Two Wall Street Journal stories from this past summer.
The first report said OpenAI fell behind Anthropic. Sam Altman staked growth on ChatGPT subscriptions. The bigger prize turned out to be coding tools sold to developers and the companies that employ them. Anthropic’s revenue growth and valuation both surpassed OpenAI’s.
The second article reported OpenAI is investing $150 million in a partner program and establishing a separate services company, DeployCo, which is seeded with $4 billion and led by TPG. OpenAI acquired two consulting firms. It now fields hundreds of forward-deployed engineers who work inside customer offices and build applications by hand. Anthropic, Microsoft and Amazon have announced comparable programs.
The frontier labs have not settled on a business model. They are testing whether to sell software, sell people, or both, in public, at a cost of billions.
The Returns Are Real and Highly Concentrated
McKinsey’s State of AI survey found that 39% of organizations attribute any EBIT impact to AI, and most of those report figures below 5%. Nearly two-thirds have not begun scaling across the enterprise. About 6% qualify as high performers.
BCG’s value-gap research put 5% of companies in its “future-built” category.
Accenture’s May 2026 study of executives at 2,000 companies found 7% with the data readiness to scale advanced AI.
Three research programs using unrelated methods converge on the same band. Roughly one company in fifteen is getting what the technology promised.
The counterevidence deserves weight. BCG’s AI Radar 2026 reports that nine in ten CEOs see initial value, and 64% of Accenture’s respondents say they are past pilots. Companies are achieving real gains in individual use cases. Very few see those gains reach the enterprise P&L. (The widely quoted MIT finding that 95% of GenAI pilots fail rests on a small sample and has drawn methodological criticism, so I leave it out of the argument.)
Gartner forecasts that over 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls.
In my 2026 predictions last December, I expected a trough of disillusionment this year because chaining non-deterministic agents compounds errors in ways that break automation at scale. That prediction has held up.
The underlying shift is real. GenAI moved automation from executing a defined “how” to pursuing a specified “what,” and agentic systems extend that toward goal attainment. I have argued this at length in my writings, Gen AI: A Generational Paradigm Shift and Rising Above AI Slop.
That is what makes the moment uncomfortable for enterprise leaders. Wait, and competitors compound advantages you cannot buy back later. Or commit to today’s stack, and the vendor reprices; the model you standardized on gets leapfrogged; the lab you partnered with decides it is a consultancy.
The forward-deployed engineer model sharpens the problem. You are invited to redesign core workflows around one lab’s models while that lab is still deciding what business it is in. That can be a good trade. Price the switching cost before you make the switch.
Amazon’s capital expenditures reached $53 billion in the second quarter, up 69% year over year. Amazon, Google, Meta and Microsoft spent $170 billion in the quarter, up 72%, and are expected to spend $1.5 trillion on data centers this year and next. Google reported negative free cash flow for the first time since going public in 2004. The signed-contract backlog at Amazon, Google and Microsoft is close to $1.7 trillion, much of it from OpenAI and Anthropic, prompting Moody’s to describe a circular system that can mask true demand.
Oracle’s June annual report disclosed $130 billion in debt and stated it could not guarantee it would manage that debt load. The New York Times Magazine’s profile of Larry Ellison cites MacroStrategy Partnership’s estimate that the AI bubble is four times the size of the 2008 housing bubble. That multiple reflects one firm’s model rather than consensus.
The more careful version of the same worry comes from Gita Gopinath, former IMF chief economist, now at Harvard. Writing in The Economist last October and walking through the arithmetic on NPR in December, she estimated the damage from a dot-com-scale correction at roughly $20 trillion in US household wealth and another $15 trillion for investors elsewhere. Her point is about exposure, not valuation. Global holdings of US equities are at record levels. US household consumption growth would drop by about 3.5%, and US growth would go to roughly zero. She also notes that the cushions available in 2000 are thinner now, given tariffs and limited fiscal room.
This buildout is financed with debt, and much of the demand for it comes from two companies that are not yet profitable.
Some investors argue that a bubble would be healthy, pointing to the fiber overbuilt in 2000 that later carried the social web. That fiber was mostly financed with equity, which can be written off quietly. Debt has covenants and payment dates, and GPUs depreciate on a shorter timeline than fiber.
My December call on compute needs a correction. I wrote that algorithmic efficiency breakthroughs might puncture the compute-capacity assumption. Sergey Brin made a similar point recently, but the mechanism in the headlines now is credit. I flagged the fragility but missed the trigger.
The WSJ reported in August that big employers are hiring again after 18 months of treating headcount as a last resort, with initial jobless claims at their lowest level since 1969. Many of those cuts were bets on capabilities that had not yet materialized. Some are now being unwound. I wrote about the reversal on my blog.
Asked whether AI means companies need more or fewer people, MIT labor economist Paul Osterman told the WSJ that no one knows.
He was answering a question about employment. The answer holds well beyond that. The impact of AI will be significant. Its shape is still unknown.
Build What Pays off in Every Branch
When the range of outcomes is this wide, stop asking which bet pays off best. Ask which investments pay off on whichever branch you land on. Data foundations clear that bar.
If agentic adoption compounds as promised, agents are bounded by the data they can reach, interpret, and be trusted with. Accenture found 72% of organizations lack trusted data of the right quality with standardized governance to support advanced AI, and more than 80% delay, limit or alter AI initiatives because of data-related risk.
If the market corrects and budgets tighten, a clean, well-governed data estate continues to deliver value through analytics, compliance, and search. None of that depends on the AI trade holding up.
If your vendor is acquired, repriced or repositioned, the data foundation moves with you.
There is a second reason, which I set out in the whitepaper Rising Above AI Slop. Autonomous systems can be constrained through input controls, behavioral constraints, and output and action controls. Input controls are the simplest to implement and the most effective because the enterprise still owns what goes in, making those controls deterministic. The rest amounts to probabilistic controls layered on probabilistic systems. Your data governance program is your input control layer.
- Inventory. Know where sensitive data lives, both in structured systems and in the documents, contracts, tickets, and messages that hold most institutional knowledge. Unstructured data is what agents now read.
- Classification. An inventory without accurate classification cannot inform policy. Programs underinvest here and rebuild it later at higher cost.
- Access controls. Agents inherit the permissions of the person who deploys them. Over-permissioned data becomes over-permissioned automation at machine speed.
- Governance and observability. Lineage, monitoring, auditability, and a defined escalation path. This cost is structural.
BCG’s 10-20-70 principle holds that 10% of AI value comes from algorithms, 20% from data and technology, and 70% from people, processes, and cultural change. If most of the value sits in process redesign, why lead with data? Because the 70% cannot be addressed first. Redesign a process around data you cannot trust, and decisions arrive faster but land wrong. Data readiness is the precondition for the process work, and it is the piece you can start on Monday without knowing which lab wins.
The technology will keep moving faster than anyone’s roadmap. Build the part that does not depend on where it lands.
