Bad things happen when you are grounded in the real world. Ergo, in systems architecture, we constantly look for feedback loops. Negative feedback loops provide stability (like a thermostat). Positive feedback loops provide amplification, but without a governor, they inevitably lead to system runaway and collapse.
You know the story of the Golem of Prague? The Rabbi builds a protector out of clay. It works great until it doesn’t. In engineering terms, the Golem failed because it lacked a governor—a negative feedback loop. In systems architecture, positive feedback loops without governors don’t just grow; they run away and tear themselves apart. That is exactly what it appears we are building with this circular financing of AI where capital flows from Big Tech balance sheets into startups and immediately flows back as revenue.
Let’s break down the Golem legend.
Clay to Golem: Protecting the Valuation
In the legend, Rabbi Loew creates the Golem out of river clay to protect the Jewish ghetto from anti-Semitic attacks and expulsion. It was a desperate measure to ensure survival in a hostile environment.
Today, all the Big Tech players (Microsoft, Google, Amazon, Meta) have convinced themselves that they are in an AI Race. If you recall from my blog “Resource Allocation: Race vs Rate?”, when you are in a race, you pull out all the stops to win and then clean up the mess after victory. These companies are spending hundreds of billions of dollars on CAPEX—building data centers and buying GPUs & TPUs —at a pace that boggles the mind.
To justify this spend, they need to show massive cloud revenue growth now. But organic market demand (real companies solving real problems for paying customers) isn’t growing fast enough to absorb that capacity.
So, they are molding the clay.
The Mechanism:
- Big Tech invests billions of cash into AI Startups (OpenAI, Anthropic, CoreWeave, etc.).
- These startups are contractually or technically obligated to use that cash to buy Cloud Services and GPUs from the investor.
- Big Tech books this returning cash as Cloud Revenue.
From an accounting perspective, this looks like growth. From a systems perspective, it is a closed loop. The money is simply making a round trip, minus some friction, allowing the provider to claim that demand for their infrastructure is explosive.
AGI Shem
A Golem is just a mud statue until you insert the Shem (a slip of paper with the magic name of God) into its mouth. The Shem animates the clay.
For Big Tech, the Shem is AGI Artificial General Intelligence (AGI).
If you peer through the fog of hype, the current generation of LLMs are just great language processing tools (think parser+++). However, you cannot justify a $3 trillion market cap on better chatbots. You need magic. You need the promise of a god-like intelligence that changes physics and economics. Oh and now it’s also Geopolitics (e.g. race w/China) to prepare the field for a government bailout.
The concept of AGI is the Shem that suspends disbelief. It allows investors to ignore the terrible unit economics of the current generation because they believe they are funding the creation of a new species. It turns a bad business model into a civilizational imperative.
Golem Works
Initially, the Golem works perfectly. In the legend, it protects the community and does the heavy lifting.
In the market, the AI Golem is currently serving its masters well:
- Nvidia sells chips to hyperscalers.
- Hyperscalers invest in startups.
- Startups buy chips/cloud from hyperscalers.
- Revenue beats estimates. Stock prices soar.
This creates the illusion of a flywheel. It looks like organic adoption. It validates the decision to build $100 billion superclusters. Everyone looks like a genius because the capital is moving, even if it’s just moving in a circle.
Golem Grows
The tragedy of the Golem legend is that the creature lacks a soul—it has no internal governor. It continues to grow and becomes violent. It begins to destroy the very city it was meant to protect.
The risk in AI Circular Finance is the decoupling of capital from actual value.
When Microsoft or Google invests in a startup, they are essentially booking their own capital as their own revenue. This inflates the stock price, which allows them to raise more capital, to fund more startups, to book more revenue.
Eventually, the startups have to sell a product to a third party—a real business (like a bank, a hospital, or a retailer) that needs to make a profit. If those end-users don’t appear in sufficient numbers to pay for the compute costs, the startups default.
If (when?) the startups default, the Revenue stops. But the Depreciation remains.
Big Tech is left holding a mountain of silicon that is depreciating like fresh fish. The Golem turns back into mud. But unlike river clay, this mud is made of debt and server racks that cost $XX,000 a month to power.
The Fall
In the story, Rabbi Loew eventually has to remove the Shem to deactivate the Golem, causing it to crumble and crushing him under its weight.
At some point, we either deliver AGI or deactivate the Golem. The Shem of AGI is starting to lose its power as people realize that scaling laws may be hitting a plateau. If the magic word stops working, we are left with the cold, hard reality of hardware amortization schedules.
Also – I’m beginning to think that AGI is a bad idea – but that is another blog.
The Takeaway
Once again – let’s invoke the Feyman principle:
We must never confuse activity for progress or deliverables for value.
Right now, the industry is generating massive amounts of kinetic energy through financial round-tripping. Do not mistake this for a sustainable economic engine. Real value is only created when a customer pays more for a service than it costs to provide it, because that service solves a real problem. Satya Nadella said it:
“When we say: ‘Oh, this is like the industrial revolution,’ let’s have that industrial revolution type of growth. That means to me, 10 percent, seven percent for the developed world. Inflation adjusted, growing at five percent, that’s the real marker.”
This is an apt analysis but to say LLM is parser+++ is too much. If it is parser+++ it means it is powerful than any parser. All the parsers that we have for programming languages and the sub pieces that make up the compiler should then already be rewritten with much lesser effort with LLM.
Well that is a different debate, but the economic arguments make sense.
I definately didn’t mean to say that it was a parser.
Rather that it is a language processing tool LIKE a parser, or a lexer, or a regex.
It is many many orders of magnitude more capable than all of those.
I have read too many science fiction books to trust humans with AGI.
There are two other considerations being missed in all the hype, that the industry is either having a very short memory about, or are willfully overlooking-
1- Consumer interest and what they are willing to pay for- Consumers today will buy hardware (phones most frequently, computers less so) but will rarely pay for software. Gamers are an exception (Steam or the Epic Games store are the two big ones) and niche uses (someone who needs Adobe Creative Cloud, or a pro-level audio editing app), but the average person is using Gmail for free, all social media for free, and most of the OpenAI accounts are on the free tier. It’s going to be hard sell to get the average user (and sustainable economics is based on sustainable averages) to willingly pay when so many cloud services have made us accustomed to free. It’s no longer the early 2000s and fewer people need to buy locally installed software, even MS Office or the paid version of an Office 365 subscription. For services that do have a higher user count and do produce revenue beyond ads (paying for extra storage for your Google Photos, or iCloud storage for your phone backups), the value add of a higher level that includes AI won’t be there for most people.
Simply put, you’re not likely to pay for any appreciable percentage of a trillion-dollar infrastructure build out on the consumer side of the equation.
2- Service integration- Even if we focus on business / enterprise customers that do reliably shell out for licenses (Anthropic seems to be leading here among the independent AI firms, but still not profitably so), previous technology trends suggest that AI / LLMs / whatever iteration is next, is much more likely to become a *feature*, not a product. Cloud storage, online shared document editing, virtual servers in the cloud, serverless compute, cloud DB, are all features within larger service and subscription offerings, even though a few of them started off as distinct products.
Microsoft and Google are both continuing this trend with AI, with Copilot and Gemini integration as features into their existing products. Workspace and Office 365 are solid revenue streams for both companies. As capabilities were added to Google Cloud and to Azure, prices for these rose or additional subscription tiers were added, but in general we’ve been able to count on increased capabilities at nearly the same cost over time. While profitable for them (and for Amazon with AWS) the revenue from these services still has to cover current infrastructure costs, without AI.
So this is the paradox the industry has created by subscriptions over purchases, and feature integration over new products- To absorb the cost of a trillion-dollar infrastructure build out (which as you said will depreciate like dead fish) and not just absorb but make a profit, you have to break the increasing value curve of the past. That’s either a significant cost elevation of all subscriptions, or some very pricey new tiers. Just as with consumers, it’s very debatable just how many business customers will agree to pay for that, or decide they even need it.