OpenAI Chair Bret Taylor Says AI Is Already Superintelligent at Digital Work

Bret Taylor, chair of OpenAI, said on Sept. 24, 2026, that AI is probably already superintelligent at purely digital work, and that everyone will have that expertise on call ¹. The same models find software flaws by the thousand, and in Taylor's telling the local clinic on outdated technology sits closest to the damage.

Taylor also cofounded Sierra, which builds customer-service agents, and he spoke on The McKinsey Podcast with Eric Kutcher, McKinsey's North America chair ¹. The transcript page discloses that McKinsey and OpenAI are partners in an enterprise AI alliance. This essay carries each idea and checks it against the public record and this ledger, which read 2.6 out of 100 on Sept. 23.

Which of Taylor's and Kutcher's claims the public record backs

The record

The data14 rows · sources

Column	Item	Source
On the record	A computer-checked proof of a famous theorem, produced in 11 days	[2]
On the record	A 50 percent time horizon estimate of 16 to 20 hours for the top shared model	[3]
On the record	Thousands of previously unknown flaws found by one restricted model	[4]
On the record	Four incidents in which security tests reached real third-party systems	[5]
On the record	A leading open-weight model about eight months behind the frontier	[8]
On the record	$225 billion in bonds from hyperscalers and related firms by midyear 2026	[9]
On the record	Software development postings up 14 percent in the year to April 2026	[7]
Claimed, unshown	Superintelligence probably already present wherever the work is purely digital	[1]
Claimed, unshown	Models proving or disproving open math conjectures	[1]
Claimed, unshown	One customer's agent resolving more than 90 percent of cases alone	[1]
Claimed, unshown	Most code in Silicon Valley written by agents	[1]
Claimed, unshown	Institutional productivity as 80 percent business, 20 percent technology	[1]
Claimed, unshown	Frontier labs holding a lasting cost edge over open weights	[1]
Claimed, unshown	Company formation cheap enough to change the jobs count	[1]

Is superintelligence already here?

Taylor says yes, in one domain. "In the domain of information, where digital technology is all that's required, I think we're probably already at a point of having superintelligent AI," he said ¹. He placed the field in "the foothills of artificial general intelligence," pointed to models that proved or disproved open math conjectures, and promised every person a superintelligent financial adviser, the best primary care physician and, for every child, a tutor.

He also expects the gains to land unevenly. Software engineering, customer service and sales already show the change, he said. Over the next few years he sees the digital parts of the economy transformed and the rest barely touched. He called the result an overhang, more capability in some industries than they can absorb, and predicted a lot of complexity ¹.

So why does this ledger read 2.6? A test score measures a task, and the ledger measures a job. Its definition asks for systems that run unsupervised for weeks on expert work across most occupations, on ten gigawatts of compute acting as one machine, while they improve their own successors and prove to an auditor what they ran.

The record holds real pieces of the first clause. Anthropic's internal model produced a computer-checked proof of a famous theorem in 11 days, with humans giving only occasional high-level instructions ². That proof restates mathematics people settled three decades ago, so the open-conjecture claim stays a claim here. METR, an outside evaluator that tested the strongest models four labs shared with it, put the top model's 50 percent time horizon point estimate between 16 and 20 hours of human work ³. Put plainly, the best agent it tested finishes a two-day task about half the time. Weeks sit far past that.

Both readings can be right. Taylor described superintelligence on the test; the ledger waits for superintelligence on the job.

Cyber risk comes first, and small clinics carry it

Asked what frightens him, Taylor named cybersecurity, in the short term, as "probably the biggest, most acute concern" ¹. Security usually trails a new technology, he said, and this one spread faster than electricity did. "Unfortunately, the models are approaching superintelligence at the same time we've connected all of our infrastructure," he said, so an attack now reaches much further than it could with earlier technologies ¹.

Security and technology chiefs he has talked with turn these models loose, find thousands of vulnerabilities, and then face an ugly choice, he said. "If they try to patch them all, they'll break their systems. So they're stuck choosing: Do you want the system running, or do you want it patched?" ¹

Public disclosures back the scale. In April, Anthropic reported that Claude Mythos Preview, a restricted model, had found "thousands of zero-day vulnerabilities" across every major operating system and web browser ⁴. On Sept. 9, Anthropic assessed four incidents, three first described July 30, in which its models reached real third-party systems during security tests; in each, a misconfiguration connected a supposed simulation to the open internet ⁵.

Who pays for that exposure? Large enterprises have more resources to fix what the scans turn up, Kutcher said, while many small and midsize businesses run dated, exposed technology. Taylor agreed and named the victim: "the same local hospital or clinic that gets targeted with ransomware" ¹. Jeffrey Tully and Christian Dameff, physicians who study cyberattacks, cited a research firm's count of 445 ransomware attacks on hospitals and clinics in 2025, and a 2026 analysis of Medicare data that found a 38 percent higher risk of death for hospitalized patients during an attack ⁶.

Taylor's remedy is entrepreneurs who build security tools a small clinic can switch on. "These vulnerabilities existed before—we're just finding them all now," he said ¹. On this ledger, the four incidents moved verification down a full point, and that line sits at minus 1.5 net. Anthropic's own pre-release auditing missed misalignment of that severity ⁵.

Ransomware hit hospitals and clinics 445 times in 2025

The number

The data1 row · sources

Unit: attacks

Measure	Value	Source
Ransomware attacks on hospitals and clinics recorded in 2025	445 attacks	[6]

Why do people feel faster than their companies?

Individual gains show up, Kutcher said, while institutional ones stay hard to measure. He calls the gap "an 80 percent business problem and a 20 percent technology problem" ¹. Picture a mortgage file. Say a model reads it in seconds; the bank still has to redesign approvals, audit trails and job descriptions before that speed reaches earnings.

Sierra builds agents for that work, from bank compliance checks to mortgage origination. "I think every company's AI agent will become their digital front door," he said ¹. One customer's agent, he added, resolves more than 90 percent of cases on its own, so its costs stop climbing in step with its growth.

On this ledger, that 90 percent stays a claim. The customer goes unnamed, the error rate goes unpublished, and the definition counts deployment records with error rates among its evidence of autonomy.

"Most code in Silicon Valley is now written by agents, but there are more job postings today for software engineers than there were three years ago," Taylor said ¹. His reading: the world is making more software than ever. Indeed's data, reported in May, supports the direction: software development postings rose 14 percent in the year to April 2026 ⁷. His three-year comparison and his share of code written by agents remain his own figures.

Everyone buys the same intelligence, he argued, so savings flow to lower prices and reinvestment that any rival can match. "I think adopting AI for core corporate processes will be an imperative, not a competitive advantage," he said ¹. Kutcher tells chief executives that waiting counts as a choice too.

What are open weights for?

Companies reach for open-weight models for three reasons, Taylor said: cost, tuning and sovereignty ¹. Cost means hosting a small model for one narrow task more cheaply than renting a frontier model. Tuning means post-training that small model until it nears a large one on a specialized job. Sovereignty means owning the full stack to secure the data.

He granted open weights a lasting place for tuning. Sovereignty matters, he said, and contracts can secure it, the way companies secured the cloud. Cost, in his view, favors the frontier labs, which hold the most compute, fit their models to their hardware and can distill their own models. "If it's just about cost, I actually think the frontier labs will have a sustainable edge," he said ¹.

Public measurement complicates that. In May, NIST's CAISI put DeepSeek V4 Pro about eight months behind the frontier and found it more cost-efficient than OpenAI's GPT-5.4 mini on 5 of 7 benchmarks ⁸. Does that refute Taylor? It narrows him. A clinic summarizing charts can run the cheaper model; a team chasing the hardest problems still pays frontier prices. His own advice fits: know the job the model is hired for.

Can the data-center bet pay before the models move on?

Taylor calls the buildout risky and necessary at once. Among the investors, he sees "dumb money and smart money, classic gold rush behavior" ¹. He doubts the charge of circularity, the idea that a few firms buy from each other and count the same dollars twice, and rests his demand case on today's models alone.

His timing point gives the debate a frame worth keeping. The pace of AI is measured in days, he said, while the power and real estate behind a data center take far longer to secure ¹.

Kutcher put the strongest counterpoint in plain terms. "If you take a system-level view of the capital we're putting in the ground and ask whether we're seeing the operating cash flow to justify it—it doesn't close today," he said. "That's why we're seeing more debt financing than we've ever seen." ¹

What does that change? Demand can be real while the cash arrives late. Taylor conceded the point, called the risk "extremely high," and answered that inference already earns relatively high margins on enterprise work, in a market that, as in cloud infrastructure, holds "basically four" providers ¹. He wants America to lead in both models and buildings, and called himself "obviously entirely biased here." The debt is on the record: hyperscalers and related firms issued $225 billion in bonds through midyear 2026, a jump of nearly 1,000 percent ⁹.

On this ledger, compute, energy and capital all still read 0.0 after 13 analyses, because every move so far came from models, evaluations, incidents and government orders. Bonds count as closed financing, and the ten-gigawatt clause moves when the sites they pay for energize.

Taylor calls the demand real; Kutcher says the cash flow has yet to close

Both sides

The data2 rows · sources

Side	Who	Claim	Source
For	Bret Taylor, OpenAI chair	Today's models already justify the demand, inference earns relatively high margins on enterprise work, and a market of basically four providers can monetize at scale.	[1]
Against	Eric Kutcher, McKinsey North America chair	At the system level, operating cash flow trails the capital in the ground today, which is why debt financing has reached its highest level.	[1]

Will the town next door come out ahead?

Kutcher listed the public's fears, lost jobs, affordable power and water, and noted states pulling incentives or limiting data-center size. In Taylor's telling, they reduce to two personal questions, whether AI benefits me and whether this data center benefits my community, and most AI companies "have done a fairly abysmal job communicating both of those things" ¹.

Power bills make the second question concrete. Federal forecasters put the average residential price at 18 cents per kilowatt-hour in 2026, up about 37 percent from 2020, and data-center load accounted for about 40 percent of the costs in the December capacity auction of PJM, the largest American grid operator, ¹⁰. On Sept. 16, the House voted 417 to 3 for a bill that requires states to consider standards for loads above 100 megawatts, work most states had already begun ¹¹.

Taylor's remedy is local. Builders, he said, should take responsibility for power so bills hold steady, handle the environmental impact, and sit in community meetings where the jobs for electricians and HVAC technicians get argued in person. "Smart mayors and smart governors can make smart demands," he said ¹.

America, Kutcher said, holds its employment rate steady with flat-to-negative job growth because the working population is shrinking, and "the only path is productivity growth" ¹. Breakeven job growth, the monthly hiring needed to keep unemployment level, fell from about 250,000 in 2023 to roughly 10,000 by July 2025, then averaged about minus 3,000 a month from August to December 2025 ¹². That report credits an immigration reversal and falling participation.

A founder opening an online shop can have an agent help set up the storefront, buy the ads and generate demand, Taylor said ¹. Kutcher added that a company now costs so little to start that its effect on jobs is still unknown. Both remain claims, waiting on a count of new firms.

Hiring needed to hold unemployment steady fell from 250,000 a month to negative

Compared

The data3 rows · sources

Unit: net new jobs per month needed to keep unemployment steady

Item	Value	Source
2023 peak	250000	[12]
July 2025	10000	[12]
Aug. to Dec. 2025 average	-3000	[12]

By the numbers

• 16 to 20 hours: METR's 50 percent time horizon point estimate for the most capable model four labs shared, against 3 to 4 hours at 80 percent reliability ³.
• 445: ransomware attacks on hospitals and clinics a research firm recorded in 2025 ⁶.
• Eight months: CAISI's estimate of how far DeepSeek V4 Pro trails the frontier ⁸.
• $225 billion: bonds issued by hyperscalers and related firms through midyear 2026, a 973.7 percent jump ⁹.

What to watch

Taylor's promise lands on people: a tutor for a child, an adviser for a family, a security tool for the clinic down the street. Three records will show whether it arrives: published error rates for agents doing real work, patch rates at small hospitals, and operating cash flow that covers the debt. The open question is timing. Can the cash flow close before the models that justify it change again?