commitment identically.
The near-term catalysts are AEP’s fall 2026 capital-plan update and subsequent disclosures on energization milestones, customer contributions and financing. The longer-duration effect extends through the 2026-2030 investment program. The high-conviction portfolio implication is that grid-equipment demand and regulated-utility growth can both be attractive, but their cash-flow timing is different. Equipment bookings provide evidence of investment activity; they do not by themselves establish when the utility will begin collecting the associated operating cash flows.
Show more
AI CONFERENCES AND INVESTOR DAYS, TUESDAY, OCTOBER 27, 2026
TODAY
DCD INVESTMENT FORUM MENA CONFERENCE PRESENTATION
Tue Oct 27, Dubai, United Arab Emirates
DCD Investment Forum MENA (DatacenterDynamics), October 27, 2026, Grand Hyatt Dubai, Dubai. Co-located with DCD Connect MENA. Part of the DCD Investment series.
No public webcast. In-person event. Registration:
• Setup: A one-day executive forum with 35+ speakers, convening C-suite operators, developers, lenders and investors active in MENA digital infrastructure deal flow. This is the capital side of the MENA buildout, as sovereign AI programs turn announced gigawatts into financed projects
• The number: 2,500+ senior professionals expected across the co-located events, against a regional pipeline of 2.6 GW in development and 13.8 GW planned on roughly 1.6 GW live
• Why it matters: Who finances the Middle East pipeline, and on what terms, decides how much of it gets built
• Debates: Sovereign capital versus international project finance, power and grid access as the gating item, and phased 100 MW to 200 MW delivery versus gigawatt announcements
• Watch: The opening macro panel on where MENA digital infrastructure heads next, and any named financing or partnership announcements
CELESTICA INC INVESTOR DAY
Tue Oct 27, 8:00 AM ET
Investor and Analyst Day call, hosted by Chair and CEO Rob Mionis, with an in-person Lunch Forum following in Midtown Manhattan for institutional investors and sell-side analysts.
Webcast:
• Setup: Q3 results, reported after the close Monday October 26, beat the top of guidance: revenue rose 83% year over year to $5.83B against a $5.25B to $5.55B guide, adjusted EPS rose 109% to $3.30, adjusted operating margin hit a record 8.5%, and CCS revenue rose 102% to $4.87B. Management raised the full-year 2026 outlook to $21.0B of revenue and $11.55 of adjusted EPS, cut free cash flow guidance to $460M from $600M for targeted working capital investment ahead of 2027, and gave an initial 2027 outlook of $40.0B of revenue, up 90%, with a 9.0% adjusted operating margin and $21.50 of adjusted EPS on capex of $1.5B to $2.0B
• The number: The $40.0B 2027 revenue target, well above the $35.7B analyst estimate reported ahead of the print, and whether the day substantiates it with contracted demand and 1.6T program detail
• Why it matters: Celestica is the purest listed read on hyperscaler AI hardware demand outside the hyperscalers themselves, and the day is the first long-form test of the 2027 frame
• Debates: Customer concentration, the cash cost of growth as working capital builds, and whether 9% margins hold as capacity capex rises
• Watch: Q4 guidance of $6.275B to $6.575B of revenue, 2028 demand commentary (management said customer demand forecasts are meaningfully higher than the 2027 outlook), and 1.6T program updates
MACRONIX INTERNATIONAL CO LTD INVESTOR DAY
Tue Oct 27, 4:00 PM Taiwan time (4:00 AM ET)
Third-quarter 2026 results webcast, hosted by Chairman and CEO Miin Wu and Vice President Paul Yeh, per the company investor relations calendar.
Webcast via Macronix investor relations. Direct link unconfirmed.
• Setup: The NOR flash maker is in a steep memory upcycle. September sales rose 188% year over year to TWD 8.37B and third-quarter revenue reached TWD 24.24B, up 27% sequentially
• The number: Nine-month 2026 revenue of TWD 53.83B, up 155% year over year, already above the prior full-year record of TWD 50.57B set in 2021, with Q2 gross margin at 64.4%
• Why it matters: Macronix is a clean read on legacy memory pricing power and how far the upcycle converts into durable margins
• Debates: Whether pricing holds into 2027, and the return on the TWD 15.81B capacity budget approved in July, with new capacity due in 2027
• Watch: Q3 gross margin against the 64.4% Q2 mark, Q4 sales commentary, and capex timing
GUIDEWIRE SOFTWARE INC INVESTOR DAY
Tue Oct 27, 7:30 PM ET, Las Vegas, United States
Analyst Day held inside Connections 2026, October 26 to 29 at Fontainebleau Las Vegas, with more than 3,800 attendees expected.
Webcast:
• Setup: Fiscal 2026 revenue rose 23% to $1.48B and ARR rose 19% to $1.24B, with cloud ARR up 35% to 84% of total ARR and 62 core cloud deals signed in the year. The stock sold off on fiscal 2027 guidance in September and has been rebuilding into the day
• The number: Fiscal 2027 ARR guidance of $1.45B to $1.46B, and whether the fiscal 2028 ARR figure near $1.7B is reaffirmed as a floor
• Why it matters: Guidewire is the core-system standard for P&C insurers, and migration of its installed base to Guidewire Cloud is the growth engine
• Debates: Roughly 125 on-premises customers still to migrate, the pace of cloud cohort expansion, and how fast agentic AI products monetize
• Watch: Detail on the top 10 US carrier signed September 4 for Guidewire Cloud and PolicyCenter, set to become a top 10 customer by fully ramped ARR over a 12 to 18 month deployment, plus any multi-year target updates
TOMORROW
FIVE9 INC CONFERENCE PRESENTATION
Wed Oct 28, Amsterdam, Netherlands
ServiceNow World Forum Amsterdam, RAI Amsterdam, 8:30 AM to 6:00 PM. Free, in-person only, no digital edition and no public webcast.
• Setup: Five9 and ServiceNow expanded their partnership in November 2024 around a turnkey AI solution joining ServiceNow Customer Service Management with the Five9 Intelligent CX Platform. Five9 exhibits and hosts a session, showing Five9 Fusion for ServiceNow, which puts voice and digital on one agent screen, with real-time transcription feeding Now Assist
• The number: One agent screen for voice and digital inside ServiceNow, the core Fusion claim the team is demoing on the floor
• Why it matters: Amsterdam is a live test of whether the ServiceNow channel converts into European contact-center pipeline for Five9
• Debates: AI agent deflection versus seat-based pricing pressure across CCaaS, and how much of the partnership is product versus go-to-market
• Watch: Five9 session attendance and any named customer or expansion announcements out of the forum
ORACLE CORP INVESTOR DAY
Wed Oct 28, 10:55 AM CT, Las Vegas, United States
Investor Day Meeting at Oracle AI World, live via the Oracle investor relations homepage.
Webcast:
• Setup: Q1 fiscal 2027, reported September 10, delivered revenue up 30% to $19.3B and cloud infrastructure revenue up 121% to $7.4B, and management raised the fiscal 2027 revenue outlook to at least $90B
• The number: Remaining performance obligations of $664B, up $209B year over year and $26B in the quarter, against quarterly capex of $28.5B and free cash flow of negative $5.4B
• Why it matters: This is the first long-form chance to test how the backlog converts, with 300,000 GPUs and 850 MW delivered in Q1 and utilization near 98%
• Debates: Capacity delivery pace, margins and cash flow through a $90B to $95B fiscal 2027 capex plan, and customer concentration in the AI backlog
• Watch: Any update to the capex or funding plan, RPO conversion commentary, and applications growth against the infrastructure story
LTD INVESTOR DAY
Wed Oct 28, New York, United States
Investor Day at Elevate '26, the flagship conference, October 28 to 29 at the Javits Center, with a private investor Q&A with management. The opening keynote is livestreamed free at 11:00 AM ET.
Keynote livestream:
• Setup: Q2 revenue rose 22% to $365M and the company reconfirmed full-year 2026 revenue guidance of $1.466B to $1.474B, even as it cut roughly 20% of headcount to fund the AI pivot and the stock sold off on the print
• The number: AI ARR doubled from Q1 to Q2 and drove 17% of net new ARR, against overall net dollar retention of 109% with guided pressure in the second half
• Why it matters: Elevate is where the AI workspace claim meets customers, around 250,000 of them, with every attendee building a working AI agent on site, across 200+ sessions and 150+ speakers
• Debates: Seat and credit pricing for AI versus seat erosion, the enterprise mix shift, and the reported exits of AI, R&D and product leaders mid-pivot
• Watch: Keynote product announcements from co-CEOs Eran Zinman and Roy Mann and CPO Daniel Lereya, any AI monetization metrics beyond the 17% figure, and investor Q&A tone on fiscal 2027 growth
REMAINING THIS WEEK
HON HAI PRECISION INDUSTRY CO INVESTOR DAY
Fri Oct 30, Taipei, Taiwan
Hon Hai Technology Day (HHTD26), October 30 to 31, 2026, with live streaming on the official HHTD site.
Event site:
• Setup: Foxconn is repositioning from assembler to AI platform company. Q2 revenue rose 41% year over year to TWD 2.53 trillion with net profit of TWD 60.0B, September revenue rose 38% to TWD 1.16 trillion on AI server demand, and third-quarter revenue, reported October 5, rose 47% year over year to TWD 3.03 trillion, above the TWD 2.83 trillion LSEG SmartEstimate, ahead of full Q3 results on November 12
• The number: Taiwan's largest GPU cluster, due in 2026 from as the group's NVIDIA Cloud Partner AI Factory, is the centerpiece claim to test
• Why it matters: HHTD is where the AI server story gets product proof: the MODEL A launch, the IBM AI Factory partnership for Taiwan, the Foxconn and OpenAI work on US AI supply-chain manufacturing, plus SiC, LEO satellites and FoxBrain
• Debates: How much value Foxconn captures beyond assembly in AI servers, EV timelines, and margin conversion as cloud and networking posts high double-digit growth
• Watch: Partner appearances on the opening day stage, MODEL A specifics, and any capacity or US manufacturing commitments
LATER
• Tue Nov 3: DCD Connect Virginia, Leesburg, VA, Nov 3 to 4
• Wed Nov 4: DCD Connect Brasil, Sao Paulo, Nov 4 to 5
• Wed Nov 4: DCD Investment Forum Brasil, Sao Paulo
• Wed Nov 4: Xcelerated Compute Show Brasil, Sao Paulo, Nov 4 to 5
• Thu Nov 5: Box Investor Day, San Francisco
• Fri Nov 6: Global Unichip at TSMC 2026 Japan OIP Ecosystem Forum, Tokyo
• Mon Nov 9: Entegris Investor Day, 1:00 PM
• Tue Nov 10: Cohu Investor Day, New York
• Tue Nov 10: Flex Investor Day, Austin
• Tue Nov 10: WESCO International Investor Day, Data Center Deep Dive Day, 10:00 AM
• Wed Nov 11: Munters Group Capital Markets Day, Stockholm, 8:00 AM
• Thu Nov 12: Global Unichip at TSMC 2026 Taiwan OIP Ecosystem Forum, Hsin-chu
• Sun Nov 15: Penguin Solutions at SuperComputing (SC26), Chicago
• Sun Nov 15: Supercomputing Conference 2026 (SC26), Chicago, Nov 15 to 20
• Sun Nov 15: Vertiv at SuperComputing (SC26), Chicago
• Tue Nov 17: Global Unichip at TSMC 2026 China OIP Ecosystem Forum, Beijing
• Tue Nov 17: TomorrowX Summit 2026, Austin, Nov 17 to 18
• Tue Nov 17: XTI Aerospace at 2026 ElevateUAV Summit, Florida City
• Thu Nov 19: KVH Industries at CSN Greece ICT Conference, Athens
• Tue Nov 24: Global Unichip at TSMC 2026 Europe OIP Ecosystem Forum, Munich
• Mon Nov 30: NVIDIA GTC Washington DC, Nov 30 to Dec 3
• Thu Dec 17: Methode Electronics Investor Day, New York, 8:30 AM
• Thu Dec 17: Onto Innovation Capital Markets Day, New York, 9:00 AM
• Wed Dec 30: at AI Impact Summit 2026
• Wed Feb 3: AI Cities Zaragoza, Spain, Feb 3 to 4, 2027
Show more
$AMKR KEY READ-THROUGHS FROM AMKOR TECHNOLOGY’S Q3 2026 EARNINGS CALL
Amkor’s results indicate an increasingly divided semiconductor cycle, rather than a broad-based recovery or an emerging AI slowdown. Advanced computing and automotive electronics are supporting higher-value manufacturing activity, while memory costs, component shortages, and changing production schedules are constraining smartphones and PCs. The most actionable negative signal is that these constraints now extend across both Android and iOS, including premium smartphones and wearables. The strongest positive signals concern advanced-packaging economics, inspection and metrology demand, and semiconductor content growth in vehicles. However, customer-funded capacity expansion should not be confused with immediate production availability, new equipment orders, or guaranteed investment returns. Several developments corroborate trends already disclosed by the affected companies; the incremental information lies principally in their persistence, geographic allocation, and conversion timing. Countries below identify corporate home markets rather than stock-exchange locations.
SMARTPHONES AND CONSUMER ELECTRONICS
PREMIUM SMARTPHONES ARE NOT INSULATED FROM THE COMPONENT CONSTRAINTS AFFECTING ANDROID (READ-THROUGH 1)
Affected company: Apple (AAPL: United States). Directional impact: moderately negative for near-term hardware production and revenue timing, with a smaller and less certain effect on consolidated earnings.
The critical commentary is broader than weakness in inexpensive Android devices. Amkor stated that “Component shortages affecting premium-tier smartphones and IoT wearable devices created an additional headwind” during Q3 and expects those conditions to persist in Q4. Asked to distinguish the ecosystems, CFO Megan Faust said, “we do see Android being down in Q4 compared to Q3. And we also see iOS down.” Communications revenue is expected to decline by a high-single-digit percentage sequentially, while Consumer revenue is expected to decline by a mid-teens percentage, worse than normal seasonality.
The transmission mechanism is constrained production of finished devices. A shortage of a required component can limit shipments of an otherwise complete smartphone or wearable, regardless of demand for the finished product. Higher memory costs create a separate pressure through the bill of materials. Apple can mitigate those costs through procurement, product mix, and pricing, but those responses do not make physical component shortages irrelevant. Simultaneous pressure in smartphones and wearables also reduces the likelihood that strength in adjacent hardware categories fully offsets constrained iPhone production.
The important distinction is between supply pressure and demand deterioration. Amkor’s revenue is also affected by its own SiP relocation, customer build schedules, and packaging content. Its Communications decline cannot be translated directly into an equivalent decline in Apple shipments. Apple’s previously reported June quarter was strong, including record June-quarter iPhone, Mac, and Services revenue. The high-conviction conclusion is therefore that premium hardware faces persistent supply and cost constraints, not that Apple has already experienced an equivalent deterioration in consumer demand.
The near-term catalyst is Apple’s November 2 earnings report, particularly December-quarter production availability, component costs, and hardware guidance. The trading implication is greater downside sensitivity to an outlook that assumes uncomplicated holiday-quarter fulfillment. Over a longer horizon, altered launch schedules could redistribute manufacturing activity between H1 and H2 without changing annual demand. Amkor explicitly declined to identify a substantial packaging-content benefit from new form factors, weakening any argument that those form factors automatically offset production constraints across the supply chain.
ANDROID CHIP SUPPLIERS FACE A PERSISTENT UNIT-ECONOMICS SQUEEZE, NOT SIMPLY TEMPORARY PACKAGING DISRUPTION (READ-THROUGH 2)
Affected companies: MediaTek (2454: Taiwan) and Qualcomm (QCOM: United States). Directional impact: negative for handset revenue and near-term earnings estimates; moderate potential consolidated impact, partly offset by non-handset businesses.
Amkor identified memory pricing and availability as constraints on both smartphones and PCs. CEO Kevin Engel stated that “the low-tier type phones had a harder time absorbing those cost changes than the premium tier” and added, “I think that will continue.” The expectation of another sequential Communications decline in Q4 indicates that the pressure is persisting beyond the initial disruption.
For smartphone chip suppliers, the mechanism operates through handset manufacturers’ production decisions. Higher memory costs increase the cost of a complete device. Manufacturers must absorb those costs, raise selling prices, reduce specifications, or produce fewer devices. Absorption compresses OEM margins and increases pressure on other component suppliers; price increases can reduce unit demand; production cuts reduce application-processor, modem, connectivity, and related chip purchases. A semiconductor supplier can successfully raise its own prices while still facing weaker unit volumes.
This is corroboration of an established problem, not an entirely new downturn signal. MediaTek’s Q2 prepared remarks reported a 20% year-over-year decline in Mobile Phone revenue, explicitly attributing weakness to increasing smartphone bill-of-materials costs. Mobile Phone represented 41% of its revenue. Qualcomm’s latest reported handset revenue also declined 20% year over year and represented approximately 51% of consolidated revenue. Amkor’s incremental contribution is evidence that the pressure continues into the calendar year-end production period and is not confined to a particular Android customer.
The exposure is financially meaningful. As an illustrative sensitivity, a 5% handset-revenue shortfall, holding other businesses unchanged, would reduce consolidated revenue by approximately 2.1% at MediaTek and 2.6% at Qualcomm using their latest disclosed mixes. These are exposure calculations, not forecasts derived from Amkor’s decline.
MediaTek’s October 30 earnings call and Qualcomm’s November reporting cycle are the immediate catalysts. The relevant questions are whether flagship product ramps offset weaker lower-tier demand, whether component-price increases are successfully recovered, and whether handset inventories remain aligned with actual device sales. Longer term, automotive, edge computing, and data-center diversification become more valuable. Qualcomm’s strong automotive business is an important offset, making a blanket negative conclusion on the entire company less defensible than a negative conclusion on its handset exposure.
ADVANCED PACKAGING AND SEMICONDUCTOR EQUIPMENT
PROCESS CONTROL HAS A STRONGER FUNDAMENTAL READ-THROUGH THAN AN UNDIFFERENTIATED PACKAGING-EQUIPMENT BASKET (READ-THROUGH 3)
Affected companies: Camtek (CAMT: Israel) and Onto Innovation (ONTO: United States). Directional impact: positive, with high business-level relevance for Camtek and meaningful relevance for Onto’s advanced-packaging operations; insufficient evidence for an immediate incremental earnings upgrade.
Amkor’s flagship HDFO CPU process involves more than 300 manufacturing steps. Computing revenue increased 25% sequentially, another hyperscaler HDFO CPU program began ramping in Q3, and 2 new networking and data-center 2.5D programs are expected to enter production in Q4. Management also expects Computing revenue to grow approximately 30% or more in 2027.
The transmission mechanism combines manufacturing capacity expansion with greater process-control requirements. Complex redistribution layers, interconnects, bump structures, and multi-die integration increase the importance of detecting defects before expensive downstream processing is completed. Higher wafer and package volumes support tool demand, while additional control points and tighter tolerances can increase inspection and metrology requirements per production flow. More than 300 process steps does not mean more than 300 inspection steps, and the call does not support a proportional tool-count calculation.
The exposure is independently supported by the equipment suppliers’ disclosures. Camtek’s October investor presentation identified high-end advanced packaging as approximately 60% to 70% of its business and described applications in fan-in, fan-out, interposers, and advanced bump metrology. Onto’s Q2 call reported more than $200 million of Dragonfly orders from an unnamed OSAT, mostly for delivery in 2027, and raised its advanced-packaging growth outlook. Neither disclosure identifies Amkor as the customer, so those orders cannot be assigned to Amkor.
The most important qualification is the composition and timing of Amkor’s spending. Its 2027 facility framework can reach approximately $2.25 billion to $2.30 billion before additional equipment spending, but buildings are not inspection-tool orders. Some equipment for repurposed Korean space is already on order. The current call therefore supports the durability of the investment cycle without proving that new orders should immediately be added to supplier forecasts.
Onto’s November reporting cycle and Amkor’s February 2027 capital-budget disclosure are near-term catalysts. Camtek had already presented an initial 2027 growth outlook exceeding 35% in October, making Amkor’s commentary corroborative rather than a new forecast. The longer-duration positive is continued spending on process capability through the Korean and Arizona buildouts. The call does not justify assuming that all packaging-equipment categories, particularly technologies not specified in these programs, benefit equally.
SCARCE PACKAGING CAPABILITY SUPPORTS BETTER COMMERCIAL TERMS, BUT NOT AUTOMATICALLY BETTER CASH RETURNS (READ-THROUGH 4)
Affected company: ASE Technology Holding (3711: Taiwan). Directional impact: positive for advanced-packaging pricing, utilization, and customer visibility; moderate positive consolidated read-through, tempered by capital requirements and its broader business mix.
Amkor’s strongest pricing commentary was that customers recognize the need for returns on significant capital investment. Engel stated, “the margin profile and the return is not a major issue. They are more focused on making sure they secure the capacity.” Management also described a limited number of qualified providers and said customers are securing capacity into 2027 and 2028.
The mechanism is a shift in negotiating leverage toward manufacturers with qualified advanced-packaging processes, scale, and reliable execution. When packaging constrains a customer’s product roadmap, the cost of unavailable capacity can exceed the benefit of obtaining a modestly lower packaging price. Earlier technical engagement and longer capacity commitments can improve loading visibility and support contracts that better compensate manufacturers for equipment, engineering, and operating costs.
ASE is a relevant peer through its assembly, testing, and materials operations. Its Q2 release already showed strong ATM growth and a 27.3% ATM gross margin. Amkor’s commentary supports persistence of favorable industry conditions rather than establishing an undisclosed ASE order or proving an incremental price increase. ASE’s electronics-manufacturing-services operations also dilute the consolidated sensitivity to advanced-packaging economics.
The distinction between cost recovery and genuine pricing power remains essential. Amkor said it generally passes higher substrate and material costs to customers. Such recovery can preserve profit dollars without improving margin percentages. The more valuable signal is customer willingness to support returns on scarce manufacturing capability. The call provides qualitative evidence of that willingness but no contractual margin floor or quantified industry-wide price increase.
ASE’s late-October reporting cycle is the near-term catalyst, particularly the relationship between advanced-packaging growth, realized margins, equipment spending, and cash generation. The longer-term positive is a more strategic, less transactional role for qualified packaging providers. The negative is that attractive accounting margins can coexist with weak free cash flow during a prolonged capacity expansion. Amkor’s investment peak extending into 2028 and 2029 is a warning against valuing the entire OSAT sector as though improved EBITDA immediately becomes distributable cash.
AI ACCELERATORS AND FOUNDRY SUPPLY CHAINS
CUSTOMER PREPAYMENTS IMPROVE NVIDIA’S SUPPLY SECURITY WHILE MOVING CASH COMMITMENTS EARLIER (READ-THROUGH 5)
Affected company: NVIDIA (NVDA: United States). Directional impact: positive for medium-term supply availability and manufacturing resilience; negative for the timing and flexibility of cash commitments, but not evidence of an immediate material earnings charge.
Amkor’s October 21 announcement established a new $1.3 billion NVIDIA agreement supporting South Korean capacity and stated that the companies’ agreements across Arizona and South Korea total $3 billion. On the earnings call, management described customer advances as a major source of expansion funding and identified a late-2028 ramp for the NVIDIA-related Arizona opportunity.
The economic benefit to NVIDIA is access to qualified production capacity ahead of competing demand. Advance funding can enable suppliers to purchase equipment, build facilities, and align process development with product roadmaps. Geographic diversification also reduces dependence on a narrower set of manufacturing locations. The benefit is principally a reduction in future supply-execution risk, not an immediate increase in current-quarter GPU shipments.
The offset is earlier cash deployment and less flexibility if product requirements, demand, or manufacturing architectures change. Prepayments economically accelerate cash outflows ahead of future purchases. They are not necessarily an immediate expense, and the disclosed commitment should not be treated as though its full amount has already been paid. Amkor’s disclosure does not reveal all refund, cancellation, volume, or pricing provisions.
The scale mismatch is important. A $3 billion customer commitment is highly significant relative to Amkor’s investment program but much less consequential relative to NVIDIA’s financial scale, which included $96.2 billion of revenue in its latest reported quarter. The high-conviction negative is increased capital commitment and reduced optionality, not a claim of financing stress at NVIDIA.
The agreement was announced before the earnings call and should not be counted again as a fresh order. The call adds useful sequencing: Korea is the nearer-term expansion platform, while the discussed Arizona production contribution is later. NVIDIA’s November earnings commentary on supply commitments and subsequent disclosures of Amkor funding receipts are near-term catalysts. The longer-term positive depends on successful qualification and production loading during 2027 and 2028. Nothing in the call supports assigning the unnamed flagship CPU program to NVIDIA or converting Amkor’s Computing growth directly into a NVIDIA revenue revision.
ARIZONA PACKAGING IS COMPLEMENTARY TO TSMC’S U.S. FABS, BUT SUPPLY-CHAIN LOCALIZATION REMAINS INCOMPLETE (READ-THROUGH 6)
Affected company: Taiwan Semiconductor Manufacturing Company (2330: Taiwan). Directional impact: strategically positive for its U.S. manufacturing proposition; low immediate earnings impact, with benefits predominantly dependent on 2028-and-later execution.
Amkor reiterated that Arizona Phase 1 construction is scheduled for completion at the end of 2027, with Phase 2 expected to break ground in late 2027. The planned campus represents approximately $12 billion of investment and more than 90,000 square meters of cleanroom space. These plans sit alongside the June 2026 agreement under which TSMC can procure advanced packaging and testing services from Amkor through a 10-year collaboration framework.
The mechanism is complementary rather than necessarily competitive. Domestic wafer fabrication does not create a fully domestic semiconductor supply chain if advanced packaging and testing must occur elsewhere. Qualified Arizona packaging can make TSMC’s U.S. wafer output more useful to customers seeking geographically integrated manufacturing, reduce cross-border production transfers, and strengthen the commercial proposition of its regional ecosystem.
This does not establish that Amkor is taking packaging revenue from TSMC. The disclosed framework explicitly involves procurement and cooperation. Nor should Amkor’s $12 billion campus be added to TSMC’s capital budget or treated as a quantified TSMC revenue opportunity. The allocation among customers, package types, and production periods remains undisclosed.
The timing creates an important negative portfolio implication. Building completion is not equivalent to qualified high-volume output. Customer qualification, equipment installation, yields, and utilization remain necessary milestones. The late-2028 NVIDIA-related Arizona ramp illustrates why meaningful Asian production dependence persists during the intervening period. Geographic diversification announcements therefore do not justify immediately removing Taiwan-related supply-chain exposure from AI and semiconductor portfolios.
TSMC had already reported its September-quarter results on October 15. Amkor’s call is not a leading indicator for that completed quarter; it provides additional information about regional manufacturing integration. The next relevant milestones are the 2027 construction and equipment-installation updates, followed by customer qualification and production loading in 2028. Management’s acknowledgement of higher U.S. labor and building costs also means localization should be valued as resilience and customer access, not automatically as a source of higher consolidated margins.
PCS AND ENTERPRISE CLIENT HARDWARE
AI SERVER STRENGTH IS MASKING A WEAKER PC ECONOMIC ENVIRONMENT (READ-THROUGH 7)
Affected company: HP Inc. (HPQ: United States). Directional impact: negative for PC unit demand and Personal Systems margins; moderate consolidated earnings relevance, with printing and company-specific mitigation providing offsets.
Amkor explicitly separated data-center growth from client-computing demand. Engel stated, “PC is relatively soft, but the Data center has offset that.” He also observed that memory is a sizable component of PC costs and that Amkor had seen corresponding market impacts. Its record Computing revenue therefore cannot be used as evidence of a broad PC recovery.
For HP, the transmission mechanism is the interaction among input costs, selling prices, and units. Higher DRAM and storage costs can compress system margins if not fully recovered. Passing those costs to customers can protect profit per unit but reduce affordability and shipment volumes. Premium mix can raise reported revenue even while the underlying unit market deteriorates.
HP’s latest reported results demonstrate that distinction. Personal Systems revenue increased 18% year over year to $11.8 billion, while total units declined 16%. Segment operating margin was 4.6%. Amkor’s commentary corroborates the persistence of cost and volume pressure rather than predicting a deterioration that had not previously appeared in HP’s results.
The earnings sensitivity is meaningful because system margins are relatively thin. At the latest reported quarterly revenue, a 50-basis-point Personal Systems margin change represents approximately $59 million of operating profit, or about 11% of the segment’s reported operating profit. This is an illustrative sensitivity, not an estimate revision derived from Amkor.
The counterevidence is material. HP reported improved memory availability and fulfillment, raised its annual guidance, and generated strong quarterly cash flow. Amkor’s comments do not invalidate those company-specific measures. The more defensible negative conclusion is that a rapid, uncomplicated recovery in PC unit demand or margins remains difficult to underwrite.
The near-term catalyst is HP’s fiscal year-end reporting and initial fiscal 2027 outlook, particularly whether pricing continues to offset component inflation without further unit weakness. The longer-duration implication is a greater divergence between premium commercial systems and price-sensitive consumer PCs. Amkor’s separate observation that Arm migration continues should not be confused with growth in the overall PC market: architecture gains can coexist with a contracting unit pool.
AUTOMOTIVE AND INDUSTRIAL SEMICONDUCTORS
VEHICLE COMPUTING CONTENT AND MAINSTREAM INVENTORY NORMALIZATION ARE IMPROVING TOGETHER (READ-THROUGH 8)
Affected companies: NXP Semiconductors (NXPI: Netherlands) and Qualcomm (QCOM: United States). Directional impact: positive, with substantial relevance to NXP’s end-market mix and a meaningful but smaller consolidated contribution at Qualcomm.
Amkor’s Automotive & Industrial revenue increased 37% year over year and 7% sequentially, reaching a record. Management identified growth in ADAS, infotainment, industrial applications, and mainstream automotive packages. Engel stated that “the mainstream legacy type packages going into cars are also rebounding” and described customer inventories as being under better control. Q4 revenue is expected to grow further, albeit at a slower low-single-digit sequential rate.
The most valuable signal is that structural content growth and cyclical normalization are occurring together. More sophisticated ADAS, vehicle networking, and centralized or zonal computing increase processing and connectivity requirements per vehicle. At the same time, improving inventory positions can allow semiconductor purchasing to track actual production more closely instead of being suppressed by destocking. These mechanisms can support chip demand even without strong growth in vehicle units.
NXP has directly relevant exposure through S32 automotive processing, vehicle networking, microcontrollers, and industrial edge processing. Its latest reported mix was approximately 55% Automotive and 22% Industrial & IoT. Its own Q2 commentary had already described accelerating S32-related design activity and improving inventory metrics. Amkor’s observations therefore provide independent corroboration of the existing recovery rather than evidence that NXP should replicate Amkor’s 37% growth rate.
Qualcomm’s latest reported automotive revenue increased 61% year over year to approximately $1.59 billion. Amkor’s emphasis on infotainment and advanced in-car computing supports the durability of that demand channel. Nevertheless, this positive exposure must be weighed against Qualcomm’s larger handset business. The same call provides both a favorable automotive read-through and an unfavorable smartphone read-through for Qualcomm.
NXP’s October 27 earnings report is the most immediate catalyst, particularly whether automotive processing, industrial demand, and inventory normalization remain consistent with its prior outlook. Qualcomm’s November results provide the next test of whether automotive growth continues to offset handset pressure. Over a longer horizon, rising semiconductor content per vehicle is the more durable driver than any short-term restocking cycle.
The conclusion should not be extended indiscriminately to all automotive exposures. Amkor did not provide evidence of stronger vehicle unit sales, improved automaker margins, a broad EV-demand acceleration, or a comparable recovery in silicon-carbide power devices. The highest-conviction benefit is concentrated in automotive processing, connectivity, and selected mainstream semiconductor demand, not the entire automotive value chain.
Show more