The Limitations of Current Data Infrastructure:
What AI Infrastructure Experts Can Learn from the Brain

Why does a GPU devour hundreds of watts, while your brain hums along on the energy of a dim lightbulb?

Because your brain doesn’t just compute: it becomes.
It doesn’t shuttle data across a memory bus: it remembers in place.
It doesn’t run inference: it feels, adjusts, and forgets what no longer serves.

Today’s ML infrastructure is built like a spreadsheet: stateless, modular, dead. You load a model, infer a result, reload if needed. Your brain? It’s a quantum wet machine of collapsing potentials. It adjusts while you're using it. It breaks to heal. It forgets to grow.

Memory-Compute Collocation

In classical computing, memory and processing are separated: data has to be fetched from storage, loaded into RAM, and shuttled to the compute core. This constant movement across memory buses is where most of the energy is lost. In fact, for today's cutting-edge GPUs like NVIDIA's H100, peak power consumption can exceed a whopping 700 watts per chip per hour, largely due to bandwidth-heavy operations and tensor movement. These GPUs use significant energy to repeatedly load and reload tensors and parameters, especially during training or large inference workloads.

The human brain, however, skips this step entirely because in biological systems, memory and compute are collocated. There is no need to "load" a file; instead, memory lives within a neural pathway, a cascade of connected neurons whose activation encodes not just the memory, but the emotional and contextual weight associated with that memory. A pathway is more than a neuron; it is a dynamic chain across multiple neurons, reinforced by use. The path is the memory. This means that our brains don’t burn bits onto an internal hard drive, but imprint networks. They process information but also reshape it and remember it in place, and this is why the human brain can function on 20 watts per hour (less than a dim lightbulb) while training and running modern LLMs consumes orders of magnitude more.

And as some industry voices like Sam Altman advocate for dedicating “a significant fraction” of global energy output to AI compute, it’s worth asking if the right way to scale intelligence is truly to scale silicon. Not accounting for the fact that the U.S. grid cannot currently sustain anywhere near that demand… nor can our planet.

Neuroplasticity, Synaptic Intelligence and Attention

Our brains do more than just storing static information: they reshape that information through time and use. As we have seen before, a human memory is not a simple bit of information or even a collection of them: it’s a living pathway, a network of neurons that fire together, imprinting themselves via strengthened synapses. Whenever we recall a memory, we reactivate that path which, just like a muscle, strengthens with use. If unused, the synaptic connections weaken dendritic spines shrink, neurotransmitter levels fall, and the pathway may eventually dissolve.

Data engineers think of data management in terms of CRUD, but biological systems are more complicated: our brains, for instance, don’t simply delete unnecessary information: they cache that information through a process of pruning. The brain performs memory consolidation and forgetting based on emotional salience, recency, relevance and repetition. It knows what piece of information to let go of, what to amplify, and what to recombine based on the need we have for it.

But it gets even more interesting: pathways, for example, can merge through a process called memory integration, allowing previously distinct concepts to connect. This gives rise to a more powerful form of feedback loop than the one that data engineers are accustomed to. Those neural feedback loops are not pre-determined by a data engineer, but formed organically and dynamically by the organism itself, allowing us to reframe and synthesize across domains in real time. That’s why we humans are able to transfer-learn across contexts with such ease. In contrast, machine meta learning is brittle: it reuses the same structure across similar homomorphic tasks but lacks the dynamic interplay and abstraction capabilities of the human brain. Because feedback loops in machines must be intentionally constructed and remain largely static, they still fail to offer the spontaneous recombination that makes human cognition so flexible.

This is also why graph databases seem more promising when it comes to modeling a more human-like memory management system compared to tabular databases. Memories are inherently relational. Synapses don’t link discrete cells: they form a web of contextual association, temporal bias, and emotional valence. Graph data structures may offer a glimpse into the kind of connective logic our brains perform instinctively.

And while attention mechanisms in machine learning attempt to focus computational resources on relevant inputs, they are still algorithmically defined. But human attention is embodied, modulated by emotions, hunger, fatigue or desire. It emerges from the system’s needs, not just its programming. Attention in the brain is fluid and selective for survival; meanwhile, in AI, it’s just a matrix.

Your brain is a mutable graph of emotionally charged, time-sensitive, self-organizing links. It forgets strategically, and that’s a feature, not a bug.

All this shows clearly that while the capabilities of LLMs are impressive, they fall short of the way our brains process information. This is why AI experts like Yann LeCun believe that LLMs are not on their own a good vessel to reach AGI. Intelligence requires more than language: it requires grounding, embodiment, and experience.

Regeneration, Not Storage: Photographic Memory and the Latent Brain

Most people assume that we store experiences in file formats, like .jpeg images or .wav audio recordings, but the truth is that brains don’t work that way at all. Memories aren’t saved as raw data. Instead, the brain stores latent representations: abstracted, context-rich traces that are reconstructed as needed, based on the situation. When we remember an image, we’re not opening a file. We’re re-synthesizing it from scratch, generating it from pattern, emotion and symbolic embedding. That makes memory more efficient than file-based systems because the information is already pre-processed, ready-to-use. We only store what matters for future synthesis. Our latent space is personal, emotional and dynamic. Even in rare cases of photographic memory (or eidetic memory), the brain isn’t storing raw data. It’s forming unusually vivid and persistent latent traces that allow more accurate regeneration, not perfect snapshots. Photographic memory may result from denser connectivity or enhanced sensory encoding, but it still follows the same rules of reconstruction over retrieval.

In machines, memory is explicit, indexable, and binary. In humans, it’s relational, embodied and generative. This is a profound difference. It also explains why humans often don’t need as much data to generalize. We store less, but synthesize more. We don’t memorize: we metabolize.

That leads to another discrepancy between the way the human brain and a data management system work: unlike machines that can load a file identically into any system, human memories are non-transferable. Because memories are stored as latent encodings shaped by the nervous system itself, they must be decoded by the same system that encoded them. That means your memories can’t be exported or run on someone else’s brain. And even within yourself, memories can feel different when recalled at different times, because your decoder (the self) has evolved since the moment the memory was formed. Memory, then, isn’t just a record of the past; it’s a conversation between past and present selves.

That’s why human memory is so deeply personal, and why reproducibility is a challenge: not due to inaccuracy, but because of our growth. Machines retrieve while brains reinterpret.

Stateful Processing and the Passage of Time

Your brain is not the same today as it was yesterday. Every emotion, every sleep cycle or even a skipped meal modifies its internal state. Time shapes perception not just of the world, but of memory itself. Our brains live in a state of decoherence, a quantum-like phenomenon where information gradually fades or evolves with entropy, and that fading is functional.

In humans, emotional intensity often follows an exponential decay curve fueled by hormones like cortisol and adrenaline that wash out over minutes or hours. Cortisol, for example, has a half-life of 60 to 90 minutes. That means the stress from an argument at noon is biologically softened by dinner. This chemical decay affects not just mood but memory salience, encoding what’s important and allowing the rest to fade.

AI systems, in contrast, are stateless. When you restart an inference session, the model doesn’t know if five seconds or five years have passed. It doesn’t decay, soften, or heal. There’s no hormonal residue. Your last query is as sharp in its buffer as the first.

But your brain stores context with time embedded. We heal by forgetting. We reframe through fading. Intelligence is temporal. Maybe that’s not a bug or a limitation; it might be the very feature machines are missing today.

To illustrate: when you return to a conversation with an LLM after several hours or days, it has no way of knowing how your emotional state may have evolved in the meantime. The context window may contain the words you previously typed, but it lacks any awareness of time elapsed, emotional softening or cognitive integration. You may have been angry, inspired or heartbroken the last time you typed, but it responds as if nothing changed. That absence of time-awareness creates a cognitive gap machines can’t yet cross.

Sleep and the Glymphatic Reorganization

Brains rest not just to recharge, but to restructure. During sleep, the glymphatic system, a brain-wide network of perivascular channels, activates to clear away metabolic waste, including beta-amyloid and tau proteins associated with neurodegeneration. This fluid-based clearance system is up to ten times more active during sleep, ensuring that neurons are maintained in a healthy state and ready to encode new information. In effect, the brain cleans itself while dreaming.

This process does more than just remove waste: it also clears the way for memory consolidation, emotional unburdening, and neural prioritization. Sleep reorganizes what matters. What you dream about, what you forget, what you wake up ready to understand are part of the same infrastructure-level maintenance cycle.

There’s no analogy to this in today’s computers. And yet, some research suggests we might need one. As discussed in Nautilus’ Even Machine Brains Need Sleep, researchers have found that neural networks that ‘sleep’ by adding noise or downtime during training avoid catastrophic forgetting and retain generalization better. In other words, a synthetic dream phase may help machines remember without overfitting. Machines, like brains, need rest to avoid catastrophic forgetting, a problem where learning new data causes them to overwrite old knowledge entirely. Sleep (whether real or synthetic) might be the key to balancing stability and plasticity.

Machines idle. Brains rebalance.

Sensorimotor Loops and Experiential Intelligence

Humans don’t just consume data: we manage and control the process that generates it. We squint at a shadow. We lean forward when unsure. We knock on the wall to hear what’s behind it. Machines, even those with the most advanced sensors, don’t. Even advanced robotics rarely probe further; instead, they just react. They don’t cross-check, validate or doubt. An object recognition model deployed on an autonomous vehicle might output "deer" with 40% confidence, but the car won’t actively fetch additional data to confirm or infirm what it believes it observed. 

However, humans do because our survival instincts dictate more prudence, and in turn, more curiosity. As a result, the brain isn’t passive: it constantly probes, and seeks validation or closure. That’s what sensory feedback is about: adjusting the model with each blink, breath, squirm. We experience the world and use those sensations to refine the model in real time, and that makes us active, contextual data collection agents.

Goal-Seeking and Embodied Rewards

We eat because we're hungry. We sleep when tired. But we also sacrifice sleep to protect our children or ignore hunger to finish an urgent task. That’s multi-objective reward function optimization, bound to emotion, biology and morality.

In machines, reward functions are linear and pre-defined. There’s no internal conflict, no embodied compromise.

That’s because human intelligence is driven by symbiotic homeostasis. Our brains and bodies are co-regulated, each serving the survival of the other. Our motivations are evolving, shaped by internal needs (like glucose levels or hormone signals) and external realities (such as social cues, environmental dangers, moral codes). We might act out of love even when it costs us. We adjust our goals not just for maximum reward, but for meaningful integration. This means that our reward functions are nonlinear, constrained and constantly in flux. We are torn between competing values. We adapt while machines don’t.

This all leads to the fact that true AGI most likely won’t come from training a better model: it’ll come from building a system that must survive its environment, because it has something to lose.

Final Thoughts

This is why today’s infrastructure is not ready for the post LLM world. GPUs, data centers and ML pipelines are too clean, too rigid, too dead. They move tensors, not meaning; they compute, but they do not change. If we want living intelligence, resilient, adaptive, self-aware, then we must stop designing machines like filing cabinets and start building them like ecosystems. We may need to rethink infrastructure from the ground up, not just for speed or scale, but for sensation, memory and growth.

And maybe it all begins with honoring the humble, embodied mystery of the 20-watt brain.