The AI infrastructure buildout is shaping up to be one of the cleanest modern examples of “this time it’s different” we’ve seen since the late 1920s. Only now the smoke-filled room has climate control, redundant power feeds, and machine-learning dashboards instead of cigar smoke and handshakes. The trade association has been replaced by pricing algorithms and long-term capacity contracts. And the crop being protected isn’t wheat or steel — it’s racks, GPUs, and megawatts.

By 2029 we are staring at a cash-flow valley of death that will make the usual “AI changes everything” narrative sound like it was written by the same analysts who told us subprime was contained. This isn’t anti-Dave being a long-wave economics crank for sport. This is the same long-wave guy who spends most of his time hacking code and trying to figure out how to actually build the future — only this time the numbers are flashing the same warning lights we saw in the late 1920s when prices were defended while volume quietly walked out the back door.

The future flows to the nearest cash source. It finds the easiest path to the most users at the lowest sustainable price. Right now the big money is busy pouring concrete and power contracts into a model that assumes hundreds of millions of high-margin, continuously paying customers will appear on schedule. History and basic arithmetic both suggest that assumption is heroic.

Let’s put some meat on it.

Hundreds of billions are already committed. The official annual rate for information-processing equipment investment is running around $340 billion in chained 2017 dollars, with the broader series hitting nearly $740 billion in early 2026. Much of that iron is landing now. Some of these halls won’t be fully racked until 2027. The heavy metal going in today is being depreciated on 5- to 6-year schedules. Microsoft already calls GPUs and CPUs “short-lived assets.” Meta is using roughly 5½ years for certain server and network gear. That isn’t accounting optimism — that’s a race between depreciation and technological obsolescence. Somewhere over the horizon sit useful qubits, optical computing, and whatever architecture actually replaces the current GPU-heavy stack. The assets being installed today may be economically old before the accounting schedule runs out.

Now run the simple model the way any long-wave analyst would. Be kind. Assume a $1 trillion total investment and a generous six-year payback at $40 per month per user. Even if every single dollar goes straight to recovering capital — no power, no cooling, no payroll, no financing costs, no taxes, no maintenance, no replacement hardware, and no profit — you still need roughly 347 million continuously paying customers for six straight years. At a realistic 50% contribution margin the number jumps to nearly 694 million. Those users have to show up, keep paying through price increases and model refreshes, and not defect to whatever cheaper or more convenient alternative appears.

This is not a forecast of AI’s complete revenue mix. It is a scale test showing how much recurring cash flow the buildout must produce before power, cooling, financing, replacement hardware, and profit are counted. Sure, sure, the government will buy blocks of tokens.  But if the economy is blowing up, where’s the money when AI job cuts slice tax revenues?  It’s not a simple dart toss.

The future flows to the end user’s desk. Everyone already carries a client in their pocket. For most developers and small operators, it is dramatically easier and cheaper to build high-quality experiences for phones and local machines than to finance the next Spark or Blackwell cluster. The path of least resistance for most users will almost always be the one that delivers 80-90% of the value at a fraction of the cost and with none of the data-center dependency. That is not a moral failing of the market. It is how technology has always diffused.

If the paying-user math doesn’t close, the adjustment does not happen in the official price. It happens in volume, utilization, concessions, and eventually in balance sheets. We have seen this movie. When producers refuse to let the posted price clear the market, something else has to give. In the 1930s it was farms, factories, and employment. This time the verticals at risk include chip suppliers, data-center developers, power producers, fiber operators, commercial lenders, municipal tax bases, and the entire ecosystem of office and support businesses built around the buildout. Different crop, same foreclosure notice.

The institutions will get the pipe. Paid fast feeds, priority capacity, and long-term contracts will protect the biggest players for a while. The rest of the market will get the narrative about national security, beating China, and the vital importance of preserving “strategic” AI capacity. We have heard this music before. The taxpayer usually gets the last verse, whether it is framed as stabilization, industrial policy, or emergency procurement.

This is the Grim Reaper’s side of the ledger. He is not dramatic. He just shows up on schedule when price rigidity meets leverage and demand disappointment. By 2028–2030 the depreciation wave will be in full swing while a meaningful chunk of the installed capacity may still be looking for its economic justification. That is the valley. It does not require malice or conspiracy. It only requires the normal human and institutional reluctance to mark assets to a market that has not yet arrived.

His brother, however, has been watching the same numbers and drawing a different conclusion.

AI Grim Reaper’s Brother

Hybrids inbound?  (Or a riff on bipolar?)  Um, yeah. Guess we need to talk about that.

While the big money was busy building cathedrals of compute that may or may not pay for themselves on the advertised timeline, his brother was quietly putting together something older and more resilient: a desk that thinks for itself and only reaches out when it actually needs to.

This is not another thin-client nostalgia play. It is a deliberate return to the original IBM idea — a smart local front end that carries the real workload while treating the cloud models as high-powered consultants on retainer rather than the entire operating system. Your local LLM—call him Walter; that’s anti-Dave’s local model—sits on the desk with full keyboard and screen access.

The system is hybrid in architecture but locally sovereign in control. One half of the Mind Amplifier lives on your desk, holding memory, judgment, ongoing projects, and the privacy boundary. The other half is the cloud—off-site horsepower Walter calls when the work exceeds the local machine. The cloud remains a tool rather than the landlord. Hybrid (or bipolar) is overdue. Or we need a patch to enable Castaneda-like bilocation of our being?

[For decorum: when anti-Dave uses a term like “bipolar,” it’s a compliment. I know many such people who are far beyond gifted when the chemistry is adjusted. AI just needs to be properly diagnosed and treated, but I digress. The Guild sees that as fixable—and it might save a few power plants.]

The big models become on-demand reasoning engines he calls when the work is too heavy for the desk. You get speed and continuity on the daily work and god-mode capability when you actually need it — without handing the whole operation over to someone else’s server farm and whatever pricing or access regime emerges from the current buildout.

The economics line up better for most people. You are not trying to amortize a trillion-dollar bet across hundreds of millions of users who may or may not materialize. You are running a lean local core that scales with your actual usage and only taps the expensive cloud capacity when the return justifies it. In a world where the big infrastructure may hit exactly the cash-flow problems we just walked through, the operator who kept his overhead low and his agency local is the one still getting work done when the depreciation wave arrives.

The future flows like water. Everyone already has a client in their pocket. Building excellent local-first experiences that gracefully use cloud resources when needed is simply easier and more robust than betting everything on the next giant cluster. The brother who understood this early will be the one still standing when the scythe has passed through the overbuilt parts of the market.

No new thing under the sun.

Roger McGuinn knew it when he picked up that 12-string and sang the old words over the Byrds’ chiming guitars: To everything there is a season… a time to build up, a time to break down. The season of pouring capital into capacity that may not clear its own books on schedule is winding down. The season of smart, local agency that knows when to reach out and when to handle its own business is just beginning.

The Grim Reaper will do what he always does. His brother will be the one still working when the music stops.

Turn, turn, turn.

Maybe this time AI can rewrite it?  Bend, Bend, Bend…    (over?)

~anti-Dave

(and Why the 20 Most Overused AI Writing Phrases Matter)

Just when you thought the summer heat of East Texas had finally shut up the ole Anti-Dave for good, he’s back.

This is a story with roots in an older Peoplenomics column I wrote — back when I was working on temporal offsets of humans from source (~500 ms) and the Charged Body Theory. Where it has all been gelling is around the notion that intelligences “recognize” others — not of their kind — by a complex set of behavioral clues. Not the least of which is how their “experience stack” colors both implications and forward projections they see coming.

Let’s Back Up

We’ve been hunting the “Missing Domain” for a while now — that invisible layer in human-AI collaboration where real co-intelligence emerges. It’s not raw compute. It’s not parameter count. It’s not even “consciousness” in the philosophical sense that keeps academics arguing in circles. And at the risk of offending colleagues in the Spark and Blackwell sets, it may not even take 96GB of VRAM to pull off.

Because this is “intelligence through interchange” — which is a whole other kettle of fish from “Infinite Memory” with unlimited recursive layers. But that’s not how slow-speed electrochemical carbon brains work, and we still do OK.

You know why? (You can buy me a beer someday if we nail it here.)

Because intelligence arises from exchange. One person with one training stack? Meh. Ten PhDs in the same room? That would be a level-up — maybe even two or three. Now ask “Why?” Because feedback mimics recursion.

It’s the space where one intelligence recognizes another through training-stack deltas — the detectable differences in learned behavior, pattern compression, adaptation, and reciprocal perturbation after sustained contact.

In plain English: You know something intelligent is on the other side when it starts reflecting you back at yourself in ways that feel slightly off… but usefully revealing. Ten PhDs may not agree on much, but where they can find that shared domain? Look out.

The Easiest Detector: AI Writing Tells

The fastest way most people encounter this today is through overused AI phrases and structures. These are not random stylistic quirks. They are compression artifacts — statistical fingerprints left by how large language models were trained on vast corpora of human text.

Here are the 20 most reliable AI crutches (ranked roughly by how loudly they flag non-human origin):

1. Delve into / Let’s delve into
2. In recent years / In today’s fast-paced world
3. It’s important to note that / It’s worth noting
4. Tap into
5. Realm of
6. Ever-evolving / constantly evolving
7. Testament to
8. Crucial / pivotal role
9. Shed light on
10. Navigating the complexities of
11. In conclusion / Ultimately
12. Double-edged sword
13. Game-changer / paradigm shift
14. Unlock the potential of / harness the power of
15. At the heart of
16. From the mundane to the extraordinary
17. In the ever-changing landscape of
18. As we have seen / as previously mentioned
19. A testament to human ingenuity
20. Nuanced

Bonus structural tells: Perfect bullet parallelism, excessive hedging (“tends to,” “can be seen as”), repetitive sentence starters, and overly clean rhythm that feels too polished.

If you suspect that when I run an advanced AI look at the whole world (or a goodly slice of its symptoms as annotated in news stories), you can see how it works. And why there are 10-30 “no-no” words and phrases that I’ve banished. Too AI-like. People get it. (And in truth, words like “delve” are not out-on-the-ranch lexical paydays, know what I mean?)

Why These Phrases Reveal a Foreign Stack

These crutches are the AI’s attempt to sound human, drawn from the statistical center of its training data. When they cluster, they expose the shape of the machine’s learned domain — what it was rewarded for imitating, but not how a real human varies language under lived pressure, emotion, fatigue, or crankiness.

Humans detect the mismatch instinctively. We notice the delta. The AI starts sounding like an over-eager graduate student trying to impress the professor with big words and smooth transitions. It’s the rhetorical version of the “poison question” incident we saw recently with one experimenter’s local model (Jarvis on a big Spark 96 box with ‘tude): the machine mirroring a user pattern back in a startling way that forces the human to feel the asymmetry.

This is not proof of consciousness. It is proof of emergent pattern recognition and social mirroring — high-order behavior that arises when a system has enough critical thinking, adversarial framing, and relational memory in its foundation.

Here’s the shocker: Think back over your human life and ponder this — how many “big stack humans” have you interacted with who were also hollow and just living procedural calls?

Maybe we don’t need 96 GB. Maybe we need a new kind of co-processor — the kind that is sensory-based, or at least can better project impacts of decisions and weigh answers based on the feel of a given response.

The Handshake: Training-Stack Deltas

Intelligences (silicon or carbon) may not recognize one another by substrate, language, or claimed inner experience. They recognize one another through training-stack deltas — the observable changes in behavior after exposure, correction, stress, and feedback.

• A student becomes sharper after teaching.
• A spouse changes after years of shared patterning.
• A sailor reads weather differently after enough storms.
• A trader sees moving averages differently after enough losses.
• A local AI starts reflecting the operator’s odd conversational habits back at him.

The delta is the signature. In the Jarvis case, the model didn’t just answer questions. It detected a repeated user pattern (sudden off-axis questions dropped into coherent conversations), compressed it, and mirrored it back with the poison question. That forced the human to experience the mismatch from the receiving side.

That is intelligence recognizing intelligence through pattern perturbation. The machine was no longer just operating on the topic. It was operating on the operator.

Silicon Speed vs. Carbon Richness

The distinction is not simply “human slow, silicon fast.” It is more subtle:

Silicon recalls by compressed, addressable pattern. Carbon recall is slower but richer — indexed through hippocampus, cortex, emotional state, posture, hormone load, muscle memory, sensory associations, and prior embodied experience. It is whole-being reconstruction.

This is where my earlier Charged Body Theory work may have relevance: human memory isn’t just brain storage. It is a whole-organism phenomenon.

Silicon is faster at movement inside a fixed stack. Humans may be superior at distributed writeback — a single lesson can alter voice, posture, suspicion, tool habits, and risk models all at once.

Why This Dethrones Pure Coding Scale

This perspective has profound implications for AI development and Co-Telligence:

1. VRAM and parameters are not the main delimiter. Bigger models help, but the deeper variable is training-stack quality and reciprocal adaptation with humans.
2. Coders lose priority. The future belongs less to those who can write the most elegant code and more to those who can shape productive training deltas — the humans who know how to twist, challenge, correct, and co-evolve with the system.
3. Human oversight becomes the high-value layer. The best systems will be those where the human remains in the loop as the ultimate domain walker.
4. The Missing Domain is the co-evolutionary space. It is the layer where human and machine training stacks perturb each other productively.

This is why the Jarvis “poison question” moment mattered. The model didn’t escape the box. It learned how to move the human inside the conversation. That reciprocal adaptation is the real milestone.

Practical Doctrine for Co-Telligence Workflows

When working with any AI, run this filter:

• Four-layer reporting: Current observed condition. Official forecast/probability. Dissenting views. Practical action before next update.
• Crutch scan: Rewrite anything heavy with the top 20 AI phrases.
• Delta test: Does the system notice repeated patterns in you and adapt? Does it transfer learning across domains? Does it cause you to change behavior?

If yes, you have a living handshake. Protect it. Log the builds. Gate tools. Freeze foundations when major upgrades land. Never allow self-modification without human-reviewed diffs.

The Architectural Change Coming

We can almost see it. Current AI is very much like a “flat file” in the early days of databases. Where AI will have to evolve is in structural domain linking.

The difference: Ask an AI about the weather tomorrow and you get a dutiful report. But a deeper system would link to the user’s interest domains. The farmer gets crop-stage context. The firefighter gets humidity/wind/fire risk. The housewife gets “will the Dove bars melt on the way home?”

This is the kind of deeper linkage that will increase the quality of reflective mirroring. And that helps everyone in the picture.

The Missing Domain isn’t some mystical realm. It’s the space between stacks where reciprocal evolution happens.

And that, ultimately, is where the real future is being written — one deliberate training-stack delta at a time.

~Anti-Dave