THE FRAGILE AI STACK
The modern AI stack is both complex and fragile. It’s built from a patchwork of third-party models, open-source frameworks, cloud services, and compute constraints, all orbiting around a single gravitational force: data. Whether you believe language models are “stochastic parrots” or “blurry JPEGs of the internet,” the next leap in AI will come from high-quality multimodal data, the kind AI models need to see, hear, read, and infer, not the chaotic exhaust of the public web. Sadly, it’s also the only way corporations will achieve the efficiencies they so desire when laying off people.
That makes data the linchpin of AI success, but its misuse leads to failure. Yet the traditional mechanisms we use to protect data do not meet the requirements of this responsibility. LLM monitoring, dynamic masking, standard encryption techniques, and even access controls all fall short in an AI-driven ecosystem where models need context to reason. If you remove too much information, the model’s intelligence collapses. Add too much, and you create an entirely new class of security risks. In many ways, an LLM is simply a probabilistic version of our own brain, and we’re still trying to secure it with tools meant for a filing cabinet.
FROM PROBLEMS TO MANIA
With these obsolete traditional mechanisms in mind, it becomes obvious that businesses will retreat to the one thing they can control: flooding capital into the construction of one might call artificial brains. AI tech. Massive data centers. Hyper-scale compute clusters. It’s reminiscent of Tulip mania, a bubble in the 17th century where the price of tulip bulbs soared to extreme levels before collapsing, hence the term, Data Center Mania.
Examples:
- Record-breaking financing for Meta’s Hyperion Data Center.
- Stock market volatility due to AI circularity (see: Oracle)
- Elon Musk’s vision of an inferencing Tesla fleet producing 100 Gigawatts
- Walmart planning flat headcount despite growth, as AI reshapes roles
What looks like mania is actually the collision of two old forces: greed (chasing outsized returns) and exhuberance (belief tha the future can be engineered faster). So, will these new “neurons” of data centers democratize intelligence? Or empower state-sponsored cyber-offensive and accelerate a dystopia? Probably both. What’s certain is that Metcalf’s Law is in full swing, and these new neurons will unfold, like a flower, into our work, companies, and daily lives.
Why did I choose the phrase “new neurons”? There’s a reason. One way to understand Data Center Mania is by first looking at how AI compares to the human brain.
NEURAL NETS AND THE VISUAL CORTEX
In a recent keynote, Larry Ellison compared the evolution of a Convolutional Neural Network (CNN) to that of the human visual cortex, with its distinct areas called V1, V2, and V4. V2 extracts edges, depth, texture, and simple patterns. V4 goes further, assembling color, curvature, and complex shapes into objects, the moment vision becomes meaningful. CNNs follow the same path as they learn to see.
While CNNs follow a similar path to early visual processing, there is a newer architecture that gives us a twist. Vision Transformers (ViTs), an alternative neural network architecture that excels in computer vision, break images into patches and excel at complex visual tasks. They resemble the inferior temporal cortex in the brain, the region involved in processing complex visual information. ViTs, like our inferior temporal cortex, can perform facial recognition.
Beyond Larry’s keynote, things become clearer when we take a look at how Transformers work under the hood, where we find that Transformers are composed of encoders and decoders – the encoder processes the input text (i.e., the prompt) and generates a sequence of hidden states that represent the meaning of the input text. The decoder uses these hidden states to generate the output text.
Transformers are only one chapter of the story. A Probabilistic Neural Network (PNN) is a type of neural network used for classification that works by estimating the probability of an input belonging to a class or group. Basically, it guesses the way we humans do. And once these systems are scaled, emergent properties appear that are not programmed. They appear as models get larger and learn more, sometimes in sudden jumps.
FLOP, CLOPs, & THE NATURE OF COMPUTATION
As larger models gain these emergent behaviors, computation becomes a proxy for intelligence. This is where Floating-Point Operations Per Second (FLOPS) becomes a representation of intelligence. And if we take this to the next step in computation, Circuit Layer Operations Per Second (CLOPs) enter the picture. They are the way quantum computers are benchmarked, measuring how fast a quantum processor can run circuit layers. They also exhibit similarities to the brain. For example, when you see an object, light enters your eyes (the input state of a circuit). It gets converted to electrical signals, which pass through layers of retinal neurons (literal biological neural circuit layers).
Between FLOPs and CLOPs sits analog hardware, Resistive Random-Access Memory (ReRAM) chips. ReRAM is a type of analog chip that stores data by changing the electrical resistance of a material, rather than the digital state (1s and 0s). One of the vendors behind ReRAM is Extropic, which is developing hardware called Thermodynamic Sampling Units (TSUs), where their inherent probabilistic nature, or “error,” is a feature. Basically, they treat randomness as a feature. “Don’t let the facts get in your way,” a friend once told me, and these chips do just that for AI, letting randomness reign, just like what happens in real life.
But if this is true, how much will this affect the fortunes being poured into massive data center projects?
GPUs VS TSUs vs THE HUMAN BRAIN
The architecture of a GPU is ideal for matrix operations (mathematical calculations involving large grids of numbers, like matrix multiplication). A GPU performs billions of matrix operations to approximate sampling from a probability distribution. In complex systems like AI, calculating exact answers is often computationally infeasible or impossible, and instead use approximation. A TSU from Extropic skips all of that and samples directly. It avoids energy-intensive matrix multiplication by using the inherent “noise” of physics for computation, which is a direct physical process rather than a mathematical approximation. Simply put, some are saying that these chips are 1,000x to 10,000x more efficient than current GPUs. A Medium article explained it’s the difference between aimlessly walking around a territory (GPU method), in contrast to navigating with a map (TSU method).
On the other end of the spectrum, there’s the human brain itself. It turns out that neuroscientists have done a lot in trying to understand how the brain works. The human brain learns by shifting the strength of connections between neurons. It also turns out that scientists have measured the brain’s consumption of energy in terms of wattage, which is about 20 watts of power, yet it packs roughly 80 to 100 billion neurons and about 100-500 trillion synapses. That’s absurdly efficient, and a worthy benchmark for both intelligence and energy use, offering a different view of data centers themselves.
TODAY’S DATA CENTERS
Getting back to Larry’s keynote, he told the audience that Oracle is building a 1.2-billion-watt brain, with half a million Nvidia GPUs. That’s enough energy to power a million homes. I recently read a post from Ann Davis Vaughan where she mentions that some in the data center industry have begun joking about “bragawatts,” gigawatts not just as infrastructure, but as a form of signaling. Scale itself has become a proxy for seriousness, ambition, and inevitability, even as the industry is still proving that denser clusters reliably produce better intelligence.
But Larry isn’t alone. What he didn’t say in his keynote is that AWS, Meta, Microsoft, Lambda, Nvidia, Blackstone, and many others are also building data cetners. In fact, according to a study by Goldman Sachs, by 2027, AI data centers will use 84 gigawatts, representing up to 12% of the entire nation’s energy capacity.
And what’s also notable about this is that the same fortunes building AI data centers are the same fortunes building quantum data centers. Quantum systems are potentially 13,000x more efficient than classical data-brain data centers. In fact, the two need to come together: Classic holding onto data, from which quantum can process. This coalescence is no better represented by a recent announcement about Nvidia’s NVQLink, which binds the two, acting as what Nvidia’s CEO Jensen Huang called “the Rosetta Stone of the quantum era,” enabling every future Nvidia GPU supercomputer to operate as a hybrid quantum system.
A TIMELINE OF ASCENDING INTELLIGENCE
So, putting this all together, I asked my quantum group to put a timeline together, using FLOPs and CLOPs as measurement sticks for this ascendant intelligence. The point is to show how quickly compute is scaling, and why the mania is only accelerating.
If we go from left to right, the Mac Mini came out in 2005, and if you string together about 600 to 1000 of them, using some HPC compiler, you could match something like a dog’s intelligence (some breeds require more than others…). With the new M5 coming out, that number would be reduced to half. And, just to note, a human brain would be about 10 Peta FLOPs.
Moving on, Oracle’s Abilene Data Center opened this year, producing 10 exaflops, which is more than enough to be equivalent to a person. Next year, Oracle is building a zetta-scale supercluster, which is equivalent to 1000 human brains. Then there are the other guys, like Meta, Google, MS, and Amazon, building their zetta-scale data centers.
All together, we have the following so far:
Oracle data center: 10 million trillion FLOPs = 10 exaFLOps
Oracle super cluster: 16 billion trillion FLOPs = 16 zettaFLOPs
PAST THE UNIVERSE’S LIMITS
Sometime in the next 10+ years, the advent of quantum computing may give us more computation capacity than a universe-scale classical computer. That is, there are about 2^265 atoms in the visible universe. So, even if you built a classical computer using all the atoms in the universe, it would only be able to simulate the interactions of 265 quantum mechanical particles. A quantum computer with just 265 noiseless, error-free qubits can do this directly.
So, to put it into perspective, quantum computers will have more computational capacity than even a classical computer built from all the matter in existence. So, yeah, THAT.
MANIA, BABEL, OR REAL ESTATE?
With compute scaling toward the unimaginable, we’re left with a question about intent. Is the mania around data centers more like the story of Babel than Tulips? Or to put it another way, it is about reaching for the stars or merely a real estate opportunity. Babel wasn’t built by fools or fanatics, it was built by ambitious people who were motivated by the idea that scale could collapse uncertainty.
As one firm put it, “data centers provide an essential service, in this case digital interconnectivity…” But at the same time, the same firm says, “data centers meet the definition of real estate because they are physical buildings… and offer stable rental incomes, because they are leased to a third-party tenant.”
To the latter, the ROI is currently roughed out between 12% to 18%, requiring careful financial planning. To the former, Henry Fielding puts it best: “Happy the man and happy he alone, he who can call today his own.”
mh