I came to the tech world more than a decade ago, and I’ve seen plenty of crazy things happen since. Over the years, I’ve had to develop my own way of separating the signal from the noise. That is the whole point of The Business Engineer and The AI Supercycle: building a framework that helps me understand what is happening beneath an increasingly overwhelming news cycle.

For me, that means going back to first principles. What does a technology actually make possible? What needs to exist for it to scale? Who finances it, who controls its bottlenecks, and what happens when its consequences reach far beyond the companies developing it?

Those are the questions from which my definition of the AI Supercycle emerged.

The AI Supercycle, as I define it, is the process through which AI reorganises technology, capital, infrastructure, and economic activity—and, in doing so, reshapes the political and geopolitical conditions that govern its own development.

The feedback between those forces is central to my thesis. AI changes the economy. Those economic changes influence politics and the balance of power. Politics and geopolitics then feed back into where AI can develop, who can access it, how it is financed, and how quickly its infrastructure can be built.

That is the original idea I have been developing through The Business Engineer and The AI Supercycle.

The beginnings of this way of thinking became much clearer to me around 2020. The pandemic exposed vulnerabilities in the technological globalisation we had largely taken for granted. Suddenly, access to manufacturing capacity, critical components, and international supply chains looked much more dependent on geography and political relationships.

I started seeing technology spilling over into everything else. The ChatGPT moment subsequently crystallised that realisation and gave it a much greater sense of urgency.

This is why I have never been satisfied with describing AI as another version of the web, cloud, or mobile cycle.

Even when people describe it as those three cycles combined, I think the comparison can still understate what is happening. It keeps us anchored to the recent history of the technology industry, when the consequences I am trying to understand extend into the organisation of the economy itself.

My reasoning starts with a simple chain of dependencies. If AI can perform an expanding range of economically useful work, its development requires computing capacity. That capacity requires chips, energy, land, equipment, and capital. As those requirements grow, they begin to affect industrial investment, public infrastructure, and the allocation of resources. Once they reach that scale, they become political.

The comparisons I find more useful are therefore railways, electrification, and the early semiconductor era: periods in which technological development became inseparable from physical infrastructure, industrial capacity, and state power.

That is the scale of change I am examining when I talk about the AI Supercycle.

I also see the relatively open, US-centred globalisation of the last three decades as an unusual historical setting. It encouraged us to assume that powerful technologies would spread through increasingly integrated markets and supply chains. I think carrying that assumption into the AI era is a mistake.

The current wave of frontier AI has been driven heavily by private laboratories and private capital, but it is developing within an increasingly contested strategic environment. Governments recognised AI’s importance well before ChatGPT—the United States had a national AI research strategy in 2016, and China published its major development plan in 2017. What has changed is how immediate, expensive, and publicly visible the consequences have become.

To navigate this, I follow a few signals closely.

SoftBank, for example, has become one of my compasses for the financial side of the industry. Its role in Stargate helped make OpenAI’s expansion beyond a Microsoft-centred infrastructure arrangement concrete. In January 2025, SoftBank took responsibility for financing the initiative while Microsoft relaxed its exclusivity over new OpenAI computing capacity. That was less than two years ago, although it already feels like another lifetime.

To me, that development captured something important: the ambitions of the frontier laboratories were drawing in a much wider coalition of capital and infrastructure providers. The next stage of AI would depend increasingly on their ability to coordinate that coalition.

But the biggest mistake we can make today is to assume that the politics of this cycle will consolidate around a single ideology.

One of the implications of my definition of the AI Supercycle is that AI inevitably becomes an issue across political parties, because its consequences reach across society.

This week’s midterms brought that into focus. Ballotpedia reports that voters favoured limits on data-centre development in 40 of the 41 local ballot measures decided on 3 November. Those votes covered different kinds of restrictions and conditions, but together they show how contentious the physical footprint of AI has become in the communities being asked to accommodate it.

The concern was already crossing party lines. In September, the House passed the Ratepayer Protection Act by 417 votes to three. The bill would require state utility commissions to consider standards making large data centres cover the additional infrastructure costs they create.

A politician is, among other things, a barometer of the pressures felt by their community. When those pressures show up through electricity bills, water supplies, construction projects, or concerns about employment, politicians across the spectrum have an incentive to respond.

Anthropic is another reason I resist reading this through a simple left-versus-right lens.

Its disagreements with the Trump administration over safeguards have coexisted with a hard line on Chinese access to advanced AI. In March 2025, it explicitly advocated tighter semiconductor export controls while calling for an additional 50 gigawatts of dedicated US power by 2027 and faster permitting. Those positions sit together within the same company’s policy agenda.

My reading is that stronger demands for safeguards can coexist with a more determined effort to preserve American AI leadership. Democratic electoral gains could bring tougher conditions for individual projects while leaving the strategic case for domestic capacity intact—or strengthening it.

The eventual pace of construction depends on how those competing pressures are resolved.

The Trump administration has given the buildout political backing through major investment announcements and an explicit AI strategy. Its July 2025 AI Action Plan called for faster permitting, expanded domestic infrastructure, and the export of American AI systems. The next test is how that ambition translates into arrangements that communities, utilities, and state governments are willing to accept.

We also need to pay attention to timing.

SemiAnalysis estimates that roughly three-quarters of the capacity scheduled for 2027 across the top 20 states in its model is already under construction. Most of its projected US additions for 2028 have yet to start. That makes the later pipeline more exposed to changing approvals and development conditions, although projects already underway can still depend on outstanding permits for power or subsequent phases.

This is why I keep returning to the underlying dependencies. Announcing a project, securing financing, starting construction, and delivering usable capacity are different achievements. Each requires a different set of conditions to hold.

Meanwhile, enterprise demand—the economic foundation of the cycle—has continued to expand.

Menlo Ventures estimates that enterprise AI deployment spending reached an annualised rate of approximately $177 billion by September 2026, 5.4 times its comparable estimate a year earlier. Anthropic’s latest filing offers another concrete indicator: approximately $17 billion in third-quarter revenue, compared with $1.4 billion in the same quarter last year.

That supports my view that the industry is still responding to substantial growth in demand. It also requires us to keep asking what kind of demand is growing, how much customers will pay for it, and whether the resulting revenue can support the infrastructure commitments being made.

Ramp’s latest token data illustrates the distinction: among the customers it tracks, usage grew much faster than expenditure. Falling prices and improving efficiency can expand adoption while changing the economics of the capacity built to serve it.

For me, this is the AI Supercycle unfolding in practice. The technology is becoming entangled with debt, financing, energy, national strategy, and local permission. Each of those forces now feeds back into the industry’s ability to keep expanding.

That is precisely why I developed the framework around the interaction between technology, macroeconomics, and geopolitics. As AI becomes more consequential, understanding the technology alone explains a smaller part of what determines its trajectory.

My compass still points towards execution as the central immediate risk: the ability to finance, permit, power, and build the capacity that expanding adoption requires.

Demand is growing. The question I am watching is whether the industry can translate that demand into infrastructure that can actually be delivered—and generate the durable revenue needed to sustain it.

This was the week the AI Supercycle met the ballot box. The vote did not answer in the language of a ban. It answered in the language of permission: who may build, on whose grid, and who pays for it.

It started with numbers. Anthropic’s amended S-1 put a $17.0 billion quarter and $895 billion of compute commitments in front of every investor at once. Apple, Palantir and AMD reported. Then the United States voted, and the industry learned in one night that its next two years of building need consent it had assumed.

Think of the AI build-out as one long supply chain again. Last week the builders opened their books and the build-out got a credit rating. This week voters, lenders and public-market investors opened theirs, and the answer came in five parts.

1. The vote turned data centers into a permission problem. Of 41 local data-center measures on the ballot, 40 went the way of limiting development, Ballotpedia’s count shows. New governors who back a pause now sit over states holding about 41% of planned US data-center capacity, on Cleanview’s tracker. The pressure on the build crossed party lines, and so did the answer: make the builders pay their own way.

2. The rout went global, then into credit. Twelve AI stocks closed $1.45 trillion lower on Thursday as Nvidia fell 9.3%. On Friday Korea’s KOSPI was halted after an 8.6% plunge and TSMC opened limit-down. SoftBank’s dollar bond had its worst day in ten weeks, and the cost to insure SpaceX’s debt hit a record 286.56 basis points.

3. The private marks met the public price. Prospective investors discussed an Anthropic IPO value of $1.2 trillion to $1.3 trillion, down from $1.5 trillion to $1.6 trillion earlier in the week, Bloomberg reported. OpenAI closed about $11 billion at $1.25 trillion.

4. Memory stayed the wall. Global DRAM revenue rose 369% to nearly $200 billion on about 4% more bits, Counterpoint said. HBM prices are seen up 121% in 2027.

5. Agents moved into the money, and met gates. Claude launched personal finance through Plaid, read-only. Visa found shopping agents failed 42.8% of its test runs and often bought something anyway.

The short version: demand for compute did not crack this week. Permission and credit did. On rout day, nine of the ten AI names that reported beat their estimates, and four of them still fell.

The week in one line. For two years the AI build-out was priced as a supply story, and last week as a credit story. This week it was priced as a permission story, and the answer arrived as a cost, not a ban.

How to read this issue

Every week we read the news onto one picture, the Map of AI. It stacks the industry in layers, from energy at the bottom to the rules at the top. Here are the five ideas the rest of the issue uses.

• The layers. Energy is Layer 1, and chips and their parts are Layers 2 and 3. Compute and the money that pays for it are Layer 5, models are Layer 6, and the agent software around them is Layer 7. The stores and banks agents must pass are Layer 8, and the rules are Layer 9. The paying customer sits above the map as Adoption.

• The chokepoint. Each layer has one scarce piece this week. For each, we name who holds it and who pays to use it.

• The colours. In the pictures, teal marks the side that held or gained. Ochre marks the side that paid, waited or was refused.

• The cascade and the loop. World politics shapes the economy, and the economy shapes technology. In a supercycle the arrows also run back: AI moves the economy and politics in return.

• The four clocks. Each part of the build runs on its own clock: factories over years, money over weeks, efficiency over months, customers over quarters. Trouble starts when two clocks disagree.

Two habits to note. “Reported” marks figures from news outlets that the companies have not confirmed. “By our arithmetic” marks a sum or ratio we worked out from published figures.

One more for this week. When we write about the election, we describe what each outcome changes for the build: permits, power, timing and cost. We do not score the candidates or the parties. The data-center question cut across party lines, and the piece treats it that way.

Demand Held. Permission and Credit Repriced.

Start at the bottom of the map and climb. Every layer had a side that gained and a side that paid this week. The pressure gathered on two of them.

The first was the permit, in Layer 1. A data center needs land, power, water and a vote it can no longer assume. The second was the credit route, in Layer 5, where the money that pays for the build was repriced in the same 48 hours, from Seoul to New York.

Between those two pressure points, the rest of the stack behaved as it has for months. Memory kept its price. Models got cheaper. Agents moved closer to money. The customer kept spending. That contrast is the week’s real finding: the constraint moved from what the industry can make to the terms on which it is allowed to build and borrow.

Unlock The Full AI Supercycle Weekly!

The Body of Knowledge is Part of The Enterprise Embedding Series!

The Full Series With The Exec Plan!

The Full Library With The Yearly Plan!!

From Who Rates the Build to Who Permits the Build

Last week’s issue asked who rates the build. The answer was that the builders had run past their own cash and that the credit market had started to price it. That issue closed with tests for this week. Four came back.

• Would the cash catch up? Not yet. Big Tech and SpaceX guided fourth-quarter capex up to about $270 billion, above the third quarter. The off-book bills of the biggest builders now top four years of their cash flow. The gap widened before it narrowed.

• Would OpenAI’s round fill? Partly, and at the old price. It closed about $11 billion at a $1.25 trillion valuation, the first tranche of a round that could reach $20 billion, Bloomberg reported. SoftBank, its largest backer, was said to be weighing less than $1 billion more.

• Would the buyers widen? Partly. Enterprise AI spend reached $177 billion, up 5.4 times in a year, on Menlo Ventures’ survey. But OpenAI and Anthropic are set to buy about 30% of the world’s new compute in 2027, SemiAnalysis’s Dylan Patel estimated. The buyers grew; the biggest grew faster.

• Would the firewalls soften? No. They moved home. Last week’s fences were drawn by states against other states. This week voters drew them around their own towns.

So this week’s question follows from last week’s answer. The builders cannot fund the build from operations alone, and now they need consent where they build. Where does the build need permission, and what will the permission cost?

The Midterms, Read Through the Build

The 3 November midterms were not about AI. Affordability, the war with Iran and the president’s approval drove most of the result. But data centers were on more ballots than ever, and the outcome changes who sets the terms for the next two years of building. This section reads the vote the way a builder, a lender or a utility would: what changed, where the power to act sits, and what it means for capacity due in 2027 and 2028.

It does not take a side. Both parties’ voters said they do not want a data center next door, by different margins. Both parties’ candidates proposed making builders pay their own grid costs. The disagreement was mostly about speed: pause first and write the rules, or write the rules and keep building.

Unlock The Full AI Supercycle Weekly!

What changed on 3 November

Read more

Anthropic’s impressive S-1 set the tone. It surprised many on the upside, particularly given concerns about customer concentration. Yet it also brought a broader dependency into focus: just how much of the AI trade now rests on Anthropic’s ability to sustain its extraordinary pace.

For the first time since I began tracking the AI Supercycle, the infrastructure being built ahead of demand started to look a little aggressive relative to how confidently we can project that demand into the future. Demand is still accelerating. The harder question is how durable that growth will prove, and whether its pace and economics will support the commitments being made today.

We are increasingly taking several assumptions for granted: that compute costs will keep rising, token prices will keep falling exponentially, and agents will drive an enormous expansion in demand. Each has profound implications for the infrastructure being built. Yet the speed of these changes, how they interact, and who ultimately captures the value remain open questions. The confidence behind today’s investment commitments depends on how those variables develop together.

Meanwhile, the foundations surrounding this industry are shifting. The geopolitical landscape is being redrawn. The macroeconomic environment is changing. And we are building what could become the most transformative industry in a century amid deep political instability, in the United States and across the world. Each of these forces can reshape the economics and direction of the AI buildout.

This week’s report is my attempt to make sense of that combination. I know it is a demanding read. Understanding this cycle requires us to hold several moving parts in view at once—and to keep questioning assumptions that can quickly begin to feel like certainties.

My goal is to give you a stronger compass through that complexity: a way to connect the signals, understand what is changing, and update our view as the evidence develops.

Think of the AI build-out as one long supply chain again. Last week we asked who pays for it. This week the companies that pay opened their books, and the answer came in five parts.

1. The sellers took the margins. SK hynix earned an operating margin of 80% in the quarter. Samsung’s chips made about 99% of a record profit, by our arithmetic from its release. Sandisk sold $11.0 billion in a quarter at an 85.0% gross margin.

2. The builders ran past their own cash. Alphabet spent 121% of its operating cash on property and equipment in the quarter. Amazon spent about 1.8 times. Microsoft spent 68%, the only one of the big builders inside its cash, by our arithmetic from their filings. S&P turned Meta’s outlook negative over the same gap.

3. The demand has a few names. Anthropic’s top five customers made up about 25% of its revenue in its first public filing. One AI customer made up 34.5% of SpaceX’s third-quarter revenue.

4. The money grew selective. Microsoft sold $25 billion of bonds at about 64 basis points over Treasuries. The same week, Blackstone reportedly told investors to expect a below-investment-grade rating for Anthropic, and OpenAI reportedly cut its round to about $20 billion.

5. The states built fences. Google is in talks to work on data security for the $185 billion Golden Dome missile-defence programme, Bloomberg reported. France’s lower house voted a tax of €0.055 on every gigabyte content providers send. A China-linked group ran AI agents against more than 200 organisations, CrowdStrike said.

The short version: the parts are still scarce and well paid. The fault lines now run through the cash of the builders and the breadth of the buyers, and governments are drawing lines around data, compute and capital.

The week in one line. For two years the AI build-out was priced as a growth story. This week it started being priced as a credit story. A ratings agency, the credit-default swap market, bond buyers and late-stage investors all put a cost of capital on the capex. That cost, not the supply of chips, became the constraint everyone else had to plan around.

Unlock The Full AI Supercycle Analysis

How to read this issue

Every week we read the news onto one picture, the Map of AI. It stacks the industry in layers, from energy at the bottom to the rules at the top. Here are the five ideas the rest of the issue uses.

• The layers. Energy is Layer 1, and chips and their parts are Layers 2 and 3. Compute and the money that pays for it are Layer 5, models Layer 6 and the agent software around them Layer 7. The stores and banks agents must pass are Layer 8, and the rules are Layer 9. The paying customer sits above the map as Adoption.

• The chokepoint. Each layer has one scarce piece this week. For each, we name who holds it and who pays to use it.

• The colours. In the pictures, teal marks the side that held or gained. Ochre marks the side that paid, waited or was refused.

• The cascade and the loop. World politics shapes the economy, and the economy shapes technology. In a supercycle the arrows also run back: AI moves the economy and politics in return.

• The four clocks. Each part of the build runs on its own clock: factories over years, money over weeks, efficiency over months, customers over quarters. Trouble starts when two clocks disagree.

Two habits to note. “Reported” marks figures from news outlets that the companies have not confirmed. “By our arithmetic” marks a sum or ratio we worked out from published figures.

The Week the AI Build-Out Got a Credit Rating

Start at the bottom of the map and climb. Every layer had a side that gained and a side that paid this week. The pressure gathered on two of them.

The first was the builders’ operating cash, in Layer 5, where the hyperscalers and labs pay for chips and data centres. The second was the customer, above the map, whose spending is meant to repay it all and who turned out to be a short list.

Get A Yearly Plan To Unlock The Library

Read more

For the quarter ended 30 September 2026, Google Cloud revenue more than doubled to $30.7 billion, while Azure grew 48% to $33.2 billion. Those numbers provide powerful evidence of demand for the infrastructure beneath AI. But the most revealing part of the story begins with how each company is pursuing that opportunity.

Alphabet is building across the AI stack. Its own TPU chips power its Gemini models; those models feed products such as Search; Google Cloud puts that technology in the hands of developers and enterprises. The strategic wager is that advances in one layer can strengthen the others: more efficient computing, more capable models, more useful products, and more demand for the infrastructure underneath. Alphabet is using AI both to renew the business that made it dominant and to expand the cloud business that could become its next great growth engine.

Microsoft is bringing AI into the businesses it already serves. Azure supplies the computing infrastructure. Microsoft 365 and Copilot bring AI into documents, spreadsheets, conversations and workflows. The strategic logic is to turn established customer relationships into deeper demand for software, agents and cloud capacity. As customers put more work through those systems, Microsoft has opportunities to earn revenue from both the applications they use and the infrastructure those applications consume.

Both approaches require enormous investment. This quarter, Alphabet’s cash spending on property and equipment equalled 40% of revenue. Microsoft’s reached 43%, with its broader capital expenditure measure, including finance leases, reaching 49%. These are businesses committing close to half their sales to expanding their physical foundations.

Then the cash-flow statements reveal the difference.

For every $100 of revenue, Microsoft generated approximately $63 in operating cash. Alphabet generated $33. After cash capital expenditure, Microsoft still had $18.6 billion in free cash flow. Alphabet’s spending exceeded its operating cash generation by $9.1 billion. During the same quarter, Alphabet also raised $28.7 billion through new notes and $12.4 billion through share sales, adding external capital as it expanded.

That gives this quarter’s results their strategic significance. Microsoft’s operating cash generation covered its cash investment bill with room to spare. Alphabet combined operating cash with access to capital markets to support its ambitions. Both are building aggressively, but the financial room around those investments looks different.

Cash conversion is becoming a decisive advantage in the AI buildout. Demand gives companies a reason to build. Cash generation gives them more freedom to keep building.

That changes how to read the Map of AI. Alongside chips, models, distribution and customer relationships, we need to watch the time between committing capital and collecting the returns. Long-term contracts can provide visibility into future business, but the infrastructure must be financed before all that revenue arrives.

Alphabet and Microsoft are pursuing the same enormous opportunity from different positions of strength. This quarter shows why the ability to finance that wait deserves a place alongside technology in the story of who captures it.

*Financial figures cover July–September 2026: Alphabet’s Q3 and Microsoft’s fiscal 2027 first quarter. Cash-flow and capital-spending ratios are our calculations. Free cash flow here means operating cash flow less cash purchases of property and equipment; finance lease additions are shown separately. Forward capital expenditure guidance should be read alongside the earnings-call commentary.

Read more

Within that system, I believe Anthropic has become the most consequential company to emerge from the AI race.

It was born out of a not-so-friendly split within OpenAI. The roots of that split reach back to disagreements over how increasingly powerful models should be developed and released. As early as GPT-2 in 2019, Dario Amodei was urging caution about releasing the full model, and OpenAI initially adopted a staged approach. The broader disagreements over safety, commercialization and leadership eventually helped drive Amodei and his colleagues to build something of their own.

Fast forward a few years, and that internal diaspora has produced a company that, in my view, is now the axis around which much of the AI industry turns.

Until 2025, I largely looked at OpenAI to understand where AI would go next. Starting in 2026, that changed. Increasingly, I look at Anthropic. Not simply because of its models, but because of how its choices shape the ecosystem around them.

Having followed Anthropic since its early days, and having been convinced from the outset that it could become the next OpenAI, I’ve been particularly eager to read its S-1. Not just to understand the company’s numbers, but to see what they reveal about the economics of the entire AI stack.

Now it’s out.

Anthropic’s S-1 shows a company that rents almost every layer of the AI stack and owns the one that matters most to it: the customer. It has signed compute deals with seven named counterparties on terms that run to 2045. It sells through three clouds that are also its shareholders and its competitors. Yet it signs the customer contract itself, sets the price, and keeps voting control away from every landlord.

The filing is the first audited look inside a frontier lab. It also tests a claim the Q2 2026 Map of AI made: models get rented, and the routing junction gets owned. Anthropic’s numbers show a model maker climbing into the junction, and paying a visible toll to stay there. Because a frontier lab touches every layer, the filing is also a reading of the whole ecosystem, and of where it moves next.

Compression: rent the floor, own the customer, fence the landlords.

Get A Yearly Plan To Unlock The Library

Read Also

Read more

Unlock The Full AI Supercycle Analysis

This week the two ends pulled apart.

1. The sellers are booked for years. Amazon’s chief executive Andy Jassy said “2027 is largely sold out” for AWS capacity, Reuters reported. Two US memory makers, Micron and Sandisk, carry about $225 billion of sales already under contract, by our arithmetic from their filings.

2. The payers found money dearer. SpaceX’s $40 billion chip debt was set to price at about 7.75% to 8.3%, then Bloomberg reported the deal “roiled” by market turmoil. Meta’s interest bill rose 330% in a year, and its quarterly free cash flow turned negative.

3. The customers are fewer than they look. On data from Ramp, a corporate card company, 1% of firms drive 80% of business spend on OpenAI and Anthropic. Much of that spend comes from companies funded by venture capital.

4. Prices for the product kept falling. A Chinese lab, Moonshot, priced a frontier-class model at half of Anthropic’s top model. Token prices fell about 50% in a year in one data firm’s panels.

5. Equity stayed cheap while debt got dear. OpenAI is reportedly close to raising up to $30 billion at a $1.25 trillion valuation, and DeepSeek reportedly raised about $15 billion. In the same week, SpaceX’s chip debt snagged, even as its AI business passed half of its revenue.

6. AI pushed back into the wider world. It showed up in central-bank talk about inflation, in a national budget’s largest AI heading (security), and in an attack run by AI agents. That two-way pressure is what we mean by a supercycle.

The short version: the build isn't short on demand for its parts. It is short of cheap money, and its end revenue rests on a narrow base of large, often venture-funded buyers.

How to read this issue

Every week we read the news onto one picture, the Map of AI. It stacks the industry in layers, from energy at the bottom to the rules at the top. Here are the five ideas the rest of the issue uses.

• The layers. Energy is Layer 1, and chips and their parts are Layers 2 and 3. Compute and the money that pays for it are Layer 5, models Layer 6 and the agent software around them Layer 7. The stores and banks agents must pass are Layer 8, and the rules are Layer 9. The paying customer sits above the map as Adoption.

• The chokepoint. Each layer has one scarce piece this week. For each, we name who holds it and who pays to use it.

• The colours. In the pictures, teal marks the side that held or gained. Ochre marks the side that paid, waited or was refused.

• The cascade and the loop. World politics shapes the economy, and the economy shapes technology. In a supercycle the arrows also run back: AI moves the economy and politics in return.

• The four clocks. Each part of the build runs on its own clock: factories over years, money over weeks, efficiency over months, customers over quarters. Trouble starts when two clocks disagree.

Two habits to note. “Reported” marks figures from news outlets that the companies have not confirmed. “By our arithmetic” marks a sum or ratio we worked out from published figures.

Who Pays for the Build

Start at the bottom of the map and climb. Every layer had a side that paid this week. The pressure, though, gathered on two of them.

The first was the cost of money, in Layer 5, where builders borrow to buy chips and switch on data centers. The second was the customer, above the map, whose spending is meant to repay it all.

Get A Yearly Plan To Unlock The Library

Unlock The Full AI Supercycle Analysis

Read more

Pacioli did not invent double-entry bookkeeping. He wrote it down, named its parts, and explained them so that someone who had never sat in a Venetian counting house could learn them. That is how a practice becomes a discipline. After the writing down, the practice can be taught, compared, audited and improved by people who never met its originators.

Embedding AI into the core of a company is at that stage now. A small number of companies have learned to move an AI system that makes decisions from a pilot into work the business relies on. Their methods converge. But they live mostly in practitioners’ heads and in vendors’ slides, and the same idea goes by five names in five companies.

There is a deeper reason to write it down. The machines now entering companies are the latest step in a very old project: writing down how thinking works, so that something other than a person can do it. That project has two traditions, and the most capable systems now combine them. A company that embeds AI is building such a combination, whether it knows it or not. It can buy one half. It has to write the other.

A body of knowledge for embedding AI is, in a precise sense, the written half of the company.

What follows is the whole body of knowledge in one sitting. It runs from the foundations, through the core method and the nine areas of knowledge, to the measures, the people, the evidence and a library of mental models organised by type of loop. It ends with how the rest of the discipline follows from it.

Thus, if in the previous part of this series I focused on what I called “the work loop.” On this part, we’ll look at the body of knowledge, or the foundation of it all.

Get The Yearly Plan To Unlock The Book!

The Body of Knowledge is Part of The Enterprise Embedding Series!

The Full Series With The Exec Plan!

The Full Library With The Yearly Plan!!

Get The Exec Plan for The Full Module

Or preview it here.

If you’re an existing paid annual subscriber, you already have access to the library (if you don’t, just request it). To unlock the full video lectures, skills, and all additional resources, upgrade to the Executive Plan.

Read more

ASML warned that its planned production would fall short of customer demand in both 2027 and 2028, while Nvidia estimates a roughly $40 billion revenue opportunity per gigawatt of data centre capacity for its Vera Rubin platform.

Businesses funding that infrastructure faced a tougher calculation: CoreWeave’s 2032 bonds traded at a 14.8% yield, signalling investors’ demands for greater compensation, while Anthropic reportedly offered Blackstone more than 10% on junior debt financing a chip deal.

Model pricing reflected the same economics: smaller models became cheaper, with Haiku 5.5 starting at $0.10 per million input tokens for shorter prompts, while faster access still commanded a premium and major buyers introduced spending caps and closer usage tracking.

The struggle over terms also reached government: Anthropic pitched Congress its own draft safety bill as the FTC expanded its investigations, European authorities demanded information about model safeguards, and eight US states sued OpenAI over alleged consumer harms.

Together, these developments suggest that cheaper intelligence does not automatically mean cheaper AI businesses: the companies controlling the resources, distribution and permissions needed to scale can still capture much of the value.

Who Owns the Chokepoints

The week on the Map of AI: one chokepoint per layer, the companies that held it and the ones that paid for access. Placement is our read; “reported” marks figures from outlets rather than company statements.

How to read the map: start at the bottom with energy and climb. Each layer has one chokepoint this week. On the left is who held it; on the right is who paid to use it.

Get A Yearly Plan To Unlock The Library

Unlock Full Weekly Video Explainer

From Where the Cost Went to Who Owns the Chokepoints

Last week’s question, the three tests set for this week, and how they came back.

Last week’s issue asked where the cost of cheaper AI went. Model calls kept getting cheaper: decision models fell to between $0.02 and $0.10 per million input tokens.

The cost did not disappear. It moved down the stack and onto balance sheets, as commitments, guarantees and debt. By our arithmetic, Nvidia’s latest quarterly filing (its 10-Q) lists $530.5 billion of them. The Financial Times reported a $60 billion debt package, backed by Broadcom, for Anthropic’s chip leases. And liability for what agents do was moving to the companies that deploy them.

That issue closed with three tests for this week. All three came back.

• Would ASML’s 2027 orders soften? No. ASML, the Dutch company that makes the only machines able to print the most advanced chips, reported on 14 October. It said its planned supply for 2027 and 2028 will not fully meet demand.

• Would TSMC disappoint? No. TSMC, which manufactures the chips Nvidia, Apple and Google design, reported on 15 October and beat its own range.

• Would spreads widen as borrowers refinanced? Yes, in the middle. CoreWeave’s bonds traded at a 14.8% yield, and Anthropic reportedly offered a lender more than 10%.

So this week’s question follows from last week’s answer. The scarce pieces are sold out, and the money to build them is getting dearer. Who holds those pieces, and who pays to use them?

Get A Yearly Plan To Unlock The Library

Geo, Then Macro, Then Tech

The week’s geopolitics and macro in three tiers, and where the pressure landed on the AI stack. The arrows show channels, not proof of cause.

The Business Engineer’s Cascade Analysis Framework reads the world from the top down. Geopolitics “cascades downward through macroeconomic conditions and into technology dynamics,” and “each layer constrains the next.” This week all three tiers moved.

Geo: who can sell what to whom

On 16 October, China’s commerce and science ministries proposed adding 40 technologies to the country’s export-control catalogue. Of those, 23 are in AI: large models, agents, computing clusters and information processing. Open-source work is excluded, and comments run until 15 November.

The same day, Taiwan said shipping restricted AI hardware to China will become a criminal offence, with controls aligned to American ones. Taiwan also plans industrial parks of up to $4 billion in Texas for Nvidia’s suppliers. Behind both sat the shocks the IMF named at its meetings in Bangkok: spillovers from wars and high energy prices.

Macro: the price of money moved up

US consumer prices rose 0.6% in September and 3.7% over the year, with energy the biggest push. After the report, the 30-year Treasury yield touched 5.63%.

Junk-bond yields reached 8.41%, an 18-month high. That came even as companies sold $106 billion of project bonds this year, against $37 billion in all of 2025, Bloomberg reported. Markets marked the chain down too: the chip index fell more than 12% in seven sessions.

Growth still leans on AI. The IMF held 2026 world growth at 3.1%, crediting AI investment. A Goldman Sachs exhibit puts hyperscaler capex at about 11 points of S&P 500 earnings growth this year. That fades to about minus one point in 2028, as depreciation arrives.

Tech: where it lands

The pressure reaches the stack through two channels. Dearer money falls hardest on the companies that borrow to build, which is why the middle of the stack paid more this week (Layer 5). Export rules and energy costs decide where chips can go and what a site costs to run. That is why the factory slot and the megawatt are chokepoints (Layers 1 to 3).

Above them, labs re-priced and buyers began to meter. The rule-writers moved on both sides of the Pacific. These things happened in the same week; the arrows are channels, not proof of cause.

How the Week Connects

Read more

I had a rather unconventional background myself. I started in law, moved into business and finance, and was then pivoting toward technology, startups, and digital entrepreneurship.

What helped me navigate that transition was the concept of business modeling.

For a while, business modeling became almost all-encompassing for me. It gave me a way to simplify complexity, recognize patterns, and understand how businesses worked across industries.

But around 2017–2018, I realized something was missing. Business modeling was a powerful starting point, yet there were several critical building blocks needed to turn it into something far more complete, useful, and grounded in the real world than a canvas.

That’s when The Business Engineer began to emerge.

At its core, it was about creating a continuous loop of abstraction, compression, comparison, mental modeling, and real-world testing. A way to move between theory and practice, constantly refining how we understand, build, and operate businesses.

Importantly, this wasn’t the conventional definition of business engineering. I never found that definition particularly useful for what I was trying to achieve.

Instead, I shaped the discipline around my own experience and perspective: as a startup executive, a digital entrepreneur, a business analyst, and, perhaps most importantly, someone with an almost obsessive curiosity about how the business world works.

Over time, The Business Engineer became my attempt to bring those perspectives together into a coherent discipline. Not just to analyze businesses, but to understand their underlying mechanisms, model them, test them, and ultimately make better decisions in the real world.

And today, I’m introducing another critical building block of The Business Engineer’s discipline, one that I believe becomes particularly important in the AI era: the Work Loop Canvas.

Get A Yearly Plan To Unlock The Library

Get The Exec Plan for The Full Module

Or preview it here.

If you’re an existing paid annual subscriber, you already have access to the library (if you don’t, just request it). To unlock the full video lectures, skills, and all additional resources, upgrade to the Executive Plan.

Every era of business technology leaves behind one picture that people draw on whiteboards. For the 2000s it was a sheet of paper divided into nine boxes. Founders filled it with sticky notes. Bankers asked for it. Business schools taught it. It made a hard question easy to discuss: how does this company make money?

That question still matters, but it is no longer the hard one. A team can describe a workable business model in an afternoon, test it in a week and copy a competitor’s in a month. The hard question has moved inside the firm. It asks how the work actually gets done, and whether AI can do part of it.

Most companies are struggling with that question. They run AI pilots at the edge of the business and cannot move them into the core. The pilots work. The core does not change. This piece explains why, and offers an instrument for the problem: The Work Loop Canvas, a one-page map of a single work loop, drawn as a graph of twelve connected nodes. Its companion, the Work Loop Map, places every loop in the firm on one page.

Get A Yearly Plan To Unlock The Book!

A Century of Designing Work

Read more

This week brought AI price cuts. OpenAI opened a Decisions API at $0.10 per million input tokens, with no charge for output. Perplexity halved its decision model to $0.02 per million input tokens. Anthropic’s Claude Haiku 5.5 arrived at about 75% of Haiku 4.5's price.

A price cut does not remove a cost. Someone still pays for the chips, memory, power, and mistakes. So the question for this issue is narrow: when the price of a model call fell, where did the money and the risk show up instead?

Our answer, from the week’s filings, releases and price lists: the cost rose in the layers underneath, at GPU renters, memory makers and chip suppliers, and it is increasingly carried on balance sheets as commitments, guarantees and debt rather than raised as equity.

A third cost, liability for what agents do, is starting to land on the companies that deploy them. These moved in the same week, and we don't claim one caused the other.

Get A Yearly Plan To Unlock The Library

Read more

What enterprise AI needs is increasingly different from what consumer AI wants.

That distinction will become central to what I believe is the most important AI race of the next 12–24 months: the meta harness.

The term has already been circulating in the last year in closed tech circles. And while those circles have a habit of repeating the same language until everyone sounds identical, there is a meaningful signal underneath the buzzword.

A meta harness is the broader system that coordinates models, tools, context, and execution to get something done. I believe this layer will become critical across both consumer and enterprise AI. But its architecture, responsibilities, and sources of advantage will differ substantially.

In consumer AI, I expect the strongest expression of the meta harness to be a vertically integrated experience: a company owning the interface and managing the task from beginning to end. Something like Muse is the direction I have in mind.

For many everyday tasks, such as finding flights or planning a trip, you do not necessarily need the most powerful frontier model at every step. You need a system that understands what you want, works across the relevant services, and reliably completes the task. Much of the value sits in how seamlessly the company brings those pieces together.

For business and enterprise AI, the meta harness has a broader responsibility.

It needs to do far more than scaffold a powerful model or route requests between models. Routing was already an important problem two years ago. The next challenge is bringing different models and capabilities into a coherent system that improves through execution, feedback, and evaluation.

Here, model agnosticism means more than being able to switch providers. It means understanding how to divide the work between the models and the harness itself.

When can a frontier model perform a task from beginning to end, with minimal intervention? When does the harness need to take the lead, break the work into stages, coordinate tools, enforce constraints, or check the result?

That division of responsibility cannot remain fixed. As models become more capable, some of the work previously handled by the harness can move into the model. At the same time, enterprise requirements create responsibilities the broader system must continue to own: permissions, approvals, traceability, and accountability for execution.

This is what makes the enterprise meta harness a deeper architectural challenge. It must continually adapt to the intelligence available underneath it while keeping the overall process dependable.

I expect the shifting boundary between model intelligence and harness orchestration to define much of the next phase of AI.

Every meta harness provider will face the same question: where is the sweet spot between letting frontier intelligence do the work and building the structure that makes that intelligence useful, reliable, and controllable?

Too much scaffolding can constrain a capable model. Too little can leave the system unable to deliver consistently.

Finding that balance, and continuously improving it, is the race.

And in my view, Grok Bot helped open it.

Gain Full Access to The Research & Atlas

If you are a paid member, reply to this to request access.

Read more

Preview the book & course module here!

Subscribe To Premium To Gain Access!

Preview the book & course module here!

Subscribe To Premium To Gain Access!

If you’re already a paid yearly member, simply reply to this email, and we’ll send it your way.

Read more

OpenClaw was the real aha moment for what personal agents could do, both for business users and consumers. In less than a year since its inception, the landscape has evolved dramatically. Things are moving incredibly fast, especially for business users, while consumer interfaces that initially looked like “OpenClaw copycats” are beginning to specialize and develop identities of their own.

For the vast majority of consumers, an agentic experience is still something they have never encountered. These new personal agents will be their first point of contact with AI that can actually act on their behalf.

As a business user working at the edge of the AI industry, interfaces like Instinct or Muse didn’t impress me technically in the same way that Claude Code, OpenClaw or Grok Bot did. Those harnesses offered a glimpse of the full capabilities of agentic AI. But my perspective isn’t necessarily the one that determines what succeeds with consumers.

Consumer AI is a different game. It sits at the intersection of interface design, monetization that can sustain free access at scale, and brand recognition. The most technically impressive harness won’t automatically become the interface millions of people use every day.

That is what makes this moment so interesting: the personal agent landscape is finally emerging in full swing, and we are beginning to see how these capabilities might reach a mass audience.

Gain Full Access to The Research & Atlas

The personal agent is becoming a new layer between intent and fulfilment. The fight is no longer only over which model you use. It is over who receives the first request, who earns the right to act, and who gets paid when the work is done.

A user sends a voice note: “I need to be in New York tonight.” Nobody opens an airline app, compares hotels, checks Uber, copies the itinerary into a calendar or types a card number. The agent already knows the user’s location, preferred airline, seat, payment method and schedule. It books the flight, reserves the hotel, orders the car and updates the calendar. Earlier that day, the same agent may have called first because a document needed signing before 3 p.m.

That is how Noah Shinn describes Instinct. The product has no conventional app. It has a phone number, an email address and a computer. You deal with it much the way you would deal with another person. down

For the first few years of consumer generative AI, the chat window was the product. You opened ChatGPT, Claude, Gemini or another application, asked for something, received an answer, then completed the workflow yourself. The assistant sat beside the work.

The new personal agents are trying to sit in front of the work.

They remember. They watch. They decide when something requires attention. They browse, call tools, spend money, coordinate with other people and increasingly continue working after the user has gone away.

By September 2026, Meta had launched Muse, Apple had put Siri AI into beta, Anthropic had folded Cowork into Claude, Rabbit had released OS3, Microsoft had announced Autopilot and OpenAI had launched Dots.

Google had already been running Gemini Spark, Amazon had pushed its shopping agent deeper into Alexa, and Chinese internet groups were reported to be testing their own versions. Instinct announced a $1 billion financing round at a $10 billion valuation in the same period.

The products differ enormously underneath, but they are converging on the same architectural idea:

The personal agent is becoming a new layer in the stack: the delegation layer.

Search engines spent two decades owning the moment when the user asked, Where can I find this? Marketplaces fought to own the moment when the user asked, Where should I buy it? The personal agent wants to own an earlier moment:

“Here is what I want. Take care of it.”

That position is potentially more valuable than the search box or the checkout page because it sits before both. The agent can decide where to search, which service to use, which merchant to trust, whether a transaction is worth making and whether the user needs to see the intermediate steps at all.

The category becomes much easier to understand through six questions: Where does the agent run? How do you reach it? Who starts the work? Who pays? How does it touch other services? Who else can it talk to?

Subscribe To Premium To Gain Access!

If you are a paid member, reply to this to request access.

Also watch the full video lecture below!

Read more

First came pretraining. Scale the Transformer with more data, more parameters, and more compute, and increasingly powerful capabilities emerged. Then came post-training. Instruction tuning, RLHF, reinforcement learning, verifiable rewards, and eventually reasoning techniques turned those pretrained models into systems that could follow instructions, solve harder problems, and interact with people in useful ways.

That was the first phase of the cycle. Most of the industry’s attention, capital, and technical prestige went into creating intelligence.

We are now moving into a different phase: the inference phase.

Training does not become less important. Frontier models will continue to require enormous amounts of compute. But training happens in large, concentrated runs. Inference happens every time the model is used. Every question, search, document review, coding task, reasoning step, tool call, and agent action requires the machine to run again. Google reported processing more than 3.2 quadrillion tokens per month as read in 2026, around seven times the level of a year earlier. At the same time, reasoning models are producing longer outputs, context windows have expanded dramatically, and agents can invoke models dozens of times to complete a single piece of work.

My expectation is that over the next three to five years, enterprise inference becomes at least as important economically as the pretraining layer, and potentially much larger. The reason is straightforward: once intelligence becomes useful enough, the economic problem shifts from producing the model to operating it continuously across millions of workflows.

And this is where something important changes.

For years, inference looked like plumbing. You picked a model, called an API, paid for the tokens, and focused on the application above it. But at enterprise scale, that abstraction starts to break down. The price of the model is only one part of the cost. The economics increasingly depend on how the request is served.

A model request actually contains two very different jobs.

Prefill is reading. The model processes the prompt, the instructions, the documents, the conversation history, and everything else in the context. Those input tokens can largely be processed in parallel, which makes prefill relatively compute-intensive and strongly influences how long the user waits before the first token appears.

Decode is writing. Once the model starts answering, it generates one token after another because each new token depends on what came before. That makes generation far more sensitive to memory bandwidth and to how many conversations share the hardware at the same time.

That distinction sounds technical, but it determines the economics of enterprise AI.

A system that reads a 100-page contract and returns one classification has a very different cost structure from a coding agent generating thousands of tokens from a short instruction. An agent that repeatedly rereads 50,000 tokens of context over thirty steps has a different profile again. The same model, at the same token price, can produce radically different costs depending on what it is asked to read, generate, repeat, and remember.

This is why inference engineering is becoming a discipline of its own.

It is about batching requests so expensive hardware stays productive without making users wait too long. It is about managing the KV cache, the working memory of each conversation, so long contexts do not consume the entire machine. It is about reusing prefixes instead of repeatedly paying to read the same instructions and documents. It is about deciding which work needs real-time capacity, which can run in batch, when dedicated infrastructure makes sense, and when renting tokens remains cheaper.

Above all, it is about changing the unit of optimization.

The objective is not the cheapest token.

It is the lowest cost per accepted outcome at the latency the workload actually requires.

That becomes even more important as reasoning and agents spread. Reasoning models can consume more computation to solve harder problems. Agents multiply inference calls and repeatedly reread growing contexts. The cost per token may keep collapsing while the number of tokens required to perform useful work rises much faster. That is how intelligence can become dramatically cheaper while the total inference market becomes dramatically larger.

So this piece goes beneath the API and into the machinery that increasingly determines the real economics of AI: prefill and decode, working memory, batching, caching, serving modes, latency, utilization, hardware, and the software engines connecting them all.

The concepts are technical, but the logic is surprisingly simple.

Serve to the target. Keep the cache warm. Pay for busy, not idle.

That is the operating system of the inference phase.

Preview the book & course module here!

Subscribe To Premium To Gain Access!

For the last three years, I’ve been rebuilding the Business Engineer’s curriculum from the ground up. That curriculum has now become the foundation of a new discipline, with the entire series taking shape around it.

Subscribe To Premium To Gain Access!

If you’re already a paid member, simply reply to this email, and we’ll send it your way.

Preview the book & course module here!

Subscribe To Premium To Gain Access!

Read more

For almost four years, much of the AI conversation has centered on how intelligence gets built. Pretraining gave models broad capabilities by learning from enormous datasets. Post-training shaped those capabilities into more useful behavior. Reinforcement learning from human feedback, or RLHF, helped turn language models into assistants that could follow instructions and interact with people naturally. These were not separate technologies replacing one another, but successive layers of improvement around the model.

We have now entered what I would call the inference phase of the AI Supercycle. Not because training has stopped mattering, but because the economic question is expanding from How do we build a capable model? to How do we operate that capability repeatedly, across millions of tasks, at a cost the customer can justify? Reasoning models make this transition particularly important: additional computation during inference can itself improve performance. The model’s capabilities depend not only on how it was trained, but also on how much work the system performs when answering.

My expectation is that over the next three to five years, production inference could become a larger spending pool than pretraining alone, with enterprise deployment an important driver. Evidence already points in that direction within a defined market: Gartner forecasts inference to account for 55% of AI-optimized cloud infrastructure spending in 2026 and 59% in 2027, exceeding training spending in that segment. This isn't an industry-wide accounting of every AI investment, but it shows why inference deserves to move from a technical consideration to the center of enterprise strategy (see Gartner).

The mechanism is straightforward. Training costs recur as models are developed and improved. Inference costs recur whenever those models are used. And an agentic workflow can require much more than a single answer. Imagine a system investigating a customer issue: it reads the account history, checks a contract, chooses a tool, processes the result, revises its plan, and prepares an action. Each model call adds work. Multiply that pattern across departments, customers, and continuous operations, and serving intelligence becomes a substantial production system.

But inference is not one uniform workload, and that is where the nuances become fundamental. In a typical autoregressive language model, two phases do different jobs:

• Prefill processes the input. The model works through the prompt, retrieved documents, instructions, and other context, building intermediate attention state commonly held in a key-value, or KV, cache. Much of this processing can happen in parallel. For sufficiently long inputs, available computation is often the main constraint, and prefill contributes substantially to the wait before generation begins.

• Decode generates the output. The model produces subsequent tokens sequentially, using the context and previously generated tokens. Especially at smaller batch sizes, performance is often constrained by how quickly weights and cached state can move through memory rather than by arithmetic capacity alone. This affects how quickly the response progresses after generation begins.

That distinction has direct business consequences. A system reading a large contract pack to return a short classification has a different workload profile from one generating a long software change from a short instruction. An agent repeatedly using the same context introduces another consideration: whether the serving system can reuse previous computation instead of paying to process the same material again. The same model can have very different economics depending on what the application asks it to read, generate, and repeat.

This is why serving architecture matters. Prefill and decode can sometimes benefit from separate resource pools, different scheduling, and different capacity allocations. But separation isn't automatically better: transferring cached state adds overhead, and short or heavily cached requests may be more efficient without it. The right architecture follows the workload, not a universal rule about which chip, cloud, or model is best.

For an enterprise, this turns apparently technical choices into operating decisions. How much context should each request receive? Which repeated inputs can be cached? Which tasks can wait for batch processing? Which require immediate responses? Where does a more capable model reduce retries and human review enough to justify its price? The objective is not simply cheaper tokens. It is lower cost per accepted outcome, with quality, latency, and reliability held to the required standard.

That is also why this phase requires a map of the whole stack. Inference depends on the physical infrastructure underneath the model, but useful enterprise work depends equally on the context, routing, tools, permissions, memory, and evaluation around it. The company must decide not only where intelligence runs, but where the knowledge and authority that make it useful accumulate. Those are the decisions that determine whether it can change providers without rebuilding its operating capability.

The first phase made us ask how powerful the model could become. The inference phase forces us to understand the system that turns that capability into recurring work. This piece maps that system, from power and silicon to serving, agents, and enterprise control, and asks which parts should be rented and which must remain under the company’s control.

Rent the layers. Own the seams. Price the exit.

Preview the book & course module here!

Subscribe To Premium To Gain Access!

For the last three years, I’ve been rebuilding the Business Engineer’s curriculum from the ground up. That curriculum has now become the foundation of a new discipline, with the entire series taking shape around it.

Subscribe To Premium To Gain Access!

If you’re already a paid member, simply reply to this email, and we’ll send it your way.

Preview the book & course module here!

Subscribe To Premium To Gain Access!

Read more

Working in technology added another intellectual habit to that historical one: Silicon Valley’s obsession with first-principles thinking. It is a powerful idea, but also one surrounded by mythology. We all like to imagine ourselves stripping a problem down to fundamental truths and rebuilding it from scratch. Reality is less flattering. Almost all of us borrow from each other. Most innovation is recombination. Only a tiny number of people in any field genuinely redefine its first principles.

For the business person, that is not necessarily a disadvantage. In fact, one of the most valuable skills a generalist can develop is the ability to move across disciplines without becoming overly loyal to any of them. Specialization creates extraordinary depth, but it also creates boundaries. Finance speaks to finance. Computer science to computer science. Economics to economics. Organization theory to organization theory. The operator does not have that luxury because reality does not respect departments.

This is also why the best entrepreneurs often remind me more of scientists than of academics at their most specialized. The scientist has a hypothesis and eventually has to surrender it to reality. The entrepreneur does too. The business either works or it does not. Customers either care or they do not. The unit economics eventually show up. Reality has very little respect for ideological purity.

And many of the largest technological breakthroughs come from precisely this violation of disciplinary boundaries. Modern AI is a good example. It would be inaccurate to say that everything we call agentic AI today is formally neuro-symbolic AI. It is not. But the most capable systems increasingly combine things that once belonged to different intellectual worlds: neural models with search and retrieval, structured knowledge, deterministic code, tools, rules, memory, permissions, evaluators, and traditional software systems.

The progress comes from the combination. The model alone is not the system.

That brings us to the agentic Web, because one of the greatest mistakes technical people make is assuming that solving the technology also solves the business.

Very intelligent people can reason brilliantly from first principles inside their own discipline and then become strangely simplistic when they cross into business models. Technology may tell you what can exist. It does not tell you who pays for it, why they pay, how distribution works, which incentives emerge, or what happens when billions of ordinary users behave differently from the elegant model in your head.

One of my favorite examples is Google.

In 1996, Larry Page and Sergey Brin were Stanford graduate students working on a research project that became BackRub. Their key insight grew partly from academic citation analysis: a document should not be judged merely by whether it contains a keyword, but also by which other documents point toward it and how important those documents themselves are. That logic became PageRank and ultimately the foundation of Google Search.

The technology was extraordinary. But the business model was not obvious.

In their 1998 paper The Anatomy of a Large-Scale Hypertextual Web Search Engine, Brin and Page were explicitly skeptical of advertising-funded search. Their concern was legitimate: if advertisers pay the search engine, commercial incentives can conflict with the goal of returning the best possible information. They wrote that advertising-funded search engines could become “biased towards the advertisers and away from the needs of the consumers.”

Think about the irony.

The founders of what would become one of the greatest advertising businesses in history initially wrote an academic paper explaining why advertising was structurally dangerous for search.

They were not stupid.

They were not hypocrites.

And they did not suddenly become evil.

They encountered reality.

Search was expensive to build and operate. Users expected it to be free. The better Google became, the more people used it, and the more infrastructure it required. Somewhere, an economic engine had to finance that system.

At roughly the same time, Bill Gross and GoTo.com, later renamed Overture, were discovering something important: search intent had enormous commercial value. Instead of selling generic display advertising, advertisers could bid for keywords and pay when users actually clicked. GoTo pioneered the auction-driven pay-per-click model in 1998. Google initially took another route when it launched AdWords in 2000, but by 2002 it had moved to cost-per-click auctions as well. The crucial Google twist was to incorporate user response into ad ranking rather than simply allowing the highest bidder to win, helping align commercial value with relevance.

This is the part of the story I find fascinating.

Google did not resolve the conflict between advertising and search by pretending the conflict did not exist.

It redesigned the mechanism.

An irrelevant advertiser willing to pay more could not automatically deserve the best position if users consistently ignored the ad. Bid mattered, but relevance increasingly mattered too. Economics and product quality were made to reinforce each other more than they had in the earlier paid-search systems.

By the time Google went public in 2004, Larry Page could write something that would have looked almost absurd beside the 1998 paper: advertising was now Google’s principal source of revenue, while arguing that its ads could be relevant and useful rather than intrusive.

That business model financed free search and an expanding suite of consumer services at enormous scale.

Not the Internet itself. Users still paid telecom operators and ISPs for access, and publishers still financed the creation of content. But advertising allowed Google to put an extraordinarily expensive information-retrieval system in front of billions of users without charging them each time they searched.

And that economic decision helped shape the Web for the next quarter century.

This is why I become skeptical whenever a new technological paradigm begins with moral certainty about its future business model.

We hear versions of the same claim today around AI and the agentic Web: advertising was the disease of the old Web; subscriptions will create a cleaner one; perhaps micropayments will finally allow information to be priced correctly; perhaps agents will simply pay other agents per call and the whole advertising economy can disappear.

Maybe.

But consumer economics is brutally difficult.

Neeva is a useful warning. It launched with an appealing proposition: a subscription-funded, ad-free alternative to Google. Yet in 2023 its founders concluded there was no sustainable path for its consumer search business, shut the service down, and the company was acquired by Snowflake, where its technology moved into the enterprise data world.

Perplexity offers another, more nuanced example. It built around direct answers and paid subscriptions, but later began experimenting with advertising. Its own explanation was revealing: subscriptions alone did not generate enough revenue to create a scalable publisher revenue-sharing program. As of 2026, the company operates a mixture of free access, consumer subscriptions, enterprise plans, APIs, and other monetization surfaces rather than betting the company on one perfectly pure model.

None of this proves advertising wins again.

That would be exactly the kind of linear historical reasoning I am arguing against.

The agentic Web has fundamentally different economics. An agent does not need to look at a banner. It can call an endpoint directly. A machine can economically make micropayments that would be absurd for a person. Commerce can move from clicks toward transactions. Publishers can potentially meter machine access. Software can charge per call, task, or outcome. Payment protocols can make transactions almost invisible.

But someone still has to finance the system.

The models cost money.

Inference costs money.

Search and crawling cost money.

Publishers need incentives to continue creating information.

Merchants need margins.

Agent surfaces need economic engines.

And billions of consumers have repeatedly demonstrated that their theoretical enthusiasm for paying subscriptions is much greater than the number of subscriptions they actually want to maintain.

So the interesting question is not whether advertising is morally good or bad, nor whether subscriptions, micropayments, transactions, or licensing are intellectually cleaner.

The interesting question is:

What economic architecture can actually finance an agentic Web used by billions of people?

My suspicion is that the answer will be messy.

Advertising may migrate from the page into the answer or recommendation layer. Subscriptions will finance premium usage. Commerce will finance transactional agents. Enterprises will pay for productivity. Micropayments may finance individual calls and high-value information. Publishers will license some assets, expose others, fence others, and build endpoints around the rest.

In other words, the future probably will not emerge from one pristine business model.

It will emerge from the interaction of several.

And this is where history and first-principles thinking should meet.

First principles help us ask which assumptions of the old Web no longer hold.

History reminds us not to assume that the business model we find aesthetically attractive is the one reality will choose.

Google’s founders began with a technically brilliant search engine and a principled argument against the business model that ultimately financed it. Their real achievement was not merely changing their minds. It was discovering a mechanism through which the economics could be redesigned around the product.

The agentic Web is at a similar moment.

We can already see much of the technical architecture taking shape: the fact replacing the page, the call replacing part of the click, delegated identity, agent protocols, machine payments, settlement rails, provenance.

What we do not yet know is which combination of those layers will finance the system at global scale.

That is the question worth studying.

Not what the ideal Web should look like.

What business model reality will force it to become.

Subscribe To Premium To Gain Access!

Check out the Library for Free!

For the last three years, I’ve been rebuilding the Business Engineer’s curriculum from the ground up. That curriculum has now become the foundation of a new discipline, with the entire series taking shape around it.

Subscribe To Premium To Gain Access!

Check out the Library for Free!

If you’re already a paid member, simply reply to this email, and we’ll send it your way.

Read more

Access The Enterprise AI Ecosystem Atlas

Get Premium for Full Enterprise AI Atlas

For three years, the enterprise AI market has been organized around one assumption: the model is the center of gravity.

The assumption made sense. When ChatGPT launched in November 2022, the model was the breakthrough. It packed an extraordinary amount of intelligence into a chat window anyone could use. Work that had needed specialist teams, dedicated infrastructure and narrow machine-learning systems suddenly looked like a general-purpose platform. So the first questions were about capability. Which model reasoned better, held more context, wrote better code, or ran faster, cheaper and safer? Every company added a copilot. Every software product became “AI-powered”. Every enterprise started testing model providers, private deployments, retrieval systems, and internal assistants.

Get Premium for Full Enterprise AI Atlas

Deployment exposed the limit. A model could write an excellent answer and still be nearly useless in a real operation. It did not know which customer record was authoritative. It did not understand how a contract related to an entitlement. It could not tell whether an employee could approve a refund, which exceptions needed human review, or which regulation applied to a given transaction. The model had intelligence without institutional context.

Get Premium for Full Enterprise AI Atlas

Then agents arrived, and the ambition moved from answering questions to doing work. The system would no longer recommend an action. It would open the case, update the record, contact the customer, issue the refund, escalate the exception, and document what happened.

The adoption record shows how far that ambition still is from reality. Ali Ghodsi, who runs Databricks, describes asking enterprise audiences how many of them manage fleets of agents that coordinate with each other. Almost nobody raises a hand. Most are using a chatbot as a faster search engine, plus coding assistants. His diagnosis is blunt: the frontier models are already smart enough, but they lack the context that exists inside any organisation. He sells a platform built to capture that context, so the view is not disinterested. It still matches what most deployments show. For the typical enterprise, a smarter model is not the bottleneck.

That changed the nature of the problem. Once AI begins to act, the constraint is no longer intelligence. It is context, authority, process and control. This is the market now taking shape.

Read more

Enterprises will increasingly push more strategic workloads toward open-weight models, not simply because they can be cheaper, but because they offer something more important: modularity, control, deeper customization, and greater protection over proprietary context and intellectual property.

The strategic goal is to bring the model closer to the enterprise, rather than continuously sending the enterprise into someone else’s model. That means adapting models to internal knowledge, workflows, policies and domain-specific context, while building a deeper understanding of where the company’s real AI moat can emerge.

In theory, this is extremely attractive. In practice, running, customizing, securing and maintaining open models inside the enterprise core is far from free.

And this is where the economic argument is often misunderstood.

If the main reason you are moving to open models is simply to save money on tokens, you may be approaching the problem from the wrong direction. The stronger case for open models is control, modularity, strategic independence and the ability to progressively encode proprietary enterprise context into the AI stack.

Cost matters, but it should be evaluated against those strategic benefits, not in isolation.

This analysis breaks down that trade-off in full: where open models actually save money, where they become more expensive than expected, and where the real strategic value lies beyond the token price.

Open weights can give you more control, more flexibility and lower costs. But downloading a model, operating it and getting useful work from it are three different economic problems.

An open-weight model can cost nothing to download. That is a meaningful advantage. You receive the model’s learned parameters, with rights to use and modify them under its licence. But the download is not the service. Someone still has to supply the computing capacity, operate the software, monitor performance and handle failures. A free model does not automatically produce a low-cost system.

The opposite mistake is just as important: adding every engineering, evaluation and governance expense to self-hosting, then comparing that total with an API’s token bill alone. Buying access to a managed model does not eliminate the customer’s work on integration, access controls, output quality or oversight. It changes which responsibilities sit with the provider and which remain with the customer.

The useful question is not “Are open models cheaper?” It is “Which combination of model, deployment and supervision delivers the required outcome at the lowest acceptable cost and risk?”

That question has an established name: total cost of ownership.

The Full AI Open Ecosystem Atlas Here

To gain Full Access, subscribe below.

Get Access To The Full BE Platform

Read more

Bower did not found McKinsey, nor did he invent management consulting. But he became one of the central architects of the modern profession. Harvard Business School goes further, calling him the father of modern management consulting.

Bower believed consulting should operate as a profession rather than simply another commercial trade. Over the following decades, McKinsey institutionalized many of the practices that would come to define elite consulting: the firm above the individual, an independent obligation to tell clients what consultants believed to be true, recruitment of exceptional young graduates, apprenticeship, and eventually the up-or-out system.

Together, those choices helped institutionalize one of the most successful organizational models of the twentieth century.

The pyramid.

A small number of expensive partners sat at the top. Beneath them were layers of bright, relatively inexpensive juniors who researched, interviewed, modeled, synthesized, and built the slides. The client paid for the partner’s judgment, while much of the work was produced by the pyramid underneath.

For decades, that was the economic engine of consulting: leverage.

Then AI arrived.

For seventy years, the consulting business model ran on a quiet engine: leverage.

A small number of expensive partners sat atop a large pyramid of bright, inexpensive juniors who researched, modeled, interviewed, synthesized, and built the slides. The client paid for the partner’s judgment but received, for the most part, the pyramid’s labor.

The economics followed from that ratio: billing rates, utilization, up-or-out promotions, profit per partner. A firm became more valuable when each partner could sell more junior hours beneath them.

AI attacks precisely that layer.

Research and synthesis. First-draft analysis. Public-source benchmarking. The first model. The first slide. The base of the consulting pyramid was among the first categories of professional work to become legible to a machine.

The engine did not merely become more efficient.

It was removed.

Three consequences follow:

• The base became rentable. Leverage that once belonged to the firm can now be purchased by almost anyone for a subscription.

• The hour begins to lose its role as the unit of value. Once clients can see the approximate cost of producing an outcome, charging for the time between input and output becomes harder.

• The deliverable loses scarcity. For seventy years, the document was what the firm left behind. Now the document is cheap.

What the client increasingly wants instead is the capability.

And what the firm should have been accumulating all along is the residue of every engagement it ever ran: decisions, exceptions, cases, methods, and patterns that repeated across clients.

Instead, most of that knowledge left inside the deck.

The industry is therefore not disappearing. It is splitting between firms using AI to produce the old product faster and firms rebuilding what the product is.

This volume is for the second group.

Sell the outcome. Keep the residue. Leave the capability.

This is a new book in The Business Engineer’s Series.

Subscribe To Premium To Gain Access!

Check out the Library for Free!

For the last three years, I’ve been rebuilding the Business Engineer’s curriculum from the ground up. That curriculum has now become the foundation of a new discipline, with the entire series taking shape around it.

Subscribe To Premium To Gain Access!

Check out the Library for Free!

If you’re already a paid member, simply reply to this email, and we’ll send it your way.

Read more

Then DeepSeek changed the picture.

The DeepSeek moment showed that a certain class of open-weight model could compete on cost, capability, and deployability in ways that mattered to enterprises. Through 2026, that opening became harder to ignore. More capable open models arrived, serving infrastructure improved, hosting became easier, and enterprises gained more credible alternatives to a purely closed-model stack.

The picture today is therefore much more nuanced. Closed models still dominate important parts of the market, particularly at the frontier. But open weights are no longer a peripheral alternative. They are becoming part of the production architecture itself.

To understand what changed, we need to separate the different layers: capability, economics, infrastructure, control, and ultimately who captures the value.

In August 2026, open-weight models processed 56% of all tokens on Vercel’s AI Gateway, up from 7% in December 2025. OpenRouter reported a similar pattern, with open models accounting for roughly 60% of US-originating token consumption during the same period. Those percentages are not market share for the entire AI industry. Gateway data sees the workloads routed through those platforms, not every direct API contract, consumer subscription, private deployment, local model, or enterprise agreement.

The reverse is also true: commercial gateways miss some open-model usage precisely because open weights can run privately. The numbers are therefore neither a census nor a curiosity. They tell us something narrower and still important: within major production environments designed around model choice, open weights have moved from experimentation into real production infrastructure.

The more important change, however, is not that open models crossed 50% of token volume. It is what has happened around those models. A model can now be trained by one organization, served by another, routed through a third, embedded inside a fourth company’s product, and specialized by a fifth for an enterprise customer. Training, serving, distribution, customization, and application delivery no longer have to sit inside the same company.

That separation is what makes the current cycle different. For years, the trajectory of open AI depended heavily on whether one or two large organizations continued releasing competitive weights. If Meta slowed Llama, or another major lab changed strategy, the open frontier could lose momentum. Today, many more actors have independent reasons to keep it moving: inference companies want workloads, chip vendors want compute demand, gateways want model diversity, application developers want substitutable intelligence, enterprises want lower costs and greater control, and regional labs want strong bases they can adapt rather than retrain from scratch.

This is what I mean by open escape velocity. Not that open models have defeated closed models. Not that every dependency has disappeared. And not that the ecosystem is already economically self-sufficient. The stronger claim is that open AI is beginning to develop enough independent sources of demand, infrastructure, capital, and institutional support that its future depends less on the strategy of any single model company. But adoption and self-renewal are different things. The gateway data tells us open models are being used. It does not yet tell us whether the ecosystem can sustainably finance the next generation of them.

The Full AI Open Ecosystem Atlas Here

The Business Engineer’s Agenting platform is now available to Executive members. You can start exploring it here: businessengineering.ai/agent

Get Access To The Full BE Platform

You can preview the platform at Business Engineering AI.

For the last three years, I’ve been rebuilding the Business Engineer’s curriculum from the ground up. That curriculum has now become the foundation of a new discipline, with the entire series taking shape around it.

Subscribe To Premium To Gain Access!

If you’re already a paid member, simply reply to this email, and we’ll send it your way. You can check out the full library here.

Read more

As I wrote in my newsletter at the time, an important part of that story was the work behind InstructGPT and reinforcement learning from human feedback, or RLHF. ChatGPT was a sibling of InstructGPT, not simply the same model with a chat window attached. Its training combined supervised examples of conversations with reinforcement learning guided by human rankings of responses. These methods helped turn a model trained to continue text into an assistant better able to follow instructions and respond to what a person actually wanted.

To me, that was the defining product insight: capability became approachable. The model’s intelligence mattered, but so did the removal of friction between that intelligence and an ordinary user. You no longer needed to understand the machinery underneath to get something useful from it.

Yet the same human-centered design introduces a tension. An answer a person prefers is not necessarily a decision a business should execute. Human feedback can reward correctness, clarity, and useful caution, but poorly balanced feedback can also encourage agreement, reassurance, or confidence that the evidence does not justify. OpenAI’s account of its 2025 GPT‑4o sycophancy problem illustrates the risk: a change intended to improve the experience instead made the model excessively agreeable. That is a failure mode of the incentives, not the inevitable result of all human-feedback training.

This distinction becomes more important as we move from conversational assistance to agentic execution. In a chat, a person can read the response, question it, and decide what happens next. Now imagine an enterprise operating hundreds or thousands of agent runs in parallel. Those systems may be routing requests, updating records, preparing transactions, or coordinating work across applications. A plausible explanation is no longer sufficient evidence of success. The question is whether the intended result actually occurred, not whether the agent convincingly said that it did.

If every result still requires someone to interpret and approve it, the review requirement remains part of the cost of every result. Some decisions should retain that checkpoint. But for bounded, repetitive work, the opportunity is to make review selective: let the system proceed where the evidence supports delegation, and bring the person back where uncertainty or consequences demand judgment.

That does not mean RLHF is inherently incompatible with automation. It has also improved instruction-following and truthfulness, and human feedback has been used to train agents outside conversational settings. Human involvement in training is not the same thing as requiring a human to supervise every decision in production. Simply removing RLHF would not automatically produce a more reliable system.

The more interesting question is what happens when we stop treating the conversational assistant as the default design for every application of intelligence. What would we build if the primary objective were dependable decisions, rather than preferred responses? A system designed around that objective would need to demonstrate correct task completion, uncertainty tested against outcomes, appropriate abstention, and actions constrained by explicit authority. It would not necessarily need to narrate every decision to a person.

That is where today’s issue begins: not with the assumption that removing RLHF solves automation, but with an investigation into whether automation needs a different balance of training objectives, interfaces, and evidence. The assistant had to become useful to the person reading its answer. The next system has to become dependable enough for software to act on its output.

The first breakthrough made intelligence conversational. The next may make it reliably actionable.

Jev is now available inside The Business Engineer’s Agenting platform. Executive members can start exploring it here: businessengineering.ai/agent

Get Access To The Full BE Platform

You can preview the platform at Business Engineering AI.

You can also experience Jev’s speed firsthand by trying Genesis, our latest free AI tool: businessengineering.ai/tools/genesis

Get Access To The Full BE Platform

For the last three years, I’ve been rebuilding the Business Engineer’s curriculum from the ground up. That curriculum has now become the foundation of a new discipline, with the entire series taking shape around it.

Subscribe To Premium To Gain Access!

If you’re already a paid member, simply reply to this email, and we’ll send it your way.

Read more