OpenAI has one of the most enviable problems in technology: overwhelming demand.
It also has one of the most dangerous.
Unlike traditional consumer software, artificial intelligence does not simply sit on a server waiting to be downloaded. Every response requires computation. Every reasoning step consumes accelerator time. Every agent that searches files, runs code, calls tools or revises its work can trigger dozens of additional model requests.
The result is a peculiar economic dilemma. More users create more revenue, but they also create more inference costs. More capable models attract heavier workloads, but those workloads require more tokens, longer context windows and greater computing capacity.
OpenAI’s survival challenge is therefore not merely to build the world’s most intelligent model. It must make intelligence cheap enough to produce at planetary scale without allowing the price of that intelligence to collapse faster than its own costs.
That is why OpenAI is beginning to look less like a software company and more like an industrial conglomerate.
OpenAI’s financial results show both the extraordinary strength of its business and the extraordinary cost of operating it.
Audited financial documents covering 2025 reportedly show:
The headline net loss was magnified by large non-cash fair-value adjustments associated with OpenAI’s investor and corporate structure. Nevertheless, the underlying operating picture remained severe: expenses substantially exceeded revenue, while research and development alone cost more than the company’s entire annual sales.
At the same time, OpenAI’s commercial growth has been historic. On March 31, 2026, the company said it was generating approximately $2 billion in revenue per month, equivalent to a roughly $24 billion annualized run rate. It also reported more than 900 million weekly ChatGPT users, over 50 million subscribers and an enterprise business representing more than 40% of revenue. By late July, OpenAI said its models had reached one billion active users and more than two million businesses.
This is the paradox at the heart of OpenAI: few companies have ever grown this quickly, but few software companies have required so much capital to support that growth.
Not necessarily.
The claim that every ChatGPT request costs OpenAI more than it earns is too broad. In the reported 2025 accounts, $13.07 billion of revenue exceeded the $7.5 billion categorized as cost of revenue. That indicates positive reported gross profit before research, product development, sales, compensation and other corporate costs.
The real problem is more complicated. Some workloads may be profitable, while heavy users particularly those running long reasoning sessions or autonomous agents under flat-rate subscriptions may consume far more compute than their monthly fee covers. In addition, OpenAI is spending billions on model training, talent, data centers, future capacity and products whose revenue may arrive years later.
The company is therefore not trapped simply because “every question loses money.” It is trapped because its total investment requirements are expanding faster than even its exceptional revenue.
It is tempting to summarize OpenAI’s situation by saying that the company will earn $25 billion and lose $33 billion in 2026. But those figures come from different sources and use different accounting definitions.
Sacra estimates that OpenAI reached approximately $25 billion in annualized revenue in February 2026. It also estimates that inference expenses will rise from $8.4 billion in 2025 to $14.1 billion in 2026, while cash burn could reach approximately $27 billion. These are third-party estimates, not audited results or formal company guidance.
The $33 billion figure comes from a separate external forecast of OpenAI’s 2026 GAAP loss. That forecaster gives a very wide range from $20 billion to $55 billion because stock compensation, fair-value adjustments and other non-cash items can dramatically change the final accounting result. It should therefore be described as a scenario rather than a confirmed loss projection.
The safer conclusion is still striking: OpenAI is producing tens of billions of dollars in revenue, but it remains dependent on outside capital while financing one of the largest infrastructure expansions in corporate history.
In March 2026, OpenAI closed $122 billion in committed funding at an $852 billion post-money valuation. That capital is not simply being raised to cover chatbot operations. It is financing the transition from an AI research organization into a global intelligence platform.
Classic Software-as-a-Service(SaaS) businesses become highly profitable because the marginal cost of adding another customer can be extremely low.
AI has a different cost structure.
A conventional application retrieves stored information or executes predefined logic. A generative model performs substantial mathematical computation whenever it produces an answer. Reasoning models may generate internal intermediate steps. Multimodal systems process images, audio or video. Agents can repeatedly call models, retrieve information, run tools and evaluate their own output.
OpenAI explains that a single Codex turn may inspect source code, search deployment history, read incident reports, edit files and run tests. Each step can require another request. If an agent loop makes 30 model calls, small inefficiencies are paid 30 times.
That makes inference a variable production expense. OpenAI is effectively manufacturing a customized unit of intelligence every time a user asks it to perform work.
This is why its pivot resembles industrialization. The company must optimize the entire production line: models, routing, software kernels, memory movement, networking, chips, energy, data centers and pricing.
OpenAI is often described as attempting to stop renting infrastructure and begin owning the full stack. The reality is more nuanced.
The company is certainly seeking more control, but it is not planning to own every cloud, chip factory or data center itself. OpenAI CFO Sarah Friar has explicitly described a strategy of keeping the balance sheet relatively light by partnering rather than owning, using flexible commitments across infrastructure providers and hardware platforms.
OpenAI’s emerging infrastructure portfolio includes cloud relationships with Microsoft, Oracle, AWS, CoreWeave and Google Cloud. Its hardware strategy spans Nvidia, AMD, AWS Trainium, Cerebras and a custom processor developed with Broadcom. Data-center capacity is being developed with partners including Oracle, SoftBank and others.
The objective is not total ownership. It is bargaining power, supply certainty and the ability to co-design every important layer.
That distinction matters. OpenAI is trying to become the architect of an industrial ecosystem while allowing partners and investors to finance much of its physical machinery.
The scale of those ambitions is measured in gigawatts. OpenAI says its available compute expanded from 0.2 gigawatts in 2023 to 0.6 gigawatts in 2024 and approximately 1.9 gigawatts in 2025. Over the same period, annualized revenue reportedly grew from $2 billion to $6 billion and then to more than $20 billion.
Other infrastructure announcements are even larger. The Stargate project announced an intention to invest up to $500 billion over four years in US AI infrastructure. Separate OpenAI partnerships contemplate at least 10 gigawatts of Nvidia systems and six gigawatts of AMD GPUs. These announcements overlap and should not be added together as though they were a single finalized capital budget, but they reveal the industrial scale of the company’s plans.
OpenAI’s first line of attack is algorithmic efficiency.
In July 2026, it released the GPT‑5.6 family:
As of August 8, 2026, standard short-context API list prices per one million tokens are:
Independent evaluation firm Artificial Analysis found at GPT‑5.6’s launch that Sol scored one point below Claude Fable 5 on its Intelligence Index while costing approximately one-third as much per evaluated task. Sol’s measured task cost was $1.04 in that benchmark, with Terra and Luna lower still. Those numbers are benchmark-specific and predate OpenAI’s July price reductions, but they support the broader claim that OpenAI is aggressively targeting the price-performance frontier.
The most important efficiency gains may be happening below the product layer.
OpenAI says GPT‑5.6 Sol helped its engineers rewrite and optimize production kernels i.e. the low-level software used to execute model operations on accelerators. According to the company, those changes helped reduce end-to-end serving costs by 20%. Sol also designed and ran experiments that increased token-generation efficiency by more than 15%.
This creates an unusual feedback loop: OpenAI’s models are now being used to optimize the systems that run those same models.
Software optimization has limits. At sufficient scale, hardware economics become decisive.
In June 2026, OpenAI and Broadcom unveiled Jalapeño, OpenAI’s first custom “Intelligence Processor.” Unlike a general-purpose GPU, Jalapeño was designed specifically around large-language-model inference: model kernels, memory movement, networking, scheduling and the serving patterns observed across ChatGPT, Codex and the OpenAI API.
The widely circulated claim that Jalapeño will reduce cost per token by 50% requires caution.
Broadcom CEO Hock Tan has reportedly described early testing as producing roughly that level of savings against current-generation GPUs. However, OpenAI’s official announcement does not publish a percentage. It says only that early testing shows substantially better performance per watt and that final performance is still being measured. No independent benchmarks or detailed comparison methodology have yet been released.
Jalapeño should therefore be viewed as a potentially transformative cost lever not as proof that OpenAI has already halved inference expenses.
It is also not a complete break from Nvidia. OpenAI said in March that Nvidia remained the foundation of its infrastructure and that most training and inference still ran on Nvidia GPUs. Custom silicon gives OpenAI an alternative for predictable, high-volume inference workloads, but its broader compute strategy remains deliberately multi-platform.
The strategic benefit is as much about leverage as technology. Owning the chip architecture gives OpenAI more influence over its cost structure, supply chain and negotiations with outside hardware providers.
Reducing the cost of intelligence solves only half the problem.
OpenAI must also make each unit of intelligence economically more valuable.
That is the purpose of Codex and the broader movement toward agents. A chatbot produces a response. An agent can inspect information, operate tools, write and execute code, revise its work and continue until it produces a usable result.
By June 2026, OpenAI said Codex had more than five million weekly active users, more than six times the number following the February launch of its desktop application. Around 20% of users were knowledge workers rather than traditional developers, using Codex to create reports, spreadsheets, presentations, contracts, analyses and workflow automations.
Agent use is also becoming substantially more compute-intensive.
In May 2026, more than 70% of sampled Codex users requested at least one task estimated to require more than an hour of human work. Among OpenAI’s heaviest internal users, the 99th percentile generated more than 60 hours of agent activity per day by running multiple agents in parallel. OpenAI cautions that the human-time estimates are model-generated and should be treated as directional, but the pattern is clear: users are shifting from individual questions to long-running digital labor.
This creates both the opportunity and the danger.
An agent that builds an application, investigates an incident or prepares a financial model can create far more economic value than a chatbot answer. That gives OpenAI an opportunity to charge based on usage, outcomes or work completed rather than relying solely on flat monthly subscriptions.
But the agent may also consume dramatically more compute. It can maintain a large context window, make repeated model calls, use tools, run tests and operate in parallel with other agents.
Codex is therefore both OpenAI’s monetization engine and its compute-demand accelerator.
OpenAI cannot keep all of its efficiency gains.
Competition forces it to pass a significant portion to customers.
Only three weeks after launching GPT‑5.6, OpenAI cut Luna’s price by approximately 80% and Terra’s by 20%. The reductions reflected serving improvements, but they also arrived amid growing customer sensitivity to AI costs and intense competition among OpenAI, Anthropic, Google and lower-cost open-weight models.
This is the pricing trap:
Efficiency does not automatically become profit. OpenAI must reduce its internal cost faster than market prices fall.
It must also capture more value through enterprise contracts, usage-based APIs, commerce, licensing and eventually outcome-based pricing. OpenAI has already said it expects business models tied to intellectual property and results delivered to become more important as AI moves deeper into science, finance, healthcare and other high-value domains.
This leads to the final obstacle: Jevons’ Paradox.
The theory holds that making a resource more efficient can increase total consumption rather than reduce it. When the effective price of using a resource falls, people discover more uses for it. If demand grows faster than the efficiency improvement, total resource consumption rises.
AI already displays this pattern.
The Stanford AI Index found that the cost of querying a model with GPT‑3.5-level benchmark performance fell from approximately **$20 per million tokens in November 2022 to $0.07 in October 2024 **a decline of more than 280 times. Yet total demand for AI inference, data centers and accelerators continued to rise rapidly.
Researchers studying AI’s environmental impact have warned that efficiency improvements can produce rebound effects: cheaper models encourage more products, more users and more intensive workloads, potentially offsetting reductions in energy or computing requirements per task.
OpenAI’s own description of its business contains a version of this paradox. The company says better infrastructure lowers the cost per unit of intelligence. Cheaper and more capable intelligence supports more complex workflows. Those workflows increase usage and create demand for still more compute.
Jevons’ Paradox does not necessarily make profitability impossible. It changes the equation.
If an agent costs half as much to run but customers use it ten times more, OpenAI’s total compute expense rises. However, if the agent also produces enough value to generate proportionally more enterprise or usage-based revenue, the rebound effect becomes a growth engine rather than a financial trap. The relevant metric is therefore not simply cost per token. It is:
Gross profit per successfully completed unit of work.
An inefficient model that completes a valuable legal, engineering or scientific task may be more profitable than a cheap model answering millions of low-value questions. OpenAI’s task is to route each job to the least expensive model capable of completing it reliably and then price the result according to the value it creates.
OpenAI’s success should ultimately be judged against five measures.
Token prices can be misleading. A cheaper model may require more attempts, more tools or more human correction. OpenAI must lower the total cost of producing an acceptable result.
Lab claims will matter only when the chip runs diverse workloads at high utilization across real data centers. Power efficiency, yield, software compatibility and deployment speed will determine whether the custom-silicon advantage survives outside controlled tests.
Investors will need to understand margins after cloud payments, compute credits, depreciation, revenue sharing and other partner arrangements — not simply adjusted metrics that exclude inconvenient costs.
Flat subscriptions become dangerous when customers can run numerous parallel agents continuously. Usage-based, enterprise and outcome-linked pricing must grow alongside the amount of work performed.
OpenAI is coordinating hundreds of billions of dollars in prospective infrastructure. Its partner-financed model reduces the burden on its own balance sheet, but underused capacity, take-or-pay agreements or slower-than-expected demand could still create enormous liabilities.
OpenAI is no longer merely a model laboratory.
It has consumer distribution approaching global utility scale. It has a rapidly expanding enterprise platform. It operates developer APIs, autonomous agents and emerging commerce systems. It coordinates multiple cloud providers, accelerator vendors, data-center developers and financial partners. It is designing software kernels and now its own silicon.
That is the foundation of an AI industrial empire.
But the empire remains unfinished and heavily subsidized.
OpenAI must simultaneously build more infrastructure, make that infrastructure more efficient, defend its models against powerful rivals, reduce customer prices and persuade users to pay for increasingly autonomous work.
If model efficiency and custom silicon reduce the cost of successful tasks faster than competition reduces prices, OpenAI could generate enormous operating leverage. Every additional agent, workflow and enterprise customer would help absorb fixed investments and finance the next generation of infrastructure.
If usage grows faster than efficiency, prices collapse faster than costs, or infrastructure commitments arrive before profitable demand, the same strategy could create the largest money furnace in technology history.
OpenAI’s future will therefore not be decided solely by whether GPT‑5.6 or its successor is the smartest model.
It will be decided by whether the company can turn intelligence into an industrial commodity without commoditizing its own profits.
Disclaimer: The Views expressed here are strictly my own.
OpenAI’s Pivot: From Massive Losses to an Industrial Empire was originally published in Towards AI on Medium, where people are continuing the conversation by highlighting and responding to this story.