OpenAI Loses Over $100 Billion a Year: How the AI Cost Paradox Is Devouring Tech Giants

The "Coconut Stand" That Loses More the More It Sells

The blogger uses a vivid metaphor to describe OpenAI's business model: imagine your friend opens a coconut stand, selling each coconut for $1, with customers lining up in what looks like a booming business — but you later discover that the cost of selling one coconut exceeds $2. Labor, rent, and utilities together far outweigh the revenue from each sale. This is exactly OpenAI's current situation.

The numbers don't lie. Last quarter OpenAI recorded roughly $5.7 billion in revenue, which sounds impressive, but for every $1 it brought in, it lost even more. Wall Street uses a term to describe this state: an operating margin of -122%. Operating margin is a key metric for measuring a company's core business profitability, calculated as operating profit divided by revenue. A -122% margin means that for every $1 of revenue, the company had to spend an additional $1.22 in operating costs — in other words, to earn that $5.7 billion in revenue, the company burned through nearly $7 billion just to keep the lights on.

To understand why this figure is so alarming, you need to understand the dual cost structure of large models. Training cost is a one-time investment used to tune the hundreds of billions or even trillions of parameters in a model over vast amounts of data, costing tens of millions to hundreds of millions of dollars. It is a sunk cost that can be gradually amortized against future revenue. The underlying principle is backpropagation: the model computes prediction errors on each batch of training samples and passes gradient signals back through the network layer by layer, fine-tuning the weight of each parameter. This process requires tens of billions of iterations to converge to a usable level of model capability. The real, continuous bleeding comes from inference cost — every time a user interacts with ChatGPT, thousands of backend GPUs must be called upon to perform intensive matrix operations. A GPT-4-class large model requires an astronomical number of floating-point operations per forward pass, generating real-time electricity bills, network bandwidth costs, and hardware depreciation. Unlike training, inference has no moment of "completion" — as long as users are using the service, the billing clock keeps ticking. As the user base swells from millions to hundreds of millions, this cost grows linearly or even super-linearly, completely eating up the marginal revenue from subscription fees. This organization, which was founded with a "non-profit" mission, has now literally achieved "non-profitability."

Signing Nearly a Trillion Dollars in Purchase Orders in One Year

Over the past year, Sam Altman has kicked off a rare buying frenzy in the AI market: dropping roughly $300 billion on Oracle, securing $100 billion in investment from NVIDIA, a $90 billion AMD chip order, and roughly $250 billion worth of cloud computing resources from Microsoft. Altogether, OpenAI signed nearly a trillion dollars in various agreements within about a year.

The core targets of these purchases — GPU chips and cloud computing resources — are precisely the material foundation of AI inference costs. The GPU (Graphics Processing Unit), originally designed for graphics rendering, has thousands of parallel computing cores that happen to be well-suited to the large-scale matrix multiplication operations required by deep learning. NVIDIA established the de facto standard for AI computing power through its CUDA programming framework. Its H100/H200 series GPUs are based on the Hopper architecture and integrate Tensor Cores optimized specifically for Transformer models, with each card priced at over $30,000. The Transformer architecture is the foundational structure of nearly all mainstream large language models today. Its core mechanism, the attention mechanism, allows the model to dynamically weigh the relevance of all words in the input sequence when generating each word — this full-sequence scanning gives the model powerful contextual understanding, at the cost of computational load that grows quadratically with sequence length, making it extremely sensitive to computing demands. When running models with hundreds of billions of parameters, the model weights alone require hundreds of GB of memory, necessitating clusters of dozens to hundreds of GPUs interconnected via high-speed NVLink or InfiniBand to work in coordination. OpenAI's nearly trillion-dollar purchase agreements are essentially using capital to lock in computing power supply for years to come, in order to cope with the inference pressure brought by continued user growth. Yet such a massive capital flow yields not profit, but an even deeper black hole of losses.

The "Number Magic" in the Financial Reports

The blogger points out that Sam is extremely skilled at packaging the company to look healthy and shiny. If you ask him how the company is doing, he'll proudly tell you: revenue soared from $3.7 billion to $13.1 billion, more than tripling in a year — this is indeed a growth curve rarely seen at any company.

But he deliberately omits the details: costs also skyrocketed from $12.5 billion to $34 billion. As for the losses, the figure Sam quotes is "only $14 billion," which sounds fairly manageable. But this number cleverly ignores expenses like stock-based compensation. Once those are included, the true loss jumps to roughly $26 billion.

This style of financial presentation is nothing new in the tech industry. "Adjusted EBITDA" (a profit metric that strips out interest, taxes, depreciation, amortization, and stock-based compensation) has long been used by Silicon Valley to dress up the books, with Uber, Lyft, and Snap all relying on similar framing to tell stories to investors in their early days. To understand this trick, you need to understand the accounting nature of stock-based compensation: when a company grants options or restricted stock units (RSUs) to employees, no immediate cash payment is required, but under GAAP (U.S. Generally Accepted Accounting Principles), it must be recorded as compensation expense at fair value. The logic behind excluding it from the "adjusted" metric is controversial because stock-based compensation (SBC) is a real economic cost. It dilutes existing shareholders' equity, essentially trading future stock for present talent — just passed on to shareholders in a delayed and dispersed manner. Even more notable is that for companies like OpenAI, which are not yet publicly listed and whose employee equity has limited liquidity, both the actual incentive effect and valuation basis of SBC carry greater uncertainty, and its "fair value" itself contains a subjective element. The roughly $12 billion gap between OpenAI's $14 billion figure and its true loss of around $26 billion consists primarily of this selectively obscured compensation cost.

Even the so-called "good news" doesn't hold up to scrutiny. Sam will emphasize that margins are improving and the cost per ChatGPT message is dropping. But the blogger cuts right to the point: "Getting less bad at losing money is still, fundamentally, losing money." Growth and sustainability have never been the same thing, and Sam is betting precisely on you not noticing the difference between the two.

The Customer Side Is Equally Grim: 95% of Enterprise AI Investments Yield Zero Returns

If OpenAI's own losses can be defended as "strategic investment," the data on the customer side is even more alarming. An MIT study found that enterprises invested $30 to $40 billion in AI, of which 95% saw no return whatsoever, and only about 5% of enterprise AI pilots actually translated into profit growth. Goldman Sachs, looking at the entire U.S. economy, reached a similarly near-zero conclusion about AI's net effect.

The blogger cites some shocking spending cases: Uber's CTO admitted the company burned through its entire annual AI budget in a matter of weeks; GetSwan, a company with only 4 employees, spent over $113,000 on AI tokens in a single month — nearly $28,000 per person; and one developer burned through $1.3 million in OpenAI tokens in a single month.

Understanding these numbers requires grasping the basic logic of token economics. A token is the basic unit large language models use to process text, produced by a tokenizer that splits text into pieces — for English, roughly every 4 characters correspond to 1 token; for Chinese, roughly every 1-2 characters correspond to 1 token. Each user question and each model response consumes hundreds to thousands of tokens, and OpenAI's API bills by token usage. Output tokens (the model's generated responses) are typically 2-3 times more expensive than input tokens, because the generation process is autoregressive — the model generates only one token at a time, then appends it back to the input sequence and runs another complete forward pass to predict the next token, repeating this cycle until it generates a complete response. This means a 200-token response requires running 200 forward passes, with computational overhead strictly proportional to output length. When enterprises embed AI into automated workflows, it's very easy to trigger the "hidden scale trap": the cost per call seems tiny, but if each business record triggers one call, a daily volume of a million records means tens of thousands of dollars in spending per day. Even more dangerous, multi-turn conversations carrying historical context accumulate input tokens linearly with each turn, causing single-session costs to grow geometrically. The staggering bills at companies like GetSwan are precisely the price of rushing deployment without any unit economics calculations: cost stress-testing wasn't done during product feature design, and only after launch did they discover that every additional active user was digging deeper into their profits.

It should be noted objectively that some critics argue the six-month window used by the MIT study is too short and may underestimate slower-to-materialize benefits. This criticism has some validity. But as the blogger puts it, when 19 out of 20 pilots produce nothing, you don't need a ten-year study to figure out where things went wrong.

The Four-Step Vicious Cycle

The blogger summarizes a script that plays out repeatedly across industries, outlining the complete path by which enterprises fall into the AI cost quagmire.

Step One: Layoffs to cut costs. Large numbers of tech industry jobs have been eliminated, with roughly half directly attributed to AI. Standard Chartered announced the elimination of 7,000 positions in one go, betting on automation, with Meta and Amazon following suit. But once AI takes over, accidents abound: Volkswagen spent heavily building a unified AI operating system, resulting in a 20-million-line-of-code, bug-ridden mess that delayed the Porsche EV and Audi Q6 e-tron by over a year and led to 1,600 layoffs. Pizza Hut used AI to manage delivery, turning what was a 30-minute delivery into 45 minutes or more, and even attracted a $100 million lawsuit from its own franchisees. These cases reveal a common pattern: AI excels at improving efficiency in controlled, clearly bounded tasks, but often fails in real-world business scenarios requiring contextual judgment, exception handling, and cross-system coordination — and these scenarios are precisely the daily bread and butter of enterprise operations. From an engineering perspective, real-world business scenarios are full of long-tail problems — the 5% of exceptional cases often consume 50% of the manual processing time, and these low-frequency, high-variance situations are exactly what current large language models struggle most to generalize to.

Step Two: In a panic, management believes the cure is "more AI." NVIDIA CEO Jensen Huang reportedly demanded that engineers spend the equivalent of half a year's salary on AI tokens annually, or else they weren't being efficient enough. Employees fearing layoffs have no choice but to comply — Amazon employees even admitted to using AI on completely unnecessary tasks just to boost their internal usage scores. This kind of "performative usage" is known in organizational behavior as compliance theater: when management sets "AI tool usage rate" or "token consumption" as a performance metric, the rational choice for employees is to optimize those numbers themselves rather than the actual output goals behind them. This phenomenon corresponds economically to Goodhart's Law — "When a measure becomes a target, it ceases to be a good measure." Employees' behavioral goals become alienated from "solving real problems" into "satisfying visible assessment metrics," ultimately producing large numbers of formally compliant but substantively useless AI calls, further inflating token consumption data, causing management to misjudge the technology's value and pour in more investment, forming a false prosperity.

Step Three: Budgets run dry. Microsoft actually cut its use of cloud code, because the more employees used it, the higher the cost. Uber's COO also publicly stated that there is currently no evidence that spending more on AI produces a better product — which corroborates the MIT and Goldman Sachs research.

Step Four: Begging AI companies to lower prices. And this is precisely where it gets most absurd.

Cut Prices or Raise Them: A Race to the Bottom

Reports say OpenAI is considering lowering prices to compete against rivals like Anthropic's Claude. But the blogger repeatedly emphasizes: OpenAI loses money every time its product is sold or used.

This creates an unsolvable knot: cutting prices means digging your own grave faster; raising prices means customers already crushed by costs will immediately turn to Anthropic, Microsoft, or any cheaper competitor.

This pricing dilemma has deep roots in industry logic: the marginal cost of large-model API services does not approach zero, which is exactly the opposite of the rules that governed the internet software industry over the past two decades. Traditional SaaS product business models were built on the physical fact that "the cost of replicating software approaches zero" — a single codebase serving 1 user versus 10 million users has limited marginal cost difference, allowing fixed development costs to be amortized across a massive user base, with margins improving as scale expands. This logic supported the commercial myth of internet companies routinely achieving gross margins above 50% over the past two decades. Large-model inference breaks this logic: each request consumes real physical resources — chip computation, electric current flowing through wires, water evaporating from cooling systems — resources that have a clear physical cost floor and cannot disappear through "softwareization." More subtly, current mainstream large models use dense architectures, which require activating all parameters during inference — by contrast, sparse architectures such as Mixture of Experts (MoE) can activate only a subset of parameter paths, potentially reducing inference cost while maintaining model capability, but this has not yet been proven to sufficiently bridge the cost gap in commercial scenarios. The larger the scale, the higher the absolute value of losses. Zooming out, this is essentially a race to the bottom — whoever can price their service lowest temporarily wins, until their own cash flow collapses first, then they run off to Wall Street to raise capital, borrow money, and use every means available to sustain the performance.

It's worth mentioning that despite continuous losses, Sam still completed a funding round of roughly $122 billion, with the company valued at a staggering $852 billion, and has already begun preparing for an IPO at an asking price of over a trillion dollars. This valuation logic rests on one core assumption: that someday, the computing cost of model training and inference will drop dramatically due to hardware advances (such as more efficient chip architectures) or algorithmic breakthroughs (such as more streamlined model distillation techniques), at which point marginal costs will truly narrow. Knowledge distillation is one of the technical paths pinned with high hopes: using the outputs of a large "teacher model" to supervise the training of a small "student model," enabling the latter to approximate the former's capabilities as closely as possible while drastically reducing parameter count — in theory reducing inference cost by an order of magnitude. The problem is, no one knows when "someday" will actually arrive, or how much capital must be burned before then. A company that loses money every time it's used, yet whose valuation soars into the sky — this itself is the biggest footnote to this frenzy.

Data Centers: The Bill Ultimately Falls on Ordinary People

All this AI frenzy requires real computing power to support it — namely AI data centers — and the electricity and water bills often end up being paid by ordinary residents.

A hyperscale data center serving large-model inference has energy demands far exceeding traditional internet facilities. Power density is the key difference: traditional server racks have a power density of about 5-10 kilowatts, while AI racks equipped with GPU clusters can reach 50-100 kilowatts or even higher, placing multiplied pressure on power supply, cooling systems, and building structures. The overall electricity consumption of a hyperscale data center can reach hundreds of megawatts (MW), equivalent to the daily power needs of a medium-sized city. Unlike ordinary internet data centers, AI inference workloads require GPUs to maintain extremely high sustained utilization rates (typically over 80%) with almost no idle time, meaning power consumption runs at near-full capacity around the clock. At the same time, the enormous heat generated by GPU computation requires cooling systems to work nonstop. Power Usage Effectiveness (PUE) is the industry-standard metric for measuring data center energy efficiency, defined as the ratio of total facility power to IT equipment power — an ideal value of 1.0 means all power is used for computation itself; traditional data centers have a PUE of about 1.5-2.0, while top AI data centers can lower PUE to 1.1-1.2 through technologies like liquid cooling, but the absolute total energy consumption still continues to climb due to the explosive growth in computing demand. Evaporative cooling towers use the latent heat of water vaporization to dissipate heat, with some facilities evaporating 1-2 liters of water for every kilowatt-hour consumed, resulting in daily water consumption reaching millions of gallons, sparking resource competition with agricultural irrigation and municipal water supply in already water-stressed arid regions.

The blogger mentions that residents of Maryland were forced to bear roughly $2 billion in costs to upgrade a power grid supplying out-of-state data centers — data centers they themselves can't even use. This is not an isolated case, but a systemic phenomenon occurring in many parts of the U.S.: the cost of expanding grid infrastructure is spread across all electricity users in the district in the form of "electricity surcharges," regardless of whether they benefit. This mechanism is known in regulatory economics as cost socialization: the expansion profits of private enterprises are enjoyed by shareholders and investors, while the cost of infrastructure upgrades is passed on to all taxpayers and electricity users through public utility rates, creating a classic asymmetry between benefits and risks.

Public backlash is spreading: nearly 47% of Americans say they don't want data centers built near their homes; voters in Festus, Missouri recalled the city council members who approved data center construction; the Seattle City Council voted 9-to-0 to pause new large AI data centers for a year; and towns in Ohio, Indiana, Michigan, and elsewhere are passing or drafting similar bans. But this battle is far from over — tech companies have begun suing the towns that block them, and some states have even legislated to strip local governments of veto power.

If AI Is Really That Powerful, Why Force People to Use It?

This story reveals a tragic irony in the big tech industry: if AI is truly so powerful, why do so many companies need to force employees to use it? Or is this just a desperate posture, an attempt to justify the existence of funds that were poured in massively yet produced nothing?

History is full of similar tech bubble cycles: during the late-1990s internet boom, the "eyeball economy" replaced profitability as the anchor for valuation, until the Nasdaq crash in 2000; in the 2010s sharing economy wave, companies like Uber and WeWork traded cash-burning subsidies for growth data, packaging losses as "market education costs." Both bubbles shared a common feature: participants believed that when some future tipping point arrived, scale advantages would automatically make the business model viable — the internet bubble's redemption was advertising monetization and e-commerce; the sharing economy's redemption was autonomous driving reducing labor costs. The AI frenzy's presumed redemption is a technical collapse in computing costs. It's worth noting that this redemption logic isn't entirely baseless: over the past decade, GPU computing density has indeed increased at a rate roughly approximating Moore's Law, with performance per watt of floating-point operations continuously improving; but at the same time, model parameter scale and training data volume have expanded in tandem, and the scaling laws — the power-law scaling of model performance with parameter count and data volume — mean that competitors chasing stronger capabilities keep pushing the cost ceiling higher, forming an arms race with no end in sight. Whether the AI frenzy will repeat the same mistakes, or truly reach a technical cost inflection point that brings a profitability turnaround, remains undetermined. But what is certain is that the rules of this game have never been fair to ordinary participants.

When the people selling the "AI miracle drug" can't even balance their own books, why would you trust that they've balanced the books for their customers? This is the AI cost paradox that is devouring OpenAI. Whatever the final outcome, at least more and more ordinary people are soberly realizing that they are footing the bill for a party that isn't theirs.

Key Takeaways

Key Takeaways