The Gap between AI Capability and AI Absorption
The defining business opportunity of the next decade will be getting AI inside the businesses that weren’t built for it.
A few weeks ago I was at Stripe Sessions, listening to various versions of the future from the stage and the lunch tables. Most of my highlights took place in the hallways, but there was one stage moment that changed my perspective about what’s happening in the economy right now.
Patrick Collison put up a chart during his opening keynote: new business creation on Stripe, consistent and linear through most of 2025, then vertical in early 2026. He called it parabolic.
The capabilities of artificial intelligence models are moving at frontier pace. The absorption of those capabilities, and the financial gains that will accompany them, is slooow.
The companies holding the microphone at Sessions are not the broader economy. They are the leading edge. The Federal Reserve’s April synthesis puts US firm-level AI adoption at 18% by year-end 2025. MIT’s NANDA initiative reviewed 150 executive interviews, 350 employee surveys, and 300 public deployments, and found that 95% of enterprise GenAI pilots produced no measurable P&L impact. Somewhere between $30 and $40 billion has been spent and almost none of it has produced operating leverage. Wild!
The gap between capability and absorption is massive. Who’s going to usher in actual implementation? Whoever they are, they stand to make a lot of money – and by make, I mean create.
We saw this pattern with electrification
Pearl Street Station went live on September 4, 1882. Six 100-kilowatt dynamos in Lower Manhattan lit 400 lamps for 85 customers across a quarter-square-mile of the financial district. The promise of electrification had come to fruition but the ripples didn’t spread out at the speed of light (ba-dum-tss!). Forty years later, factories still hadn’t gotten much productivity out of it.
The reason was architectural. Pre-electric factories ran on line-shaft drive, so one steam engine sat in the basement, overhead shafts ran the length of the building, and belts dropped down to each machine. When electricity arrived, they pulled out the steam engine, replaced it with a giant electric motor, and left everything else in place. The result was almost no productivity gain.
The gains only came when factories rebuilt around unit drive. Every machine got its own motor, the overhead shafts came out, and the floor plan got redesigned around the new substrate. That retrofit didn’t happen until the 1920s.
The buildings just weren’t built for the technology that arrived in them.
Retrofit or renovate?
I’m gonna argue for renovation. Most firms are treating AI as something added on to an existing structure but that mentality is what generated the NANDA failure rate. The model may perform as advertised inside a pilot but then loses its edge when it has to coordinate with the rest of the business, which was never designed to receive what the model is producing. See what I mean?
Replatforming is the work of renovation, and renovation is six jobs running in parallel: a new data architecture, a new ontology, a new process design, a new set of decision rights, a new org structure, and a new equation for what a single employee actually costs once compute lives on the headcount line. You map what’s existing, redesign it, do the change management, and only then do you plug in the agents.
The NANDA study also separated success rates by who did the work. External partners succeeded at roughly twice the rate of internal builds. Renovation sits so far outside the muscle memory of an operating firm that the ones doing it alone produce most of the failures.
Why the frontier can’t bridge its own gap
Sam Altman appeared at Sessions in conversation with John Collison and named OpenAI’s role plainly. He wants OpenAI to be infrastructure, a utility that other people build on top of.
He has tried to scale OpenAI into the absorption layer anyway. The Forward Deployed Engineering organization, modeled on Palantir’s FDE, went from 2 engineers to roughly 40 in under a year. The instinct is on but the math still doesn’t work. A frontier-firm payroll can’t cover the surface area of every 200-person business in the country that needs the work done.
The skill set required to walk into one of those companies and make AI live inside it is rare: deep AI fluency, operator experience, change-management instinct, and the credibility to get past the company’s immune system. The people who have all four sit at the partner level of operating-experienced firms. I’m one of them. None of us would take a W-2 job at a frontier salary band, and the frontier cannot stand up a partner model fast enough to cover the absorption surface area of the global economy.
The frontier needs a partner category that doesn’t have a name yet.
Whoever solves pace owns the era
Two timelines are running in parallel. The absorption work itself will take decades. Electrification took forty years. Cloud is twenty years in and still half-deployed. The strategic window for category leadership is much shorter. The names attached to the cloud era (Andreessen, Bezos, Benioff) got attached during a window that lasted maybe five years.
The supply side of pace is going vertical. The humans doing the renovation work have AI doing the time-consuming parts alongside them, and the work compresses every quarter.
The demand side does not compress. Boards still meet quarterly. Procurement cycles still take six months. Cultural change still takes years. Conway’s Law still applies: any AI integration that respects the existing org structure will reproduce that org’s existing limitations, which means real rewiring requires reshaping the organization before the technology can do anything inside it. Substantial lift, but a lasting and extremely profitable one.
The binding constraint is the pace at which operating businesses can absorb structural change without breaking themselves.
The Collisons’ chart has a vertical axis labeled capability. The chart nobody has drawn yet has a vertical axis labeled absorption. It’s flat right now but it’ll become a hockey stick of its own.
The translator firms that exist in five years are being built right now, mostly in private, by operators who already see the gap.