The AI boom's biggest risk is not whether demand for artificial intelligence is real. It is whether the companies buying compute can pay the bills when they come due, according to a new analysis published on Gradient Flow by Ben Lorica.
The article, titled "The AI Boom's Hidden Payment Schedule," argues that the financing structure behind frontier AI labs is more fragile than the usual bubble debate suggests. Lorica posted the piece on September 1, 2026, and it directly references a recent article, "Why the AI Boom Is Hitting a Reset Wall," as a starting point for the deeper look at contract mechanics.
The Teaser Period and the Delayed Bill
Take-or-pay contracts are the backbone of today's compute deals. Frontier labs commit to years of compute capacity, but payments only start when the capacity is actually delivered. That delivery takes time. Building a gigawatt-scale data center can take 24 to 36 months.
Contracts signed during the 2025 and 2026 boom have what Lorica calls a "teaser period." Cash costs mostly arrive in 2027 and 2028. Once capacity goes live, the bill does not fall if utilization or revenue is below plan. The commitment is fixed, regardless of how the business performs.
This timing creates a gap. Providers can report huge contracted backlogs while buyers have not yet incurred the full cost of capacity under construction. Booked compute and billed compute are different numbers, and the distinction matters more as the delivery dates approach.
OpenAI's Compute Bill vs. Revenue
The analysis applies this framework to OpenAI, the frontier AI lab at the center of the boom. OpenAI's committed compute alone rises above 200% of revenue at the 2027 peak. That means the company's compute obligations would be more than double its revenue in that year.
To get compute spending down to 100% of revenue, revenue would need to compound at roughly 217% annually from the 2025 base. That is an extraordinary growth rate, and Lorica does not assume it will happen. Instead, he notes that the earlier cloud era had a different structure. Enterprises paid cloud commitments out of existing operating business. Today's biggest compute commitments depend on future revenue and future fundraising.
That dependence is the core vulnerability. The question is not whether AI demand is real, but whether the financing structure can handle the moment when compute commitments turn into bills. Contracts put a date on the question.
The Funding Gap
Lorica's analysis estimates about $375 billion in uncovered compute costs from 2026 through 2030. That figure covers only compute. Once the rest of the business is included, the analysis implies roughly $400 billion to $500 billion of external funding needed.
Without new capital, cash is projected to turn negative in the first quarter of 2027. That timeline is not distant. It is less than six months away from the publication date of the analysis.
Stay ahead of the AI curve
The most important updates, news, and content — delivered weekly.
No spam. Unsubscribe anytime.
The scale of the gap is striking. The $375 billion figure spans five years, and the $400 billion to $500 billion range adds operating costs on top. These are not small shortfalls. They are amounts that would require repeated, massive fundraising rounds or a fundamental change in how compute is financed.
Demand Does Not Have to Collapse
The important nuance, Lorica writes, is that demand does not have to collapse. It only has to grow more slowly than assumed when the contracts were signed. That slower growth would be enough to create stress.
Stress would probably not be a sudden default. Instead, Lorica lays out three options for a take-or-pay customer facing a shortfall: raise more capital, renegotiate, or sublease capacity. Each option has its own consequences.
Renegotiation may be the one to watch most closely. A volume deferral can be described as "capacity rephasing." A lower rate can look like a new partnership arrangement. These are not dramatic events. They are quiet amendments.
Once a major contract is amended, investors can no longer assume every dollar of contracted backlog will arrive exactly as written. The early signal may not be a headline; it may be a change in terms.
What to Watch in the Coming Quarters
The analysis suggests that the first signs of trouble will not be loud. They will appear in contract language, in announcements about capacity timing, or in partnership structures that shift payment terms.
The 2027 and 2028 period is when the teaser period ends and the real bills arrive. That is also when OpenAI's compute costs peak above 200% of revenue, and when cash turns negative without new capital. The 2026 through 2030 window contains the $375 billion in uncovered costs.
Lorica's piece is a commentary, not a prediction. It does not say the boom will end. It says the financing structure has a date with reality, and that date is approaching. Related content on the same site includes "The AI Data Center Backlash Has a Blind Spot" and "The Bear Case for AI Data Centers," both of which examine different angles on the same infrastructure buildout.
The article also includes share buttons for X, Facebook, LinkedIn, Email, and a "Like" button, suggesting it is intended for broad distribution among investors and industry observers.
The core message is simple. The AI boom is not just about technology. It is about who pays, when they pay, and what happens if they cannot.