Can OpenAI actually afford all this compute?

How much will OpenAI spend on compute by 2030?

OpenAI is expected to spend somewhere between $600 billion and $856 billion on compute and infrastructure through 2030, depending on the source. Its revenue pace was about $50 billion a year in September 2026.

The estimates have moved around. In February 2026, CNBC reported that OpenAI told investors its compute target was about $600 billion. In July 2026, the Wall Street Journal reported that the projection had risen to around $750 billion. In September 2026, the Financial Times reported a company presentation showing about $856 billion for computing power and infrastructure through 2030. The sources do not agree, and the categories may differ. Sources that disagree on the compute bill show how hard it is to price; we break down training and running costs in our report on frontier AI labs.

Project Camellia shows how far the strategy has moved. OpenAI is designing and developing the Georgia data-center project itself, with a reported $20 billion commitment and 3.2 gigawatts of power scheduled to arrive in phases between 2028 and 2032. OpenAI says it will cover the infrastructure and electricity-service costs rather than passing them to local ratepayers.

Signed capacity and working capacity are different things. Sachin Katti, OpenAI’s compute chief, said in October 2026 that the company had secured double-digit gigawatts of new capacity across sites in Ohio, Georgia and elsewhere. But Epoch AI estimated in April 2026 that only about 0.3 gigawatts of the seven US Stargate sites was operating, all in Abilene, Texas, with most other sites due in late 2028.

Execution is also under pressure. Katti will leave on November 15, 2026, after earlier exits from OpenAI’s Stargate team. The Wall Street Journal reported his exit amid concerns about Stargate execution. OpenAI also paused its proposed Stargate UK project in April 2026. In October 2026, it hired Will Friedman from AWS as its energy regulatory lead, Data Center Dynamics reported, but no replacement for Katti has been named.

The financial risk comes from the length of the commitments. OpenAI can cancel a research experiment or reduce hiring fairly quickly. Data-center leases, chip orders, power contracts and dedicated cloud capacity can run for five, eight or even twenty years. Once construction starts, stepping away becomes expensive.

What does “afford” really mean for OpenAI?

OpenAI can fund the next few years of compute today, while sustainable affordability remains unproven.

The word “afford” hides three separate questions.

The first concerns cash. Can OpenAI make the payments that come due soon? Its March 2026 funding round, the last $10 billion installments of which Nvidia and SoftBank reportedly paid on October 1, 2026, makes that likely.

The second concerns access to capital. Will investors, banks, cloud companies and chipmakers keep financing OpenAI as the infrastructure expands? They still are, but on tighter terms. In late October 2026, OpenAI reportedly lowered its new funding target to about $20 billion, and Broadcom’s chip financing has not closed.

The third question is much harder. Can ChatGPT, Codex, enterprise agents, the API, advertising and future products eventually produce enough gross profit to cover the full cost of running OpenAI?

OpenAI currently passes the first two tests and has not yet passed the third. Its own forecast, as reported by the FT in September 2026, shows negative free cash flow every year through 2030.

A rival now offers a useful comparison. Anthropic’s public IPO filing, released on October 26, 2026, lists about $265.6 billion of fixed or minimum compute commitments; Reuters puts its broader compute deals above $600 billion. The filing shows a $974 million operating profit in Q2 2026, and Anthropic reportedly told investors it earned an 11% adjusted operating margin in Q3 and about $1.70 of revenue per $1 of compute. OpenAI has not disclosed a comparable number.

Capital providers can now ask OpenAI for the same kind of proof.

How much compute has OpenAI already promised to buy?

OpenAI has publicly announced at least $710 billion of cloud contracts with four providers, and its financial statements reportedly show about $665 billion of purchase commitments.

The largest reported contract is OpenAI’s agreement to buy around $300 billion of Oracle computing capacity over roughly five years, beginning in 2027.

Microsoft is the second-largest piece. In October 2025, OpenAI agreed to buy an additional $250 billion of Azure services. Microsoft remains OpenAI’s primary cloud partner, although an April 2026 amendment lets OpenAI serve its products across other clouds. Microsoft’s October 2026 quarterly filing shows $10.0 billion of revenue from OpenAI in the September quarter, including revenue sharing. That is not a clean compute bill, but it shows the scale of one supplier’s invoices.

OpenAI has also expanded its AWS commitment to about $138 billion. The relationship began with a $38 billion agreement and later added $100 billion of AWS Trainium capacity over eight years.

CoreWeave says its agreements with OpenAI have reached as much as $22.4 billion.

Adding those four figures gives $710.4 billion. That is a simple sum, not a clean total. Some capacity may overlap, and the contracts run over different periods.

The reported $665 billion figure comes from The Information’s review of OpenAI’s financial statements, published in June 2026. It covers Microsoft, Oracle, Amazon, Stargate and Fluidstack. The figure is unaudited, and its date is unclear, so it does not exactly match the public deal list.

New leases sit on top. In August 2026, OpenAI signed a 20-year lease for an Ohio campus built by SB Energy, starting at 4.25 gigawatts with an option to reach 8 gigawatts. Hardware programs with Nvidia, AMD and Broadcom create yet another layer, but much of that hardware will sit inside these same facilities. Reuters, citing the FT, reported in October 2026 that the new funding round could come with an option to lease new computing capacity in the UAE. Its size and timing are not disclosed, so it cannot be added to any total.

OpenAI has committed to several hundred billion dollars of compute. No one has published a verified, deduplicated total.

Provider	Publicly identified commitment	Main purpose
Oracle	Approximately $300B	Large Stargate cloud deployments
Microsoft Azure	Approximately $250B (incremental, October 2025)	Primary cloud capacity and services
AWS	Approximately $138B	Nvidia and Trainium capacity
CoreWeave	Up to approximately $22.4B	Training and advanced AI workloads
Combined	Approximately $710.4B (simple sum)	Not deduplicated; excludes the Ohio lease and hardware programs

Who is paying upfront for OpenAI’s data centers?

OpenAI’s partners are carrying most of the upfront construction bill, while OpenAI commits to years of future payments.

The numbers make this clear. As of March 31, 2026, OpenAI reportedly had no debt, lease liabilities under $750 million and only $46 million of quarterly spending on buildings and equipment. The assets, and the debt behind them, sit with its partners.

Cloud providers and infrastructure developers raise the money for buildings, chips, cooling, power connections and networking. OpenAI then pays through cloud contracts, capacity reservations, leases and usage commitments.

The Ohio deal shows the newest version of this model. Nvidia agreed in August 2026 to provide residual-value guarantees on the lease, reportedly capped at about $105 billion. Nvidia pays only if OpenAI defaults on the lease or becomes insolvent, and OpenAI must reimburse it. SB Energy builds, owns and operates the campus, with capacity arriving in phases from 2028. Nvidia also becomes the site’s exclusive compute supplier.

Custom chips are harder to finance. In May 2026, The Information reported that OpenAI struggled to close about $18 billion of financing for an early tranche of its Broadcom chips. In October 2026, Broadcom was reported to be arranging more than $50 billion of private credit from lenders such as Apollo and Blackstone. Those talks were early and not closed.

Oracle shows the risk on the other side of these deals. Oracle reported $638 billion of remaining performance obligations, meaning contracted revenue it has yet to recognize. S&P Global estimates that roughly half of that backlog is tied to OpenAI.

On July 9, 2026, S&P cut Oracle’s credit rating to BBB-, the lowest investment-grade level. It now expects Oracle to burn nearly $42 billion of free operating cash in fiscal 2027. On October 8, 2026, Oracle shares fell 4.5% after a report of a possible $20 billion revenue shortfall tied to OpenAI. That report has not been confirmed.

OpenAI has persuaded suppliers to finance a remarkable share of its expansion. Their willingness to carry that risk is still one of OpenAI’s most valuable financial assets, but the cost of that willingness is rising.

How much revenue is OpenAI making right now?

OpenAI was generating close to $50 billion of annualized revenue in September 2026, according to reports from Reuters and the Financial Times.

That figure needs care. Earlier press reports put the September pace near $70 billion. The gap appears to reflect how revenue shared with cloud partners is counted, not a sudden drop in sales. OpenAI reportedly told investors it targets $70 billion by December 2026. That is a goal, not a result. Revenue counted two ways makes any year-on-year comparison hard; we go through what has changed over the past year in our frontier AI labs market report.

Annualized revenue describes the current monthly rate multiplied across a year. It should not be confused with revenue actually recognized. OpenAI reportedly recognized $5.7 billion in Q1 2026 and $6.7 billion in Q2 2026. The FT reported a forecast of $36 billion for the full calendar year.

The customer base is enormous. OpenAI said ChatGPT passed 1.2 billion weekly users on October 7, 2026, when it rolled out GPT-6 more broadly. It also reports more than 50 million paying subscribers, a figure last updated in March 2026.

Business customers contributed more than 40% of revenue in March 2026. OpenAI said in September 2026 that it serves more than 2.5 million businesses.

Advertising has grown fast. OpenAI said on August 31, 2026, that ChatGPT Ads had reached $1 billion of annualized revenue in less than 200 days, with tens of thousands of advertisers. That remains small beside subscriptions, but it shows the free audience can earn money.

The rival comparison has turned. Anthropic’s IPO filing shows about $11.6 billion of Q2 2026 revenue, against a reported $6.7 billion for OpenAI, and Anthropic reportedly reached $16.8 billion in Q3. The two companies count cloud-partner revenue differently, so the gap is not exact. But OpenAI is no longer clearly the largest frontier lab by revenue.

OpenAI has the revenue scale of a major software business. Its compute commitments belong to a much larger financial category, closer to a global cloud provider or energy company.

Has more compute really produced more OpenAI revenue?

Yes, through 2025. OpenAI’s own three-year figures show available compute rising 9.5 times while annualized revenue rose about ten times.

Available capacity increased from 0.2 gigawatts in 2023 to 0.6 gigawatts in 2024 and approximately 1.9 gigawatts in 2025. Annualized revenue climbed from $2 billion to $6 billion and then more than $20 billion.

That works out to around $10 billion of annualized revenue per gigawatt in each period.

The 2026 data point is missing. Revenue has since reached about $50 billion annualized, but OpenAI has not published how many gigawatts it had available in 2026. To keep the same ratio, it would need about 5 gigawatts in use. Epoch AI’s April 2026 estimate shows only a small part of the new Stargate capacity running so far, so most of that would have to come from Microsoft, Oracle, AWS and CoreWeave clouds.

The pattern supports OpenAI’s claim that limited compute has constrained growth. ChatGPT has faced usage caps, API customers have hit capacity limits and new products have often launched gradually.

Still, three observations cannot establish a permanent rule. One gigawatt used for mature inference may produce a very different return from one used for training. Newer chips also deliver more useful computation from the same amount of electricity.

The real test arrives from 2028, when Ohio, Georgia and most Stargate sites are due to come online. OpenAI then needs to repeat its early performance across a fleet several times larger.

Year	Available compute	Annualized revenue	Annualized revenue per gigawatt
2023	0.2 GW	$2B	Approximately $10B
2024	0.6 GW	$6B	Approximately $10B
2025	Approximately 1.9 GW	More than $20B	More than $10.5B
September 2026	Not disclosed	About $50B (reported)	Not calculable

Can OpenAI survive a $278 billion cash burn?

OpenAI can survive its current losses for a few years, but its own forecast shows it will need more outside money well before it becomes self-funding.

The September 2026 presentation reported by the FT expects about $278 billion of negative free cash flow from 2026 to 2030. Reuters coverage said OpenAI was on track to use up the cash from its March round by 2028. No verified break-even year has been published.

The history explains the scale. OpenAI reportedly produced an operating loss of around $20.9 billion on $13.1 billion of revenue in 2025, after a loss of about $8.8 billion on $3.7 billion of revenue in 2024. Revenue grew faster than the loss, but the absolute loss still rose by more than $12 billion.

The latest detailed quarter is Q1 2026, as reported by The Information. OpenAI recorded $5.7 billion of revenue, a $3.7 billion cash burn and an operating loss of about $9.3 billion including stock-based pay. Its net loss topped $21 billion, but about $12.4 billion of that was a non-cash accounting charge on convertible securities.

The gross margin was about 39%, up from about 33% a year earlier. That means OpenAI keeps around 39 cents from each dollar of product revenue after direct serving costs, before research, staff and sales. Mature software companies often keep 70% to 90%.

OpenAI’s cushion is still large. The March 2026 round brought $122 billion of committed capital at an $852 billion valuation, and Nvidia and SoftBank reportedly paid their final installments on October 1, 2026. OpenAI reported more than $73 billion of cash after Q1 and an undrawn credit facility of about $4.7 billion.

No verified Q3 2026 cash balance has been published. The burn will rise as new capacity arrives from 2028. OpenAI has runway today, but its next raise is shrinking: the reported target fell from $30 billion to about $20 billion in October 2026. A smaller round does not mean sales are falling, but it leaves less room for error.

Are Nvidia, Amazon and AMD financing their own customer?

Yes, and the loop has grown tighter in 2026. Real customer demand still gives the structure economic substance.

Amazon has fully funded its $50 billion investment in OpenAI, as its October 2026 quarterly filing confirms, while securing a cloud relationship worth around $138 billion. Amazon gains a major AWS customer, demand for Trainium chips, OpenAI products on Bedrock and custom models for Amazon services. Suppliers backing their own customer blurs the funding picture; we show who gets funded, and by whom, in our frontier AI labs business models report.

Nvidia plays three roles at once. It invested in OpenAI’s March 2026 round, reportedly completing a final $10 billion installment on October 1. In 2025, it signed a letter of intent to invest up to $100 billion as OpenAI deploys up to ten gigawatts of Nvidia systems; how much of that plan still stands has not been confirmed. And it now guarantees up to $105 billion of the Ohio lease, where it is the exclusive chip supplier.

AMD issued OpenAI warrants covering as many as 160 million AMD shares. They vest as OpenAI deploys up to six gigawatts of AMD hardware and as AMD reaches commercial and share-price milestones. The first gigawatt was targeted for the second half of 2026. We found no public confirmation that it is running. AMD reports its third-quarter results on November 3, 2026.

These arrangements give OpenAI capital or potential equity value that helps it buy infrastructure from the same companies. The suppliers gain sales, long-term customers and a stake in OpenAI’s success.

OpenAI has more than a billion weekly users, a $1 billion ads business, millions of business customers and heavy API demand. There is real activity beneath the financial engineering.

A downturn would expose the circularity quickly. Falling OpenAI demand would hit the suppliers’ investments, weaken their largest future customer and leave new data centers underused at the same time. Nvidia’s Ohio guarantee makes that link explicit: it pays only if OpenAI cannot.

Can multiple chip suppliers lower OpenAI’s compute costs?

OpenAI’s multi-chip strategy should lower long-term compute costs, provided the company can run workloads efficiently across very different systems.

OpenAI once depended heavily on Microsoft Azure and Nvidia GPUs. It now uses Microsoft, Oracle, AWS, CoreWeave and Google Cloud, alongside several chip architectures. It also plans 750 megawatts of Cerebras inference capacity through 2028.

Its announced hardware programs include up to ten gigawatts of Nvidia systems, six gigawatts of AMD GPUs, around two gigawatts of Amazon Trainium capacity and ten gigawatts of OpenAI-designed accelerators developed with Broadcom.

Those figures overlap. A Broadcom or Nvidia system can sit inside a Microsoft, Oracle or OpenAI data center. The announcements describe supplier relationships, not a clean total of separate physical capacity.

The variety still gives OpenAI stronger bargaining power. Frontier training can use the most capable hardware. High-volume inference can move to cheaper systems. Less urgent jobs can run when capacity and electricity are cheaper.

OpenAI’s Jalapeño chip is the clearest example. OpenAI unveiled the inference accelerator with Broadcom in June 2026 and expects first deployment by the end of 2026. Early lab tests reportedly show much better performance per watt than current systems, but final data has not been published. On October 27, 2026, Oracle said it plans to deploy Jalapeño to expand efficient inference capacity for OpenAI. That is a plan: Oracle gave no capacity, start date or cost savings.

The chip also shows the funding limit. The broader Broadcom program needs very large outside financing, and the $50 billion-plus credit package reported in October 2026 has not closed. Cheaper chips only help if someone pays to build them.

The challenge is software. OpenAI must keep models, kernels, networking and scheduling efficient across Nvidia, AMD, Trainium, Cerebras and its own processors. Poor utilization can erase the savings from cheaper hardware.

The strategy should reduce OpenAI’s cost per useful task over time. It also gives the company an escape route if one supplier’s prices, delivery or roadmap becomes unattractive.

Can cheaper OpenAI tokens rescue the business?

Cheaper tokens help, but agents are raising total token use faster than prices fall, and rivals are cutting prices too.

The original GPT-4 API cost $30 per million input tokens and $60 per million output tokens. OpenAI’s cheapest current model, GPT-6 Luna, launched in September 2026 at $0.10 for input and $0.50 for output.

Headline input prices have therefore fallen by about 300 times and output prices by about 120 times. The models differ in capability, but the trend is clear. On October 26, 2026, OpenAI added GPT-6.1 Terra, a capable agent model for computer use, at $1 per million input tokens and $5 for output, half the rate of GPT-6.1 Sol.

The top of the range is still expensive. GPT-6 Astra, OpenAI’s frontier model, lists at $10 per million input tokens and $50 for output. Google’s Gemini 4 Argon launched on October 13, 2026, at an introductory $2 and $10. DeepSeek’s V4.1 Flash costs as little as $0.15 and $0.60 off-peak.

That pressure is the real problem. OpenAI cannot raise prices to cover its compute bill when Google and Chinese labs offer strong models for much less. On OpenRouter, in the week of October 12, 2026, DeepSeek handled 24.4% of text requests, against 18.5% for OpenAI.

Agents push in the other direction. OpenAI has said its APIs process more than 15 billion tokens per minute; we found no newer figure. Agents read files, search, plan, call tools, check results and keep working for hours, using far more tokens than a chat reply. In late October 2026, OpenAI also doubled some Pro usage allowances, cancelled a planned cut to its Pro 200 plan and added API credits to eligible subscriptions. Customers get more work for the same fee, which likely means more compute served per subscriber.

A falling token price can therefore produce a rising total compute bill. Electricity became cheaper and society used more of it. Computing followed the same pattern for decades.

OpenAI needs the cost of completing a useful task to fall faster than the price customers are willing to pay for that task. Token efficiency alone cannot settle the equation.

Will enterprise agents make OpenAI’s compute affordable?

Enterprise agents are still OpenAI’s best route to affordable compute, because companies pay for completed work. But Anthropic leads that market today.

A business compares an AI agent with the cost of an employee or contractor, which gives OpenAI more room to charge according to value. An agent that saves an engineering team hundreds of hours can justify thousands of dollars of compute. A chatbot that answers casual questions cannot.

OpenAI’s revenue mix is moving that way. Enterprise products and APIs made up more than 40% of revenue in March 2026, and OpenAI expects business revenue to match consumer revenue. Enterprise products making up over 40% of revenue raises the question of who pays, and how much, which we walk through in our frontier AI labs business models report.

Codex is the early test. It passed 5 million weekly users in June 2026. In OpenAI’s own enterprise data that month, Codex produced 64% of the output tokens across Codex and ChatGPT. Coding agents are expensive to run, but successful tasks create far more value than chat.

New deals show traction. On October 6, 2026, Atlassian expanded its partnership so that GPT-6 models power its enterprise agents. OpenAI Frontier also lets businesses deploy agents with shared context and permissions, with AWS distributing it. In October 2026, Uber began rolling out ChatGPT Work to more than 30,000 employees after a 1,000-person pilot, Oracle started selling OpenAI’s API, Codex and Work through Oracle Marketplace in the US, and ServiceNow expanded its partnership. None of these announcements disclosed contract value or margin.

The competition is strong. In a JetBrains survey from May to July 2026, 39% of professional developers used Claude Code at work, against 16% for Codex. Menlo Ventures estimated that Anthropic held 40% of enterprise LLM API spending in 2025, against 27% for OpenAI. Most enterprises also use several model providers, which keeps prices down.

Enterprise growth improves the economics only when contract value rises faster than compute use. OpenAI has not yet shown that it can win this market on price and quality at the same time.

What happens if OpenAI’s growth slows?

OpenAI’s compute plan becomes uncomfortable if revenue growth falls much below about 75% a year, which is roughly what its own forecast assumes.

We ran a simple stress test. We started from the reported 2026 revenue forecast of $36 billion and calculated cumulative revenue from 2026 to 2030 under three growth rates.

OpenAI’s own path, as reported by the FT, goes from $36 billion in 2026 to $350 billion in 2030, or about $840 billion in total. That requires about 76% growth every year for four years.

At 50% annual growth, cumulative revenue reaches about $474.8 billion. That would be exceptional for almost any company, yet it covers well under the lowest compute estimate of $600 billion.

At 75% annual growth, cumulative revenue reaches about $739.8 billion. That is close to the WSJ’s $750 billion compute figure, but below the FT’s $856 billion, and it leaves nothing for staff, research or sales.

Doubling revenue every year produces about $1.12 trillion. Only then does revenue clearly exceed every compute estimate, and revenue still cannot all go to cloud providers.

Growth can slow fast. Anthropic’s quarterly revenue grew about 147% from Q1 to Q2 2026, then about 45% from Q2 to Q3. The FT also reported that at least 47 major US data-center projects stalled or were canceled in 2026. Capital can also tighten while sales still grow: OpenAI’s reported funding target fell by a third in October 2026.

OpenAI may delay or renegotiate parts of the plan. Some spending depends on deployment, and partners finance much of the physical assets. But leases such as Ohio run for twenty years.

Annual revenue growth	Approximate five-year cumulative revenue	What it would mean
50%	$474.8B	Massive business, still well below planned compute spending
75%	$739.8B	Roughly OpenAI’s own plan; about equal to compute alone
100%	$1,116B	Clearly above compute projections before other expenses

Can a smaller funding round or an IPO close OpenAI’s compute gap?

No. New funding can pay for the next stage of the buildout, but the current round is shrinking and an IPO is not close.

OpenAI confidentially submitted a draft registration statement to the SEC in June 2026. It has since deferred its IPO plans. Reuters reported in late October 2026 that OpenAI is pursuing a listing in 2027. No offer size, price range or date has been published.

In the meantime, the private round has shrunk. OpenAI first sought up to $30 billion at a $1.4 trillion pre-money valuation. By October 30, 2026, Reuters, citing the FT, reported a target of about $20 billion at about $1.25 trillion, and the Wall Street Journal reported about $11 billion of commitments a day later. Reports disagree on whether $1.25 trillion is pre-money or post-money. The round has not closed, and the last completed valuation remains $852 billion, from March 2026.

The scale of the compute plan limits what any raise can achieve. A $20 billion round would cover only about one-fourteenth of the $278 billion of negative free cash flow OpenAI reportedly expects through 2030. The $11 billion committed so far covers about 4%.

An IPO would be bigger but still partial. At the $852 billion valuation, selling 10% of the company would raise about $85 billion before fees. At the proposed $1.25 trillion, the same stake would be worth about $125 billion. That would be one of the largest equity raises ever, and it would still cover only part of the plan.

Repeated raises would dilute existing shareholders. Public investors would also demand clear answers on gross margins, cash burn, long-term contracts and the return on each new gigawatt. Anthropic publicly filed for its IPO on October 26, 2026, and is reportedly targeting a mid-November listing. Its filing, with dated compute minimums and quarterly operating results, already sets a public benchmark for those questions.

Funding buys time. It cannot turn low-return infrastructure into high-return infrastructure.

Can OpenAI actually afford all this compute?

Yes, OpenAI can afford its compute expansion for now, although it depends more than ever on investors and partners to carry the risk.

OpenAI has enough cash, committed capital, credit and supplier support to keep building through the next phase. An immediate funding crisis looks unlikely.

The business underneath is real. OpenAI was earning close to $50 billion a year in September 2026, serves more than 1.2 billion weekly users and has a growing ads and enterprise business. Its first 1.9 gigawatts of compute came with a tenfold rise in revenue.

The problem is scale. OpenAI expects to burn about $278 billion through 2030, and its reported purchase commitments reach about $665 billion. Its newest projects, such as Georgia and the 20-year Ohio lease, tie it to infrastructure for decades.

Partners have absorbed much of the upfront burden. Oracle, Amazon, Microsoft, CoreWeave, SoftBank, SB Energy, Nvidia, AMD, Broadcom and private lenders are financing assets OpenAI could not buy alone. But Oracle’s downgrade, Broadcom’s unfinished chip financing, a funding round cut from $30 billion to about $20 billion and the departure of OpenAI’s compute chief show the strain on that model.

That model works while OpenAI keeps growing and capital providers stay confident. It starts to crack if revenue slows, margins stay low or lenders demand stronger guarantees.

As seen above, OpenAI’s own plan needs about 76% growth a year just to match its compute bill. It also needs enterprise agents to win against Anthropic, custom chips to cut unit costs and prices to hold up against Google and Chinese rivals.

Our judgment is simple: OpenAI can finance all this compute today, while its ability to pay for it from its own business remains unproven.

OpenAI is making a survivable bet rather than a comfortably affordable one. If revenue keeps compounding near its recent rate, the compute will look visionary. A sustained slowdown would leave OpenAI and its partners with one of the largest infrastructure mismatches ever created by a private company.