What is the metric for the $278 billion cash burn reported by the media?

This is the prediction metric disclosed by media on September 20 from OpenAI's perspective: a cumulative cash burn of $278 billion by 2030. A prediction is not a fact, but it shows that even the party involved is planning based on a cash gap in the hundreds of billions. Another anchor point for the same period is Oracle's single-quarter AI infrastructure investment of $28.5 billion, where management still feels it's not enough; this is the capital expenditure metric disclosed by the media.

Why is there always ROI anxiety in corporate AI budgets?

The guest provides three reasons from the perspective of seventeen years in technology management, all based on experience rather than statistics: First, the attribution of returns is inherently ambiguous—if a customer service team adopts an AI tool and complaint rates drop, it's hard to distinguish the tool's contribution from the workflow's natural improvement. Second, there is an asymmetry between immediate expenses and delayed returns; bills come monthly, returns are calculated annually, and AI capabilities depreciate rapidly. Third, many companies treat "adopting AI" as an endpoint rather than a starting point, and the capabilities are never integrated into any revenue-generating business line. There is only one countermeasure: AI expenditures must be tied to specific metrics of a specific business line; if they can't be tied, they shouldn't be approved.

How can an individual determine whether an AI expense is a cost or an investment?

Calculate three accounts together: the explicit account (monthly fees, API fees, fixed expenses), the time account (the repetitive labor saved by the tool converted by one's own hourly cost; the guest estimates this to be three to five times the monthly fee, explicitly stating this is a personal estimate, not statistics), and the opportunity account (traffic conversion brought by the tool site matrix, treated as an option, not a wage). As long as the time account holds up among the three, the expense is an investment, not consumption. Then use three questions as a yardstick: What happens if you cut it? Who does it work for? Who is the bill addressed to?

Who Pays the AI Bill: From $278 Billion in Industry Cash Burn to a 699 Yuan/Month Personal Tool Tab

Opening: An Electricity Bill Mailed to an Ordinary Household in North Carolina

The English edition of La Noticia published a commentary on September 20, with a very direct title: Who Pays the AI Bill. The scenario is in North Carolina, USA: AI data centers are flocking to settle there, and the local utility company Duke Energy admits that the electricity demand driven by data centers accounts for a large chunk of the electricity growth from new economic projects.

Servers run twenty-four hours a day, requiring cooling, substations, and transmission lines—all of which cost money. The commentary author asks: if the project shrinks or even withdraws a few years later, who will cover the grid construction costs already incurred?

Shiwen: Why start this episode with an electricity bill?

Yongliang: Because this bill is close to everyone. The industry-level bill is counted in the hundreds of billions, the company-level bill in the hundreds of millions, but ultimately, a bill will always be handed to a specific person—it could be an electricity bill for a North Carolina household, or it could be my 699 Yuan monthly tool fee. In this episode, we will lay out the bills at all three levels one by one: who is paying, who is betting, and who is anxious. Let me clarify the rule for this episode first: all large numbers are media metrics, which I will indicate item by item; for bills involving myself, wherever I estimate, I will explicitly state that it is my estimate, not statistics. When it comes to bills, the most feared thing is ambiguous metrics.

Q1: Breaking Down the North Carolina Case First—“Whoever Creates the Bill, Pays It”

Shiwen: What is the core proposition of that commentary?

Yongliang: The one-sentence version: whoever creates the bill should pay it.

His reasoning chain is not complex. The electricity demand of AI data centers is real, and so is the investment, but the utility company has to bet heavily on expanding the grid for a super-sized customer, and the bill will ultimately be spread into electricity rates. If the project succeeds, everyone is happy; but what if it fails? Where does the cost of the idle infrastructure go? His answer is clear: it cannot be “all of us.”

He makes an analogy that I will borrow because it’s accurate: you are welcome to invest, but the rules must be established first, so that ordinary citizens don’t become “involuntary partners”—ordinary households didn’t vote to decide whether to subsidize a data center, yet they will participate in this gamble through their monthly electricity bills. His bottom line is not radical: promoting economic development is one thing, making the public underwrite the location risk of a large corporation is another; the two must not be conflated.

Shiwen: Is there a real-world basis for such concerns?

Yongliang: Yes. The risk scenario he describes is a script that has actually played out in the industry: a utility company makes a million-dollar-level infrastructure bet for a super-sized project, and a few years later the project shrinks or pulls out entirely, leaving the grid and substations as sunk costs amortized by all users. Electricity demand forecasts inherently carry uncertainty, and the question of who bears the cost of forecast errors existed before AI arrived; AI has simply raised the magnitude of electricity demand by another notch.

A caveat: this is a media commentary, not a news report; its role is to thoroughly explain the risks and ask the questions, not to provide adjudicated cases.

Q2: The Industry-Level Bill—$278 Billion, $28.5 Billion, and Builders Hitting the Brakes

Shiwen: Zooming out to the entire industry, what is the scale of this cash burn now?

Yongliang: Let’s state two numbers first, both from the media, with the metrics clearly indicated.

The first number: according to media reports, OpenAI’s prediction is a cumulative cash burn of $278 billion by 2030. This is the prediction metric disclosed on September 20—a prediction is not a fact, but it shows that even the party involved is planning based on a cash gap in the hundreds of billions.

The second number: according to reports, Oracle’s single-quarter AI infrastructure investment is $28.5 billion, and management still feels it’s not enough. To put it in a way that helps everyone grasp the magnitude: this is the capital expenditure metric disclosed by the media; this single-quarter figure already exceeds the cumulative lifetime investment of the vast majority of publicly traded companies.

Put the two numbers together: on one hand, cash burn is in the hundreds of billions; on the other, single-quarter investments are record-breaking yet still feel slow. This is the current state of the industry-level bill—no one denies the total amount, but the problem is that it is currently propped up mainly by financing and expectations, not yet caught by revenue. In other words, the payer on this level’s bill is written as “future revenue,” and whether that future is willing to arrive and when is exactly the bet of this whole gamble. The power plants have started being built, the substations are already on the blueprints; these expenses are solidly in the present; the corresponding revenue, however, lives in forecasts.

Shiwen: Even Anthropic has reportedly voiced dissent?

Yongliang: Reportedly, Anthropic’s CEO has proposed a pause in AI development. Note the metric: this is headline-level information reported by the media; I don’t have the primary text for the scope and conditions of the proposal, so I won’t detail it in the text. But even just as a signal, it’s striking—even within the most steadfast builder camp, some are starting to publicly discuss whether to hit the brakes. On the market side, investment commentaries have already discussed nuclear operator Constellation as a target under the “AI power logic”; this is a commentary metric, not news. Capital is voting with its feet on the power landscape of the next decade, and where the power comes from and who pays the bill is exactly the question asked by the commentary in Q1.

Q3: The Company-Level Bill—ROI Anxiety and the “Cost Center” Debate

Shiwen: The industry’s bill is propped up by financing; what does it look like when it lands inside a company?

Yongliang: It becomes a tug-of-war in budget meetings.

Over the past seventeen years of leading teams, peaking at seventy people, this is the table I’m most familiar with. When AI expenditures enter the budget, they always face the same question: are they a cost center or a profit center? Buying compute, models, and tools is easy; but when and how much will come back on the revenue side—no one dares to sign a guarantee. Thus, you see a widespread ROI anxiety—the bill anxiety at the company level is not about being unable to pay, but about being unable to clearly articulate the value after paying.

Shiwen: Where does the inability to articulate it clearly come from?

Yongliang: Three reasons, all from my management perspective, not statistics.

First, the attribution of AI returns is inherently ambiguous. A customer service team adopts an AI tool, and complaint rates drop—is it the tool’s credit, or the credit of the workflow improvements that came along with it? With unclear attribution, ROI only holds up in presentations, collapsing as soon as it reaches the CFO.

Second, there is an asymmetry between delayed returns and immediate expenses. Bills come monthly, returns are calculated annually; during that year-plus in between, the people spending the money naturally feel the pressure in conference rooms. What’s more critical is that AI capabilities themselves are rapidly depreciating: the flagship configuration bought this year might be on the standard configuration list next year. Expenses are rigid current amounts, while returns are depreciated future cash flows—this time gap is the true source of ROI anxiety.

Third, and most critically: many companies treat “adopting AI” as an endpoint rather than a starting point. Tools are bought, training is done, press releases are issued, and then there’s no “and then.” Of course the books won’t balance—it’s not that AI isn’t worth the money, but that the capabilities bought with this money haven’t been placed into any revenue-generating business line.

My experience is: AI expenditures must be tied to specific metrics of a specific business line; if they can’t be tied, they shouldn’t be approved. Without this step, all AI budgets are cost centers; with it, they have a chance to become profit centers.

Q4: The Personal-Level Bill—699 Yuan per Month, How to Make AI Sustain Itself

Shiwen: Finally, landing on yourself, what does your bill look like?

Yongliang: A monthly tool fee budget of 699 Yuan; this is my real situation.

699 Yuan a month, over eight thousand a year. At the company level, this amount isn’t even a fraction, but for an individual developer, it must answer one question: can you sustain yourself? My approach is: use this 699 Yuan to build a matrix of tool sites, letting the network revenue generated by the sites cover the tool fees themselves. The goal is simple—when reconciling accounts at the end of the month, the tool fee column can be offset by network revenue, and the account is considered closed.

Shiwen: How do you calculate this bill?

Yongliang: I can share the method and estimates, but I won’t report the exact numbers; these are my estimates, not statistics. My algorithm is to calculate three accounts together: the first is the explicit account—monthly fees, API call fees, rigid expenses; the second is the time account—how many hours of repetitive labor this 699 Yuan saves me, converted at my own hourly cost—the direction is clear: the time cost saved by the tools each month is conservatively three to five times the monthly fee; the third is the opportunity account—the traffic and conversion brought in after the tool site matrix is up and running. This part fluctuates the most, so I treat it as an option, not a wage.

Among the three accounts, as long as the second one holds, this expense is no longer consumption but an investment; the third determines whether it can grow into a business. The difference between an individual and a company is not as big as imagined—both must answer “what comes back when this money goes out,” it’s just that an individual’s account is simpler, so simple that you can’t use “organizational complexity” as an excuse. There’s also a counter-intuitive point: the fault tolerance of a personal account is actually higher than that of a corporate account. If a company makes a wrong decision, it could mean tens of millions in sunk costs plus a whole year of organizational morale; if I personally misjudge a tool, the maximum loss is a few dozen Kuai in monthly fees, and I can just switch it next month. Therefore, individuals are actually the best suited to practice “how to sign an AI bill”—the tuition is cheap, and the person signing is right there in the mirror.

Q5: Wrap-Up—Behind the Bill Is Choice, and Expectation Management

Shiwen: The bills at all three levels have been laid out; what is the final conclusion?

Yongliang: It’s never that no one pays the AI bill, but rather that the person paying is not necessarily the one spending the money.

North Carolina households might pay for data centers’ location risks through their electricity bills; this is the bill being quietly passed off. The financing market pays for startups’ cash gaps; this is the bill being delayed for payment. Corporate budgets pay for the posture of “having adopted AI”; this is the bill being paid in confusion. Three situations, one root cause: when spending money, no one writes the words “who pays” into the contract.

The most worthwhile attitude to take away from that North Carolina commentary is neither exclusion nor anti-investment, but keeping the accounting in the open—welcome investment, but establish the rules first. This attitude applies equally to companies and individuals: when approving AI budgets, agree on the return metrics first; before buying an AI tool, figure out who it works for. The bill itself is not scary; what’s scary is when the bill arrives, and you realize your name is in the signature line.

Finally, a word on expectation management. These three levels of bills are essentially all cases of expectations preceding the bill: the industry builds today’s power plants based on revenues ten years from now; companies approve this year’s budgets based on next year’s efficiency; I pay this month’s fees based on the potential of the tool site matrix. Expectation management is not some profound science; it’s simply asking yourself before each signature: if the returns don’t come, can I accept this loss, can I absorb it? If yes, sign; if no, cut the amount down to a range you can accept, then sign. This one question applies universally—industry, company, individual.

Epilogue

Shiwen: Summarize this episode in one sentence?

Yongliang: The scarcest thing in the AI era is not compute, but people willing to sign the bill. Industry level, company level, personal level—the bill amounts differ by nine orders of magnitude, but the format of the signature line is the same: whoever decides, is responsible.

Shiwen: I give this sentence to everyone. See you next episode.

[Deep Dive] How to Judge Whether an AI Expense is a Cost or an Investment

Many readers might think: I understand the reasoning, but how do I judge when it falls on me? I’ll lend you the yardstick I’ve used for reviewing budgets over the past seventeen years—three criteria, all based on experience, not statistics.

First, ask “what happens if you cut it?” If you cut this expense, does the business hurt immediately, hurt slowly, or not hurt at all? Immediate pain usually indicates infrastructure, slow pain usually indicates efficiency tools, and no pain—mostly something bought to follow a trend that year.

Second, ask “who does it work for?” Good AI expenses all have a specific role profile: which step and what type of labor hours it replaces. An expense that can’t articulate a role profile is highly likely to become a cost center.

Third, ask “who is the bill addressed to?” When an individual buys a tool, the bill has their own name on it, so individuals can actually calculate this bill more clearly than companies—no one is there to underwrite it for you. This also explains why I put the personal account at the finale of this episode: it’s not the amount that matters, but the clarity of the signature.

Sources for This Episode

La Noticia 2026-09-20 English edition commentary article “AI: Who Pays the Bill?” (Media commentary metric, not news reporting; the North Carolina scenario, Duke Energy’s statement on electricity share, the “involuntary partner” analogy, and the stance that “promoting economic development does not equal subsidizing corporate risk” are all paraphrased commentary viewpoints; the original text is not directly quoted in the body)
whalesbook 2026-09-20 paraphrase: OpenAI predicts a cumulative cash burn of $278 billion by 2030 (Media-disclosed prediction metric, marked as “according to media reports” in the body)
Memeburn 2026-09-15: Oracle’s single-quarter AI infrastructure investment of $28.5 billion still “not enough” (Media metric, marked as “according to reports” in the body)
Blockonomi 2026-09-15/16: Anthropic CEO proposes a pause in AI development (Headline-level metric reported by media, marked as “reportedly” in the body)
Motley Fool commentary: Nuclear operator Constellation discussed by investors as an “AI power logic” target (Investment commentary metric, used only as corroborating evidence)