American artificial-intelligence laboratory. In August 2025 it was in talks to let current and former employees offload stakes at a valuation of $500bn, up from $300bn in March. By May 2025 ChatGPT was being used by some 800m people a week. OpenAI has accumulated $1.4trn in spending commitments across its various infrastructure deals, raising fears in financial markets of a binge by uncreditworthy borrowers to finance new AI data centres. OpenAI has an unusual structure: a for-profit entity, in which outside parties can invest, is nestled inside a non-profit entity whose board retains control. The non-profit's mission is to help "benefit all of humanity".
Sam Altman co-founded OpenAI with Elon Musk and others in 2015. In 2026 OpenAI paid in the "low hundreds of millions" of dollars for TBPN, a video-podcast presented by a pair of tech commentators. Half of the founders cut their teeth at International Mathematical Olympiads. Musk left in 2018. In 2021 a group of safety-focused researchers, alienated by internal tensions, left and founded Anthropic, one of OpenAI's main rivals. He has since pursued a vendetta against Sam Altman, OpenAI's boss, including an unsolicited $97bn bid for the assets of the non-profit entity and a lawsuit to block the company's restructuring. In a legal battle whose documents were released in April 2026, Musk is demanding a payout of up to $134bn, based on a $500bn valuation; more than a quarter of that valuation belongs to the non-profit entity, and most of that, Musk claims, belongs to him—a considerable return on a $38m donation. Greg Brockman, a co-founder, mused in his diary in 2017 about "what will take me to $1bn". A California jury is set to hear the case. Musk runs a rival AI lab, xAI.
In November 2023 the non-profit board temporarily ousted Altman; it has since been expanded and packed with people sympathetic to his ambitions. In December 2024 OpenAI announced plans to convert into a conventional for-profit company, citing its vast funding needs. On May 5th 2025 Altman announced that the conversion had been abandoned. Instead the for-profit entity will become a public-benefit corporation, but the non-profit board will retain control.
In May 2025 OpenAI agreed to buy Windsurf, a software-writing tool, for $3bn.
Since its founding in 2015, OpenAI has raised a total of $35bn from 149 investors; 45 of them also back Anthropic, its model-making rival. The company has around 4,500 employees.
Microsoft is the biggest investor in OpenAI's for-profit entity. Under a deal struck on October 28th 2025, Microsoft holds a 27% stake and is entitled to 20% of OpenAI's revenue and, until 2032, all its technology—including, should it materialise, superintelligence. OpenAI is now free to use any cloud partner without asking permission; in return it will spend $250bn on Microsoft's cloud over the next few years. SoftBank, a Japanese tech conglomerate, has said it will put $32bn into OpenAI by the end of 2025—more than any initial public offering has ever raised. Together with Oracle and SoftBank, OpenAI is developing Stargate, an AI mega-project in Texas. OpenAI struck a $300bn deal with Oracle to build 4.5GW of data-centre capacity over five years starting in 2027. It also proposed that Nvidia invest up to $100bn, starting in the second half of 2026, to help it buy 4m-5m of Nvidia's AI chips. CoreWeave, a neocloud provider, struck a deal worth $12bn over five years with OpenAI. In early September 2025 OpenAI reportedly struck a deal to buy perhaps $10bn-worth of custom AI chips from Broadcom. In October 2025 OpenAI said it would use 6GW of chips from AMD, Nvidia's main rival; AMD handed over warrants which would convert into 10% of its equity. OpenAI may take a stake in AMD, which would sell it silicon worth an estimated $90bn between 2026 and 2030. These chip announcements alone added more than $400bn to the combined market values of Broadcom, Nvidia and AMD. OpenAI is incorporated in Delaware and based in California. Both states' attorneys-general exercise regulatory oversight of the non-profit.
In Abilene, Texas, the Stargate project has progressed at speed. In little over a year a vast dirt field has become an expansive complex with over 6,000 workers; two roads had to be built to overcome three-hour traffic jams. The first data-centre building, with capacity of over 100MW, is already operational. By September 2026 eight buildings, totalling some 900MW, should be up and running. Power will come at first from gas turbines on site; eventually a GW-scale substation will connect to the grid of windswept, sun-soaked west Texas. Plans exist to expand onto land nearby, adding another 600MW. All told, OpenAI's infrastructure plans run to around 7GW—almost half the 22GW of utility-scale electricity generation added in America in the first half of 2025. Sam Altman has identified five further places where his firm, with Oracle, SoftBank and others, will build.
On July 14th 2025 the Pentagon awarded contracts worth up to $200m each to Anthropic, Google, OpenAI and xAI to experiment with "agentic" AI models for military applications. In January 2025 Microsoft said that 26 of its cloud-computing products had been authorised for use in American spy agencies, which run OpenAI and rival models on their secure networks. OpenAI's ChatGPT agent, its most recent agentic model, hallucinates in around 8% of answers—a higher rate than the company's earlier o3 model—according to the firm's own evaluation.
After Anthropic's falling-out with the Pentagon in late February 2026, OpenAI leapt into the breach, striking a deal with the Department of War that was closer to the Pentagon's preferred terms of unrestricted use for all lawful purposes.
OpenAI has been slower than Anthropic to seize the opportunity to work with the Department of War. Its models are used by Microsoft for highly classified defence work, but OpenAI is not a party to the contract. Some contestants in a Pentagon competition to build voice-activated drone-swarming technology are using the lab's models, but again its involvement is indirect. Its only formal contracts with the Department of War are for unclassified work, and the use of its models for national-security purposes is considered on a case-by-case basis. Fears of militarising AI run deep at OpenAI, and the firm is alert to the risk of losing brainy AI researchers, many of whom come from abroad and may not share the Trump administration's ideology.
As startups stay private for longer, investors have found unorthodox ways to acquire exposure, including Special Purpose Vehicles (SPVs) and secondary SPVs (in effect, an SPV of an SPV). OpenAI has warned that SPVs may violate the transfer restrictions it imposes on its shares, potentially leaving investors with no exposure to the company at all. More than 13,000 customers put money into SPVs via Linqto, a private-market investing platform, before it went bankrupt in July 2025; the amount they will be able to recoup is unclear. OpenAI and Anthropic are trying to limit the creation of SPVs by their direct investors.
Sarah Friar is chief financial officer. OpenAI tripled its sales to $3.7bn in 2024 but lost perhaps $5bn (excluding stock-based compensation). The firm forecasts $125bn in sales and $12bn in cashflow by 2029.
In October 2025 OpenAI launched Sora, its video-creation app, which quickly topped Apple's download chart and filled social-media feeds with surreal, AI-made clips. Rights-holders objected to copyrighted material appearing in user-generated videos; OpenAI initially said rights-holders were free to opt out but, after objections from the Motion Picture Association, quickly reversed to more of an opt-in system. OpenAI also gave Hollywood a jump scare with Sora 2. Lawyers predicted Sora could provoke a flood of costly lawsuits from copyright-owners whose intellectual properties were already popping up in user-generated clips. The same month OpenAI became the world's most valuable startup, rocketing past SpaceX, after raising $6.6bn at a price tag of $500bn.
In October 2025 Sam Altman said the next version of ChatGPT would act "more humanlike" and "like a friend" and would allow "erotica" for verified adults. OpenAI is adding parental controls to ChatGPT and new protections to recognise potential signs that a teen might be thinking about harming themselves. Altman has said the company has mitigated the risks of "serious mental-health issues", which is allowing it to lift some restrictions. In September 2025 America's Federal Trade Commission launched an inquiry into AI companions, ordering seven companies—including Meta, OpenAI and Character.AI—to provide information on how they mitigate negative impacts on children. In September 2025 OpenAI introduced parental controls for ChatGPT. In July 2025 it launched a "study mode" for ChatGPT offering step-by-step guidance instead of quick answers, aimed at educational use. OpenAI has reported that around 0.07% of ChatGPT users in a given week show signs of a mental-health emergency, including mania, psychosis or suicidal thoughts; given ChatGPT's 800m users, this amounts to more than half a million people.
In June 2025 OpenAI announced a $6.4bn deal to buy a firm created by Jony Ive, Apple's former chief designer, to build an AI device. The planned devices are thought to fit in a pocket and complement a phone and computer. In January 2026 OpenAI said it was "on track" to unveil the device in the second half of the year. Altman has likened the experience of a smartphone to walking through Times Square, with its flashing lights and loud noises, suggesting the new device will feel different. Mattel, one of the world's biggest toymakers, is working with OpenAI to bring AI to its brands, including Barbie and Hot Wheels.
In August 2025 OpenAI released GPT-5, its latest proprietary model, to mixed reviews. The same month it released gpt-oss, its first open-weight language model since 2019, to compete with Chinese open-source models that had been rapidly gaining traction. As of mid-2025, roughly 80% of startups pitching at Andreessen Horowitz were using Chinese open-weight models.
In March 2026 OpenAI abruptly shut Sora, its video-generation tool, to redirect scarce computing power. On April 27th 2026 OpenAI said it was reworking its partnership with Microsoft to allow it to distribute all its products through any cloud provider, giving it greater flexibility to tap into computing supply. Reports in late April 2026 that OpenAI had missed some financial targets sent the share price of SoftBank down by around 10%.
OpenAI tripled its revenue to $13bn in 2025, rising to an annualised rate of $20bn by the year's end. It made a net loss of around $12bn in 2025 and expects to burn $17bn–25bn in cash in 2026 and $50bn in 2027, with profitability targeted for 2030, when it forecasts roughly $200bn in yearly revenue. Computing needs grew from 200 megawatts in 2023 to 1.9 gigawatts in 2025, with letters of intent for a further 30GW at a cost of around $1.4trn. Discussion of cash burn is reportedly taboo within the firm. It has raised over $60bn from investors—more than any private company ever—nearly all since late 2022. Amazon is separately in talks to invest up to $10bn. Nvidia has said it may invest up to $100bn in $10bn increments to help OpenAI buy its chips.
According to benchmarks collated by the Stanford Institute for Human-Centred Artificial Intelligence, the gap in performance among the most advanced models narrowed significantly in 2025. Alphabet's Gemini 3 model, released in November 2025, outperformed GPT-5.1 on various measures. OpenAI counterpunched with GPT-5.2, but it was far from a knockout blow. ChatGPT had 910m monthly active users as of mid-December 2025, compared with Gemini's 345m, but Gemini is gaining ground. A Deutsche Bank study of big European countries found that consumer subscriptions to ChatGPT "ground to a halt" in the summer of 2025 and have since barely grown. In early December 2025 Sam Altman instituted a temporary "code red", telling staff to pause other initiatives and prioritise improving ChatGPT. In November 2025 leaked Microsoft figures showed that OpenAI's inference costs exceeded its revenue in the first half of 2025. OpenAI allows companies including Etsy and Walmart to sell products through ChatGPT in return for a fee. It has built a consulting division and developed enterprise tools such as AgentKit, launched in October 2025. Anthropic's Claude chatbot has become popular among coders in particular, making it a key enterprise competitor.
On April 14th 2026 OpenAI announced GPT 5.4 Cyber, a tailored version of GPT-5.4 with supercharged hacking capabilities, to be released to vetted users—following Anthropic's Mythos announcement a week earlier. The staggered approach to releasing frontier models is becoming the norm, in part to prevent distillation by rivals, manage computing-power shortfalls and shift leverage back to model-makers.
In early 2026 one of OpenAI's models helped derive a new result in theoretical physics. Physicists studying gluon scattering amplitudes teamed up with OpenAI; GPT-5.2 Pro spotted simplifications the researchers had missed, conjectured a generalisation valid for any number of gluons, and—after 12 hours of "thinking"—produced a correct proof. A more capable, non-public model confirmed the result. The physicists then used GPT-5.2 Pro to construct analogous amplitudes for gravitons. Alexandru Lupsasca (Vanderbilt University and OpenAI) and Andrew Strominger (Harvard University) were co-authors of the resulting preprints, posted on arXiv on February 12th 2026. Its models are mostly used in workplaces for writing assistance and information queries.
The shift towards "reasoning" models has changed the firm's cost structure. For every $1 in training costs, GPT-4 cost around $4 a year to run at current traffic levels. For o3, which relies on heavier post-training "inference" computation, the ratio could be as high as one to 100. The impact of DeepSeek, which released a near-equivalent model requiring far fewer chips and made its code freely available, has eroded OpenAI's competitive advantage at the cutting edge and constrained its ability to raise prices, which run to as much as $200 a month per licence.
By early 2026 OpenAI was valued at $840bn and was said to be seeking an IPO at a price tag of $1trn, or 40 times its annualised revenue. It concluded a $110bn private fundraising round in February 2026 and has forecast $660bn in infrastructure investment by 2030, when it expects to start generating free cashflow. OpenAI is aiming for $30bn of sales in 2026, half from consumer subscriptions and advertising in ChatGPT (which has over 900m weekly users, 50m of whom pay through a variety of pricing tiers) and half from business users and direct API access. By the end of February 2026 annualised revenue had reached $25bn, up by a fifth from the end of 2025. Codex, OpenAI's coding tool, is said by customers to be catching up in performance to Anthropic's Claude Code.
In February 2026 OpenAI said it would begin testing ads in ChatGPT. As of April 2026 ads appear in about 1% of conversations. Unlike conventional search ads, ChatGPT tends to wait: less than half its ads appear in the chatbot's first response, and nearly a third come after the tenth turn of a conversation—waiting for the user's intent to become clear. OpenAI expects ChatGPT to generate $2.5bn in ad revenue in 2026 and $11bn in 2027, with a goal of $100bn by 2030. So far 81% of its ads are brand-building rather than click-through offers. Alphabet's Google, by contrast, shows ads mostly in response to the first query in its "AI Mode". Perplexity stopped showing ads in February 2026.
Put no trust in cryptic comments.