Can OpenAI ever become profitable?
In our generative AI market deck, you will find everything you need to understand the market
SUMMARY
Yes, OpenAI can become profitable, but not while research and infrastructure spending keep rising almost as fast as revenue.
The business is already commercially enormous. OpenAI generated about $13.1 billion of revenue in 2025 and is now running at roughly $2 billion per month, yet its cost base was still large enough to produce an operating loss of nearly $21 billion.
The frightening $39 billion net-loss headline exaggerates the immediate cash damage because about $30 billion came from a non-cash revaluation tied to the former investor structure. The cleaner underlying-loss estimate was closer to $8 billion, which is better but hardly comfortable.
OpenAI’s strongest positive signal is gross margin. Revenue grew faster than direct delivery costs, lifting the estimated gross margin from roughly 28% to 43%, so the core economics are improving even as total losses remain huge.
Consumer demand is not the main problem. More than 50 million subscribers could already support a major software company; the harder question is whether OpenAI can serve free users, agents and heavy professional workloads without letting compute consume the subscription revenue.
Enterprise is the clearest profit engine because companies pay for security, integration and completed work rather than raw model access. The catch is that early deployments still require engineers, consultants and support, so OpenAI has to make each rollout more repeatable.
Model prices will keep falling. OpenAI therefore needs to own the workflow around the model through products such as ChatGPT Enterprise, Codex, Frontier and agents, where switching involves more than replacing one API call.
Agents create the biggest upside and the most dangerous cost curve. A resolved support case or finished coding task can be worth far more than the compute behind it, but flat subscriptions break quickly when one user can trigger dozens of expensive model calls.
Frontier research remains the largest obstacle. OpenAI spent about $19.2 billion on research and development in 2025, more than its entire revenue, and the model race gives management a reason to reinvest nearly every commercial gain.
The infrastructure plan raises the stakes again. A projected compute program of roughly $750 billion through 2030 can lower supplier dependence and secure capacity, but it also requires extraordinary utilization and leaves little room for demand misses.
On the latest cost base, ordinary operating profit starts to look realistic at roughly $40 billion to $60 billion of annual revenue, depending mainly on whether gross margin reaches 60% or stays near 43%. At the current revenue pace, that threshold is reachable before 2030.
Our conclusion is that sustainable profitability is plausible near the end of the decade. OpenAI has already proved that customers will pay; it still has to show that research, agents and infrastructure can grow more slowly than the money coming in.
What would profitable actually mean for OpenAI?
OpenAI would count as truly profitable only when its normal business can fund model operation, research, staff, sales and infrastructure without another giant financing round.
A profitable quarter would tell us very little if it came from delaying a model launch, receiving cloud credits or recording an accounting gain. We are looking for recurring operating profit and positive cash flow after the company pays the real cost of serving customers and developing future products.
OpenAI's unusual accounts make that stricter definition necessary. Microsoft has historically provided computing support. Employees receive large amounts of stock compensation. OpenAI's former corporate structure also created accounting charges that looked like operating losses even when no cash left the company.
Three simple tests get us closer to the truth. Does each customer generate more revenue than the direct cost of serving that customer? Can the gross profit cover research, staff and sales? And can the remaining cash finance OpenAI's share of the infrastructure it has promised to use?
The company reaches meaningful profitability only when the answer to all three stays yes for several years, rather than for one carefully managed reporting period.
If you want more recent data on this point, please see our latest generative AI market report.
Why is OpenAI losing so much money today?
OpenAI is currently losing billions because it is running a mass-market app, an enterprise software company and a frontier research lab at the same time.
Audited figures reviewed by the Financial Times showed about $13.1 billion of revenue in 2025 against roughly $34 billion of costs and expenses. Research and development consumed around $19.2 billion. Direct delivery costs were about $7.5 billion, sales and marketing approached $5.7 billion, and administration added roughly $1.6 billion.
The research bill explains most of the gap. OpenAI spent about $1.47 on R&D for every dollar it earned. Free ChatGPT users add cost, although research accounts for far more of the shortfall.
OpenAI keeps funding several costly activities together. It trains frontier models, tests new architectures, hires scarce researchers, operates products for hundreds of millions of people and builds an enterprise sales operation. A normal software company usually commercializes a stable product before expanding the next one. OpenAI is commercializing today's model while financing tomorrow's model and reserving the hardware for the model after that.
Management could reduce the loss quickly by limiting free use, cutting experiments and slowing model development. That would also increase the risk of falling behind Anthropic, Google, Meta or a new open-model competitor. For now, OpenAI prefers leadership over profit.
This market map, featured in our generative AI market deck, highlights top companies and startups in the generative AI market
Are OpenAI's losses really as bad as the headlines say?
The $39 billion headline loss overstates the cash damage, while the underlying business still burned roughly $8 billion.
Most of the reported net loss came from a non-cash revaluation tied to OpenAI's former investor structure. As the company's valuation increased, accounting rules forced it to record a much larger liability for investor rights. The Financial Times reported that this charge was about $30 billion and should not recur in the same form after the corporate restructuring.
Even after removing that charge, the accounts remain unhealthy. OpenAI recorded an operating loss of nearly $21 billion. A person familiar with the figures estimated the underlying loss at around $8 billion after also adjusting for stock compensation, Microsoft compute credits and other non-cash items.
Each measure answers a different question. The net loss describes the official accounting result. The operating loss shows how far ordinary expenses exceeded revenue. The adjusted figure comes closer to the cash strain, although excluding stock compensation and subsidized compute makes the business look cheaper than it would be on fully commercial terms.
Read the figures plainly: the headline is inflated, and the remaining loss is still enormous.
| Measure | Approximate result | What it tells us |
|---|---|---|
| Revenue | $13.1B | OpenAI already has a very large commercial business |
| Total costs and expenses | $34.0B | Spending reached about 2.6 times revenue |
| Operating loss | $20.9B | Ordinary accounting expenses remained far above sales |
| Reported net loss | $38.5B | The former investor structure heavily distorted the result |
| Underlying loss estimate | About $8B | Cash economics looked better than the headline, but remained deeply negative |
Is OpenAI's revenue growing fast enough to catch its costs?
For now, OpenAI's revenue is growing faster than its costs in percentage terms, although the dollar gap remains brutal.
Revenue more than tripled in 2025, while total costs increased by about 174%. That is genuine progress. Each new dollar of sales was beginning to carry slightly less expense than before.
The absolute loss still widened because OpenAI started from such a high cost base. Imagine revenue rising by $9 while costs rise by $22. The percentage growth can look better, yet the bank account still moves in the wrong direction.
The latest revenue pace is much stronger than the full-year figure suggests. OpenAI says it is now generating about $2 billion per month, equivalent to roughly $24 billion a year if that pace holds. At the end of 2024, it was generating about $1 billion per quarter. Monthly revenue has therefore reached roughly twice the old quarterly level in a little over a year.
Recent internal misses deserve attention, but they need context. The Wall Street Journal reported that OpenAI fell short of some user and revenue targets, raising concern inside the company about future compute bills. A business can miss an aggressive internal plan and still grow at historic speed. OpenAI's problem is that its spending commitments were designed around even faster growth.
The revenue engine looks strong enough to support a path toward profit. It does not yet look strong enough to support every infrastructure promise OpenAI wants to make.
As this chart shows, and as featured in our generative AI market deck, search interest in LLMs has surged
Does ChatGPT have enough paying users?
Yes, ChatGPT already has enough paying users to support a huge profitable company, while most of its audience still pays nothing directly.
OpenAI says ChatGPT has more than 900 million weekly users and over 50 million consumer subscribers. A rough comparison puts subscribers at about 5.6% of weekly users. The two figures cover different populations, so the ratio is only approximate, but it shows the size of the remaining monetization opportunity.
Fifty million subscriptions can generate serious money. At an illustrative average of $15 per month, they would produce $9 billion a year. The real average will differ because OpenAI sells several plans at different prices, and some customers pay through workplaces rather than consumer accounts.
The free audience gives OpenAI another route. Its early advertising pilot crossed $100 million in annualized revenue within six weeks, according to the company. That launch was fast, but $100 million remains tiny beside OpenAI's overall revenue. Ads are an experiment with promise, not a solution to the profit problem.
The broader opportunity includes shopping commissions, paid actions and premium tools. ChatGPT users often arrive while researching a purchase, comparing services or deciding what to do next. OpenAI can monetize those moments without persuading every user to buy a subscription.
The danger is obvious. Too many ads or commercial recommendations could weaken the trust that made ChatGPT valuable. OpenAI needs to earn more from free users while keeping the answers useful enough that they return.
| Current measure | Approximate figure | What it means |
|---|---|---|
| Weekly ChatGPT users | More than 900M | OpenAI has one of the largest consumer audiences in technology |
| Consumer subscribers | More than 50M | Direct paid demand is already substantial |
| Rough subscriber-to-user ratio | About 5.6% | Most weekly users remain outside the paid base |
| Illustrative revenue at $15 per month | $9B yearly | The subscriber base alone could support a major software company |
| Early advertising run rate | More than $100M | Fast initial demand, but still small beside the core business |
Can enterprise customers become OpenAI's profit engine?
Enterprise is currently OpenAI's best route to profit because businesses will pay more for completed work, security and integration than for a clever chat window.
OpenAI says enterprise activities already produce more than 40% of its revenue. It serves over one million business customers, and its APIs process more than 15 billion tokens per minute. Those figures show that workplace use has moved well beyond small pilots.
Its recent launches show where it expects the money to come from. Frontier helps companies connect agents to internal data, permissions and workflows. DeployCo sends specialists into organizations to turn models into working systems. OpenAI also committed $150 million to a partner network that includes consultancies and implementation firms.
Three examples point in the same direction. Accenture is rolling ChatGPT Enterprise across tens of thousands of employees. Snowflake signed a $200 million agreement to bring OpenAI models into customer data environments. OpenAI models and Codex are now available through AWS, letting companies buy them through infrastructure and procurement systems they already use.
These contracts can be hard to replace once they are working. An AI system connected to customer records, support tools, software repositories and approval rules cannot be swapped out casually. The customer pays for the whole working setup rather than a single model call.
There is a cost attached. Deployments require engineers, consultants and support. OpenAI may look more like Palantir during the early rollout and more like a high-margin software platform after the work becomes repeatable. Enterprise can carry the company toward profit, provided each new contract eventually needs less human help than the previous one.
If you want more recent data on this point, please see our latest generative AI market report.
This chart, featured in our generative AI market deck, shows annual VC investment in generative AI startups
Is OpenAI already losing the enterprise AI race?
OpenAI's share of enterprise model spending is shrinking, while its broader workplace business keeps expanding.
Menlo Ventures estimates that OpenAI's share of enterprise spending on large language models fell from about 50% in 2023 to 27% in 2025. Anthropic reached roughly 40%, with an even stronger position in coding. The estimate comes from a venture firm's market study rather than audited industry accounts, but the direction is hard to dismiss.
Companies now mix models. A bank might use Claude for coding, GPT for customer interactions and Gemini inside Google Workspace. Procurement teams can route each task toward the model that offers the best mix of quality, speed and price. That behavior weakens the loyalty OpenAI enjoyed when GPT was the obvious default.
OpenAI's own enterprise business is still expanding quickly. Workplace products account for a large share of its revenue, and the company serves more than a million business customers. Codex has also reached about four million weekly users, according to OpenAI.
OpenAI is trying to sell the whole workplace system instead of only the underlying model. Frontier, Codex, ChatGPT Enterprise and DeployCo combine models with permissions, data access, interfaces and implementation. A customer can swap an API model more easily than a system embedded across its daily work.
OpenAI's enterprise position will weaken if buyers continue treating models as interchangeable and choose another company for the surrounding software. It can still win the broader market by becoming the place where companies manage AI work, even when every task does not run on the same model.
Can OpenAI keep its margins while AI models get cheaper?
OpenAI can protect its margins only by selling outcomes and workflow control, because token prices will keep falling.
The price drop has already been dramatic. The original GPT-4 API charged about $30 per million input tokens and $60 per million output tokens. OpenAI's current GPT-5.6 Terra tier charges $2.50 for input and $15 for output.
Input pricing has fallen by roughly 92%, while output pricing has fallen by 75%. The newer model also handles a far larger context and more complex work. Customers are receiving much more capability for much less money.
Efficiency explains part of the decline. Chips improved, software became better optimized, prompts can be cached and smaller models can handle work that once required the flagship. Competition explains the rest. Anthropic, Google, Meta, open-weight developers and lower-cost Chinese labs prevent OpenAI from keeping every saving.
A model API alone will struggle to preserve software-like margins. Price comparisons are easy, switching costs can be low and customers increasingly use several providers. OpenAI needs revenue from products where the model represents one piece of a larger service.
Subscriptions, enterprise agents, coding systems, advertising, commerce and outcome-based pricing all move in that direction. The valuable product becomes a resolved support case, finished software task or completed purchase. Customers care far less about the number of tokens when the final result saves real time or money.
If you want more recent data on this point, please see our latest generative AI market report.
This chart, featured in our generative AI market deck, looks at OpenAI’s strategy in generative AI
Will AI agents make OpenAI richer or poorer?
AI agents can become OpenAI's highest-value product and its fastest way to waste compute.
A normal chatbot might answer after one model call. An agent can search files, browse the web, use software, check its work and try again after a failure. One request may trigger dozens of model calls before the user sees a result.
Here is the pricing problem. A fixed monthly subscription works well when usage stays predictable. It becomes dangerous when one active customer can consume more compute than the monthly fee covers.
The upside is much larger than chat. A company may argue over the price of a million tokens, yet happily pay more for an agent that closes a support ticket, reviews a contract or repairs a software bug. OpenAI can charge for the value of the completed task rather than the raw computation behind it.
Codex offers the clearest early test. OpenAI recently reported about four million weekly users, and companies are expanding it through partnerships with Accenture, PwC, Infosys, Dell and Cisco. Coding has attractive economics because a useful piece of software can be worth far more than the compute used to create it.
The same logic fails for many casual requests. Spending $10 of compute to replace $100 of paid work can create strong margins. Spending $20 to satisfy a user on a $20 monthly plan leaves nothing for research, staff or infrastructure.
OpenAI will need strict usage limits, model routing and pricing that rises with the amount of work performed. Agent demand can transform the business, but only when revenue follows the compute bill.
Can inference get cheaper faster than usage grows?
OpenAI has already improved its delivery economics, but cheaper inference keeps encouraging heavier use.
Direct cost of revenue rose from about $2.65 billion in 2024 to $7.5 billion in 2025. Revenue grew faster over the same period. Using the reported figures, gross margin improved from roughly 28% to 43%.
This margin gain is one of the strongest reasons to believe profitability is possible. OpenAI kept a larger share of each revenue dollar after paying the direct cost of serving customers, even while usage expanded sharply.
The next stage will be harder. Reasoning models spend more time thinking. Agents call tools and repeat steps. Voice, image and video products use more expensive infrastructure than plain text. Every technical saving creates room for a more demanding product.
OpenAI's current model lineup shows how it plans to control the bill. GPT-5.6 Sol targets the hardest professional work, Terra balances capability and cost, and Luna handles price-sensitive high-volume tasks. Routing a simple question to Luna instead of Sol can cut the input price by 80%.
The company can also use caching, batching, shorter contexts and custom hardware. Those tools reduce the cost of each unit of work. Profit arrives when OpenAI keeps some of the saving rather than spending all of it on longer answers, more reasoning and lower prices.
This chart, featured in our generative AI market deck, shows annual funding in generative AI startups
Will frontier research keep eating the profits?
Yes, unless OpenAI eventually decides that another model leap is less valuable than keeping part of the cash its products generate.
Research and development reached around $19.2 billion in 2025, equal to approximately 147% of revenue. OpenAI could have removed several other expenses and still lost money because research alone cost more than the entire business earned.
OpenAI chose a race with no natural finish line. Training runs require chips, power, data and specialist teams. After a successful model launches, the company immediately begins work on the next one. Anthropic, Google, Meta and well-funded new labs make a long pause dangerous.
Microsoft, Alphabet and Meta can fund frontier research with profits from cloud computing, advertising and mature software. OpenAI asks its AI products to finance the research needed to defend those same products. The loop becomes painful whenever model costs rise faster than commercial revenue.
Lately, the company has shown some willingness to concentrate its spending. The Financial Times reported that it shelved costly side projects and refocused on ChatGPT and enterprise products. The new push into Frontier, DeployCo and business partnerships also points toward work with a clearer route to revenue.
The number to watch is simple: how much of each revenue dollar still goes back into research. OpenAI could spend $25 billion a year on research and still become profitable if revenue reaches $70 billion with healthy margins. Profit remains distant if research climbs almost dollar for dollar with revenue.
Has OpenAI's infrastructure spending gone too far?
OpenAI's latest compute plan is now large enough to threaten the profitability of an otherwise excellent product business.
The Wall Street Journal reports that projected computing spending has risen to about $750 billion through 2030, up from roughly $600 billion earlier in the year. The revision came as OpenAI expanded cloud contracts and began taking more direct control of data-center development.
One new Georgia project illustrates the shift. OpenAI plans to commit about $20 billion and has contracted 3.2 gigawatts of power for the site. The company is also involved in Stargate developments, large cloud agreements and a Broadcom partnership covering ten gigawatts of custom accelerators.
We cannot add every announcement together. Some plans overlap, some money will come from partners, and several commitments depend on future deployment. The headline figures describe available or planned capacity more than a single bill payable by OpenAI tomorrow.
Even so, the direction is clear. OpenAI is moving from buying cloud services toward designing chips, reserving power and developing infrastructure. Custom hardware could lower Nvidia dependence and remove part of the supplier margin. It also gives OpenAI another expensive asset to fill.
Utilization now becomes the central test. A costly cluster running paid workloads around the clock can generate valuable capacity. The same cluster becomes a financial trap when demand misses the forecast or customers shift toward cheaper models.
OpenAI's product business can support enormous infrastructure. A $750 billion plan demands extraordinary utilization, pricing power and growth for several years. That standard is much harder to meet.
If you want more recent data on this point, please see our latest generative AI market report.
This chart, featured in our generative AI market deck, compares the main business model options for generative AI SaaS platforms
Does Microsoft make profitability easier or harder?
Microsoft made OpenAI's growth possible, but the updated deal still sends money and future purchasing power back to Microsoft.
Microsoft supplied early capital, cloud infrastructure and global enterprise distribution. Without Azure supercomputers and Microsoft's sales reach, OpenAI would have found it much harder to train large models and put them in front of major companies.
The relationship now carries a heavy financial cost. OpenAI has committed to purchase an additional $250 billion of Azure services. Revenue-share payments from OpenAI to Microsoft continue through 2030, subject to a total cap, while Microsoft has stopped making its own revenue-share payments to OpenAI.
Microsoft also keeps a license to OpenAI models and products through 2032. The license has become non-exclusive, and OpenAI can use other cloud providers, which gives the company more freedom than before. That freedom has already led to deeper relationships with Amazon, Oracle and other infrastructure partners.
Diversification should improve OpenAI's bargaining power and reduce dependence on one supplier. It can also leave the company carrying several huge commitments at once.
Microsoft is both an accelerator and a toll collector. The partnership helped create OpenAI's current revenue, while the cloud purchases and revenue share will absorb part of the profit that revenue may eventually produce.
How much revenue would OpenAI need to break even?
Using its latest full-year cost base, OpenAI probably needs roughly $40 billion to $60 billion of annual revenue before ordinary operating profit becomes realistic.
As seen above, the reported accounts leave about $26.5 billion of expenses outside direct delivery costs. We can use that figure to build a simple break-even range.
At the latest reported gross margin of roughly 43%, OpenAI would need around $62 billion of revenue to cover those other expenses. A 60% gross margin lowers the requirement to about $44 billion. At 70%, the threshold falls below $38 billion.
This calculation holds operating expenses still, which will not happen. Research, sales and infrastructure teams will continue growing. It also ignores future efficiency gains, so it may understate how much better the cost base could become.
The range gives us the right order of magnitude. OpenAI currently runs at about $2 billion of monthly revenue. Reaching $40 billion would require that pace to rise by roughly two-thirds. Reaching $60 billion would require it to rise by about 150%.
Those increases are achievable for a company growing this quickly. They become much less useful when each extra dollar of sales triggers another similar increase in research and compute commitments.
| Assumed gross margin | Revenue needed to cover about $26.5B of other expenses |
|---|---|
| 43% | About $61.6B |
| 50% | About $53.0B |
| 60% | About $44.2B |
| 70% | About $37.9B |
This chart, featured in our generative AI market deck, illustrates how revenue is distributed across customer segments in the generative AI market
When could OpenAI realistically become profitable?
The end of the decade is plausible, but only if today's revenue growth survives and the compute plan stops expanding.
Start with the current revenue pace of roughly $24 billion a year. Three years of 30% annual growth would lift it to about $53 billion. Three years at 40% would produce roughly $66 billion. Both outcomes fall inside or above our broad break-even range.
The company also has time. OpenAI recently raised $122 billion in committed capital at an $852 billion post-money valuation. That financing can cover several years of heavy losses and lets management avoid an immediate retreat from research or infrastructure.
Capital does not repair weak economics by itself. A company can burn through a huge funding round when spending rises with every new ambition. OpenAI has already missed some internal user and revenue targets, and its compute forecast has continued moving upward.
A realistic route to profit requires several changes together. Enterprise products need to become the largest revenue source. Gross margin should climb toward 60%. Ads and commerce must turn part of the free audience into revenue. Agents need pricing tied to the work they perform. Research and infrastructure spending must grow more slowly than sales.
OpenAI could report a profitable quarter earlier by delaying projects or trimming experiments. Sustainable yearly profit looks more credible near the end of the decade. A prolonged model race or another major infrastructure expansion could push it well beyond that point.
Can OpenAI ever become profitable?
Yes, OpenAI can become profitable, but its current strategy may keep delaying the moment on purpose.
OpenAI already has the ingredients of a profitable company. ChatGPT has a vast audience, tens of millions of subscribers and a rapidly growing business customer base. Enterprise products are becoming a larger part of revenue. Direct delivery margins have improved. OpenAI also has several new ways to earn money from agents, coding, advertising and commerce.
The company keeps reinvesting almost without a ceiling. Higher revenue becomes a reason to reserve more compute, train more ambitious models and enter more markets. This approach may produce the leading AI company. It can also postpone positive cash flow year after year.
We see a credible profitable business inside OpenAI today. It would charge more closely for work completed, route tasks toward cheaper models, keep research spending below gross profit and build infrastructure in stages as paid demand appears.
Management is currently pursuing a more aggressive version. The latest infrastructure plan assumes extraordinary growth and leaves little room for weak pricing, slower adoption or underused capacity.
Our final judgment is clear: OpenAI can become operationally profitable, probably near the end of the decade, if revenue keeps compounding and spending discipline finally appears. Profitability remains a choice as much as a technical destination. The company has proved that customers will pay. It still has to prove that it can stop spending the money faster than it arrives.
If you want more recent data on this point, please see our latest generative AI market report.
This chart, featured in our generative AI market deck, shows how AI video generation technology has evolved over time
OUR METHODOLOGY
This analysis tests whether OpenAI can become sustainably profitable, rather than whether it can produce one profitable quarter. We looked at recurring operating economics: revenue, direct delivery costs, gross margin, research spending, sales and administration, infrastructure commitments and the cash demands that remain after accounting adjustments.
We separated reported net loss, operating loss and estimated underlying cash burn because each answers a different question. The large revaluation charge tied to OpenAI's former investor structure was treated as an accounting distortion, while stock compensation, Microsoft compute credits and other adjustments were kept in view because they still affect the cost of running the business on fully commercial terms.
We used the latest reported full-year cost base to estimate the revenue needed for break-even at several gross-margin levels. The range is a planning test, not a forecast: it holds non-delivery expenses broadly constant so readers can see how strongly the answer depends on margin improvement and spending discipline.
Consumer subscriptions, advertising, enterprise deployments, APIs and agents were assessed separately because they carry different economics. We prioritized evidence that showed paying demand, usage, pricing, implementation effort or switching costs more directly than broad claims about market leadership.
Infrastructure announcements were not simply added together. Some commitments overlap, depend on future deployment or will be financed with partners, so we used them to judge the scale and direction of OpenAI's capacity strategy rather than treating every announced dollar as an immediate bill.
Key sources include OpenAI's disclosures on scale and usage, OpenAI's financing announcement, OpenAI's work on agents, OpenAI's explanation of its commercial model, OpenAI's enterprise AI report, OpenAI API pricing, OpenAI's Codex materials, reporting from the Financial Times and The Wall Street Journal, Menlo Ventures' enterprise AI research, and company announcements from Accenture, Snowflake, Amazon Web Services, Broadcom, Oracle and Microsoft.
In our generative AI market deck, we identify pain points entrepreneurs should prioritize
Related blog posts
- Can OpenAI actually afford all this compute?
- Does OpenAI still have a moat?
- Who is paying for OpenAI’s data centers?
