As I explained in my analysis of the Open-Weight Alliance, this was also one of the main forces behind NVIDIA’s push to open up the model market. NVIDIA does not want value and control to concentrate at the model layer. It wants competition there, while the industry remains structurally locked into its infrastructure layer.

In this analysis, I break down the forces beginning to erode NVIDIA’s moat and explain how the evolving structure of the AI industry could gradually weaken its control.

Every economic transaction in AI, when you strip the wrapper off, is denominated in tokens. A token is the fundamental unit — a piece of text, a subword, an image patch, an audio frame — and everything in the industry is a mechanism for producing tokens, storing them, moving them, or charging for them. Model quality is tokens generated per query that are correct. Model speed is tokens per second per user.

Model cost is dollars per million tokens. Training economics is dollars per training-run token. Cloud gross margin is the spread between the cost of producing a token and the price of selling one. Even valuation ends up back at tokens: a $514B backlog is a promise about future tokens.

Access My Harness Via The BE Platform!

Follow me on The AI Supercycle as well!

Join The Business Engineer Premium Now!

This is not a metaphor. It is the unit. And once you accept that, the whole industry sorts itself into a small number of moving parts: mechanisms that make tokens, mechanisms that gate tokens, and mechanisms that own the customer relationship at the moment a token is priced.

The previous guides in this series — The Fed and the Mint, Kimi K3, Where Google Sits in the Map of AI, Three Tolls on One Road, What the Wallet Sees — each took one slice of the token economy and read it to the ground. This piece pulls them into one map. It is meant to be a reference: the shape of the thing, the players in each region, and the bottlenecks binding each region right now.

The organizing insight is one Andrew Feldman surfaced almost in passing on the Mad Podcast, and once you see it the whole industry looks different: a token has a lifecycle, and the lifecycle has three stages that are as different from each other as manufacturing, warehousing, and retail.

• Training. The weights are made. Sparse linear algebra at planetary scale, one time per model, amortized across every subsequent token that model will ever generate.

• Prefill. The input pass. When you type a prompt, the model has to read it — process every token of context into an internal representation the generation step can use. This step is parallelizable: all the input tokens are known, so the machine can chew through them simultaneously.

• Decode. The output pass. The model generates the answer one token at a time, each new token conditioned on all previous tokens. This step is sequential and cannot be parallelized. Token 500 cannot be computed before token 499. It is a chain, and the chain cannot be broken.

Each of the three has its own physics. Each has its own bottleneck. Each has its own set of players. Each has its own cost curve. And — this is the punchline the rest of the piece is going to unpack — the frontier is moving deeper into the stage where the bottleneck binds hardest, which is decode. Reasoning models spend most of their compute in decode. Agents multiply decode by the number of turns. RL now folds decode into training itself. The memory wall the industry hit in 2024–25 was not a temporary supply crunch. It was the frontier discovering which stage of the token lifecycle it was actually spending its money in.

Let me walk the three stages, then the players, then the economics.

Access My Harness Via The BE Platform!

Follow me on The AI Supercycle as well!

Join The Business Engineer Premium Now!

Read more

This piece answers that question in three moves.

Part I traces how enterprises get captured today, why every prior alternative to that capture has failed, and what specific market condition makes a genuine alternative possible for the first time.

Part II names the category the argument arrives at — the Independence Integrator — and specifies what it is structurally.

Part III gives the full operational playbook for the FDE motion: how to recruit the people, engage the vendors, deliver the double-objective work, run the migration test, and execute the exit.

The piece is longer than most in the series. It is the flagship for the human/organizational layer, and it needs to work both as an argument and as a manual.

Access My Harness Via The BE Platform!

Follow me on The AI Supercycle as well!

Join The Business Engineer Premium Now!

Get The Book - Premium Members Only

Read more

Twenty-five of the largest names in AI — Meta, Microsoft, NVIDIA, Palantir, IBM, Hugging Face, Mistral, and eighteen others — s…

Read more

There was one line in that same earnings call that I did not have the space to properly develop. It deserves its own piece, because it is the single most important disclosure about the shape of the next AI cycle — not the size of it, the shape — that any physical-layer executive has made in six months.

Wei, buried inside a Q&A answer:

“The emergence of agentic AI is leading to a resurgence in the role of CPUs in AI data centers... whatever CPU architecture wins — x86, Arm-based, or RISC-V — they are almost all TSMC’s customers.”

Two sentences. Forty-one words. Everyone chasing the capex line missed the composition line.

Here is what those two sentences actually say. And why do they mean the AI Supercycle just changed compute shapes without changing size?

Access My Harness Via The BE Platform!

Follow me on The AI Supercycle as well!

Join The BE At 50% Off Now!

The disclosure wasn’t a prediction. It was a description.

Six weeks before Wei made that comment, on May 31, 2026, at GTC Taipei, Nvidia had already announced — with almost no fanfare relative to the Rubin GPU headlines — that its next-generation host processor, Vera, was being explicitly positioned as “the CPU for agents.”

Not “the CPU for AI.” Not “the CPU for cloud.” The CPU for agents.

The specs matter less than the branding, but they matter too: 88 custom Olympus cores, ARM v9.2-compatible but not Neoverse-derived (Nvidia’s first fully in-house CPU core design), 1.2 TB/s of memory bandwidth, up to 1.5 TB per socket, with a memory subsystem tuned specifically for what Nvidia called “the control-heavy, data-movement-intensive work that agents generate.”

Nvidia’s own framing in the Vera launch materials — this is worth reading twice:

“The next bottleneck in AI infrastructure is not the GPU at all — it’s the host processor that runs the agent orchestration loop around it.”

Nvidia. The company whose entire market capitalization rests on GPUs being the bottleneck. Saying the bottleneck is moving.

So when Wei made his CPU comment on the earnings call this month, he was not predicting a workload phase change. He was describing the demand pattern that his largest customer had already committed silicon to. The two signals — from the accelerator king and from the foundry that manufactures for it — are converging on the same reading.

That reading is: the workload shape of AI is changing. The compute shape is following.

Compression: When Nvidia and TSMC are simultaneously repositioning around CPUs, the industry has already turned. The market just hasn’t priced it.

The workload shape has changed

Access My Harness Via The BE Platform!

Follow me on The AI Supercycle as well!

Join The BE At 50% Off Now!

Every prior compute regime was defined by a dominant workload shape.

• Training is dense, batched, matrix-bound. Every parameter update touches every core. Memory access is highly regular. This is the workload GPUs were literally built for.

• Chat inference is similar at smaller scale. A user prompt goes in, tokens come out. One long matrix multiplication per token, tiny amounts of control flow between them, no meaningful state to maintain across the request.

• Agentic execution is fundamentally different.

An agent is not a single inference call. It is a loop — a long-running orchestrated process that:

• Plans a multi-step approach from a high-level goal

• Calls tools (code execution, web browsing, database queries, retrieval)

• Waits for tool responses, sometimes seconds, sometimes minutes

• Aggregates results and re-plans

• Maintains state across dozens or hundreds of steps

• Spawns sub-agents and manages their coordination

• Handles errors, retries, and recovery

The LLM inference itself — the actual matrix multiplication that produces the next token — happens in tiny bursts inside this loop. The rest of the time, which is most of the time, the agent is doing work that looks nothing like matrix multiplication. It is control flow, state management, memory access, tool orchestration, and inter-process coordination.

That work has a name in silicon: it is CPU work.

The Exponential View State of the AI Economy report from June makes the volumetric case explicit. A single-turn code reasoning task consumes about 1,190 tokens. A multi-turn code chat consumes about 3,390 tokens. An agentic coding task using OpenHands consumes 4.17 million tokens — roughly 1,200x more than the chat equivalent.

But here is what almost every commentator gets wrong about that 1,200x number: it isn’t 1,200x more inference. It’s 1,200x more total token movement through the system. And most of that movement is orchestration — reading and writing context, passing tool responses around, maintaining state, aggregating outputs. The GPU touches a fraction of it. The CPU touches all of it.

A useful mental model: the GPU is the muscle. The CPU is now the conductor. In a chat workload, the conductor is doing almost nothing — one wave of the baton per token. In an agentic workload, the conductor is running a full orchestra with dozens of independent sections, all of which have to stay coordinated, in tempo, for the entire duration of a piece that may last minutes to hours.

You do not scale an orchestra by hiring more percussionists. You scale it by hiring a better conductor.

The industry has already pivoted. Almost every hyperscaler noticed. The market didn’t.

Access My Harness Via The BE Platform!

Follow me on The AI Supercycle as well!

Join The BE At 50% Off Now!

If you look at what silicon has actually been announced in the last six months — not marketing language, actual silicon — the pattern is unambiguous.

Nvidia — Vera. 88 Olympus cores, Arm v9.2, positioned as “the CPU for agents,” in full production Q1 2026, shipping in the Vera Rubin NVL72 rack alongside 72 Rubin GPUs and 36 Vera CPUs. The rack ratio — 2:1 GPU to CPU by socket — is worth noting. Blackwell’s Grace pairing was 2:1 as well by unit count, but the CPU per rack was less capable and the software framing was “accelerator host.” Vera Rubin flips the framing. The CPU is now equal-status silicon.

AWS — Graviton5. 192 cores, launched June 2026, described in AWS’s own materials as rebuilt for agentic AI. AWS disclosed at the same event that its custom silicon business — Graviton, Trainium, Nitro — had crossed a $20 billion annual run rate. Graviton now accounts for more than half of new CPU capacity added to AWS in the last three years. The Trainium3 UltraServer explicitly pairs Graviton CPUs with Trainium accelerators. The bundle isn’t accidental. It’s the architecture.

AMD — EPYC Venice. 256 Zen 6 cores, TSMC 2nm, paired with the Instinct MI400/MI450X in AMD’s Helios rack-scale platform for multi-gigawatt H2 2026 deployments. AMD has already publicly named the follow-on to Venice — codename Verano, also on 2nm, explicitly optimized for agentic workloads via native LPDDR memory support. The word “agent” now appears on x86 chip roadmaps. That is new.

Google — Axion. Arm-based, deployed as the host processor for the next-generation TPU 8t (training) and TPU 8i (inference) superpods. Google’s own numbers: C4A Axion instances deliver up to 65% better price-performance and 60% better energy efficiency versus comparable x86. Google is not marketing Axion as a general-purpose cloud CPU. It is marketing Axion as the orchestration layer around TPUs.

Microsoft — Cobalt. 128 Neoverse N2 cores, deployed across Azure regions, now powering production workloads for major customers including Databricks. Cobalt is the least aggressive of the hyperscaler custom CPUs in raw specifications, but it is the most deeply integrated into Azure’s control plane — which is precisely the layer that gets stressed by agentic workloads.

Arm — AGI CPU. 136 Neoverse V3 cores on TSMC 3nm, positioned as reference silicon specifically for agentic scaling. Reported $2 billion backlog with hyperscaler design wins. Arm’s CEO Rene Haas, blunt about the reframing: “AI has fundamentally redefined how computing is built and deployed. Agentic computing is accelerating that change.”

Meta. Reportedly signed with AWS to buy Graviton CPUs specifically for its agentic and inference workloads. This is significant. Meta has historically built its own custom silicon (MTIA), and Meta buying general-purpose Arm CPUs from a rival hyperscaler is a very unusual admission that the demand curve is running ahead of internal capacity.

Every single one of those is manufactured by TSMC.

Wei’s comment on the earnings call wasn’t a prediction. It was a description of his own order book. The Vera CPU tapes out at TSMC. Graviton5 tapes out at TSMC. EPYC Venice tapes out at TSMC. Axion tapes out at TSMC. The Arm AGI CPU tapes out at TSMC. The 3nm and 2nm capacity being converted from N-1 nodes, which I discussed in last week’s TSMC piece, is being converted precisely because the CPU designs have started to compete with the GPU designs for leading-edge wafer allocation.

That is what “resurgence in the role of CPUs” actually means when spoken by a foundry chairman. It means his customers have started ordering leading-edge CPU designs at rates that require him to allocate wafers accordingly.

Compression: The industry has already spent the money. The workload shape change has already been ratified by silicon commitments. What remains is for the software layer, the analyst layer, and the market to catch up.

x86’s era just ended, and almost no one wrote about it

Buried in the June 2026 Computex disclosures — the same event where Nvidia’s Rubin platform got most of the headlines — was a data point that will look, in retrospect, like the moment a thirty-year technology regime ended.

Arm crossed 50% of hyperscaler CPU share.

That is not a projection. It is a share of the fleet, measured today. AWS runs more than half of its new CPU capacity on Graviton. Google’s next-generation TPU host is Axion. Microsoft’s Cobalt is in production. And Nvidia’s Vera — powering the flagship rack architecture that Microsoft, Google, AWS, and CoreWeave will all deploy — is Arm.

The economics of getting to 50% are structural, not cyclical. Developing a chip at 3nm or 2nm costs between $200 million and close to $1 billion, over 3–5 years of engineering. Hyperscalers have already committed those resources. Nobody is walking Graviton, Axion, or Cobalt back to x86. Ever.

What’s left for x86 is a defensive game. AMD has done a genuinely impressive job of adapting — EPYC Venice at 256 cores on 2nm is competitive on core count with anything Arm has shipped. Intel is behind but not absent, with Clearwater Forest (288 E-cores on 18A) and Diamond Rapids scheduled for later in 2026. But the interesting competitive question is no longer “x86 vs. Arm.” It is “who wins within the Arm ecosystem.”

That is a very different competitive geometry. And it maps directly onto the hyperscaler alignments that Wei described from TSMC’s side:

• AWS is going custom (Graviton for CPU, Trainium for accelerator)

• Google is going custom (Axion for CPU, TPU for accelerator)

• Microsoft is going hybrid (Cobalt for internal, plus a very large Nvidia relationship for external customer-facing GPU capacity)

• Meta is buying (from AWS, of all places — the surprise move of the year)

• Nvidia is going vertically integrated (Vera + Rubin as a single platform, sold as a unit)

The one common denominator across all five is that they are all TSMC customers. Which is exactly why TSMC is indifferent to which CPU architecture wins. It gets the wafer either way.

Compression: The CPU wars aren’t between architectures anymore. They’re between vertical integration strategies within a single architecture family.

The fleet mix problem

Access My Harness Via The BE Platform!

Follow me on The AI Supercycle as well!

Join The BE At 50% Off Now!

Here is where this becomes an AI Supercycle question, not a chip industry question.

Last week’s TSMC piece — and the Exponential View State of the AI Economy report before it — both organized their analysis around a single core question: can falling token prices move enough token volume to earn a return on the CapEx? The report’s depreciation waterfall assumed a specific mix of what gets consumed. Roughly $848B of 2026 hyperscaler + neocloud CapEx, generating roughly $111B of annual depreciation charge, requiring roughly $223B of AI revenue to clear the hurdle with 50% headroom.

Those numbers assume a fleet whose compute is being sold in a certain shape. Overwhelmingly, until now, that shape has been GPU-hours. Every unit-economic model in the industry — from hyperscaler pricing pages to app-layer per-user cost models — is calibrated to GPU-hours consumed per unit of work.

If the workload shape shifts toward agentic execution, the compute mix consumed to service a given user request changes. Roughly:

• A chat completion: 95%+ of the paying compute is GPU. CPU is orchestration overhead, negligible.

• A single-turn RAG query: ~80–85% GPU, 15–20% CPU (retrieval, embedding lookup, response assembly).

• A multi-turn agent task: possibly 40–60% GPU, 40–60% CPU, depending on the ratio of inference calls to tool calls. This ratio is the number that matters, and there is no industry consensus yet on what it is at scale.

Take that shift seriously and two things become true simultaneously.

First, the hyperscaler fleets built in 2024–2026 will be GPU-heavy relative to the workload they end up serving in 2027–2028. This is the fleet mix problem. GPUs sitting under-utilized while CPU capacity is the constraint. This is the mirror image of the situation in 2023–2024, when CPU capacity was abundant and GPU capacity was rationed. The industry solved the GPU shortage by building GPU-heavy fleets. Now it is at risk of the opposite mismatch.

Second, the per-user economics of agentic products may actually improve, but only for operators who can control the mix. CPU compute is cheaper per unit of work than GPU compute. If you can shift 40% of an agent’s compute onto Graviton or Axion instead of H100s, per-agent margins go up meaningfully — but only if the workload orchestration is fine-grained enough to let you do that split. This is why AWS’s Bedrock+Trainium+Graviton bundle is starting to make commercial sense as a vertically integrated offering, and why Nvidia is pushing Vera Rubin as a single platform rather than as separate CPU and GPU procurements. The bundling strategy is the tell. Everyone building a stack knows what mix they need. Everyone selling a stack is trying to control it.

For app-layer players — Cursor, Claude Code, Devin-class products — this is a moment of genuine strategic choice. The most naïve approach is to remain a “wrapper” and let the model provider price the mix. The most defensible approach is to co-locate parts of the agent orchestration close to the compute, take advantage of the CPU/GPU mix themselves, and start capturing margin that would otherwise flow to Nvidia.

That is exactly the “value moving up the stack” pattern the Exponential View report documented — the app + model layer grew 2.95x in deduplicated share in a year. The agentic CPU turn is one of the specific mechanisms driving that migration.

Compression: The Supercycle just went from a single-variable problem (how much compute) to a two-variable problem (how much, and in what mix). The winners will be the operators who control both.

What the market saw vs. what actually happened

Last week I noted that TSMC’s stock fell 2% on the earnings guidance. Nvidia fell 2.4%. Micron fell 5.6%. Marvell fell 8.7%. The market read raised CapEx as risk.

If you now overlay the CPU disclosure onto that reaction, the misread becomes even clearer. The market was pricing a single question — is there enough demand to justify the buildout? — and used the capex line as the answer. But there were two demand curves in the disclosure. There was a capex curve, which the market interrogated. And there was a composition curve, which the market ignored.

The composition curve is arguably the more important one. It tells you not just that AI compute demand is going to keep growing, but that the type of AI compute demand is going to change in a way that only certain vendors are positioned to serve. Vera Rubin as a unit. Graviton5 + Trainium3 as a unit. Axion + TPU 8t/8i as a unit. Helios (Venice + MI450X) as a unit.

The vendors who win the next three years will not be the ones with the fastest single-chip GPU. They will be the ones with the most coherent bundle — CPU + GPU + interconnect + memory + software — engineered specifically for the orchestration-heavy shape that agents actually have.

That is a very different competitive question than “who has the best GPU.” And it is the question that the market’s reaction to last week’s TSMC earnings suggests it has not yet learned to ask.

What to watch for over the next 24 months

Access My Harness Via The BE Platform!

Follow me on The AI Supercycle as well!

Join The BE At 50% Off Now!

If the agentic CPU turn is real — and every physical-layer signal now suggests it is — then here are the specific indicators worth tracking.

Rubin Ultra timing and CPU ratio (H2 2027). The Kyber NVL576 rack architecture takes 576 GPU compute dies per rack. Does the CPU count scale linearly, superlinearly, or sublinearly? If it scales faster than 1:1, that is Nvidia confirming the CPU-per-GPU ratio needs to keep climbing.

Graviton6 announcement. If AWS announces Graviton6 within 18 months of Graviton5 — a shorter cycle than the historical 2-year cadence — it is because the internal demand signal is running hotter than the current chip can service.

Ampere Computing’s fate. The one independent Arm server chip maker, at 192 cores, with a serious business but no hyperscaler home. Acquisition target — the question is by whom. If Nvidia buys them, Nvidia is doubling down on being the CPU platform. If a hyperscaler buys them, the winner is one that decides it needs Ampere’s IP faster than it can develop its own second generation.

RISC-V’s first hyperscaler design win. Wei explicitly named RISC-V in the same sentence as x86 and Arm — “they are almost all TSMC’s customers.” That is not accidental. The first serious hyperscaler-committed RISC-V design will signal a third leg forming in the CPU market. Watch Meta, watch Alibaba, watch Tencent.

Bedrock + Trainium + Graviton reported margin in AWS Q3/Q4 earnings. If AWS starts breaking out the agentic AI stack as a segment — or if analysts start reconstructing the margin from customer disclosures — the numbers will tell you whether the vertically integrated agent-shaped bundle actually earns economic rent, or whether it just shifts the mix without changing the total.

Meta’s next silicon disclosure. Meta buying AWS Graviton is a very odd move for a company that has historically built everything itself. Either Meta’s own MTIA program has slipped, or the demand curve got so steep that even Meta can’t build fast enough. Either interpretation is a Supercycle signal.

Key mental models

Workload Shape Drives Silicon Shape. Every compute regime is defined by a dominant workload shape. Training and chat inference are matrix-bound and GPU-shaped. Agents are orchestration-heavy and CPU-shaped. When the dominant workload changes, the silicon mix changes with a two-to-three-year lag. That lag is the fundamental clock speed of the AI industry.

The Conductor and the Orchestra. The GPU is the muscle; the CPU is now the conductor. In a chat workload, the conductor is doing almost nothing. In an agentic workload, the conductor is running a full orchestra. You do not scale an orchestra by hiring more percussionists.

Description, Not Prediction. When a physical-layer executive — foundry chairman, chip designer CEO — describes a demand shift, they are describing an order book that has already committed silicon. That is the strongest possible signal, because it comes with sunk capital behind it. Wei’s CPU comment was a description of Nvidia’s Vera, AWS’s Graviton5, AMD’s Verano, Google’s Axion, Microsoft’s Cobalt, and Arm’s AGI CPU — all of which are already in the wafer queue.

The Fleet Mix Problem. The Supercycle’s first phase mismatch. Hyperscaler fleets built in 2024–2026 are GPU-heavy relative to the workload shape that will run on them in 2027–2028. This mismatch — underutilized GPUs alongside constrained CPUs — is likely to be the first localized “bubble deflation” the AI economy experiences. It will not look like the .com bust. It will look like a phase transition inside a single stack layer.

Vertical Integration Beats Best-of-Breed. In the agentic era, the vendor with the most coherent CPU + GPU + interconnect + memory + software bundle wins over the vendor with the best single component. Nvidia understands this — Vera Rubin is sold as a rack, not as chips. AWS understands this — Bedrock + Trainium + Graviton is one story. The chip designers still competing on individual best-chip specs are already fighting the last war.

The Arm Threshold. Arm crossing 50% of hyperscaler CPU share at Computex 2026 is the moment the x86 era ended in the data center. This is a structural, not cyclical, shift — nobody is walking $200M–$1B chip investments back. Every serious hyperscaler CPU roadmap through 2030 is Arm-based. The interesting competition is now within the Arm ecosystem: custom silicon vs. Nvidia Vera vs. reference AGI CPU vs. Ampere.

Two-Variable Cycle. The AI Supercycle is no longer just about how much compute the industry can build. It is about how much and in what mix. Every CapEx, revenue, elasticity, and depreciation-headroom model built in 2024–2025 assumed a single-variable framing. Every model built from here on has to solve for two.

Bundling as the Tell. When every major vendor is trying to sell you the CPU and the GPU as a unit — Vera Rubin, Helios, Bedrock + Trainium + Graviton, Axion + TPU — the vendors themselves are telling you the mix is the strategic variable. If the mix didn’t matter, they’d sell the components separately at maximum-margin prices for each. The bundling strategy is the strategy.

The Two-to-Three-Year Lag. From foundry decision to consumer application. What TSMC and Nvidia decided in early-to-mid 2026 shows up as CPU capacity in hyperscaler fleets in 2027, shows up as agent-shaped software architecture in 2028, and shows up as end-user product economics in 2029. If you are building an agent product now, you are building for a stack that has not yet been fully deployed. That is an execution risk and a positioning opportunity.

The bottom line

Access My Harness Via The BE Platform!

Follow me on The AI Supercycle as well!

Join The BE At 50% Off Now!

Last week’s TSMC earnings had two disclosures. The one everyone covered was the CapEx. The one that matters more was the composition — the workload shape of the next AI cycle, ratified by silicon commitments across every major hyperscaler and chip designer, and confirmed on the receiving end by the foundry that manufactures for all of them.

The AI Supercycle just changed compute shapes without changing size. The bill for the next three years is still the same magnitude, or larger. But the mix of what gets built to service it — what the fleet actually looks like on the racks — is different from what any 2024 model assumed.

Nvidia knew this in May, when it branded Vera “the CPU for agents.”

AWS knew this in June, when it launched Graviton5 “rebuilt for agentic AI.”

AMD knew this before that, when it named its post-Venice x86 CPU after the word “Verano” and marketed it explicitly for agent workloads.

Google knew this when it decided the next TPU host processor would be Axion, not any general-purpose part.

TSMC knew this from the wafer allocation requests it was receiving, which is why Wei made the comment in the first place.

Which leaves the market, the analysts, and most of the AI industry commentary the odd person out — still reading the capex line while missing the mix line entirely.

The mix line is the story.

If you got the TSMC piece last week and thought the takeaway was “AI infrastructure demand is real and multi-year” — that was half of it. The other half is that the shape of that demand is turning underneath the numbers. And the turn is already priced into the silicon, already committed at 2nm and 3nm, already loaded into the wafer queues that Wei’s audit process governs.

The Federal Reserve of AI made a rate decision last week. It also, and more consequentially, made a composition decision. The market has read half of it.

The other half is the piece that will still be paying off in 2029.

Recap: In this issue!

Agentic AI is changing the architecture of compute

• The next phase of AI is defined less by larger models and more by autonomous agents.

• Unlike chatbots, agents continuously plan, execute, coordinate tools, manage memory, and adapt over long-running tasks.

• This changes not only how much compute is required, but the type of compute required.

CPUs are becoming a strategic component of AI infrastructure

• TSMC highlighted that agentic AI is driving renewed demand for CPUs inside AI data centers.

• GPUs remain the engine for training and inference, while CPUs increasingly orchestrate workflows, memory, scheduling, communication, and tool execution.

• Future AI systems will rely on balanced CPU-GPU architectures rather than GPU-centric designs alone.

The transition is already underway

Every major AI infrastructure provider has redesigned its roadmap around agentic workloads:

• NVIDIA → Vera

• AWS → Graviton5

• Google → Axion

• Microsoft → Cobalt

• AMD → EPYC Venice

• Arm → AGI CPU

These are already funded silicon programs, suggesting the industry has committed to this transition well before it becomes visible in software adoption.

AI workloads are becoming orchestration-heavy

Traditional AI applications mainly execute matrix computations on GPUs.

Agentic systems require continuous:

• Planning

• Memory management

• Tool execution

• State persistence

• Multi-agent coordination

• Error handling

As AI applications become more autonomous, orchestration becomes an increasingly large share of total compute.

The strategic question is now compute mix

The next AI cycle is no longer defined simply by expanding compute capacity.

Instead, success depends on optimizing:

• GPU utilization

• CPU orchestration

• Memory bandwidth

• Networking

• Workload allocation

The industry is shifting from maximizing compute to optimizing the composition of compute.

AI platforms are replacing standalone chips

Leading vendors are increasingly competing with integrated platforms rather than individual processors.

Examples include:

• NVIDIA Vera + Rubin

• AWS Graviton + Trainium

• Google Axion + TPU

• AMD Venice + Instinct

Competitive advantage increasingly comes from the coordination of the entire platform rather than peak chip performance alone.

Arm has become the new hyperscaler standard

• Arm now powers more than half of new hyperscaler CPU deployments.

• AWS, Google, Microsoft, and NVIDIA have all adopted custom Arm-based designs.

• The report argues the competitive landscape is no longer x86 versus Arm, but which Arm ecosystem delivers the best integrated AI platform.

AI infrastructure may face a compute imbalance

• Most AI infrastructure deployed over the past several years was optimized for GPU-heavy workloads.

• Agentic AI increases CPU demand much faster than previously expected.

• This raises the possibility that future AI data centers become constrained by orchestration capacity rather than accelerator capacity.

Vertical integration becomes the competitive advantage

Companies that control multiple layers of the AI stack—from CPUs and GPUs to networking and software—are positioned to optimize workload distribution, reduce costs, and improve system efficiency.

The report suggests the next competitive battle will be won by integrated AI platforms rather than by individual hardware components.

The indicators to monitor

The report highlights several signals that will reveal how quickly this transition unfolds:

• CPU-to-GPU ratios in next-generation AI systems

• Adoption of custom hyperscaler CPUs

• Continued growth of Arm in data centers

• Economics of integrated AI platforms

• CPU utilization within agentic workloads

The broader takeaway

The report argues that the next AI cycle is not simply about building more compute—it is about building the right mix of compute.

The industry has already committed billions of dollars toward architectures optimized for agentic AI. As autonomous systems become more common, competitive advantage will increasingly depend on the ability to integrate CPUs, GPUs, networking, memory, and software into a single optimized platform rather than maximizing GPU performance alone.

With massive ♥️ Gennaro Cuofano, The Business Engineer

In this piece, I’ll focus instead on the dilemma Google faces as it navigates that transition and evolves into a fundamentally different kind of company.

Selling the substrate for a $514 billion backlog might be the shrewdest move of the AI cycle — or the most sophisticated version of the classic incumbent’s mistake we’ve seen since IBM shipped the personal computer with someone else’s operating system and someone else’s processor. The Q2 print doesn’t tell us which. What we already know about how technological transitions unfold suggests the question is worth asking before the next four quarters make the answer obvious.

Access My Harness Via The BE Platform!

Follow me on The AI Supercycle as well!

Join The BE At 50% Off Now!

Read more

The headline is easy to state and almost beside the point. Revenue of $119.8 billion, up 24% year over year. Operating income of $40.8 billion, up 30%. Operating margin expanding to 34.0% from 32.4%. Ignore net income and EPS entirely this quarter — the $112 billion net income figure is inflated by a roughly $98 billion mark-to-market gain on Alphabet’s private-company stakes, and it tells you nothing about the operating business. Anyone quoting Alphabet’s “298% net income growth” as an AI signal is reading noise.

The real story is that Alphabet has quietly rebuilt three separate models at once:

• The operating model — what Google is and how it makes money — flipped from consumer advertising to full-stack infrastructure.

• The financing model — how it funds itself — flipped from disciplined single-source (operating cash flow) to aggressive three-source (cash + debt + equity), with free cash flow going negative for the first time in the modern era.

• The investing model — where the capital goes — flipped from disciplined capex + buybacks + M&A to a triple play of physical build-out, cross-ecosystem equity stakes, and hardware stockpiling.

Each reinvention is a real break with two decades of practice. Read together, they position Google as the only company that credibly occupies every layer of the Map of AI — a distinction that raises a question the good numbers paper over, and one Google will have to answer aloud within the next year: is this a consumer company that owns infrastructure, or an infrastructure company that owns a consumer surface?

Let me walk each reinvention, and then place Google on the map.

Access My Harness Via The BE Platform!

Follow me on The AI Supercycle as well!

Join The BE At 50% Off Now!

Read more

Curated Version Coming On AI Supercycle

Read more

On July 15, ASML raised its full-year guidance and told the market it will expand production capacity by roughly 30% in each of the next two years. Two days later, TSMC raised its capital spending, committed another $100 billion to Arizona, and announced 13 new fabs at home.

Both companies see multi-year visibility they’ve never had before. Both are raising output mid-year. Both are committing to physical expansions that only pay off at the end of the decade.

The two companies with the most reliable read on AI demand — because they carry the highest cost of being wrong — are betting on it more aggressively than at any point in their histories.

That should settle a question that’s been running loudly for six months. It doesn’t, because the AI-skepticism camp has split cleanly in two, and reading the market response to this print cycle requires knowing which camp you’re listening to.

One camp — the disingenuous one — is arguing that AI demand is fake, artificially inflated, a hyperscaler capex bubble justified by circular contracts and marketing.

This is the loudest camp on social media, and it is not what the numbers show. TSMC audits data-center construction before it commits at the foundry. ASML cross-references the whole contract chain before it commits at the equipment layer.

Both discount their customers’ forecasts and both still raised. The idea that TSMC and ASML — the two most physically exposed companies in the whole stack — are collectively hallucinating a fake demand curve is not a serious position.

The other camp — the serious one — is arguing something more interesting: that AI demand is real but the AI industry is now running on three different clocks that don’t match, and the interesting risk lives in the gaps between them. That is a much harder argument to dismiss, and the rest of this piece is really about that argument.

Something else changed in this print cycle, too, and it matters more than the numbers. The market has started to distinguish between companies that control scarce production capacity and companies that merely benefit from the AI narrative. TSMC held. NVIDIA, Micron, ARM, Marvell — all fell, some sharply. The toll booths held up better than the tourists.

That divergence is the frame for everything that follows.

Access My Harness Via The BE Platform!

Follow me on The AI Supercycle as well!

Join The BE At 50% Off Now!

Read more

The argument is about the tape, not the technology.

On July 17, the chip complex sold off hard. TSMC fell 7% the same day it posted a 77% jump in quarterly operating profit. SoftBank dropped 9%. The Nasdaq closed −1.5%. The Philadelphia Semiconductor Index pushed further into bear territory. Within hours, the tape had a name for it: the Kimi Moment. Another Chinese lab closes the gap, another repricing of the AI capex thesis.

That story is wrong on two levels. Both structural. Both worth naming.

• Level 1 — Attribution Substitution. The visible catalyst is not the causal one. The July 17 move was months of geopolitical damage to the physical stack, absorbed by the tape in a form it could name.

• Level 2 — Interpretation Inversion. Even if you granted the tape its catalyst, the bearish read is inverted on almost every claim it rests on. Cheaper models expand demand rather than compress it. Open weights push closed labs outward rather than inward. Advertised innovation is not the same as frontier innovation. Local inference does not kill capex.

Together, that’s the Kimi Delusion.

Read more

That is the current stock market valuation resting, in one form or another, on the assumption that generative AI is a real technology wave with real revenue.

It is also, until now, the number no one has been able to properly interrogate — because the demand side of AI has been fundamentally opaque. The supply side (chips, hyperscalers, energy) is well-documented.

The demand side — what customers are actually paying for, and whether the revenue is real — has been buried inside private company reports, segment totals, and circular flows between OpenAI, Anthropic, the hyperscalers, and the app layer.

A few days ago, in our AI Supercycle session, we sat down with William Gildea from the Exponential View team to walk through their new State of the AI Economy report — the most rigorous attempt I’ve seen to solve the demand-side visibility problem.

What the Exponential View team did — building bottom-up financial models for 1,000+ firms, tracing every revenue line back to primary filings, and then deduplicating the flows so a single dollar doesn’t get counted three times as it flows down the stack — is exactly the kind of work the industry needs but almost nobody had done.

Below is a compressed strategic reading of that report — organized around the five layers they use (Demand, Economy, CapEx, Tokens, Stack), along with the mental models that matter for anyone trying to navigate this cycle over the next decade.

Access My Harness Via The BE Platform!

Follow me on The AI Supercycle as well!

Subscribe Now at 50% Discount, Expiring!

Demand: Real, Big, and Faster Than Any Wave Before It

Access My Harness Via The BE Platform!

Follow me on The AI Supercycle as well!

Join The BE At 50% Off Now!

The headline number: $110 billion in trailing 12-month deduplicated GenAI revenue, now running at a $175 billion annualized pace. That gap — 1.6x between banked and run-rate — is itself the growth signal.

Three things make this number different from prior wave estimates:

It’s deduplicated. When a Cursor customer pays $100, roughly $60 of that flows to Anthropic or OpenAI as token spend, and roughly $30 of that flows to Azure, AWS, or a NeoCloud for compute. If you naively stacked those, you’d count $190 of “AI revenue” from $100 of actual customer demand. Exponential View built full P&Ls and cash-flow statements for every major player and attributed value-add layer by layer: $40 to the app, $30 to the model lab, $30 to hosting. It sums to $100, not $190. That is what “real revenue” means in this report.

It grows 3x faster than any prior IT wave. Internet (1995), mobile apps (2007), cloud (2010), GenAI (2023) — time-aligned to year zero, GenAI is on a trajectory 3x steeper than any of them. This is not a qualitative claim. It is the measurement of the first three years of revenue for each wave, adjusted for inflation.

Each new $1B of cumulative revenue is arriving 90x faster than in 2023. In 2023, it took the AI industry ~180 days to add $1 billion in cumulative revenue. In mid-2026, it takes less than two days. That is a compounding curve that the traditional “compare it to the .com era” mental model will systematically underestimate.

The reason this wave is faster isn’t primarily that AI is a “better” foundational technology than the internet was. The internet was arguably more foundational. The reason is compounding: AI is arriving in a world that already has the internet, mobile, cloud, ubiquitous payments, mature distribution, and instant global software delivery. Every prior wave is now infrastructure for this one. What took the internet a decade to distribute takes a foundation model six months.

That has two important implications the report is explicit about:

• Modeling projections against prior waves systematically understates AI. Demand projections, depreciation schedules, and CapEx payback timelines built on internet-era analogies will all be too slow.

• Society and institutions have far less time to adjust. The internet took a decade to reshape work, media, and family life — and we were still caught off guard. GenAI is being given twelve to twenty-four months. The adjustment gap is going to be far more violent.

Meanwhile, the compute supercycle is on. Global semiconductor revenue reached $792B in 2025, projected to hit $1.51T in 2026 — nearly double in one year, when the industry historically grew 5–8%. The US power sector, which had flatlined at ±0 TWh/month annual growth from 2008 to 2024, is now growing at +9 TWh/month. The largest data centers have grown ~50x in four years — from ~20 MW supercomputers to xAI’s Colossus (~300 MW) to Anthropic’s Rainier (targeting ~2.3 GW across ~30 buildings).

And each new data center dollar buys different things than it did five years ago. Chips’ share of build cost went from 40% to 60%. Memory went from a 2% rounding error to ~18%. The buildout has become vastly more silicon-heavy and vastly less concrete-and-cooling.

The Compression: Every naive analogy for AI understates it. This is a computing revolution being built on top of a compounding stack of prior revolutions.

Economy: Big Is Still Small, and Still Early

Access My Harness Via The BE Platform!

Follow me on The AI Supercycle as well!

Join The BE At 50% Off Now!

This is the section that makes the report honest.

Because for all the growth above, GenAI revenue in mid-2026 is still equivalent to just 0.42% of US GDP — versus the IT sector’s 9.4%. Corporate profits alone are 32x larger than all deduplicated GenAI revenue combined.

At the individual company level, the picture is even more disciplined. Uber’s public commitment to spend up to $1.5k per engineer per month on AI puts them in the top 10% of per-employee AI spend among 70,000 US businesses tracked by Ramp. And even at that cap, the total ($~90M/year for 5k engineers) is a rounding error against Uber’s $8.7B EBITDA and $14B in people OpEx.

That is the shape of the AI economy right now: it is growing at unprecedented speed from a tiny base. Both facts are true. Both matter.

There are two more subtle observations worth pulling out of this section:

Seven in ten corporate AI claims are still about cost, not revenue

Exponential View parsed S&P 500 earnings calls from Q4 2022 to Q1 2026 and coded every AI impact claim. The distribution is telling: 25% cost reduction, 23% time savings, 22% throughput, 18% quality, 7% conversion, and just 6% revenue gain.

Roughly 70% of corporate GenAI claims are efficiency plays. Only 6% are top-line revenue claims. This matches every prior general-purpose technology cycle — companies chase the cost side first because a $1M saving flows to the bottom line at 100% margin, while a $1M revenue gain comes with cost of sales. But it also means the growth story you hear from the tech supply side is not yet the growth story you see in enterprise P&Ls.

The consumer surplus is a $3–4B/month invisible economy

The Stanford Digital Economy Lab, led by Erik Brynjolfsson, has been running the “how much would you need to be paid to give up all AI tools for a month?” study. Applied to US consumer welfare, they estimate GenAI is producing ~$14B/month of consumer surplus against ~$11B/month of measured revenue.

That is roughly $3–4B/month of value that never touches anyone’s income statement — a 30% uplift on top of the reported revenue that GDP has no way to capture.

The historical analogies are worth naming. Between 1880 and 1920, electric light became ~99.97% cheaper. An hour’s wage bought 40,000x more light. The direct GDP impact was approximately zero, because the pricing collapsed faster than the volume expanded in the accounts. Between 2000 and 2020, free search, encyclopedias, and maps displaced paid alternatives — Google Search alone generates an estimated $17.5k/year/person of consumer surplus that shows up nowhere. GDP knows the price of everything and the value of nothing.

The Compression: A growing GenAI economy is not automatically good news for OpenAI, Anthropic, or the hyperscalers. But it is unambiguously good news for consumers — and that value is real, permanent, and mostly invisible in the official statistics.

CapEx: The Largest Buildout in Tech History Is Paying Back — For Now

Access My Harness Via The BE Platform!

Follow me on The AI Supercycle as well!

Join The BE At 50% Off Now!

Here is where the tension in the AI Supercycle lives.

Cumulative hyperscaler + NeoCloud CapEx will reach ~$2 trillion through 2026. The 2026 figure alone is ~$848B, up from $459B in 2025 and $268B in 2024. Of that $848B, roughly $535B sits above the pre-AI trend line — that is the pure AI-attributable CapEx wedge.

And every major forecaster has spent the last three years chasing the actual number higher. The report shows the analyst forecast fan-out to 2030 — every line, from Barclays to JPMorgan to Morgan Stanley to Goldman, has been repeatedly revised higher as actuals have kept exceeding the previous consensus.

The financing shift is where systemic risk enters

In 2022–2024, hyperscaler CapEx was overwhelmingly funded from operating cash flow. Hyperscalers were cash-rich. That’s changed. External capital — debt, equity, and leases — is now filling the marginal dollar. NeoCloud CapEx is primarily debt-funded. Even hyperscaler CapEx now shows a growing debt and lease component, especially at Oracle and Meta.

The report is precise about what this means: paying cash keeps a bad bet inside the firm — only denting profits. External funding moves the risk outside the firm, where third parties expect repayment. As external financing grows, the interconnectivity of the AI economy with the wider financial system grows with it. That is the mechanism by which a localized underperformance in one layer becomes a broader problem.

The depreciation waterfall is the honest test

If you want to answer “is the AI CapEx paying back?” you cannot compare this year’s spend to this year’s revenue. CapEx is a stock; you build the data center once, and depreciate it over 6 years for chips, 14 years for the building envelope. Each year, only a fraction of the spend passes through as a cost — but that fraction accumulates.

The report’s depreciation waterfall shows this beautifully. The 2026E annual AI CapEx depreciation charge reaches ~$111B — but it took cumulative CapEx of ~$1 trillion to get there. And to justify that depreciation with 50% infrastructure margin headroom, the industry needs ~$223B/year of revenue supporting it.

Here’s the crossing point that matters: In Q4 2025, for the first time, quarterly AI revenue exceeded quarterly AI CapEx depreciation. By Q1 2026, hyperscaler + NeoCloud AI revenue has 19% headroom over depreciation. Across all GenAI revenues (including the app and model layers), the headroom is 32%.

That is the punchline of the CapEx chapter, and the reason the subtitle is “paying back (for now)“. The current buildout is being covered by the current revenue run rate. What’s not yet covered is the cumulative bill — the accumulating depreciation stack from prior years — plus the additional capacity that hasn’t yet come online.

Two mechanical levers can rescue the math

• Longer chip useful lives. Depreciation schedules assume 6 years for chips. If actual useful life is 8 years — supported by the fact that 8-year-old T4 GPUs still earn 42% gross rental yields — Q1 2026 infrastructure headroom moves from 19% to 36%. Every additional year of chip life materially compounds the payback.

• Rising rental rates. H100 1-year rental contracts collapsed from a $3.05/hour launch premium to $1.70/hour peak-overbuild fear in mid-2025. In the last two quarters they’ve climbed back to $2.40/hour, driven by surging inference demand. That is what utilization catching up to supply looks like in real time.

The Compression: The AI CapEx race is not currently in bubble territory. It is in “revenue must keep compounding, or the depreciation stack catches up” territory. Those are very different states.

Tokens: The Unit of Value, and Why That Framing Is Almost Right

Access My Harness Via The BE Platform!

Follow me on The AI Supercycle as well!

Join The BE At 50% Off Now!

Jensen Huang says the input is electrons and the output is tokens. Sundar Pichai calls tokens the fundamental unit of AI. It is the closest the industry has come to a shared unit of account.

The Exponential View position: sort of. Tokens are AI’s billing metric — but not yet a proper unit of value.

The numbers are staggering. Global inference token volume exceeded 30 quadrillion per month in mid-2026, growing 14x year-on-year. Blended API prices collapsed from ~$17 per million tokens to ~$2 over 18 months — a 90% decline for a given intelligence level. Meanwhile the Epoch Capabilities Index moved from 112 to 158 for frontier models. Cheaper AND better, at the same time.

Elasticity is what makes this cycle work

Every 10% price cut generates 12–18% more tokens. This is measured price elasticity of 1.2 to 1.8 — well above 1.0, which means total token spend keeps rising even as prices collapse. Sundar Pichai’s I/O 2025 quote nails it: Google’s monthly token processing went from 9.7T to 480T tokens — a 50x rise against a 97% price decline. Similar patterns hold at ByteDance and OpenAI.

The elasticity is the mechanism that lets the industry serve a compounding depreciation stack with collapsing unit prices. Volume compounds faster than price falls. So far.

The chat-to-agent transition is a token amplifier

A single-turn code reasoning task consumes ~1,190 tokens. A multi-turn code chat consumes ~3,390 tokens. An agentic coding task using OpenHands consumes ~4.17 million tokens — roughly 1,200x more than the chat equivalent. As users shift from chat to agents, per-user token consumption isn’t rising linearly. It’s rising by three orders of magnitude.

That is why agent coordination density (measured as % of prompts that result in a tool call) has tripled since January 2026, from ~5% to ~15% on OpenRouter. Every meaningful improvement in agent capability structurally multiplies inference demand.

Quality-adjusted output tokens: the metric to internalize

This is the methodological contribution from the Exponential View team that I think will become standard.

Raw token counts are deceptive. You can burn arbitrary reasoning tokens to marginally improve output quality. What the user actually receives is only the output tokens, not the input or the reasoning tokens. And output from a better model is worth more per token.

So: quality-adjusted output tokens = output tokens × model capability score.

Between January 2025 and April 2026, raw total tokens grew ~39x. Raw output tokens grew only ~30x (because more of the growth was in input and reasoning). But quality-adjusted output tokens grew 33x to 59x depending on which capability index you use (Epoch, METR, or Artificial Analysis).

The gap between raw and quality-adjusted growth tells you something about where the actual value is being created. It also solves the “how do we measure output when reasoning tokens inflate the counts?” problem in a clean way.

Data-center-scale monetization is climbing

There’s a fascinating scissor chart deep in the tokens section. Revenue per trillion tokens has fallen since its 2023 peak, mirroring the price decline. But revenue per gigawatt of data center capacity has climbed past $7B/GW — because the token throughput per GW has grown far faster than the per-token price has fallen. Same physical asset, dramatically more monetization per unit of capacity per year.

The Compression: Cheaper tokens plus better models plus agentic workflows equals an inference demand curve that outruns the price collapse. The economics work as long as the elasticity holds.

Stack: Where the Value Is Captured — and Where It’s Moving

Access My Harness Via The BE Platform!

Follow me on The AI Supercycle as well!

Join The BE At 50% Off Now!

This is the section that anyone building or investing in AI needs to read carefully.

Today, revenue is concentrated: NVIDIA alone captures a run-rate ~$300B in Q1 2026, dwarfing every other player in the stack. TSMC is at ~$90B. The hyperscalers, model labs, and app layer combined are still a fraction of the chip layer.

But the deduplicated share is moving. Between Q1 2024 and Q1 2026:

• Hosting layer: 89% → 82%

• Foundation model layer: 8% → 11%

• App layer: 3% → 7%

Deduplicated app + model revenue is growing at 2.95x per year — nearly 3x annually. Value is migrating up the stack, exactly as every prior tech wave eventually did.

Pricing power follows competition, not stack position

The Herfindahl-Hirschman concentration analysis is the sharpest single insight in the whole report. Concentration by layer:

• Chips: HHI 4,775 — highly concentrated (NVIDIA)

• Hosting: HHI 2,589 — moderately concentrated

• Apps: HHI 2,597 — moderately concentrated

• Foundation Models: HHI 1,594 — least concentrated when you measure by token share, because of the open-weight competitors

That last bullet is crucial. Foundation model revenue is concentrated in OpenAI and Anthropic. But foundation model tokens served — the actual production workload — is far more contested because open-weight models (DeepSeek, Kimi, MiniMax, Qwen) are consuming a growing share of inference volume.

The frontier premium is real but time-limited

Frontier Labs can charge a premium for the current-quarter frontier. GPT-4.5 launched at $100/M tokens. Opus 4.1 at similar levels. But last year’s frontier commoditized fast. GPT-4-level capability that cost $30/M tokens in 2023 costs ~$0.10 today. Anthropic’s Opus 3 pricing at launch → GPT-5-nano pricing today: a ~300x price collapse for equivalent capability.

The line to watch: labs earn margin on the rate of frontier improvement, not on holding any static position. The moment the open-weight tier catches up to a given capability tier, the closed-weight premium at that tier collapses.

The OpenRouter signal — quiet but important

Among OpenRouter’s self-selecting “model-routing” users — the sophisticated buyers who explicitly choose which model to route each request to — Google + OpenAI + Anthropic’s combined token share fell from 72% to 33% over the last year. DeepSeek, MiniMax, Tencent, Xiaomi have picked up the difference.

This isn’t representative of the whole market. It is representative of what happens when a buyer is willing and able to compare across all model options. It suggests the closed-weight premium is more fragile than the revenue numbers alone imply.

Labs are responding — vertically

Under pricing pressure, both major labs are pushing into apps (Claude Code, Codex, Claude for Legal, Codex for Legal, Claude for Education) and into infrastructure (Anthropic’s $50B American AI infrastructure commitment with Fluidstack; OpenAI’s Stargate — $500B over four years, $100B deploying immediately). They are trying to occupy both the layer above and the layer below simultaneously, because staying in the middle is where the margin gets squeezed.

Four end-state scenarios worth internalizing

The report closes with four scenarios for how the stack economics could play out. Compressed:

In every scenario, consumers win. In some scenarios, producers also win. In none of the scenarios does today’s revenue distribution across the stack persist unchanged.

That is the underlying structural reality of the AI Supercycle: the stack is not a steady state. It is a distribution being actively contested from both ends.

The Compression: No one in this stack is a static actor. Every layer is trying to eat the layer above it and the layer below it simultaneously. In the AI Supercycle, vertical integration is the default response to competitive pressure — from both directions.

The single sentence that closes the report

Access My Harness Via The BE Platform!

Follow me on The AI Supercycle as well!

Join The BE At 50% Off Now!

Exponential View ends the whole 60-slide deck with one line, and it’s the right one:

“AI demand is more revenue-validated than any prior platform shift. The investment case comes down to whether falling prices can move enough token volume to earn a return on CapEx.”

That is the AI Supercycle in one sentence. Real demand. Real payback pressure. Collapsing unit prices. Rising volume. Nobody sitting still. The math works — but only if the elasticity holds and the volume keeps compounding.

If you take one thing away from this session, take that sentence.

What This Means for Operators — My Reading

Access My Harness Via The BE Platform!

Follow me on The AI Supercycle as well!

Join The BE At 50% Off Now!

Zooming out from the report and into what I think this means practically for anyone building in or around AI right now:

1. The bear case (”it’s all circular financing”) is dead in its strongest form, alive in a narrower form. Deduplicated revenue is $110B trailing. That’s real external demand. But the depreciation stack is still growing faster than revenue in absolute terms, and external financing is a rising share of the funding mix. The system is finely balanced, not collapsing.

2. Consumer surplus is the story nobody is telling. $3–4B/month of value is being delivered outside any income statement. Even if every AI-native company failed, that surplus is real and permanent. You cannot uninvent knowledge worker productivity.

3. Value is moving up the stack, but the ceiling is competition. The app layer will earn margin — for the specific players who can defend proprietary data, domain-specific workflows, and increasingly, their own models. Generic wrappers are being squeezed from above by Claude Code / Codex-style lab-native apps, and from below by cheaper open-weight tokens. The middle is the hardest place to stand.

4. The bubble, if it deflates, will not deflate the way the .com bubble did. This is a stack of overlapping S-curves — chips, energy, models, apps, agents — each with its own bottleneck, its own capital demand, and its own timeline. What we’ll see is a sequence of smaller, localized re-pricings, not one big one. Which means the survivable strategy is to be in multiple layers, not just one.

5. Watch the elasticity number. The 1.2–1.8 elasticity is the single most important economic variable in the whole cycle. If it stays above 1.0, volume growth outruns price collapse and the industry economics work. If it drops below 1.0 — because agentic workflows saturate, or because open-weight competition compresses the frontier premium too fast — the depreciation math gets very hard, very quickly.

Recap: In This Issue!

AI demand is real and accelerating

• Deduplicated GenAI revenue reached $110B TTM, running at an annualized pace of roughly $175B.

• Revenue is growing significantly faster than previous technology waves, including the internet, cloud, and mobile.

• AI adoption is compounding on top of existing infrastructure (cloud, internet, payments, software), dramatically compressing adoption timelines.

Revenue quality is stronger than many assume

• The report removes double-counting across the AI stack (applications, model providers, cloud infrastructure).

• Instead of counting the same dollar multiple times, it attributes revenue to where value is actually created.

• This challenges the narrative that AI revenue is largely circular spending within the ecosystem.

The AI economy is still in its infancy

• Despite explosive growth, GenAI revenue remains a very small share of the broader economy.

• Most enterprise AI spending is still focused on productivity improvements rather than revenue generation.

• The report suggests the current opportunity is much larger than today’s measured revenue implies.

Consumer value exceeds measured revenue

• AI is already creating substantial consumer surplus that does not appear in GDP or corporate financial statements.

• Users receive significant value from AI tools without directly paying for it, similar to what happened with Google Search and other internet services.

• The economic impact of AI is therefore larger than current financial metrics suggest.

The AI infrastructure buildout remains justified—for now

• AI infrastructure investment continues at an unprecedented pace.

• Current revenue generation is now covering annual depreciation costs for infrastructure, indicating that the investment cycle remains economically sustainable.

• The key risk is whether demand continues to compound fast enough to absorb future capacity.

External financing is becoming more important

• Early AI infrastructure expansion was largely funded by hyperscaler cash flows.

• Increasingly, debt, leases, and external capital are funding incremental investment.

• This creates greater financial interconnectedness and increases systemic risk if AI demand slows unexpectedly.

Tokens have become the economic engine of AI

• Token volumes continue to grow exponentially while token prices continue to fall.

• Demand has remained highly elastic: lower prices generate proportionally more usage.

• As long as token consumption grows faster than prices decline, industry revenues can continue expanding.

Agentic AI changes the economics

• Autonomous agents consume orders of magnitude more tokens than traditional chatbot interactions.

• Multi-step reasoning, planning, memory, and tool use dramatically increase inference demand.

• Agentic AI is therefore expected to become a major driver of compute consumption over the next several years.

Value is gradually moving up the stack

• Today, NVIDIA and infrastructure providers capture most of the economic value.

• Over time, a growing share of value is shifting toward foundation models and application developers.

• This mirrors previous technology cycles, where value progressively migrated from infrastructure to software.

Competition is increasing across every layer

• Open-weight models are rapidly eroding pricing power for older frontier models.

• Frontier labs are responding by expanding both upward into applications and downward into infrastructure.

• Every participant is attempting to vertically integrate to defend margins.

The critical variable is elasticity

The report concludes that the AI investment thesis depends on one central question:

• Can falling token prices generate enough additional demand to offset the enormous capital invested in AI infrastructure?

If token demand continues growing faster than prices decline, the economics of the AI supercycle remain intact. If elasticity weakens materially, pressure will build across the entire AI ecosystem.

With massive ♥️ Gennaro Cuofano, The Business Engineer

But the function is the same.

This Thursday, C.C. Wei delivered TSMC’s second quarter results:

• Revenue of $40.2 billion, up 33.7% year-over-year

• Fifth consecutive record quarterly profit

• Full-year growth revised up to “slightly above 40%” — from the previous “above 30%”

• An additional $100 billion commitment to Arizona, bringing US commitment to $265 billion

• 13 more fabs planned in Taiwan

• Capital spending raised for the third time this year, now $60–64 billion

• Forward guidance that the next three years of investment will be significantly higher than the past three

TSMC’s stock fell 2%. Nvidia fell 2.4% alongside it. Arm fell 5.4%. Micron fell 5.6%. Marvell fell 8.7%.

The market read this as an overspending signal. That reading is wrong — and understanding why it’s wrong requires seeing TSMC as what it has actually become: the central bank of the AI industry.

The rate decision no one is calling a rate decision

When the Federal Reserve meets, it doesn’t just report the economy’s condition. It sets the price of money — which then transmits through banks, credit markets, business investment, employment, and eventually prices for everything else. The meeting is important not because of the reporting but because of the decision embedded inside it.

TSMC’s quarterly earnings do the same thing for AI compute. The numbers are the reporting. The capex line, the node roadmap, and the wafer allocation are the decision. And the decision has become so consequential that no other single event in the AI industry — not an Nvidia earnings call, not an OpenAI product launch, not a hyperscaler capex announcement — carries as much information about what the next three years of AI look like.

Consider what the market actually learned this week.

The wafer count barely moved. TSMC shipped 4,336 thousand wafer-equivalents in the quarter, versus 4,174 in the prior quarter — a 4% sequential increase. But revenue grew 12%. The eight-point gap is the average selling price of a wafer, which climbed from roughly NT$272,000 to NT$293,000 in a single quarter.

That’s not price extraction; TSMC explicitly refused to raise prices. It’s customers migrating up the technology ladder, and TSMC letting them do it at the margins the ladder produces.

Read as a central bank action, that’s the equivalent of the Fed shifting the composition of its balance sheet toward longer-duration securities without changing its overall stance. The base rate — the wafer count — is stable. The mix is where the policy lives.

At the same time, TSMC’s capex path this year has moved three times:

• January 2026: $52–56 billion

• April 2026: closer to $56 billion

• July 2026: $60–64 billion — plus $100 billion additional Arizona commitment, plus 13 Taiwan fabs

That is forward guidance in its purest form. And Wendell Huang, the CFO, went further in the Q&A: the next three years of investment will be “even more significantly higher” than the past three.

That is the equivalent of a central bank publishing not just a rate decision but a five-year path. And it’s being delivered by the only institution in the AI industry whose commitments cannot be reversed. Concrete doesn’t uncure. Fabs don’t relocate. Nodes don’t undevelop. Every dollar of TSMC’s capex is a physically irreversible bet on what AI demand will look like in 2028.

Reading the dot plot

What did the guidance actually say?

Read more

The Five C’s Reread established the architectural principles for building an AI stack that resists lock-in to any specific vendor.

This piece answers the operational follow-up: given that no enterprise can reliably predict which coalition wins, how does an enterprise actually build so that its leverage and its compounding surface survive whichever one does?

The core insight is that the winning-coalition question is the wrong question for a buyer to try to answer. A buyer who bets correctly on Coalition 2 winning still ends up locked into whichever Coalition 2 vendor they picked, and that lock-in eats their leverage over the 5-10 year enterprise horizon. A buyer who bets wrong — betting on Coalition 1 while Coalition 3 emerges as the dominant substrate, for example — ends up rebuilding their AI stack in year 4. Both bets have the same structural problem: the enterprise’s fate becomes coupled to a single coalition’s success, and the enterprise pays the switching cost when the coalition landscape evolves.

The right posture is different. Architect so the same underlying stack works whether Coalition 1, Coalition 2, or Coalition 3 dominates. This is what “regime-agnostic” means. Not neutral — the enterprise still buys products, still signs contracts, still adopts vendors — but structurally arranged so that the coalition warfare above the enterprise’s architectural layer does not translate into forced migrations, hostage renegotiations, or catastrophic re-platformings. Coalition warfare becomes market volatility the enterprise participates in but does not depend on.

This piece walks through what regime-agnostic architecture looks like concretely: the survival playbook for each coalition winning scenario, the portfolio approach to distributing coalition exposure, the specific components worth owning versus outsourcing, the enterprise-size gradient of pragmatic implementations, and a concrete action timeline for the next 18 months.

Read more

And once you do understand the dynamics behind the Enterprise AI Coalitions.

That is how you can really grasp the Five Cs.

Control, Capability, Choice, Cost, and Compound are architecturally right. They correctly identify the load-bearing components of enterprise AI. They correctly diagnose the leaks that default configurations produce. They correctly assert that partial adoption is not partial protection.

But the reading matters more than the doctrine.

Read naively — as a vendor product specification — the Five C’s produce a subtle disaster: they migrate enterprise lock-in from the model layer to the platform layer.

The prison walls move. The enterprise is still inside them. An enterprise that adopts the Five C’s as most Coalition 2 vendors prescribe them — buying Control as a platform’s eval registry, Capability as a platform’s adaptation runtime, Choice as a platform’s orchestration API, Cost as a platform’s routing feature, Compound as a platform’s memory graph — has decoupled from a specific frontier lab and coupled itself to a specific enterprise platform.

Three to five years from now, that enterprise will be structurally married to Microsoft, Palantir, or Databricks, with the same permanence it once had with Oracle, SAP, or IBM.

Read structurally — as vendor-neutral architectural principles — the Five C’s produce something different: real enterprise AI independence, calibrated against the honest constraint that total independence is impossible in an ecosystem this interwoven.

Compute lives on hyperscalers. Frontier models come from a small number of labs. Frameworks come from open communities and vendors. No enterprise can escape all of these. What is achievable is maximum practical independence — distributed dependencies such that no single vendor’s leverage over the enterprise is structural.

This piece dissects each C in full: what it structurally defends against, what the vendor trap for that C looks like, what real ownership requires, how the C connects to the routing layer, and what practical implementation looks like at enterprise scale.

It then walks through the interlock (why the C’s must hold simultaneously), the realistic limits of independence (what enterprises must accept), and the portfolio and implementation gradients that turn the framework into an operating manual.

Read more

Now, I’ll take a step beyond it to explain which coalitions these are, why they are developing, and why this potentially matters to you, especially if you work, advise, or build for Enterprise AI.

The enterprise AI market is not a diffuse competition of hundreds of vendors. It is a three-way alignment problem in which the industry is quietly resolving into three distinct coalitions, each with a coherent theory of where the moat should live, each with a shipping product signature, each with a structural incentive that makes its argument non-negotiable to its participants.

This is what markets do when a new infrastructure platform matures. In every prior enterprise IT era — mainframe, PC, internet, mobile, cloud — the same pattern emerged: an initial period of vendor proliferation, then a coalition-formation phase in which the market consolidates around two or three structurally opposed theses about where value lives, then a settling phase in which one coalition captures the majority of enterprise budget and the others compress to niche.

Enterprise AI is now in the coalition-formation phase. The vendor proliferation is largely over. The signaling of coalition alignments has begun. The settling has not started.

This piece dissects the three coalitions structurally. For each: who is in it, what its underlying thesis is, what product signature it is shipping, what structural incentive keeps its members aligned, and what vulnerability could fracture it.

The goal is to make the industry’s organization legible — because enterprises architecting AI investments in 2026-2027 are implicitly picking a coalition, and the pick will define what their AI stack looks like for the decade after.

Read more

Satya Nadella’s Reverse Information Paradox essay, just came out, and it argues that in the AI age, buyers pay twice for intelligence — once with money, and again with the proprietary knowledge they must reveal to make that intelligence useful.

Every prompt, correction, eval, and trace becomes institutional know-how that leaks upward to the model provider. Nadella’s prescription is a “trust boundary” inside which the enterprise’s evals, memory, adapted weights, and orchestration accumulate untouched — and a set of five principles enterprises must own: Control, Capability, Choice, Cost, Compound.

Embedded halfway through the essay is a quote from Palantir CEO Alex Karp: “What the technical customers want is control over their compute, their models, their data stack, and their alpha. They want to know they own the means of production, and it’s not being transferred to someone else.”

That quote is not rhetorical borrowing. It is the two most important enterprise infrastructure players in the market publicly agreeing on where the AI moat lives next — and telling every buyer, analyst, and competitor to plan accordingly.

Read the essay that way and every paragraph clicks into place. Nadella’s five C’s are not principles. They are the requirements document for a class of infrastructure that Microsoft is building through Foundry, Azure AI, and Copilot Studio, and that Palantir has already shipped as AIP, Ontology, and Evolve. The essay’s closing call to “distribute the learning infrastructure to every firm” actually distributes something else: a shared strategic thesis between two of the enterprise’s most important vendors. The moat is migrating from the model to the tenant boundary that owns the enterprise’s learning loop.

Recent analysis in these pages — The Routing Paradigm for Enterprise AI — traced how price discovery entered the AI stack at the harness layer (see AI & The Harness Theory for the underlying mechanic) and cascaded up and down. Nadella’s essay is what that cascade sounds like when it reaches the C-suite. It names the compounding surface — evals, memory, traces, adapted weights, feedback — and asserts that this is the asset the enterprise cannot afford to give away.

That naming matters. Because as long as the moat was the model, Microsoft and Palantir were playing different games. Once the moat migrates to the tenant boundary, they are playing the same game — and they need the same argument to win it.

That argument now has a name. Call it the Tenant Boundary Doctrine: whoever operates the trust boundary within which enterprise evals, memory, and adapted weights compound — operates the enterprise’s learning loop, and therefore the enterprise’s AI-era moat.

Read more

Access My Harness Via The BE Platform!

Follow me on The AI Supercycle as well!

Subscribe Now at 50% Discount, Expiring!

Not a company. A position. The role Anthropic occupies today — the owner of the scaffolding around AI models where all the real value is compounding — is about to be re-occupied at the enterprise level by something open. Open weights. Open framework. Open runtime. Owned by the enterprise, not rented from a lab.

This piece is the argument for why that has to happen.

To follow it, you need three ideas up front. I will explain each one properly in the sections that follow — this is just so nothing catches you off guard.

Idea one — the harness. Every useful AI system today is a model plus a lot of scaffolding around it. The model is the brain. The scaffolding is everything else: the loop that decides what to do next, the memory system that keeps track of context, the tools the model can call, the guardrails that keep it safe, the runtime it executes in. That entire wrapper around the model is called a harness. When people talk about Claude Code or Codex, they are talking about harnesses, not models.

Idea two — where the value is. The models themselves are becoming commodities. Open-weight models are now nearly as good as the closed frontier at a fraction of the cost. What still creates lasting value — the moat, the lock-in, the compounding improvement — is the harness around the model, not the model itself. This is settled ground at this point. Every serious operator agrees.

Idea three — the split I am making. So far, the best harnesses have been closed. Anthropic built Claude Code around Claude. OpenAI built Codex around GPT. The lab owns both the model and the harness, and the tight integration between them is the edge. My claim is that this edge is about to move. The winning harness for the enterprise will not be closed. It will be open — because of what an enterprise structurally is, and what it structurally cannot outsource.

That is the whole piece in advance. Now let me build the case.

Read more

AGaaS breaks that vocabulary because AGaaS sells something structurally different. AGaaS — Agentic-as-a-Service — charges for the execution of outcomes by agents, not for access to tools operated by humans.

That single reframing is the disruption. It rewires what gets sold (outcomes, not access), who does the buying (business budgets, not IT budgets), where margin lives (substrate, not interface), and which metrics still predict the future (consumption, not seats). Every SaaS moat was built for a world where the seat was the atomic billing unit. In AGaaS, the seat is gone. The agent is the atomic unit, and the moats have to be rebuilt from scratch around it.

Three sequential inversions define the transition, and understanding them is prerequisite to understanding what defensibility looks like inside the shift:

• The operator inversion. The agent, not the human, operates the workflow. The billing unit stops tracking headcount and starts tracking agent activity. Most enterprise software companies have completed this inversion architecturally by mid-2026, even if they have not repriced.

• The buyer inversion. The person authorizing the purchase is the executive who owns the outcome (the CRO, the COO, the general counsel), not the IT function that owns the tool stack. Procurement, budget line-items, and the sales motion all rearrange. Most vendors are still mid-flight on this one.

• The margin inversion. Gross margin structure shifts from “software plus support” (high fixed margin, low variable cost) to “consumption minus inference” (variable margin dominated by token cost). The P&L reshapes. This inversion is deferred at most incumbents through mechanisms like internal absorption (”Customer Zero”), buybacks, and reclassification, but it becomes visible once consumption revenue scales past a threshold.

The AGaaS transition is what these three inversions look like when they run through the enterprise software category. It has a canonical form, and its canonical form has been mapped in the prior work of this analytical arc:

Access My Harness Via The BE Platform!

Follow me on The AI Supercycle as well!

Subscribe Now at 50% Discount, Expiring!

• “From SaaS to AGaaS: The Full Cascade” (March 2026) mapped the underlying value-layer inversion — value migrating from the interface layer, where SaaS captured it, down to the substrate layer, where AGaaS captures it — and the five-layer cascade this inversion produces across the enterprise software stack. It also established the three canonical AGaaS pricing models: outcome-based, consumption-based, and hybrid.

• “The Four Frictions of AGaaS” (April 2026) mapped the frictions that slow the transition — financial, operational, competitive, commercial — and why they multiply rather than add.

• “The Tell: Salesforce Q1 FY27” (May 2026) used the largest pure enterprise-software incumbent as an instrument to locate where the transition actually sits. The verdict: architecturally migrating, mechanically still SaaS. Operator inverted, buyer not, margin deferred.

• “The Four Intelligence Moats” (June 2026) — the direct parent to this piece — mapped how the entire AI stack builds defensibility across paradigms: corpus, verifier, harness, container. That piece established the machine-side moat map.

This piece completes the arc by mapping defensibility inside the transition itself. If old SaaS moats no longer apply and machine-side moats alone are not sufficient, what moats does an AGaaS company actually build? The answer is six — three inherited from the machine-side (verifier, harness, container), three new to the buyer relationship (trust, integration, feedback). Together they define what AGaaS defensibility looks like once the transition matures. They also determine, mechanically, which of the three AGaaS pricing models a company can actually charge — a link the closing section of this piece will make explicit.

The order in which the six moats are built matters more than the number of them. They are not parallel. They are serial. Each one unlocks the next. A company that tries to build them out of order finds itself with moats that do not compound — and, more consequentially, with pricing power that stalls at whichever inversion its missing moat was gating.

Access My Harness Via The BE Platform!

Follow me on The AI Supercycle as well!

Subscribe Now at 50% Discount, Expiring!

Read more

• The inversion: routers turn models from products into suppliers, migrating pricing power from the model layer to the allocation layer — whoever makes the purchasing decision captures the leverage.

• The pincer: labs compress inference costs from the supply side (compute multipliers) while routers compress demand from the buy side; the gains land in different pockets.

• The cascade: routing reshapes every layer beneath it — extending GPU fleet life, redirecting silicon roadmaps toward inference-specialized chips, and turning energy allocation into a per-query decision.

• The barbell: optimizing allocators hollow out the mid-tier model market, splitting demand between irreplaceable frontier intelligence and near-free commodity models.

• The meta-moat: evaluation data — knowing empirically which model is good at what — is the routing layer’s defensible asset; the scoreboard beats the switch.

The routing paradigm is not a product story. It is the moment the AI stack acquired an internal market — a mechanism that prices intelligence continuously and transmits that price signal down to the physical world and up to the regulatory one. Whoever operates that mechanism operates the stack’s capital allocation.

The Architecture: A Market Grows Inside the Stack

Access My Harness Via The BE Platform!

Follow me also on The AI Supercycle!

Subscribe Now at 50% Discount, Expiring!

Strip away the product names and a router is one thing: a capital allocation engine for intelligence.

Every query entering an AI system carries two hidden variables — the difficulty of the task and the value of getting it right. For three years the industry priced neither. Users manually selected a model, defaulted to the most expensive one out of habit or brand trust, and paid frontier prices for tasks a model one-tenth the cost could handle. Structurally, this was a market without price discovery: heterogeneous demand met heterogeneous supply through a single crude interface — the model picker dropdown — and the mismatch was absorbed as waste.

Routers are the price discovery mechanism arriving. They estimate task difficulty, match it against a live map of model capabilities and prices, and dispatch each query to the cheapest supplier that clears the quality bar. The Information’s reporting this week catalogued the forms this now takes, and the taxonomy is worth holding onto because each form represents a different owner of the allocation decision:

Lab-internal routing — OpenAI’s GPT-5 switching between sub-models under the hood. The lab keeps the allocation decision in-house and arbitrages its own portfolio. The customer never sees the spread between what they paid for and what actually ran.

Neutral marketplace routers — OpenRouter, which raised $120 million on the thesis, plus Martian and Not Diamond underneath it. These sit outside any lab and optimize honestly across the full supplier population.

Platform gateways — Databricks’ Unity AI Gateway, Palantir’s Evolve. Enterprise infrastructure players bundling routing into the data and workflow platforms enterprises already run on.

Agent-level routers — Cognition’s new sidekick architecture, which delegates easier subtasks to cheaper agents mid-workflow. Here routing happens inside a running agentic process, not at the query boundary.

DIY and model-as-router — IT departments vibe-coding their own switches, or developers handing a cheap model like DeepSeek a menu of models and asking it to pick. The Arcee AI approach: use the cheapest competent judgment available to allocate the expensive judgment.

Five forms, one mechanism: a decision that used to belong to the user — or default silently to the model vendor — now belongs to an intermediary layer. And in business architecture, whoever makes the purchasing decision captures the negotiating leverage.

The reported numbers establish that this is not marginal optimization. Palantir says Evolve cut computing costs for one task by 97% by swapping a frontier OpenAI model for its Nano sibling. McCarthy Building, a $9 billion construction firm, cut token consumption 60% year-over-year — partly through model swapping, partly because the router rewrites prompts to burn fewer tokens for the same outcome, which is a second and underappreciated function: routers optimize not just which model but how the model is asked.

Cognition’s router reportedly matched Anthropic’s Fable 5 on a coding benchmark at 35% lower cost. Sakana claims a coordinated multi-model system competitive with frontier flagships by routing math to one lab and science to another.

When an intermediary reliably delivers frontier-equivalent outcomes at a third to a thirtieth of the cost, it is not a feature. It is a new layer of the stack asserting itself. Which raises the structural question this piece exists to answer: where exactly does it sit, and what does it do to everything above and below it?

Read more

Let me start by saying, this is not a framework.

It is a complete grand-strategic apparatus, built over the last decade, that stacks four layers on top of each other: a philosophy about what reality is and how it can be known, a theory about how complex systems in the AI business era actually behave, a method for producing structural understanding of specific domains, and a set of frameworks that are outputs of the method rather than the substance of the practice.

Most of what gets called strategic analysis in the AI economy occupies only the fourth of those layers — a framework, imported from somewhere, applied to whatever domain the analyst is looking at. The Business Engineer occupies all four, and the value comes from the stack being complete.

This piece is the description of that stack. What sits at each layer? Why does the stack have to be complete for any layer to work? And why it matters that the whole apparatus operates at the grand strategy altitude — reading the territory — rather than at the strategy or tactics altitudes, where most business analysis lives.

Access My Harness Via The BE Platform!

Follow me also on The AI Supercycle!

Subscribe Now at 50% Discount, Expiring!

Grand strategy, strategy, tactics — and why the confusion is expensive

Access My Harness Via The BE Platform!

Follow me also on The AI Supercycle!

Subscribe Now at 50% Discount, Expiring!

I have written about this before and I will repeat the frame here because it is foundational for everything that follows.

Grand strategy is reading the territory. It is understanding the underlying structure of the reality you are operating in — the distributions, the forces, the constraints, the dynamics that determine what is possible and what is not. Grand strategy is upstream of everything else. It is the layer where you look at the world and try to see what it actually is, before you decide what to do about it. Without grand strategy, everything downstream is guessing.

Strategy is the map. It is what you draft after you have read the territory — a working representation of the terrain that you can use to plan movement. The map is provisional; it is corrected when the territory turns out to be different from what the map showed. But the map is always derived from the territory. Any map that is not derived from a proper reading of the territory is not a strategy — it is a fiction being applied to a situation that does not match it.

Tactics are the routes. Given the territory and the map, tactics are the specific paths you can take. Many paths are possible. Some lead where you want to go. Most of what people call “strategy” in business is actually tactics — the selection of specific paths — dressed up in strategic language.

The confusion between these three altitudes is one of the most expensive mistakes in business, and it is the specific mistake most strategic analysis is currently making in the AI era.

People are producing tactical outputs, calling them strategy, and skipping grand strategy entirely — which is to say, they are drawing maps and selecting routes without ever having read the territory.

In a stable environment, this can be gotten away with; the territory doesn’t move fast, and old readings still roughly hold. In the AI era, the territory is moving faster than any map can keep up with, and any strategy or tactics not derived from continuous fresh reading of the territory becomes obsolete before it can be executed.

The Business Engineer is, first and foremost, an apparatus for reading the territory. That is what makes it grand-strategic rather than strategic or tactical. Everything else — the strategies it derives, the tactics it enables — is downstream of this. The apparatus is designed to do the grand-strategic work that most business practice skips, and to do it repeatedly, at scale, against the specific class of complex systems that define the AI era.

The philosophy

Access My Harness Via The BE Platform!

Follow me also on The AI Supercycle!

Subscribe Now at 50% Discount, Expiring!

Every apparatus that reads territory operates on a philosophy — a set of underlying commitments about what reality is, what can be known about it, and what counts as knowing. Most strategic frameworks skip the philosophy layer entirely, which is why they produce outputs that are internally consistent and externally wrong when the underlying reality doesn’t match the framework’s implicit metaphysics.

The Business Engineer has an explicit philosophy, and it is worth naming because everything else derives from it.

Reality is structural and structurally knowable. The world is not a collection of narratives, opinions, or perspectives. It has underlying architecture — distributions, forces, positions, constraints — that determines what is possible in it. This architecture is knowable, not in the sense of being fully specifiable, but in the sense of having foundational properties that can be identified, tested, and used to make decisions. Structural knowability is the first commitment. Without it, no analytical apparatus can be evaluated as right or wrong, only as more or less persuasive.

Reality is authoritative over the analyst. When the analyst’s model and reality disagree, reality wins. This sounds trivial and is not, because most analytical frameworks are structured in practice to be defended against disconfirming evidence rather than corrected by it. The Business Engineer’s second commitment is that the apparatus exists to be corrected by reality, not to survive against it. Any framework that cannot be wrong, cannot be right either; falsifiability is the property that makes structural knowledge distinguishable from ideology.

Complexity is a property of reality, not a failure of analysis. Most classical business analysis treats complexity as noise to be filtered out on the way to a clean model. The Business Engineer treats complexity as the substance of the phenomenon. Power-law distributions, fat tails, non-stationary dynamics, opaque causal chains, unbounded competitive fields — these are not analytical inconveniences. They are properties of how the AI economy actually behaves, and any apparatus that filters them out to produce a clean model has produced a clean model of something that isn’t there. Complexity-as-substance is the third commitment. It changes what “understanding” means.

Compression is possible without loss of foundational structure. Even in complex domains, it is possible to identify the small number of properties that carry the weight of the phenomenon — the distributions, the positions, the constraints that determine the domain’s behavior. The rest can be discarded. This is what distillation means, and it is a philosophical commitment before it is a methodological one: you have to believe that reality has this property — that its complexity is not uniformly load-bearing, that some structural properties matter much more than others — in order to build an apparatus around finding them. If everything about a complex domain matters equally, no apparatus can produce a portable, actionable understanding of it. The Business Engineer’s fourth commitment is that this is not the world we are in. Complex domains have concentrated structural mass, and finding that mass is the analytical project.

Structural understanding must be portable and actionable, or it is not understanding. An analytical output that cannot be held in a working mind, cannot guide a specific decision, cannot survive the analyst leaving the analysis behind — that output has failed regardless of how comprehensive it is. Portability is the fifth commitment. It rules out the class of analytical outputs that look rigorous, take three hundred pages to communicate, and produce no consequential action.

These five commitments together — structural knowability, reality’s authority, complexity-as-substance, concentrated structural mass, portable understanding — constitute the philosophy of The Business Engineer. Every commitment is defended against a specific failure mode in classical business analysis. Every commitment shapes what the apparatus can and cannot do at the layers above. This is the deepest layer of the stack and the one most other frameworks don’t have at all.

The theory

Access my harness with the exec plan!

Follow me also on The AI Supercycle!

Subscribe Now at 50% Discount, Expiring!

On top of the philosophy sits a theory — a set of foundational propositions about how the specific reality of AI-era complex systems actually behaves. The theory is not the philosophy; the philosophy is about reality in general, the theory is about this reality specifically. And the theory is not the method; the theory tells you what kind of world you’re in, the method tells you what to do about it.

The Business Engineer’s theory, compressed to its essentials, holds that:

• AI-era domains are power-law and fat-tailed by structure, not by accident. The concentration of value at the top of every distribution — labs, platforms, models, capabilities, returns — is not a temporary condition of an immature market. It is a structural property of how systems built on massive fixed capital costs, learning curves, and network effects distribute their outputs. Any analysis that treats these distributions as normal, or as convergent-toward-normal, is making a category error.

• Categories in the AI economy are dissolving and reforming continuously. The industry boundaries that used to organize strategic analysis — software, media, financial services, retail, consulting — are becoming accidents of historical accretion rather than meaningful strategic units. The unit of analysis has shifted from industry to capability layer, and any framework that operates on the old units produces increasingly ceremonial outputs.

• Competition is unbounded and multi-scalar. The relevant players in any domain include the labs upstream, the platforms downstream, the specialized incumbents in the vertical, the horizontal AI-native firms that could enter tomorrow, and the customers who might build their own and stop buying. Competitive analysis that identifies “the players” is producing a snapshot of who is visible today, not who will be competing tomorrow.

• Causal chains are opaque in real time and legible only in retrospect. In the AI economy, the drivers of outcomes often become identifiable only after the outcomes have occurred. Cause and effect happen at inference speed. Strategic planning that assumes causal legibility in advance is assuming a property the domain does not have.

• Capability migrates outward from cores to peripheries at fractal scales. This is the specific claim the Harness Trilogy developed — that in the AI era, the locus of value shifts from the model to the harness around it, and the same shape appears at the individual, organizational, and societal scales. This is a theoretical claim about how AI-era systems evolve over time, and it has predictive content that the last eighteen months have validated.

• Structural transformations run on multiple nested clocks. The AI Supercycle framework specifies this: there are cycles within cycles within cycles, each running on its own clock, and the interactions between the clocks determine what looks like a bubble, what looks like a supercycle, and what actually is happening structurally. Any theory that operates on a single timescale misses the interaction effects that produce the phenomena everyone is trying to explain.

Together these propositions constitute the theory of how AI-era complex systems behave. They are not exhaustive. They are the foundational propositions — the ones that survive against reality repeatedly, produce non-obvious predictions, and structure how the method should be applied. The theory is what makes the method appropriate to this reality specifically, rather than to reality in general. A method without a theory is a technique; a theory without a method is an interpretation. The Business Engineer has both, and they are calibrated to each other.

The theory is also falsifiable, per the philosophy. Any of its propositions could be wrong. Any of them could be found not to hold in some specific domain. When that happens, the theory is corrected, and the method is recalibrated. This is what it means for the theory to sit on the philosophy: it is defended by the philosophy’s commitment to reality’s authority rather than by rhetorical defense against critique.

The method

Access My Harness Via The BE Platform!

Follow me also on The AI Supercycle!

Subscribe Now at 50% Discount, Expiring!

On top of the theory sits the method — the specific practice for producing structural understanding of a given domain. The method is what most other analytical frameworks think they are; The Business Engineer’s method is one layer of a four-layer stack, not the whole thing. The method’s power comes from being anchored in the philosophy and calibrated to the theory. Detached from either, it becomes a technique that can be misapplied.

The method is what I have previously described as distillation of a domain via contextualization, and I will restate it here compactly because the previous piece walked through it in detail.

• Contextualize. Identify the local distribution of the domain (power-law, fat-tailed, bimodal, normal — where does the mass sit, where are the tails) and identify where the domain sits in the broader architecture (upstream or downstream, substrate or superstrate, early or mature, regulated or unregulated). This is the grand-strategic reading of the specific territory. Everything downstream depends on getting it right.

• Abstract. Extract the foundational structural properties from the context. Discard the surface. Identify what makes the domain behave as it does at the level of underlying structure.

• Model. Build a working model of the domain’s dynamics that captures the foundational properties and is small enough to reason with. Not comprehensive. Distilled.

• Act. Use the model to make specific decisions with real stakes and real consequences. Predictions without decisions produce no feedback.

• Refine and revalidate. When outcomes arrive, feed them back. Adjust the model where the model was wrong. Adjust the abstraction where the abstraction was wrong. If the outcomes contradict the original context framing, the framing itself was wrong — go back to step one and rebuild. This is the step most methods skip, and it is the step that keeps the apparatus honest.

This is the method. Six operational moves, applied repeatedly, tightening with use. The method is what makes the apparatus a practice rather than a theory — it is what someone actually does when they are being The Business Engineer, as opposed to what they are thinking about.

But — critically — the method is not the whole apparatus. The method calibrated to a different philosophy would produce different outputs. Applied to a domain the theory doesn’t fit, it would produce wrong outputs. The method’s validity depends on the layers below it, which is why the frameworks it produces cannot be extracted from the stack and used alone. This is the specific way in which most attempts to copy analytical frameworks fail: the framework is extracted from its philosophical and theoretical context and applied in a different context, where it produces outputs that look right and are structurally wrong.

The frameworks

Access my harness with the exec plan!

Follow me also on The AI Supercycle!

Subscribe Now at 50% Discount, Expiring!

At the top of the stack sit the frameworks — the specific structural readings The Business Engineer has produced by applying the method to specific domains. These are what most readers encounter first, and they are the least fundamental layer of the practice. They are outputs, not the substance.

The frameworks include, among others:

• The AI Supercycle — the reading of the AI economy as a whole, in terms of three nested cycles, three phases, and the specific ways in which the aggregate looks like a bubble and behaves like a structural transformation

• The Map of AI — the reading of the AI stack as nine layers with different structural properties, distributions of return, and positional dynamics

• The AGaaS Trilogy and its extensions — the reading of the shift from selling access to selling outcomes as a structural transformation, with four intelligence moats and a set of derived mental models

• The Harness Trilogy — the reading of capability migration from model to harness as a fractal phenomenon at three scales, with four repeating patterns and one bedrock

• The Routing Paradigm — the reading of agentic infrastructure as a routing problem with specific structural properties

• The Five Postures — the reading of firm positioning in the agentic era as a set of five stable configurations with specific tradeoffs

Each of these is a strategic map — in the grand-strategy sense of the term — derived from a grand-strategic reading of a specific territory. Each is provisional; each is corrected when the underlying territory is found to have moved. Each is portable, actionable, and load-bearing in the specific senses the philosophy requires.

But these frameworks are not The Business Engineer. They are what The Business Engineer produces. Anyone who reads the frameworks without understanding the philosophy, theory, and method beneath them is holding the output without the apparatus that produced it — which means they cannot update the frameworks when the territory changes, cannot produce new frameworks for domains the existing ones don’t cover, and cannot evaluate whether a framework applies in a given case. The frameworks are the visible surface of the practice, and the invisible layers underneath are where the actual work lives.

This is why “framework” was the wrong word for the whole apparatus. Frameworks are outputs of the apparatus. The apparatus itself is the four-layer stack.

Why the stack has to be complete

Access My Harness Via The BE Platform!

Follow me also on The AI Supercycle!

Subscribe Now at 50% Discount, Expiring!

The stack could theoretically be broken. Someone could hold the philosophy without developing a theory. Someone could develop a theory without a method. Someone could apply a method without frameworks. Each partial version fails in a specific way, and enumerating the failures is worth doing because it clarifies why the complete stack is what the practice actually is.

• Philosophy without theory produces a general commitment to structural reading of reality without any specific claims about the reality being read. This is philosophy of science, not applied analysis. It is not wrong, but it does not produce guidance for consequential decisions in a specific era.

• Theory without philosophy produces claims about how AI-era systems behave that are not anchored in commitments about how such claims should be evaluated. This is what most AI-era analysis currently does: it makes claims about power laws, network effects, capability migration, and so on, without an underlying commitment to falsifiability and reality’s authority. The claims float, get defended, and become ideology when they should be corrected.

• Method without theory produces a technique that can be applied to any domain, without a way of knowing which domains the technique is appropriate for. The method’s contextualization step becomes generic — “understand the domain” — rather than specifically targeted at the distributions and positions the theory identifies as foundational.

• Method without philosophy produces a technique that can be executed but not evaluated. Outputs are produced; whether they are right or wrong becomes a matter of persuasion rather than validation. The revalidation step of the method depends on the philosophy’s commitment to reality’s authority; without it, revalidation becomes rhetorical rather than structural.

• Frameworks without method produces static representations that cannot be updated. The Business Engineer’s frameworks are all provisional; they are corrected when the territory moves. Without a method that continuously reapplies contextualization and revalidation, the frameworks become fossils.

• Frameworks without theory produces representations that cannot be evaluated for appropriateness to specific domains. Every framework has an implicit theory it depends on. Users who don’t hold the theory cannot know whether the framework applies in their case.

• Frameworks without philosophy produces representations that are treated as doctrine rather than as tested structural readings. This is where most strategic framework use ends up: the framework becomes a template, the template is applied without regard to whether the underlying reality matches, and the outputs are treated as strategic even though they are barely tactical.

Every partial version of the stack is a specific failure mode. The complete stack is what avoids all of them. This is why The Business Engineer is a grand-strategic apparatus rather than a strategic framework. It has to occupy all four layers to do what it does. The frameworks it produces are strategic maps; the apparatus that produces them is grand-strategic; and the philosophy underneath is the deepest layer that makes anything above it possible.

What the apparatus demands of its user

Access My Harness Via The BE Platform!

Follow me also on The AI Supercycle!

Subscribe Now at 50% Discount, Expiring!

An apparatus is only as good as the user operating it. Some frameworks are designed to be operated by anyone — that is one of their attractions and one of their limits. The Business Engineer is not designed to be operated by anyone. It is designed to be operated by someone willing to hold the whole stack, and holding the whole stack is itself a discipline.

Specifically, the apparatus demands:

• A commitment to structural reading over narrative interpretation. The user has to be willing to look for underlying architecture rather than for stories, personalities, or plausible-sounding explanations. This is a temperament as much as a skill, and it is not the temperament most business analysis rewards.

• A tolerance for being wrong. The philosophy’s commitment to reality’s authority means the user will be corrected by outcomes, sometimes publicly, sometimes at cost. Users who cannot tolerate being wrong will defend their frameworks against correction, which converts the apparatus into ideology.

• A capacity for compression. The apparatus produces distillations, not comprehensive reports. Users who cannot compress will produce outputs that miss the foundational structure while covering everything else. Compression is a discipline that takes years to develop and is rarely explicitly taught.

• A willingness to hold multiple altitudes simultaneously. The apparatus operates at grand strategy, produces strategy, enables tactics. The user has to be able to move between altitudes without confusing them, which is the specific skill the grand-strategy essay identified as most valuable and most rare.

• A refusal of imported frameworks. Every time the user is tempted to reach for Porter, or SWOT, or disruption theory, or any of the standard business-school toolkit, the apparatus requires them to stop and ask whether the imported framework’s implicit theory matches the domain. Almost always, it does not. The refusal is what keeps the apparatus honest; the temptation to import is constant.

These demands are why the apparatus produces the outputs it produces. They are not casual requirements. They are the specific dispositions the practice requires, and users who cannot meet them will produce outputs that look like Business Engineer work and are not.

What this means for the AI era

Access my harness with the exec plan!

Follow me also on The AI Supercycle!

Subscribe Now at 50% Discount, Expiring!

The AI era is going to reward the users of grand-strategic apparatuses and punish the users of tactical frameworks-dressed-as-strategy. This is a claim about the structural properties of AI-era decision-making, and it follows directly from the theory.

When the territory is stable, tactical thinking can be gotten away with; the underlying map doesn’t change, so old routes still work. When the territory moves, tactical thinking fails, because the routes lose their referent. The AI era is a moving-territory era on every dimension that matters. The distributions are shifting. The categories are dissolving and reforming. The competitive fields are unbounded. The causal chains are opaque. Under these conditions, tactical thinking without grand-strategic anchoring produces confident nonsense.

The apparatus described in this piece is one instance of a grand-strategic anchoring practice. It is not the only possible instance; there are others, some articulated, some tacit. What matters is that anyone making consequential decisions in the AI era needs some grand-strategic anchoring practice, because tactical execution alone will produce systematic errors of a kind that compound. The Business Engineer is what I have built. Whether readers use this apparatus, build their own, or find another that works — the requirement itself is not negotiable in the era we are entering.

That is what The Business Engineer actually is. A grand-strategic apparatus for reading the territory of the AI era, composed of four layers stacked completely: a philosophy of structural reality, a theory of AI-era complex systems, a method of contextualized distillation, and a set of frameworks that emerge as outputs of the method rather than as the substance of the practice. Everything I have written under this banner in the last decade is an application of this stack to a specific domain. Everything I will write is the same. The stack is the practice. The frameworks are downstream.

The correction from framework to apparatus is not cosmetic. It changes what the reader is being asked to hold, and what the practice is claiming to be. Held complete, the stack does the grand-strategic work that most business analysis skips. That is the value. That is what distinguishes it from the tactical frameworks that dominate the field. That is why the practice has held up for a decade and why it will continue to hold up as the territory keeps moving.

Key Takeaways & Mental Models

Access My Harness Via The BE Platform!

Follow me also on The AI Supercycle!

Subscribe Now at 50% Discount, Expiring!

• Framework Is the Wrong Word — The Business Engineer is a grand-strategic apparatus with four layers: philosophy, theory, method, and frameworks. Frameworks are outputs of the apparatus, not the apparatus itself. Most analytical practices occupy only the fourth layer, which is why they produce tactical outputs dressed as strategy.

• Grand Strategy Is Reading the Territory — Not planning, not framing, not selecting routes. The specifically upstream work of understanding what reality is before deciding what to do about it. The Business Engineer operates at this altitude by design.

• The Five Philosophical Commitments — Structural knowability of reality; reality’s authority over the analyst; complexity as substance not noise; concentrated structural mass in complex domains; portable and actionable understanding as the criterion of success. Every commitment is defended against a specific failure mode in classical analysis.

• The Theory of AI-Era Complex Systems — Power-law distributions by structure; dissolving categorical boundaries; unbounded multi-scalar competition; opaque causal chains legible only in retrospect; fractal capability migration from cores to peripheries; multi-clock structural transformations. These are the load-bearing propositions the method is calibrated to.

• The Method Is One Layer, Not the Whole Practice — Contextualize, abstract, model, act, refine, revalidate. Six operational moves. The method’s validity depends on the theory it is calibrated to and the philosophy it is anchored in. Detached from either, the method becomes a technique.

• Frameworks Are Outputs, Not Doctrine — The AI Supercycle, the Map of AI, the AGaaS Trilogy, the Harness Trilogy, the Routing Paradigm, the Five Postures — each is a strategic map derived from grand-strategic reading of a specific territory. Each is provisional. Each is corrected when the territory moves.

• The Stack Has to Be Complete — Every partial version of the stack fails in a specific way. Philosophy without theory is philosophy of science. Theory without philosophy is defended ideology. Method without either is a misapplied technique. Frameworks without any are fossils. The complete stack avoids each failure specifically.

• The Apparatus Demands a Specific Discipline — Structural reading over narrative interpretation, tolerance for being wrong, capacity for compression, ability to hold multiple altitudes, refusal to import inappropriate frameworks. Users who cannot meet these demands will produce outputs that look like Business Engineer work and are not.

• Moving Territory Requires Grand-Strategic Anchoring — When reality is stable, tactical thinking can be gotten away with. When reality moves, tactical thinking produces confident nonsense. The AI era is moving-territory on every dimension that matters. Grand-strategic anchoring is not a luxury; it is what keeps analysis oriented as the territory shifts.

• The Correction from Framework to Apparatus Is Structural — Not cosmetic. It changes what the reader is being asked to hold and what the practice is claiming to be. Held complete, the stack does the grand-strategic work most business analysis skips. That is the value the practice claims and the value it can be tested against.

• The Practice Is Repeatable and Portable, Not Personal — The apparatus is not a set of doctrines dependent on the person who built it. It is a repeatable stack that anyone willing to meet the demands can operate. The specific frameworks I have produced are one path through the apparatus. Others exist. The apparatus is the invariant.

With massive ♥️ Gennaro Cuofano, The Business Engineer

The frame to hold onto: we are inside a 30-to-50-year supercycle, not a news cycle. Everything feels fast because we are living it in real time. But we have time to understand it, adjust to it, and position around it. So before the panic — the mechanism.

Why the Map Had to Be Redrawn

Back in March I added two layers to the Map of AI: a harness layer and a governance layer. This week validated both, hard.

Read more