The 960-Formula Artifact

Six days ago, I read a Deloitte research article about tokenomics—the economics of AI measured in tokens. Within 48 hours, I had built a comprehensive financial model of hyperscaler AI economics with 960 formulas across 10 tabs, using Claude for Excel and Microsoft Garage's Excel Agent. The model covers token pricing trajectories, edge computing cannibalization, depreciation feedback loops, stranded asset risk scenarios, and competitive dynamics.

But the 48-hour build was the culmination of months of preparation. The frameworks, the research threads, the theoretical foundations—those had been accumulating through dozens of conversations with Claude over many months. The Deloitte article was the spark. The model was the artifact. The thinking had been developing since I first encountered Jeffrey Ding's work on technology diffusion.

This article traces that arc—from research to framework to model—because I think it reveals something important about how AI changes learning, analysis, and work itself.

The timing is particularly resonant because I'm currently teaching an Intelligent Automation Immersive program at Divergence Academy, where we're applying these exact skills—Claude, modern Excel capabilities, systems thinking frameworks—to help students build their own analytical artifacts. I'm learning by teaching, teaching by building, building by learning. The recursion is the point.

The irony isn't lost on me: I used AI in the flow of work to build a model analyzing AI economics. The tool and the subject are the same. And that recursive quality—using AI to learn about AI, to model AI, to understand AI—turns out to be the most important insight of all.

The journey began months ago with something mundane: a LinkedIn post about Excel.

Someone had shared that Excel 365 had quietly become "Turing-complete" through the addition of LET and LAMBDA functions. These functions, introduced in 2020 and 2021, allow users to define variables and create reusable custom functions directly in formulas—capabilities that previously required VBA programming. The post also mentioned new possibilities: Jaccard similarity calculations, lexicon-based semantic analysis, and integration with generative AI models.

The phrasing caught my attention: "not your father's spreadsheet program." Most people would scroll past. But I'd been developing a framework about technology diffusion—how capability spreads through organizations—and this felt like an example of the pattern.

So I asked Claude to explain it. Not a web search. A conversation.

What came back wasn't just an explanation of LET and LAMBDA. It was an analysis connecting Excel's evolution to the very diffusion framework I'd been developing. Claude observed that the underlying algorithms—Jaccard similarity, sentiment analysis—weren't new. Data scientists had been using them for years. What was new was their accessibility at the point of work. The finance analyst, the HR manager, the operations lead could now apply computational techniques without waiting for data science team bandwidth.

"This is your GPT skill infrastructure argument in miniature," Claude wrote. "The value comes not from inventing the algorithm but from widening who can apply it."

That observation landed differently than a typical search result. It connected a technical capability to a theoretical framework. It didn't just answer my question—it extended my thinking.

That Excel conversation connected to research I'd been absorbing for months—Jeffrey Ding's work on technology diffusion and great power competition.

Ding, a political scientist who won the 2025 Lepgold Book Prize, upends conventional wisdom about technological advantage. His central finding: where innovations are adopted more effectively is more important than where they are first introduced. The true determinant of competitive advantage isn't innovation capacity—it's diffusion capacity. The ability to spread technology throughout an economy, not the ability to invent it.

The evidence is striking. The Soviet Union, by 1970, led the world in R&D spending as a percentage of GNP. They pioneered continuous casting steel technology in the mid-1950s—five years before Japan adopted it. By 1980, only 10.7 percent of Soviet steel used continuous casting. Japan, adopting the technology later, produced 59 percent of its steel with the technique.

General Secretary Brezhnev acknowledged the problem explicitly in 1971: "The links connected with the practical realization of scientific achievements and their adoption in mass production are the weakest." He was describing, with unusual precision for a Soviet leader, a diffusion deficit—a structural gap between the capacity to invent and the capacity to spread.

The pattern repeats across industrial revolutions. Germany dominated the frontier technologies of the late 19th century—90 percent of global synthetic dyes, roughly 50 percent of world electrical equipment exports. Yet the United States became the preeminent economic power. The difference? The 1862 Morrill Act financed land-grant colleges dedicated to mechanical arts. Engineering schools multiplied from 6 in 1862 to 126 by 1917. By 1900, 88 percent of mechanical engineering students were in land-grant colleges emphasizing practical, applied training—not elite research. America's advantage came not from breakthrough inventions but from what British observers called "adaptation of special apparatus to a single operation in almost all branches of industry."

Innovation without diffusion equals stagnation. The Soviets won every innovation metric and lost the Cold War.

I'd written about this in an article called "24th Place"—named for Microsoft's January 2026 finding that the United States, despite leading the world in AI infrastructure and frontier model development, ranked 24th in AI usage among its working-age population. Behind Ireland. Behind Spain. Behind Belgium.

The diffusion thesis wasn't academic for me. It was becoming central to how I understood everything happening in AI.

Around the same time, I encountered Deloitte's research on "The Pivot to Tokenomics," identifying a fundamental shift in how organizations need to think about AI costs. Their core insight was straightforward but important: AI spend is structurally different from traditional IT spend.

Traditional SaaS is predictable—per-seat licensing, stable monthly costs, budgetable annual commitments. AI is none of these things. Costs spike with usage. Complexity varies by task. Success breeds more usage, which breeds more cost. A customer service AI that resolves 70% of inquiries doesn't just save money—it creates a feedback loop where improved resolution rates drive more customer interactions, which drives more AI usage, which drives more cost. The economics are volatile, nonlinear, and resistant to traditional budgeting.

Deloitte's framework identified tokens as the new atomic unit of AI economics—the currency in which AI value is now denominated. Input tokens (what you send to the model) and output tokens (what the model generates) have different costs. Output tokens cost 3-5x more because they're more compute-intensive. Token prices are declining roughly 30% annually. The mix of model tiers (frontier vs. mid-tier vs. open source) dramatically affects unit economics.

This was useful research. But reading it through Ding's diffusion lens created a different question: How do these token economics interact with everything else happening in AI infrastructure? What happens when you combine declining token prices with edge computing migration, hyperscaler depreciation assumptions, neocloud price competition, and Chinese efficiency breakthroughs? How do these dynamics feed back on each other?

That question launched what would become months of model building.

Before the model could be built, I needed to understand something more fundamental. An earlier conversation had led me down an unexpected path: into statistical mechanics, thermodynamics, and the question of how order emerges from chaos.

I had been developing a framework about systems thinking—a hierarchical progression from Patterns to Frameworks to Systems to System of Systems to Meta-systems, and finally to what I called the "Cognitive Core"—an AI infrastructure for diffusing reasoning capability throughout organizations. But a colleague had challenged me: In an extremely random world, what's the point of seeking patterns? Maybe randomness goes all the way down.

So I asked Claude. What emerged was a synthesis of physics I hadn't fully understood before.

Statistical mechanics shows that individual events can be irreducibly random—each gas molecule bouncing unpredictably—yet aggregate behavior becomes utterly reliable at scale. This isn't averaging away the randomness; it's randomness resolving into pattern through what physicists call the Law of Large Numbers. A single roll of dice is random. A million rolls produce an exact probability distribution.

Prigogine's dissipative structures showed something even more surprising: entropy—the universal tendency toward disorder—can actually create order. When energy flows through systems far from equilibrium, complex structures spontaneously emerge. A hurricane is more ordered than calm air. A living cell is more structured than its components. Disorder doesn't fight order; disorder produces order under the right conditions.

Edge of chaos dynamics from complexity science demonstrated that optimal information processing occurs at phase transition boundaries—not in rigid order, not in complete chaos, but at the critical threshold between them. Systems at the edge of chaos can propagate information across long distances, store historical memory, and perform complex computations.

These physics insights weren't just intellectually interesting. They provided the theoretical foundation for understanding what AI actually does at organizational scale. AI doesn't eliminate randomness—it finds the scale at which randomness resolves into pattern. AI doesn't impose order—it enables the dissipative structures through which order emerges from information flow.

This matters because it changes what you're trying to measure. You're not measuring whether AI "works"—you're measuring whether the organization has crossed the threshold where coherent pattern-recognition emerges from distributed AI interactions. The phase transition is the thing.

Around the same time, I was thinking about Demis Hassabis and a debate he'd been having with Yann LeCun about whether "general intelligence" exists. The conversation led to AlphaGo—and a structural insight about how AI systems actually work.

AlphaGo didn't just beat humans at Go; it demonstrated a fundamental architectural pattern. DeepMind divided the problem into two computational spaces: a policy network that generates candidate moves (divergent capability) and a value network that evaluates board positions (convergent capability). This two-space architecture—separating "what are the options?" from "which option is best?"—appears across AI systems: actor-critic methods in reinforcement learning, generator-verifier patterns in reasoning, exploratory-evaluative dynamics in search.

The insight: this architecture solves computational intractability by operationalizing the divergence-convergence dynamic. You can't evaluate every possible move. But you can generate promising moves and evaluate those. Divergence without convergence is chaos. Convergence without divergence is stasis. The two-space architecture enables both.

This connected to another research thread I'd been following: MIT's work on Sequential Monte Carlo (SMC) methods in Large Language Models. A LinkedIn post had described how standard LLMs use "greedy decoding"—selecting the statistically best next word at each step. This often produces locally optimal but globally suboptimal text. SMC explores multiple paths simultaneously, like "50 scout cars" that communicate and prune bad routes while amplifying promising ones.

The SMC paper formalized something important: the difference between local and global constraint following. Standard LLMs optimize locally at each token. SMC reweights early tokens based on downstream consequences—evaluating trajectory coherence rather than step-by-step optimality.

When I read that, I realized it described most enterprise AI adoption. Organizations optimize locally at each department, each use case, each pilot. They use greedy decoding at the organizational level. What they don't do is evaluate trajectory coherence across distributed decisions. They're missing the meta-layer that could turn local AI wins into system-level intelligence.

The Cognitive Core concept I'd been developing—an AI infrastructure that diffuses reasoning capability throughout organizations—started to look like an organizational implementation of SMC. Not just deploying AI in each function, but orchestrating AI across functions to evaluate global coherence.

All these threads—diffusion theory, physics of emergence, two-space architecture, local vs. global optimization—converged when I started looking seriously at hyperscaler AI economics.

The mathematics are stark. Hyperscalers (Microsoft, Google, Amazon) are collectively spending $400-600 billion annually on AI infrastructure. Current AI-specific revenue is roughly $25 billion. To justify the investment at reasonable hurdle rates, AI revenue needs to grow roughly 80x in five years—to around $2 trillion by 2030. Even optimistic projections (Bain estimates $1.2 trillion) leave a 40% shortfall.

This creates what I started calling the "$800 billion justification gap"—the distance between what hyperscalers are spending and what AI revenue would need to be to justify it through direct returns.

But here's where Ding's diffusion thesis becomes essential. The hyperscalers aren't betting on AI revenue alone. They're betting on diffusion—that AI will transform every workload, that AI capability will diffuse throughout cloud services, that organizations will restructure around AI the way factories eventually restructured around electricity.

The electricity parallel is instructive. The first commercial electric dynamo appeared in the 1880s. Electrification's boost to manufacturing productivity materialized about five decades later—only after factories had restructured their layouts entirely. For decades, factories wired for electricity mimicked the layouts designed for steam power. They put electric motors where steam engines had been, ran central shafts through the building, and saw minimal improvement. The transformation came when factory organization reconstituted around electricity's unique properties.

Today's enterprise AI adoption looks similar. Organizations are installing AI where analysts used to sit. They're adding AI copilots to existing workflows. They're substituting AI into industrial-era organizational structures and wondering why the productivity revolution hasn't materialized. They're electrifying the steam factory.

Meanwhile, another dynamic was unfolding. In January 2025, DeepSeek demonstrated that Chinese AI labs could achieve frontier-competitive performance with dramatically fewer resources—and released those capabilities as open-source models. This wasn't just a technical achievement; it was a strategic move that accelerated the depreciation clock on hyperscaler infrastructure and intensified competitive pressure on pricing.

The model needed to capture all of this: token economics from Deloitte, depreciation dynamics from hyperscaler financials, edge migration projections from NPU capability trajectories, competitive pressure from neoclouds and Chinese efficiency breakthroughs. And critically, how these dynamics feed back on each other.

The model's intellectual foundation came from a different direction entirely—Jeffrey Ding's research on technology diffusion and great power competition.

Ding, a political scientist who won the 2025 Lepgold Book Prize, upends conventional wisdom about technological advantage. His central finding: where innovations are adopted more effectively is more important than where they are first introduced. The true determinant of competitive advantage isn't innovation capacity—it's diffusion capacity. The ability to spread technology throughout an economy, not the ability to invent it.

The evidence is striking. The Soviet Union, by 1970, led the world in R&D spending as a percentage of GNP. They pioneered continuous casting steel technology in the mid-1950s—five years before Japan adopted it. By 1980, only 10.7 percent of Soviet steel used continuous casting. Japan, adopting the technology later, produced 59 percent of its steel with the technique.

General Secretary Brezhnev acknowledged the problem explicitly in 1971: "The links connected with the practical realization of scientific achievements and their adoption in mass production are the weakest." He was describing, with unusual precision for a Soviet leader, a diffusion deficit—a structural gap between the capacity to invent and the capacity to spread.

The pattern repeats across industrial revolutions. Germany dominated the frontier technologies of the late 19th century—90 percent of global synthetic dyes, roughly 50 percent of world electrical equipment exports. Yet the United States became the preeminent economic power. The difference? The 1862 Morrill Act financed land-grant colleges dedicated to mechanical arts. Engineering schools multiplied from 6 in 1862 to 126 by 1917. By 1900, 88 percent of mechanical engineering students were in land-grant colleges emphasizing practical, applied training—not elite research. America's advantage came not from breakthrough inventions but from what British observers called "adaptation of special apparatus to a single operation in almost all branches of industry."

Innovation without diffusion equals stagnation. The Soviets won every innovation metric and lost the Cold War.

Reading Ding's research while thinking about Deloitte's tokenomics created an obvious connection: the hyperscaler AI bet is fundamentally a diffusion bet.

Consider the mathematics. Hyperscalers are collectively spending $400-600 billion annually on AI infrastructure. Current AI-specific revenue is roughly $25 billion. To justify the investment at reasonable hurdle rates, AI revenue needs to grow roughly 80x in five years—to around $2 trillion by 2030. Even optimistic projections (Bain estimates $1.2 trillion) leave a 40% shortfall.

This creates what I started calling the "$800 billion justification gap"—the distance between what hyperscalers are spending and what AI revenue would need to be to justify it through direct returns.

But the hyperscalers aren't stupid. They're not betting on AI revenue alone. They're betting on diffusion—that AI will transform every workload, that AI capability will diffuse throughout cloud services, that organizations will restructure around AI the way factories eventually restructured around electricity.

The electricity parallel is instructive. The first commercial electric dynamo appeared in the 1880s. Electrification's boost to manufacturing productivity materialized about five decades later—only after factories had restructured their layouts entirely. For decades, factories wired for electricity mimicked the layouts designed for steam power. They put electric motors where steam engines had been, ran central shafts through the building, and saw minimal improvement.

The transformation came when factory organization reconstituted around electricity's unique properties—when power could be distributed to individual machines, when layouts could optimize for workflow rather than power transmission, when the entire production system reorganized around what electricity made possible rather than substituting electricity into a steam-era architecture.

Today's enterprise AI adoption looks similar. Organizations are installing AI where analysts used to sit. They're adding AI copilots to existing workflows. They're substituting AI into industrial-era organizational structures and wondering why the productivity revolution hasn't materialized. They're electrifying the steam factory.

When the Deloitte article provided the final piece—a concrete framework for token economics—everything clicked. I had Claude for Excel and Microsoft Garage's Excel Agent available. I had months of accumulated frameworks. I had the ICAEW Financial Modelling Code structure for professional auditability. It was time to build.

With these frameworks in mind—diffusion theory, phase transitions, two-space architecture, local vs. global optimization—I started building what would become a 960-formula model of hyperscaler AI economics. The goal was to capture dynamics that typical financial analysis misses: the feedback loops, the nonlinear thresholds, the interactions between different economic forces.

The model evolved through rapid iteration with Claude. Each conversation surfaced new dynamics that needed modeling. Each modeling exercise revealed gaps in understanding that required more research. Within 48 hours, the model was complete.

The Token Economics layer started from Deloitte's framework. Token volumes are exploding—current estimates show roughly 100 trillion tokens processed annually, growing to over 600 trillion by 2030. But prices are declining 30% annually. The math creates scissoring revenue dynamics: volume growth partially offset by price compression. The model calculates blended rates across model tiers, projects price trajectories under different scenarios, and feeds results into margin analysis.

The Edge Cannibalization layer came from tracking NPU development in devices. Apple's M3 chips have 40+ TOPS (trillion operations per second) of AI processing capability. By 2027, devices will have 100+ TOPS—enough to run capable models locally without cloud round-trips. The model uses an S-curve function to project what percentage of inference workloads migrate to edge, what revenue that puts at risk, and what mitigation strategies might offset the migration.

The S-curve captures the phase transition insight from complexity science: adoption doesn't proceed linearly. It's slow at first (below threshold), accelerates rapidly (crossing threshold), then saturates (above threshold). The formula in the Edge Cannibalization tab implements this:

That formula—and the 959 others like it—embody decisions about how reality works. The steepness parameter controls how rapidly adoption accelerates. The midpoint determines when the transition is halfway complete. The base and final rates set the bounds. Each parameter is an assumption that can be challenged, tested, and refined.

The Depreciation Feedback layer emerged from a realization that accounting assumptions are strategic bets. Hyperscalers depreciate AI hardware over 6 years. But economic useful life may be 3-4 years given technology advancement. This creates a feedback loop: longer accounting life enables more aggressive investment (lower annual depreciation expense frees margin capacity for capex), but builds future write-down exposure if actual useful life proves shorter.

The model calculates write-down exposure under different scenarios—gradual recognition, delayed-then-sudden recognition, and market-forced recognition (when one hyperscaler's adjustment forces industry-wide response). At $400 billion+ cumulative AI capex, the exposure could exceed $100 billion.

The Competitive Dynamics layer models how neocloud price pressure, Chinese efficiency gains, and hardware advancement interact. Neoclouds offer 50-70% discounts on comparable GPU capacity. DeepSeek demonstrated frontier-competitive models at a fraction of training cost. These forces create competitive pressure that feeds back into margin trajectories.

The Dynamic Feedback layer—the largest at 285 formulas—integrates everything. It resolves the circular dependencies between depreciation assumptions, competitive pressure, write-down scenarios, and revised investment returns. It answers questions like: If depreciation exposure forces write-downs, how does that affect margin capacity for further investment, which affects competitive positioning, which affects depreciation exposure?

Here's where this becomes more than a financial modeling story.

The model wasn't built in the traditional way—a financial analyst in Excel, painstakingly constructing formulas cell by cell over weeks or months. It was built through conversation with an AI, iterating through concepts, challenging assumptions, generating code, testing outputs, and refining based on results.

A typical exchange looked like this:

Then we'd build both versions, compare outputs, discuss which assumptions were more defensible, and choose.

Or this exchange about depreciation dynamics:

This collaborative iteration produced something neither I nor Claude would have built alone. I brought the strategic questions and the judgment about what mattered. Claude brought the ability to translate questions into formula structures, test edge cases, and spot inconsistencies across tabs. The conversation wasn't just prompting—it was collaborative thinking that surfaced assumptions, tested frameworks, and resolved tensions between approaches.

And this is exactly what Ding's diffusion framework predicts should happen. The value isn't in the AI having superior analytical capability—it's in the AI making sophisticated analytical processes accessible to people who can apply judgment at the point of decision.

After the initial model was built, I learned about the ICAEW Financial Modelling Code—the professional framework that structures financial models for auditability. The Code recommends a standardized tab structure: Cover → Inputs → Scenarios → Timing → Calculations → Financial Statements → Outputs → Checks → Documentation. Color coding should use blue for hard-coded inputs, black for formulas, green for inter-sheet links.

This was exactly what a professional-grade model needed. So I asked Claude to rebuild the model using ICAEW standards—and to incorporate the new Excel capabilities I'd learned about from that LinkedIn post.

The result was a v4 model with 13 tabs, proper color coding, integrated checks, and formulas using Excel's LET function for readability. The LET function allows defining named variables within formulas:

This formula—calculating hyperscaler market share response to competitive pricing—would have been nearly unreadable in traditional Excel syntax. With LET, each intermediate calculation has a name. The logic is visible.

The model now integrates with Claude for Excel—the Microsoft Garage agent that allows conversational interaction with spreadsheets. You can ask "Explain the S-curve formula in the EdgeCannibal tab" or "Compare NPV across all three scenarios" and get responses that understand the model structure.

This is diffusion in action. The ICAEW structure existed. The LET function existed. Claude for Excel existed. What was new was combining them into a tool that works the way I work—conversationally, iteratively, at the speed of thought rather than the speed of spreadsheet construction.

The timing of this model-building exercise wasn't accidental. I'm currently teaching an Intelligent Automation Immersive program at Divergence Academy, where students are learning to apply these exact capabilities—Claude, modern Excel features, systems thinking frameworks—to build their own analytical artifacts.

Teaching creates its own learning loop. When you have to explain why the S-curve steepness parameter matters, you understand it more deeply. When students ask why depreciation assumptions create strategic vulnerability, you're forced to articulate the feedback mechanisms clearly. When you demonstrate building a model through conversation with Claude, you see the process through fresh eyes.

Divergence Academy's logo incorporates the mathematical divergence operator (∇·F)—a symbol for diffusion, for spreading capability outward. That's not a coincidence. The academy exists to diffuse AI capability to people who can apply it. The model I built demonstrates the same pattern the model analyzes: diffusion trumps capability.

The students in the Intelligent Automation Immersive are learning by doing exactly what I did: reading contemporary research, asking AI to explain and extend it, building artifacts that embody understanding, refining based on feedback. They're not learning about AI—they're learning through AI. The recursion is the pedagogy.

The completed model produces several findings that traditional analysis misses:

The $800 Billion Justification Gap. Hyperscalers are betting the company on AI. Combined annual AI capex of $400-600B against roughly $25B in AI-specific revenue means they need 80x revenue growth in five years to justify the investment through direct returns. The optimistic case ($1.2T by 2030) still leaves a 40% shortfall. This isn't necessarily irrational—they may be betting on AI transforming their core cloud business—but it's a bet, not a sure thing.

The Depreciation Trap. Hyperscalers' 6-year depreciation assumptions create a reinforcing loop that builds future vulnerability. The accounting choice enables aggressive investment today but accumulates write-down exposure that could exceed $100 billion industry-wide if technology cycles prove shorter than assumed. The trigger dynamics matter: market-forced recognition (when one competitor adjusts) can cascade through the industry in quarters.

The Edge Cannibalization Timeline. Inference workloads—the volume business in AI—face significant edge migration pressure. The model projects 70%+ of inference could shift to edge devices by 2030, leaving cloud primarily for training and complex reasoning. The S-curve dynamics mean this migration will be slow-then-fast—unnoticeable until suddenly obvious.

The Stranded Asset Cascade. Different stranded asset scenarios have dramatically different timing. Gradual recognition spreads pain over years. Delayed-then-sudden creates cliff risk. Market-forced recognition can compress years of adjustment into quarters. Organizations need to model which scenario they're in, not assume they can see it coming.

The Chinese Efficiency Factor. DeepSeek-style efficiency breakthroughs multiply competitive pressure. If Chinese competitors can achieve frontier-equivalent performance at 30-50% of US compute costs, the effective competitive landscape shifts dramatically. The model shows how efficiency advantages compound through margin pressure and accelerated obsolescence.

The Feedback Integration. Most important, the model reveals how these dynamics interact. Depreciation assumptions affect write-down exposure which affects margin capacity which affects investment which affects competitive position which affects stranded asset risk which feeds back to depreciation. The Dynamic Feedback tab integrates these loops to show revised returns under different scenarios.

None of these insights are individually new. Industry analysts have discussed each of them. What's new is the integration—seeing how they interact, quantifying the feedback effects, testing sensitivity to assumptions. That integration required 960 formulas. And those 960 formulas were constructed through a process that itself demonstrates how AI changes analytical work.

Step back from the specific model and consider what the process represents.

From research to framework: Deloitte's tokenomics research, Ding's diffusion framework, physics insights about phase transitions and emergence, the SMC paper on local vs. global optimization—these provided conceptual foundations. These weren't discovered through AI; they were human research that the AI helped me access, synthesize, and apply.

From framework to model: The translation from conceptual frameworks to quantitative models happened through conversation. Each conversation surfaced assumptions that needed explicit treatment. Each assumption required research to calibrate. The AI accelerated iteration, but judgment remained human.

From model to artifact: The final Excel model is a tangible artifact that can be examined, challenged, and extended. It embodies decisions. The ICAEW structure makes those decisions auditable. The formula documentation makes them traceable. The model isn't just output—it's crystallized thinking.

This is what "AI in the flow of work" actually looks like. Not AI replacing human judgment, but AI accelerating the cycles through which human judgment gets exercised. Not AI generating answers, but AI helping formulate better questions.

And this connects back to Ding's diffusion thesis. The value of AI isn't in frontier capability—it's in spreading sophisticated analytical capability to the point of decision. The model I built isn't beyond what a skilled financial analyst could create. But it's beyond what I could create at the speed I created it, which means the analytical capability diffused to where I needed it.

There's something recursive about this entire project that matters.

I used AI to build a model analyzing AI economics. The tool and the subject are the same. And in that recursion lies the central insight: learning about AI now requires using AI, and using AI is itself a form of learning about AI.

The physics concepts—phase transitions, emergence, dissipative structures—came to me through conversations with Claude, not through reading textbooks. The financial modeling patterns—S-curves, feedback loops, scenario analysis—got implemented through iterative dialogue, not through Excel courses. The research synthesis—connecting Ding's work to Deloitte's to the hyperscaler financials—happened at conversation pace, not research-library pace.

This isn't just efficiency. It's a different kind of learning. When you have to decide on a steepness parameter for an S-curve, you're forced to think about what drives adoption rates. When you have to model feedback between depreciation and investment capacity, you're forced to understand how accounting creates strategic constraints. When you have to integrate five different risk scenarios, you're forced to consider how they interact.

The model became a learning artifact. Building it was the learning.

This suggests something important about how AI changes learning itself. Traditional learning is absorptive—reading, listening, internalizing. AI-assisted learning can be generative—building, testing, iterating. You understand something when you've modeled it, because modeling forces you to make every assumption explicit.

This is what we might call "learning in public with an AI"—creating artifacts that embody your understanding, testing them against reality, refining based on feedback. The artifact becomes proof of understanding, but also the mechanism of achieving understanding.

What does this process reveal about how AI will actually diffuse through organizations?

First, the value is in the integration. Individual AI capabilities are impressive but often disconnected. The value emerges when AI accelerates complex processes that integrate multiple capabilities—research, analysis, modeling, documentation. This is why enterprise AI adoption is harder than consumer adoption: enterprise value requires integration across domains.

Second, judgment remains central. Every parameter in the model represents a judgment call. The AI helped me iterate through those judgments faster, but it didn't make them for me. Organizations that treat AI as judgment replacement will get worse outcomes than organizations that treat AI as judgment accelerator.

Third, artifacts matter. The model is a tangible output that can be examined, challenged, and extended. This is different from AI-generated text that gets read and forgotten. Artifacts persist, create accountability, and enable building on previous work. AI that helps create artifacts is more valuable than AI that generates ephemeral outputs.

Fourth, the process is the learning. Building the model taught me more than reading about hyperscaler economics ever could. AI-assisted building is AI-assisted learning. Organizations that use AI for building will develop deeper understanding than organizations that use AI for answering.

Fifth, diffusion trumps capability. The AI capabilities used to build this model aren't cutting-edge. GPT-4 class models have been available for over a year. Excel's LET and LAMBDA functions have been available since 2020-2021. The innovation existed. What changed was my ability to access and combine these capabilities for a specific purpose. Diffusion, not innovation.

Sixth, frameworks travel. Ding's diffusion thesis, developed to explain national technological competition, turned out to explain individual tool adoption. The physics of phase transitions, developed to explain matter state changes, turned out to explain organizational AI adoption thresholds. Good frameworks travel across levels of analysis. AI helps you find which frameworks travel where.

I've started publishing these analyses in a newsletter I call The Scaffold—a reference to the temporary structure that enables building something permanent.

The scaffold metaphor captures what AI can be in analytical work. Not the building itself, but the structure that makes construction possible. Not the insight, but the process that enables insight. Not the answer, but the framework for developing better questions.

The 960-formula model is one building. This article is another. The whitepapers on diffusion theory, coherence emergence, and organizational intelligence—those are constructions that the scaffold enabled.

What emerges isn't AI-generated insight. It's human insight, developed faster and more comprehensively than would otherwise be possible. The AI diffuses analytical capability to the point of need. The human applies judgment at that point.

That's not the future of AI. It's the present—for those who've learned to use the scaffold.

And the path to learning how to use the scaffold is... to use the scaffold. Read something interesting. Ask the AI to explain it. Connect it to something else you're thinking about. Build something that embodies your understanding. Test it. Refine it. Share it.

Repeat until you've built something that didn't exist before.

The 960 formulas are just one instance of this pattern. The pattern itself is what matters. And the pattern is available to anyone willing to learn by building.

This article is part of ongoing work developing frameworks for AI diffusion and organizational intelligence. The Hyperscaler AI Economics Model v4, built using ICAEW Financial Modelling Code standards with Claude for Excel integration, is available for examination and extension—including the complete Formula Documentation tracing all 960 formulas.