Diffusion speed and the driveshaft : the only question that may matter about AI in operations (and the world economy, maybe)

On May 4, 2026, OpenAI closed a $10 billion joint venture with TPG, Brookfield, Advent, and Bain Capital. Nineteen investors. A 17.5% guaranteed annual return over five years. Minutes later, Anthropic announced $1.5 billion with Blackstone, Goldman Sachs, and Hellman & Friedman.

It follows the Prometheus project reportedly raising more than $10 billion at a $30+ billion valuation to just do “AI for physical”. The bet of Jeff Bezos, the man who could do anything, is that the gains will be made by a few select winners.

$11.5 billion in a single day following $10 billion for a single company. To push existing models into totally new companies, or companies that cannot figure out how to use them on their own, apparently.

I’ve been thinking about what that number actually means for two weeks now, and I keep coming back to a paper written by a Stanford economist in 1990, a survey of 300 manufacturing professionals published in January 2026, and a factory I visited eighteen months ago where a planning manager told me, half-apologetically, that she overrides the AI’s scheduling recommendation most mornings because “it doesn’t know what it doesn’t know.” She was wrong about the AI. She was right about the organization around it.

78% of AI projects do zero gains

Redwood Software’s 2026 manufacturing outlook gives us a number that should anchor every conversation about AI in industry. Ninety-eight percent of manufacturers are exploring AI. Twenty percent are prepared to deploy it.

Seventy-eight points.

That gap is doing a lot of work, and most commentary I’ve read treats it as a technology adoption curve that will close naturally over time, the way smartphones went from novelty to ubiquity. I think that reading is wrong. The gap is structural, and there’s a real possibility it persists for a decade or more.

Fivetran surveyed 400 data professionals across the US, UK, EMEA, and Asia-Pacific for their Agentic AI Readiness Index, published in May 2026. Only 15% of organizations are ready for agentic AI in production. Sixty percent are investing millions. Gartner expects more than 40% of agentic AI projects to be abandoned by 2027. When HFS Research asked executives what’s blocking them, the top answer from a third of respondents was unprepared business processes. Not model quality. Not cost. Not talent. The way work is organized.

And the Redwood data makes it tangible: seven in ten manufacturers have automated 50% or less of their core operations. Seventy-eight percent have automated less than half of their critical data transfers. These are companies that bought the software, installed it, ran pilots, got promising results, and then watched the results die somewhere between the proof of concept and the daily production meeting.

I’ve seen this happen enough times to have a theory about why. But first, a detour through the 1890s.

The driveshaft in the factory floor, 1895

When electric motors first became commercially available in the 1880s, factory owners did the obvious thing. They pulled out the steam engine and dropped in a dynamo. Same building. Same driveshaft running the length of the floor. Same belts connecting each machine to the shaft.

Nothing happened. Productivity barely moved. For forty years.

Paul David, the Stanford economist, spent a career studying why. His 1990 paper “The Dynamo and the Computer” is nine pages long and contains, I think, the single most important idea for understanding what’s happening in manufacturing AI right now. Here’s the core of it.

A steam-powered factory was designed around the physics of mechanical power transmission. You had one enormous engine. Its rotational energy traveled through a central driveshaft. Belts connected machines to the shaft. Everything about the layout followed from that architecture: heavy machines close to the shaft, lighter ones farther away, multi-story buildings to keep everything within reach of the power source, small windows because the walls were load-bearing. The factory was a physical expression of the steam engine’s constraints.

Electricity eliminated every one of those constraints. You could put a small motor on each machine. You could arrange machines in the order of the actual workflow, not by their weight. You could build single-story buildings with skylights. The assembly line, the thing we associate with twentieth-century productivity, was impossible under steam. It required distributed electric power.

But nobody saw this for decades. Why would they? The people designing factories had spent their entire careers optimizing steam-era layouts. They were good at it. They understood driveshafts, belt tensions, load calculations. The electric motor looked like a better power source for the same system. So they used it that way. Bolted it on and kept everything else.

David’s phrase has stayed with me: implementation on a wide scale required building up a cadre of experienced factory architects and electrical engineers familiar with the new approach to manufacturing. A cadre that didn’t exist yet. A new generation had to come up, people who had never worked with driveshafts, who could look at a factory and see the workflow instead of the power transmission system. Only then did the floor get redesigned. Only then did productivity explode.

Forty years, from the availability of the technology to the organizational redesign that released its value.

The modern driveshaft

I think about this every time I sit in a board meeting where a portfolio company presents its “AI roadmap.”

The roadmap usually looks like this: pilot in Q1, scale in Q2, full deployment by year-end. The pilot works. The scaling stalls. The full deployment gets pushed to next year. The board hears about data quality problems, integration challenges, change management. All real. All symptoms.

The disease is the driveshaft.

What does a driveshaft look like in a 2026 manufacturing company? It’s the centralized information architecture. Data flows through human layers the way mechanical power flowed through the shaft: from a central source, through rigid channels, to endpoints arranged by organizational hierarchy rather than by the logic of the work. An AI system recommends a scheduling change. That recommendation has to travel through a planning manager, a production supervisor, and a shift lead before it reaches the floor. Each of those people was hired, trained, and incentivized to make decisions in a pre-AI world. Their job descriptions assume they are the decision layer. The AI bypasses none of them.

That planning manager I mentioned? She overrides the AI because her bonus is tied to on-time delivery, and the AI’s recommendation involves a short-term disruption to the schedule that should improve throughput over the following week. “Should” is not a word that plays well against quarterly metrics. So she clicks “override,” the system logs a human intervention, and the AI’s recommendation evaporates. Multiply this by a hundred decisions a day across a factory, and you have the 78-point gap in microcosm.

The technology works. The organization around it was designed for something else.

What $11.5 billion buys

The PE partnerships suddenly make sense through this lens.

OpenAI’s Deployment Company isn’t a sales channel. It’s a transformation vehicle. PE firms take board seats, commit $4 billion of their own capital, and get influence over how AI is deployed across their portfolio companies. Anthropic’s venture, backed by Blackstone and Goldman, will develop deployment “templates” — replicable playbooks for AI integration that can be copied across firms. Nicholas Lin, Anthropic’s head of product for financial services, told Axios the quiet part clearly: there’s a big gap between what AI can do today and the value the market is getting from it.

Fortune reported last November that 85% of PE buyers now factor AI-enabled capabilities into company valuations. If you don’t integrate AI, you get penalized at exit. The arithmetic is straightforward. Buy a company stuck on the old driveshaft. Force the redesign. Capture the productivity gain. Sell at a higher multiple.

PE firms are structurally suited for this. Their entire business model is forcing operating change on companies that can’t or won’t do it themselves: new management, restructured processes, revised incentives, hard deadlines. They hold for three to five years. They have done this before with lean manufacturing, ERP implementations, and offshoring. AI is the next lever of uncomfortable implementation.

But here’s what gives me pause. David’s electricity transition took forty years. PE holds for five, maybe seven. The deployment templates Anthropic is building assume that organizational redesign can be copied and pasted across companies. I’m skeptical. The 10/20/70 rule, cited by Microsoft in its March 2026 manufacturing report, says roughly 10% of successful AI deployment is algorithms, 20% is technology, and 70% is people and processes. That 70% is culture. Culture does not paste.

What I’ve seen work, in the few cases where it has worked, is not a template. It’s a specific combination of a CEO who personally understands what the AI does, a willingness to reorganize the decision layer (not just add AI to it), managerial courage to go through it all, technical acumen to take the right calls, and enough time for the new way of working to produce visible results before the skeptics kill it. And finding all of that in a PE-owned mid-market manufacturer is rare.

The $11.5 billion bet might pay off. But it’s a bet on compressing a generational transition into a fund cycle, and the base rate for that is not encouraging.

The size of the wager

Here’s the part that doesn’t get discussed enough, even in the financial press. The AI bet isn’t a sideshow in the economy. It IS the economy. Or rather, it’s become the part of the economy that’s growing, while much of the rest stands still.

In Q1 2026, AI-related capital expenditure accounted for roughly 75% of all US GDP growth. Strip it out and the US economy is basically flat. Pantheon Macroeconomics put it directly: without AI spending, US corporate capex would be negative. That sentence should alarm anyone who thinks the AI transition is a tech-sector story.

The numbers are hard to process. Amazon, Alphabet, Microsoft, and Meta are on track for somewhere between $630 billion and $700 billion in combined capex for 2026. Ferguson Wellman notes that figure rivals Sweden’s entire GDP. Morgan Stanley’s estimate for the top five hyperscalers (adding Oracle) is $805 billion, up from a previous estimate of $765 billion. Apollo Global Management calculates that hyperscaler capex alone now represents approximately 2% of US GDP. UBS projects global AI capex reaching $571 billion in 2026 and $1.3 trillion by 2030, a 25% compound annual growth rate over five years.

What makes these numbers different from, say, the dot-com capex boom is the concentration. A handful of companies account for nearly all of it. And they are financing it increasingly on debt. Introl estimates $108 billion in AI-related debt issuance in 2025, with a projected $1.5 trillion cumulative. Meta’s capex intensity has hit 54% of revenue. Microsoft is at 47%. Alphabet at 46%. Amazon’s free cash flow collapsed to $1.2 billion as a $59.3 billion surge in infrastructure spending consumed nearly everything. Google issued a 100-year bond. These are not normal numbers.

To be fair, there are signs the investment is producing returns. Google Cloud revenue grew 63% year over year in Q1 2026 to $20 billion. Microsoft’s AI business runs at $37 billion annualized, up 123%. But capex-to-revenue ratios remain at levels that cannot be sustained indefinitely, and the gap between what’s being spent and what’s being earned has to close eventually, in one direction or the other.

The US (and hence the West) economy is running on the premise that AI investment will generate sufficient productivity gains to justify itself. David Sacks, the White House AI czar, has been explicit about this: AI capex is a 2% tailwind to GDP growth. That’s not a prediction about eventual returns. That’s the stimulus effect of the spending itself, money moving through the system. If the driveshaft doesn’t come down — if the organizational redesign stalls and the productivity gains never materialize — the spending eventually slows. And when it slows, it doesn’t just remove a tailwind. Given how concentrated US growth has become in AI-related capex, it creates a headwind that could tip the economy into recession.

Moody’s currently puts the probability of a US recession at about 42%. Mark Zandi, their chief economist, said it plainly: “Nothing else can go wrong. We’re pretty much on the edge.” Bloomberg’s consensus sits at 30% for 2026. Kalshi prediction markets show 17.5% for 2026 but 41% for 2027. The pattern in those numbers is suggestive. Markets believe 2026 is fine because the spending is still accelerating. They’re less sure about what happens when the bill comes due.

The diffusion question and the recession question are the same question, asked at different scales. If AI gains diffuse broadly and fast, the investment pays off, productivity grows, and the economy lands softly. If gains stay concentrated and slow, the investment doesn’t pay off, capex contracts, and a recession becomes likely. The eight scenarios I’ll lay out next aren’t just abstract exercises in futurism. They’re descriptions of different macro-economic outcomes, each carrying a different recession probability.

What we can see from here

Where do we actually stand, mid-2026?

The capital is concentrating. European AI startups raised $9.2 billion in Q1 — over 50% of total venture funding for the first time ever — but deal volume fell 40% year over year. Seed deals down 44%. More money, far fewer companies. The four largest European rounds in Q1 were all AI. France and the UK captured $10.3 billion of the $17.6 billion total.

Adoption is sector-dependent. IDC projects 40% of manufacturers with scheduling systems will upgrade to AI-driven capabilities by end of 2026. Scheduling is a sweet spot: contained scope, rich data, clear ROI. Move into procurement, product design, or supplier management and adoption drops off fast.

VivaTech’s 2026 Top 100 report captures something I find genuinely interesting: the era of generic AI is ending, and Europe is winning through vertical specialization across 28 sectors. The report also notes a return to hardware and physical industry after a decade of pure-software dominance. That matters because the diffusion question is fundamentally about whether AI crosses the boundary from the digital economy into the physical one. So far, the crossing is patchy.

And the workforce side is real. That planning manager isn’t going to retrain herself out of a job. The ISG finding that 79% of organizations have deployed AI agents in some form but only 31% have one running in production tells you exactly where the friction lives. It’s at the point where the technology touches the human decision layer.

Eight futures

Predicting the path from here with any precision would be dishonest. But we can map the space of plausible outcomes.

Three variables. 
1.Who captures the gains: AI-native companies built from scratch, or legacy companies that successfully adapt? 
2.How fast the displacement happens: 20% of jobs restructured in five years, or ten years for the same effect? 
3.How evenly the gains spread: concentrated in a few sectors or distributed across the whole economy?

These axes form a cube. Eight corners. Some of them feel realistic. A couple feel like they belong in a different century. All of them are possible.

The Gilded Algorithm — AI-natives win, fast, in concentrated sectors.

A handful of industries restructure completely while the rest watches from the parking lot. GDP grows, but median wages don’t, and the growth is fragile because it depends on a narrow base. Tech hubs boom. Industrial regions stagnate. The Gilded Age with better lighting. This is the scenario European policymakers fear most, and it’s the one most likely to trigger aggressive regulation that slows adoption further — a vicious cycle. Macro outlook: moderate recession risk (30–40%) within five years, because the non-adopted sectors eventually drag down demand. The hyperscaler capex can sustain the economy for a while, but consumer spending erodes as wage growth stays flat outside the winning sectors. A tariff shock or energy crisis accelerates the split violently. Soft landing probability: low.

Creative Destruction — AI-natives win, fast, everywhere.

Schumpeter’s dream. New entrants replace incumbents in every sector simultaneously — agriculture, construction, healthcare, manufacturing, all at once. Displacement is high but temporary because new companies hire as fast as old ones shrink. I find this the least plausible corner. You can build an AI-native software company in eighteen months. An AI-native steel producer needs physical capital, regulatory clearance, and supply chain trust that take years to earn. The physical world imposes a speed limit. Macro outlook: paradoxically turbulent — short, sharp recession (20–30% probability) from the displacement shock, followed by rapid recovery as new companies scale. Think of it as a controlled demolition. The economy dips and rebounds. If you can survive the transition quarter, the other side looks excellent. No government has the political stomach for this.

The Slow Fracture —

AI-natives win, slowly, concentrated. Legal services, accounting, portions of finance, and software development hollow out over a decade. No single year feels dramatic. No quarter triggers a panic. But cumulative displacement in the affected sectors is severe, and workers have nowhere to go because the rest of the economy hasn’t changed. Think of what e-commerce did to retail, stretched over fifteen years, applied to white-collar work. Macro outlook: extended stagnation rather than outright recession. GDP grows at 1–1.5% for a decade. The hyperscaler capex boom deflates gradually as ROI plateaus in the adopted sectors without spreading to the rest. No crash. No boom. Just drift. Recession probability in any given year: 25–30%, but a cumulative multi-year malaise that feels worse than a clean recession because there’s nothing to recover from.

The Long Replacement —

AI-natives win, slowly, everywhere. Incumbents lose ground across the board, slowly enough that society keeps up. Workers retrain. Policy adapts. Communities adjust. In retrospect it would look like analog-to-digital media: painful for incumbents, manageable for everyone else. This requires patience from investors and tolerance from policymakers, two currencies currently in short supply. Macro outlook: the closest thing to a genuine soft landing (recession probability under 20%). AI capex transitions gradually from hyperscaler-driven to broad-based corporate spending as adoption diffuses. The economy doesn’t spike but doesn’t crack either. The risk is political — ten years of slow displacement generates populist backlash that disrupts the process before it completes.

The PE Bet —

Legacy wins, fast, concentrated. This is the future $11.5 billion was raised to build. PE firms force organizational redesign in their portfolio companies — healthcare services, business services, consumer products, pockets of industrial manufacturing. Those sectors see real productivity gains. Everything PE doesn’t touch (education, government, construction, agriculture) stays on the driveshaft. A dual economy forms. Impressive fund returns on one side. Growing inequality on the other. Macro outlook: the most unstable scenario. Short-term GDP boost (capex pays off in the adopted sectors), followed by a second-order recession when the non-adopted half of the economy can’t absorb the price and wage divergence. Recession probability: 35–45% by 2029–2030, precisely when the early PE funds start exiting and the structural imbalance becomes visible. The recession hits the wrong people — the ones who never got the technology in the first place.

The Golden Path - Legacy wins, fast, everywhere.

Incumbents across every sector reorganize around AI within five to seven years. The driveshaft comes down globally, compressed by software’s speed and PE’s discipline. This is what the 1920s factory redesign would look like at modern tempo. GDP accelerates broadly. Displacement is real but offset by productivity gains and wage growth. I want to believe in this one. History tells me the timeline is too aggressive. The 70% that’s people and culture does not compress as smoothly as the 30% that’s technology. Macro outlook: genuine soft landing, possibly a boom. Recession probability under 15%. AI capex pays off, the hyperscalers’ debt gets serviced by rising cloud revenue, productivity growth lifts wages, consumer spending stays healthy. This is the scenario where the $600–800 billion annual capex bet looks like genius rather than hubris. The Roaring Twenties analogy works here — and the worry is what comes after.

The Driveshaft Persists - Legacy adopts, slowly, concentrated.

Automotive, pharma, and aerospace modernize. The rest of manufacturing, plus services, construction, agriculture, government — nothing. The 78-point gap narrows to maybe fifty over a decade. Commentators in 2035 write confused articles asking why AI shows up everywhere except in the productivity statistics. Paul David’s scenario applied to the present. Historically, this is what happens to general-purpose technologies in their first two decades. Macro outlook: the highest recession risk of any scenario (45–55% by 2028–2029), because the economy has poured trillions into AI infrastructure without broad productivity returns. The hyperscaler capex-to-revenue ratios become unsustainable. Capex cuts remove the tailwind that was carrying 75% of GDP growth. The recession, when it comes, is blamed on AI “failing” — when what actually failed was the organizational redesign. A few companies endure the pain and go through the other side reaping most of the rewards. Think amazon-like for commerce.

The Long Grind — Legacy adopts, slowly, everywhere.

Everyone gets there eventually. Nobody fast enough to gain a competitive edge. AI becomes like ERP in the 2000s: a cost of doing business, not a source of differentiation. The world gets marginally better at everything, dramatically better at nothing. Macro outlook: a long soft landing that never quite feels like one. GDP growth stays at 1.5–2%, recession probability in any given year sits at 20–25%, and the economy muddles through. The hyperscaler capex normalizes over time as AI becomes infrastructure rather than speculation. No crash. No boom. The most boring outcome, and arguably not the worst one.

Where the buck is headed to

If I had to place us on the cube today, I’d put the pin between The Driveshaft Persists and The PE Bet. Legacy companies are beginning to move. The pace is slow. The gains are concentrating where capital and pressure are highest. Recession probability, on my read of the current trajectory: 35–45% by 2029, driven mostly by the risk that AI capex contracts before broad productivity gains materialize.

The $11.5 billion is an attempt to drag us toward The Golden Path. I hope it works. I’m not sure it can. Not because the technology is lacking, but because the 70% that’s people and process and culture requires something that money can’t buy and templates can’t replicate: a willingness to look at your factory floor, your information architecture, your meeting structure, your incentive system, and admit that it was designed for a different world.

The stakes are no longer abstract. Somewhere between $600 billion and $800 billion is being poured into AI infrastructure this year alone, and that spending is carrying the economy. If it pays off, we get the best outcome in a generation. If it doesn’t, we get a recession triggered not by a financial crisis or a pandemic but by an organizational failure — by the driveshaft refusing to come down.

The factories that redesigned around distributed electric power in the 1920s didn’t just buy new motors. They hired architects who had never thought in terms of driveshafts. They built new buildings. They promoted a different kind of manager. The word “efficiency” changed meaning.

Something equivalent is required now. And the honest answer is that we don’t yet know how long it takes.