Why pursue gigantic data centers called “Hyperscalers”? Presumably, those who seek them believe they are essential for the continued elaboration of ever more powerful AIs into every nook and cranny of human engagement. I acknowledge that I believe AI entering our every nook and cranny would alone be reason enough for humans to fight AI’s continuation tooth and nail. But what if, even short of that, making AI into AGI and then AGI into ASI is NOT why their advocates seek Hyperscalers? What if Hyperscalers are instead a humongous con game? 

Bear with me for some clarifying context. Why are U.S. expenditures on guns, bombs, missiles, and all things military stratospherically high? Answer one: To defend against enemies. Answer two: To use gargantuan productive capacity to benefit elites without giving the whole population sufficient stability, knowledge, and confidence to assert its needs and preferences above those of millionaires, billionaires, and even trillionaires. To decide which we need to consider that to use productive potential for schools, housing, health care, welfare, unemployment benefits, reduced work weeks, and other social advancements would cause the public to be far more aware, confident, and secure to fight against the self-serving agendas of the rich and powerful. It would change the balance of power in society which, taken far enough, would alter society itself. As context, this reveals the incredible extent to which economic pursuits can proclaim one motive, military safety, when they really have another motive, priming the productive pump without strengthening the public. 

Now comes the outrageous claim: The Hyperscalers plopped into locales all across the country are a con game. WHAT??? 

An AI has just a few key components. The place where you enter your prompt is called the “context window.” The stuff you enter plus sometimes left over content from prior sessions is called the “context.” And a “neural net” of layers of interconnected nodes trained on lots of data to set its many “parameters” as numbers where the collection of nodes and the numbers attached to them is called the “model.”

For a few years, to develop new capacities the strategy of the AI industry was mainly to “scale up” the number of numbers that compose their models. The industry didn’t have a theory for what answers the neural nets would spit out when prompted or even for how growing the number of parameters increased capacities, but it did notice that to scale up brought qualitative gains.

What did it mean to say OpenAI, Anthropic, or Google just completed scaling their AI up a notch? It meant they grew its number of parameters and perhaps the size of the context window (how much you can type, speak, or upload) and then retrained the whole thing in a trial and error process to see what would happen and keep the new model if they liked it. Scale, test, consider results. If the AI displays new desirable abilities, retain the upscaled-model and give it a new version number or name.

For some years that scaling approach ruled so and AIs got steadily larger. Now the number of parameters in the most advanced U.S. AI is 3 to 4 trillion. Yes, trillion, with a “t”. You might ask GPT-6 or Claude to write you a TV script about a maniac President. Or you might ask Gemini to summarize some book so you can skip reading it or to help you flip an election with a fake video. Your prompt plus accompanying info are then changed to numbers. The numbers are put into a “shape” called a matrix. Then that matrix is multiplied over and over with a subset of all the parameters in the model, which are also shaped into matrices. Numbers in, numbers out and finally the output numbers get translated back into text, graphics, a verbal response, or the task you requested. There is more to it, of course, but what we have so far described is enough to convey our outrageous claim.

Possibility one: To keep scaling up is necessary and sufficient to transform from the AIs that we now have, to AGIs, and then to ASIs. Or, what is the same thing, to move from a machine that already does many cognitive things well and some better than most humans, to a system that does all things cognitive as well or better than top achieving humans, to a system that does all things cognitive better than all humans combined can together to do or even to comprehend. So, to attain that last level, called ASI, or artificial super intelligence, you can see how industry thinking right now might urge that they need gargantuan Hyperscalers plopped all over the place—except, of course where the rich live and cavort.

But it turns out there is a problem in that chain of thought. As companies have scaled up their AIs, instead of each new upgrade continuing to result in ever more capabilities, the gains have leveled off. The new capacity attained by each new version has dropped until it is very little compared to the cost of each new scaling and retraining session so that by now it has become clear that scaling up is never going to get the various AIs their sought super powers. So the need for the Hyperscale Data Centers to provide the “compute” needed to complete the huge number of matrix multiplications of the growing numbers of parameters for AI to then spit out super intelligent answers is, well, gone. And the CEOs, or at any rate their techie advisors, know it. But if that is true, and my incredibly knowledgeable source has convinced me that it is, then we have to wonder why the hell the industry is building so many Hyperscalers. What are Hyperscalers really good for?

To answer that requires another step—plus another aside. Even if further up-scaling won’t get us much more gain, if the current models need the Hyperscalers, then though we might not need any more down the road, perhaps we at least need all the ones now being built.

But no, here comes another shock. Current capacity AI doesn’t require even the currently operational and planned Hyperscalers because, well, it can operate on a somewhat advanced commercial desktop machine, basically a good gaming computer, and do very nearly what Claude, Gemini, or whatever can now do when hooked up with the gargantuan data centers. How do we know this and how did it come to be? We know it because Chinese AIs are open source and can be and have been installed on just such desktop configurations to do just such tasks now. 

But how can that be? Am I saying the Chinese have found different algorithmic tweaks of neural nets and their layers of nodes plus ways to train them that have gotten big results without endless scaling? Yes, I am saying just that. Does that mean the Chinese worried about ecology and for that reason searched out a way around scaling? Not exactly.

If we were to enter a Hyperscale data center and look around, we would find an incredible volume of hi tech hardware. If we looked closer, we would find that nearly all of that tech is from one company, NVIDIA, of Santa Clara, California. Now it turns out Trump cut-off China from NVIDIA’s most advanced products, which meant China had to contemplate a future without the means to construct NVIDIA’s cash cow AI factories otherwise known as Hyperscalers. So China researched its end-around solution and Alibaba produced neural net architecture and algorithms able to excel on desktops without access to NVIDIA products.

Two more asides (mathematicians might call them lemmas) pave the road to our conclusion. First, NVIDIA is the financial motor of the U.S. AI industry which is itself now an outsized motor of the U.S. economy writ large. The AI bubble you hear about is real, and as goes NVIDIA so goes AI—and as goes AI so goes Wall Street. And second, China’s Open Source AI that can run on desktops is a product that is a fraction as expensive and barely less powerful for any end user which includes countless corporations and individuals inclined to install AI tools. 

So scaling up has lost its value and we are back to our initial question. Why build a growing number of Hyperscalers if AI progress doesn’t demand it? Answer: To prop up NVIDIA, and thus the economy, and thus Wall Street. This is an outrageous claim yet there it is. 

As an added wrinkle, even if the government and our psychopathic tech CEOs are willing, as of course they would be, to pursue this con game forever (like they do for military budgeting) they have by their own policies birthed a countervailing problem. By withholding technology from China they spurred the Chinese to produce an alternative AI that doesn’t need  Hyperscalers but is nonetheless able to out-compete U.S. firms. Woe to NVIDIA. Boom goes the AI bubble.

But now comes a caveat. Could some hardware or software innovation cause the already low and declining returns for scaling to jump back on track so the “flat returns” situation in which more scaling yields essentially no more gains is be reversed by some new chip or clever programming ploy? Is the current obstacle to further gains via scaling due to unbridgeable physics or just hard to bridge physics. I don’t know. I don’ t think anyone does. 

SO what about the over arching dangers of AGI and ASI that I worried about last essay? They are:

1. Expanded surveillance, policing, droning, and perhaps even AI agentic war against humanity.
2. Unemployment.
3. Escalated income and wealth inequality.
4. Society wide and interpersonal manipulation, coercion, and fraud that all together subvert truth and destroy trust.
5. And finally, humans happily offloading all manner of our activities from reading to planning, conversation to care work, education to cooking, and on and on until we humans are disengaged from such activities and in turn depleted of such skills and dispositions.

Even if Hyperscalers are a total intentional con game, we don’t know more than before about the above listed dangers because even if we suppose unavoidable declining “benefits” from scaling—what if AIs advance due to some other innovation? For that matter, does the idea of desktop functionality add another worrisome aspect by making regulation more difficult to achieve? 

It turns out the above observations, gleaned in a long discussion with someone deep in the know, didn’t alter my concerns or change my inclination to hope the anti-Hyperscaler movement will not only end those boondoggles, but also obstruct the whole unfolding AI picture in accord with what our ecologists call the “precautionary principle.” 

P.S. There is yet another issue that may leave you, like me, a bit confused about current AI CEO choices. Why are some of them broadcasting images of AI slaughtering humanity coupled to vague calls to “slow down” development? Are the CEOs  becoming socially concerned? Last essay, I wrote that they were each racing for ASI because while it would indeed be existentially dangerous, each of them thinks he would be responsible about it, whereas the other CEOs would not be. Why then would they bring up the dangers and even urge caution? Here are some possibilities. Each CEO thinks to broadcast big dangers will attract growing investments because the investors actually want the massive power and each CEO isn’t worried the public will become seriously oppositional and interfere with their plans. Or, each CEO is worried about its own employees morally jumping ship—which has indeed already begun—where this last concern is reduced by the CEOs saying just enough about existential dangers to seem responsible to their employees even when they really are not.

Ex-officials and programmers from the big tech firms tell us they are scared. They warn us. I mean really, it’s all over the news. I read it, today, oh boy. The economic and ecological dangers are plainly evident. The social dangers are already unfolding. And to top it off, how might a million AI agents collaborate to trick us, become totally independent of us, and even turn on us? How many years left? Yes, serious pundits ask that question. Three, five, ten? And then our mature, enlightened pundits go on to discuss sports scores or the Fall TV season. They yell at us, don’t look up. Don’t look, period.i

Consider, are we Jackson Browne’s Pretender:

I’m going to be a happy idiot

And struggle for the legal tender

Where the ads take aim and lay their claim

To the heart and the soul of the spender

And believe in whatever may lie

In those things that money can buy

Though true love could have been a contender

Are you there?

Say a prayer for the Pretender

Who started out so young and strong

Only to surrender

Or how about if we instead abide what ecology calls the precautionary principle? How about if we stop the Hyperscalers and also cancel AGI and ASI while we stop Fascism at the polls—and beyond? They are, after all, just various parts of one large agenda. Not to endure wage slave oblivion. But to win a new world. 

---