Title Inflation, OpenAI's Trillion-Dollar Math Problem, and AI Utilization: Visualized.

There are a number of levers companies pull to sweeten a role — benefits package, vacation days, office perks. Two of the biggest, pay and title, are usually correlated. You get a better title and the promotion comes with it. In corporate development, the talent pool is concentrated and firms compete hard to attract and retain top performers, and they sometimes increase one without the other.

The traditional lever is salary. The firms willing to pay the most tend to attract the best candidates — but in a tightly competitive, close-knit world like corp dev, an unchecked salary war turns into an arms race. In some cases prestige — or the perceived future value of a senior-sounding title on a resume — is an even greater motivator. Sacrificing some money now can result in a more advanced career down the line.

When considering two roles at similarly viable firms, applicants are more likely to take whichever role sounds more senior. Over the past five to ten years, we've seen firms effectively "marking up" role titles as a way to attract talent — and perhaps equally importantly, retain existing talent. After maybe only six months, 'associate' becomes 'senior associate', 'corp dev manager' becomes 'assistant director', and so on and so forth, until you have relatively junior level employees occupying more senior positions, but not always with the corresponding pay increases.

Long-term, this can have a negative effect, both for employees and employers. Last week in the 2026 Salary Benchmark Report we discussed ways that professionals navigating available roles can sift through some of the uncertainty. Reporting line (who does the applicant report to), team size (does the role manage people?), deal size and volume, and integration scope are all important factors to know when applying to a new job.

Behind each growing GenAI company lies a problem — a power problem. Specifically, a power constraint. You may have heard about the immense energy requirements involved in the production and distribution of high-performing AI. To underwrite the massive user growth these companies are jockeying for, they need to be able to generate reliable power on their own terms. Alongside the customer acquisition strategy, these companies are in the middle of one of the largest infrastructure buildouts in corporate history.

What we experience as a digital product only involves a very tangible and very serious physical commitment from the companies providing it.

A "data center" in the AI context is a purpose-built facility designed around three constraints: power, cooling, and chip density.

Power comes first, and modern AI clusters are measured in gigawatts (GW) — not megawatts. A one-GW facility consumes roughly the same amount of electricity as 750,000 U.S. homes. The Oracle-OpenAI Stargate buildout is a large-scale AI data center initiative announced early last year that requires 4.5 GW of capacity — more than double the output of the Hoover Dam.

In addition to the power piece, firms need to invest massively into cooling. At full utilization, the hardware involved generates enormous heat, and firms are moving towards liquid cooling to increase power density. This typically requires access to significant freshwater, which raises environmental concerns and limits the locations that can be chosen for builds.

Finally, chips are required in huge numbers. A single Nvidia rack can cost as much as $3 million, and a 1 GW facility can hold tens of thousands of these systems.

Forecasting demand accurately is incredibly important, because these buildouts require significant capex from day one, and the contracts that finance them are not the kind you can walk away from.

At this scale, AI companies can't just buy compute as they need it. These are long-term, take-or-pay contracts — the same type used in oil and gas pipelines, LNG terminals, and nuclear site buildouts.

The core principle is an AI company commits to buy a fixed amount of compute capacity over a period of five to ten years, at a certain price per unit. The supplier (Oracle, CoreWeave, AWS), uses that contract as the financing backbone to build the facility. In order to determine how much compute they will need, AI companies need to forecast user demand out multiple years into the future — if they make mistakes in the forecast, they are liable to run into cashflow issues, and everyone along the chain of investment suffers.

This is why markets react so significantly to earnings reports of AI companies, and why so many public companies are directly impacted by the results. OpenAI's vendor stack illustrates this pattern at scale. Across seven publicly announced agreements, the company has committed roughly $1.15 trillion in compute and hardware spending between 2025 and 2035. These contracts, to providers like Broadcom, Oracle, Microsoft, AMD, and Nvidia, ramp up over time, but in general, OpenAI has to be fairly certain they will far exceed $1.15 trillion in revenue over that same period in order to justify the investment.

On April 27, that assumption took a real hit, with the Wall Street Journal reporting OpenAI fell short of both its internal revenue and user growth targets for early 2026. OpenAI publicly disputed the reporting, but the market reacted regardless, with Oracle dropping 7% in premarket.

It might be assumed that hyperscalers are paying for AI capex out of their own cash flow, but that's only partly true. While players like Meta and Google are funding their buildouts largely from cash, suppliers like Oracle and CoreWeave are heavily levered using unconventional debt structures.

The most popular structure today is the special purpose vehicle (SPV). It's a separate legal entity that owns the building, land, and power infrastructure for the data center, and the tech company then leases the assets from the SPV. Private credit and banks are financing the contracted revenues by underwriting these leases, and the result is tens of billions in debt sitting on the books of the SPV — not the tech company.

The picture this paints is more nuanced than "AI bubble" or "AI is fine". Capex funded from operating cash flow is structurally different from dot-com-era leverage, and capex-to-free-cash-flow ratios are still much more modest than they were in the dot-com peak. Real AI demand exists, and growth, in general, is impressive.

But the supplier ecosystem isn't without risk. Oracle, CoreWeave, and the regional developers are funding their buildouts through trillions of dollars of layered debt, held increasingly by pension funds, private credit, and insurance company portfolios. If AI revenue growth slows below what is projected, the contracts that anchor those structures don't disappear, and the impact is felt throughout the economy.

Most adoption frameworks assume that users overestimate their proficiency, believing they are more capable than they really are. This is called the Dunning-Kruger effect, and is a well-documented cognitive bias. With AI, we are seeing the opposite — routine users consistently underestimating their proficiency due to the sheer pace of frontier movement.

Let's take a minute to unpack some of the numbers above. 2.3 Billion people (28% of the world population) have intentionally logged on to an AI platform and tried it out. There's a large drop-off after that, resulting in only about 17% of the planet using AI on a monthly basis. Weekly users become even more concentrated — we're down to one billion people at this point.

The next bracket is what really interests me. Despite all the headlines, the noise, the pressure, and the feeling of "falling behind the curve", only about 1% of the population is on a paid AI subscription. Only 25 Million people are using AI for coding and developing their own tools. If you have a $20 ChatGPT or Claude subscription, you sit in the top 1% of AI adopters globally.

To frame this another way, roughly 40% of people who have tried GenAI have not built it into their monthly routine. Companies sit in a similar position. Most have run pilots, signed licenses, and advised staff on the benefits of using AI, but they aren't seeing the expected usage spike.

Corp dev teams are running ahead of this curve. If you're using AI for lead gen, CIM screening, drafting diligence memos, or outreach organization, you're among the 25–80 Million people making up the smallest cohorts on this graph. If you're worrying that you've already missed the bus, don't. By taking that next step you can move past 99% of the population in terms of adoption.