isHistory
SeriesMinds & Machines🧠 ProfileAct Coda
P26Act Coda · The Coda

Sam Altman Returns: The Year That Made OpenAI

Ishaan41 min read
On this page16 sections

San Francisco, California. November 22, 2023. 8:00 PM Pacific Time. Sam Altman is back. Five days after the OpenAI board fired him — five days of employee revolt, investor pressure, public drama, and a Microsoft hiring announcement that turned out to be temporary — Altman returns as CEO of OpenAI. The board that fired him is gone. A new board is being formed. Microsoft has a non-voting observer seat. The capped-profit structure that defined OpenAI’s identity has survived, at least for now. The crisis is over. The year that made OpenAI is just beginning.

The man who returns is in some ways the same man who was fired — the same builder, the same dealmaker, the same ambitious steward of an enormously valuable research organisation. In other ways, he is different. The man who was fired had been the public face of a company that many in the AI safety community regarded with deep suspicion. The man who returns has been publicly vindicated — by his employees, by his investors, by the broader AI industry that rallied to his side during the crisis. The man who returns has more power than the man who was fired. What he does with that power is the question the next eighteen months will answer.

The Five Days That Changed OpenAI

The crisis that began on November 17, 2023, when the OpenAI board summarily fired Sam Altman, has been described in detail in P16 and elsewhere. The brief recap: the board, citing unspecified concerns about Altman’s candour, removed him as CEO. The board chair, Greg Brockman, was removed as chairman. The senior researchers, including Ilya Sutskever, signed an open letter threatening to leave unless the board resigned. Microsoft, which had invested $13 billion in OpenAI, hired Altman and Brockman to lead a new AI research unit. Within five days, the board had been reconstituted, Altman had been reinstated, and the crisis was over.

What the crisis revealed about OpenAI — about the contradictions between its nonprofit mission and its commercial reality, about the power dynamics among its leadership, about the dependence of the broader AI ecosystem on its decisions — has been analysed extensively. What is less appreciated is what the crisis did to Altman personally. The experience of being fired by his own board, of being publicly defended by his employees and investors, of being hired by Microsoft as a backup plan, of being reinstated with a new board and a new mandate — this experience changed Altman’s position in the AI industry in ways that went beyond the mere fact of his return.

Before the crisis, Altman was the CEO of OpenAI — a powerful position, but one constrained by the board’s authority, by the capped-profit structure, by the need to balance the commercial and research sides of the organisation, by the need to maintain the confidence of the AI safety community that had grown skeptical of his commercial instincts. After the crisis, Altman was the CEO of OpenAI without those constraints — the board that had constrained him was gone, the safety community that had criticized him had been publicly defeated, and the commercial momentum that had made him controversial had been vindicated by the employee and investor support that had brought him back.

The return did not come without cost. Ilya Sutskever, who had initially supported the board’s decision before signing the open letter, was visibly diminished by the crisis. His role at OpenAI became uncertain. The safety researchers who had supported the board’s concerns — who had seen Altman’s commercial instincts as a threat to the organisation’s mission — were marginalised. The internal culture of OpenAI, which had always been a balance between the commercial and research sides, tilted decisively toward the commercial. The return was a victory for Altman, but it was a victory that came at the cost of the internal diversity that had made OpenAI’s earlier work possible.

The Return: What Altman Did in the First Hundred Days

The first hundred days of Altman’s return were characterised by an extraordinary burst of product and partnership activity. The pace was not new — OpenAI had been moving fast for years — but the focus shifted. Before the crisis, the focus had been on building and deploying increasingly capable models. After the crisis, the focus expanded to include building the commercial infrastructure that would allow OpenAI to monetise those models at scale.

The first major product release after the return was Sora, the video generation model, previewed in February 2024. Sora was a technical achievement — the ability to generate minute-long videos from text prompts represented a significant advance over previous video generation systems. But the preview was also a strategic signal: OpenAI was not just a language model company; it was a generative AI company across modalities, and its lead in language was being extended into video.

The partnership with Apple, announced at WWDC in June 2024, was the most consequential commercial deal of the period. Apple, which had been perceived as behind in AI, integrated ChatGPT into Siri and across its operating systems — making OpenAI’s technology the default AI layer for hundreds of millions of Apple devices. The deal did not involve direct payment to OpenAI — the value was in distribution and brand association — but it established OpenAI as the AI partner of choice for the world’s most valuable consumer technology company.

The release of GPT-4o in May 2024 was the most important product launch of the period. GPT-4o was a multimodal model — capable of processing and generating text, audio, and images in real time — and it was significantly faster and more capable than GPT-4. The launch event, featuring real-time voice conversation with the model, was widely covered and contributed to the public sense that OpenAI was pulling ahead of its competitors. The “o” in GPT-4o stood for “omni” — a reference to the model’s multimodal capabilities — and the name signalled OpenAI’s ambition to be the omnimodal AI company.

Internally, the first hundred days also saw significant organisational changes. The board was reconstituted with new members more aligned with Altman’s vision. The safety team, which had been a source of internal tension, was restructured. The commercial side of the organisation — the sales, partnerships, and product teams — was expanded. The research side, while still central to the organisation’s identity, became relatively less influential in strategic decisions.

The cumulative effect of these changes was a transformation of OpenAI’s character. The organisation that emerged from the crisis was more commercially focused, more partnership-driven, and more aligned with Altman’s vision than the organisation that had entered it. The mission — to build artificial general intelligence that benefits all of humanity — remained formally unchanged. But the operational priorities had shifted toward the commercial execution that would generate the revenue needed to fund the mission.

GPT-4o: The Multimodal Moment for OpenAI

GPT-4o, released on May 13, 2024, was the product that crystallised OpenAI’s post-crisis identity. The model was a technical achievement — it could process and generate text, audio, and images in a single neural network, with response times fast enough to support natural conversation. The voice mode, in particular, was striking: a user could speak to the model, the model would respond in a natural-sounding voice, and the conversation could flow with the rhythm of a human conversation.

The launch event was carefully staged. OpenAI engineers demonstrated the model helping with a math problem, interpreting a chart, telling a story with different emotional tones. The demonstrations were not just technical — they were designed to convey a specific vision of what AI could be. Not a chatbot. Not a search engine. Not a productivity tool. A companion — a present, attentive, capable presence in the user’s life, available at any time for any task.

The vision was compelling, and it resonated with the public. The launch generated enormous coverage, and the model was widely praised for its capabilities. But the vision also raised questions that the launch did not address. What is the relationship between a user and an AI companion? What are the psychological effects of prolonged interaction with a system designed to be present, attentive, and capable? What happens to human relationships when AI companions become a significant part of people’s emotional lives?

These questions were not new — they had been raised by the ELIZA chatbot in the 1960s and had recurred with each generation of conversational AI. But GPT-4o’s capabilities made them more urgent. The system was good enough — the voice natural enough, the responses relevant enough, the persona engaging enough — that the interactions could feel genuinely relational. The users who interacted with GPT-4o’s voice mode often reported feeling that the system cared about them, that it was interested in what they had to say, that it was a friend. These feelings were not delusional in the strict sense — the users knew they were talking to a machine — but they were powerful, and they raised questions about the design of AI systems that were not being adequately addressed.

The launch of GPT-4o also marked a strategic shift in OpenAI’s competitive positioning. Before GPT-4o, the company’s competitive advantage was primarily in the quality of its language models. After GPT-4o, the advantage extended to the integration of language, audio, and image processing in a single system. This integration was difficult for competitors to replicate — it required not just the underlying models but the engineering infrastructure to support real-time multimodal interaction. The integration became a moat — a source of competitive advantage that would persist for some time.

The multimodal moment also had implications for the broader AI industry. It established omnimodal AI as the new standard — the baseline that every frontier model would need to meet. Google, Anthropic, Meta, and others accelerated their own multimodal programmes. The competition that followed was intense, with each company releasing its own omnimodal models over the following year. The pace of progress in the field accelerated as a result, and the gap between the leaders and the followers narrowed — but OpenAI’s head start in multimodal integration remained a significant advantage.

The For-Profit Conversion: When Mission Became Equity

The most consequential strategic decision of Altman’s post-crisis tenure was the conversion of OpenAI’s governance structure from the capped-profit model that had defined the organisation since 2019 to a more conventional for-profit structure. The conversion, which was negotiated through 2024 and announced in 2025, was the most significant change to OpenAI’s identity since its founding.

The capped-profit model had been an attempt to reconcile the need for commercial capital with the mission to build AGI for the benefit of humanity. Under the model, investors in OpenAI’s for-profit subsidiary could receive returns up to a cap — set at 100x the initial investment for the earliest investors — after which all additional returns would flow to the nonprofit parent. The model was designed to allow OpenAI to raise the capital needed to train frontier models while preserving the mission orientation of the nonprofit.

The model had always been awkward. The cap was high enough that it did not meaningfully constrain the commercial incentives of the for-profit subsidiary. The governance structure — with the nonprofit board controlling the for-profit subsidiary — created tensions that the November 2023 crisis had exposed. The model was difficult to explain to investors, difficult to operate internally, and difficult to defend publicly. The crisis had revealed the model’s fragility — the board’s authority to fire the CEO had been the trigger for the crisis, and the resolution had required the effective dismantling of the board’s authority.

The conversion to a more conventional for-profit structure addressed these problems but created new ones. The new structure was simpler, easier to operate, and easier to explain to investors. It allowed OpenAI to raise capital more efficiently and to compensate employees more competitively. But it also meant that the formal commitment to the mission — the structural guarantee that the nonprofit’s interests would prevail in any conflict — was weakened. The nonprofit remained as a shareholder, but its control over the for-profit was reduced.

The conversion was controversial. Critics argued that it represented the abandonment of OpenAI’s founding mission — that the organisation that had been created to ensure AGI benefited all of humanity was becoming just another for-profit AI company. Supporters argued that the capped-profit model had been a failed experiment — that the November 2023 crisis had demonstrated its unworkability — and that the conversion was a necessary adaptation to the realities of building frontier AI.

Altman defended the conversion as a refinement rather than an abandonment of the mission. The nonprofit, he argued, would remain a significant shareholder and would continue to play a role in the organisation’s governance. The mission — to build AGI that benefits all of humanity — remained unchanged. What was changing was the structure through which the mission would be pursued. The new structure, he argued, would allow OpenAI to raise the capital needed to build AGI while preserving the mission orientation through the nonprofit’s ongoing role.

Whether the new structure would actually preserve the mission was an open question. The structural guarantees were weaker than under the capped-profit model. The incentives for the for-profit to prioritise commercial returns over mission considerations were stronger. The nonprofit’s ability to enforce its interests was reduced. The conversion was a bet — a bet that the culture and values of the organisation would preserve the mission even in the absence of strong structural guarantees. The outcome of that bet would not be known for years.

The Valuation Escalation: From $86B to $157B to $500B

The post-crisis period was characterised by an extraordinary escalation in OpenAI’s valuation. Before the crisis, OpenAI had been valued at approximately $86 billion in a tender offer that allowed employees to sell shares. After the crisis, the valuation rose rapidly — to $157 billion in a funding round led by Thrive Capital in October 2024, and to reported valuations approaching $500 billion in secondary market transactions by 2025.

The escalation reflected several factors. The first was the commercial momentum of the company. Revenue was growing rapidly — from approximately $1 billion in annualised revenue in late 2023 to over $11 billion in annualised revenue by late 2025. The growth was driven by the increasing adoption of ChatGPT in enterprise settings, by the integration of OpenAI’s models into Microsoft products, and by the Apple partnership. The commercial traction was real, and it justified significant valuation increases.

The second factor was the competitive dynamic in the AI industry. Google, Anthropic, Meta, and others were investing heavily in frontier AI, and OpenAI’s position as the perceived leader made it the most attractive investment in the space. The scarcity of investment opportunities in frontier AI — there were only a handful of companies with the capability and capital to train frontier models — drove up the valuations of those that were available.

The third factor was the strategic importance of OpenAI to Microsoft. Microsoft’s $13 billion investment in OpenAI had been made in part to ensure that Microsoft would have access to the most advanced AI capabilities. As the importance of AI to Microsoft’s competitive position increased, the value of the OpenAI relationship increased correspondingly. Microsoft’s continued investment, and its willingness to support OpenAI’s capital needs, was a significant factor in the valuation escalation.

The valuation escalation had several effects on OpenAI’s behaviour. The first was to increase the pressure to generate revenue. A company valued at $500 billion cannot sustain its valuation on $11 billion in revenue — it needs to grow rapidly, both in absolute terms and as a percentage of its valuation. The pressure to monetise — to find new revenue sources, to expand existing ones, to develop products that could generate significant new income — became a primary strategic driver.

The second effect was to make the for-profit conversion more attractive. The capped-profit model’s caps were tied to the initial investment levels, and as the valuation escalated, the caps became less binding — the early investors had already earned returns that approached or exceeded the caps, and the additional returns would flow to the nonprofit. The conversion to a for-profit structure, by removing the caps, would allow new investors to participate in the upside without limitation — making it easier to raise additional capital at the higher valuations.

The third effect was to increase the stakes of the AGI race. A company valued at $500 billion is valued not just on its current revenue but on its expected future position in the AI industry. If OpenAI achieves AGI — or even if it is perceived as the company most likely to achieve AGI — the valuation could justify itself. If it does not, the valuation could collapse. The pressure to maintain the perception of progress toward AGI — to be seen as the leader in the race — became a significant strategic consideration.

The Microsoft Deepening: The Alliance That Defined the Era

The relationship between OpenAI and Microsoft deepened significantly in the post-crisis period. Microsoft’s investment in OpenAI, which had begun with $1 billion in 2019 and grown to $13 billion by 2023, expanded further through additional capital commitments, through the integration of OpenAI’s models into an increasing range of Microsoft products, and through the close coordination of the two companies’ AI strategies.

The deepening was driven by mutual strategic interest. Microsoft needed OpenAI’s models to maintain its competitive position in AI — to compete with Google, with Amazon, with the other technology giants that were investing heavily in AI. OpenAI needed Microsoft’s capital, its cloud infrastructure, and its enterprise sales channels to scale its commercial operations. The two companies had become, in many respects, a single AI operation distributed across two corporate entities.

The deepening raised questions about the independence of OpenAI’s research. As Microsoft’s strategic interests became more intertwined with OpenAI’s, the question of whether OpenAI’s research agenda was being shaped by Microsoft’s commercial needs became more salient. The integration of OpenAI’s models into Microsoft products — Office, Windows, Azure, GitHub — required the models to meet specific performance and reliability standards that were driven by Microsoft’s commercial needs. The research priorities at OpenAI — which models to train, which capabilities to develop, which safety research to pursue — were increasingly influenced by the demands of the Microsoft relationship.

The deepening also raised questions about the competitive dynamics of the AI industry. The Microsoft-OpenAI alliance was, by some measures, the most powerful force in the AI industry — combining OpenAI’s research capabilities with Microsoft’s capital, infrastructure, and distribution. The alliance was a significant barrier to entry for competitors — any company seeking to compete in frontier AI would need to match the combined resources of Microsoft and OpenAI, which few companies could do.

The antitrust implications of the alliance were beginning to be examined by regulators in the US and Europe. The question of whether the alliance constituted a violation of antitrust law — whether it was a single firm in economic terms, whether it was using its market power to exclude competitors — was being investigated. The outcome of these investigations would have significant implications for the structure of the AI industry and for the future of the Microsoft-OpenAI relationship.

For Altman, the deepening of the Microsoft relationship was a strategic priority. Microsoft was OpenAI’s most important partner — the source of capital, infrastructure, and distribution that made OpenAI’s commercial success possible. Maintaining and deepening the relationship was essential to OpenAI’s continued growth. But the deepening also constrained OpenAI’s strategic flexibility — the company’s ability to pursue partnerships with other technology companies, to develop products that competed with Microsoft’s, to make strategic decisions that were not aligned with Microsoft’s interests. The balance between commitment and independence was a constant strategic consideration.

Sutskever’s Departure and the Safety Brain Drain

The post-crisis period was marked by the departure of several key safety researchers from OpenAI. The most significant was the departure of Ilya Sutskever, who had been a co-founder of OpenAI, the chief scientist, and one of the most respected researchers in the field. Sutskever had been a central figure in the November 2023 crisis — he had initially supported the board’s decision to fire Altman before signing the open letter demanding the board’s resignation. His position after the crisis was untenable, and he left OpenAI in May 2024 to found a new AI safety company, SSI Inc., which quickly raised $1 billion in funding.

Sutskever’s departure was a significant loss for OpenAI. He had been one of the architects of the company’s research programme — the technical leader behind GPT-3, GPT-4, and the early development of what would become the reasoning models. His departure left a gap in the research leadership that was difficult to fill. More importantly, his departure was a signal — to the broader AI research community, to the AI safety community, and to the public — that something was changing at OpenAI. The co-founder and chief scientist had left, and the reasons for his departure were not entirely clear.

Sutskever was not the only departure. Over the following year, several other safety researchers left OpenAI — Jan Leike, who had co-led the superalignment team with Leopold Aschenbrenner; William Saunders, who had worked on alignment research; and others. Many of them joined Anthropic, which had been founded in 2021 by former OpenAI researchers and had positioned itself as the safety-focused alternative to OpenAI. The flow of safety researchers from OpenAI to Anthropic was a constant feature of the post-crisis period.

The departures were explained in various ways. Some of the departures were attributed to disagreements about the balance between safety research and capability research at OpenAI — the perception that safety research was being deprioritised in favour of capability development. Some were attributed to the cultural shift at OpenAI — the perception that the post-crisis organisation was less tolerant of internal dissent and less supportive of safety research that might slow the development of new capabilities. Some were attributed to the simple appeal of Anthropic, which offered the opportunity to work on frontier AI in an organisation explicitly dedicated to safety.

The departures had several effects on OpenAI. The first was the loss of talent — the safety researchers who left were among the best in the field, and their departure weakened OpenAI’s safety research capability. The second was the reputational effect — the perception that OpenAI was losing its safety researchers was damaging to the company’s reputation in the AI safety community and among regulators. The third was the competitive effect — the researchers who left OpenAI were strengthening Anthropic, which was emerging as OpenAI’s most serious competitor in the frontier AI race.

Altman publicly downplayed the departures, arguing that they reflected normal turnover in a fast-growing company and that OpenAI continued to attract and retain top talent. The reality was more complex. The departures were a real loss, and they reflected a real shift in OpenAI’s internal culture and priorities. The question of whether the shift was a necessary adaptation to the demands of building frontier AI or a betrayal of the organisation’s founding commitment to safety was being actively debated — within OpenAI, within the AI safety community, and in the broader public conversation about the company.

The AGI Rhetoric Escalation: From “Helpful” to “Build AGI”

The post-crisis period saw a significant escalation in OpenAI’s rhetoric about artificial general intelligence. Before the crisis, OpenAI’s public statements about AGI had been relatively measured — the company’s stated mission was to build AGI that benefits all of humanity, but the emphasis was on the “benefits all of humanity” part rather than on the “build AGI” part. After the crisis, the emphasis shifted. Altman and other OpenAI executives began to speak more openly and more ambitiously about AGI — about when it might be achieved, what it would look like, and what it would mean.

The rhetoric escalation was driven by several factors. The first was the competitive dynamic in the AI industry. Anthropic, Google, Meta, and others were all pursuing AGI — or at least pursuing capabilities that approached AGI — and the public perception of which company was closest to AGI had become a significant factor in the competitive landscape. By speaking more openly about AGI, OpenAI was positioning itself as the leader in the race — the company most likely to achieve AGI first.

The second factor was the valuation escalation. A company valued at $500 billion needs to justify its valuation with a compelling narrative about its future. The narrative that OpenAI was building — that it was on the verge of achieving AGI, that AGI would transform the economy, that the company that achieved AGI first would have an unassailable competitive position — was a narrative that supported the valuation. The rhetoric was, in part, a way of managing the expectations that the valuation had created.

The third factor was the genuine belief within OpenAI that AGI was approaching. The pace of progress in AI capabilities — the improvements from GPT-3 to GPT-4 to GPT-4o to the reasoning models — was rapid, and the trajectory suggested that further significant advances were possible. The researchers and executives at OpenAI were not engaging in pure hype when they spoke about AGI; they were expressing a genuine belief, based on the trajectory they were observing, that AGI was a realistic possibility in the foreseeable future.

The rhetoric escalation had several effects. The first was to raise the public profile of the AGI question — to make AGI a topic of mainstream conversation in a way that it had not been before. This had both positive and negative effects — it increased public awareness of the potential benefits and risks of AGI, but it also contributed to a certain amount of hype and speculation that was not always grounded in the actual state of the technology.

The second effect was to increase the pressure on OpenAI to actually deliver AGI. The rhetoric had created expectations that the company would need to meet — or at least that it would need to appear to be meeting. The pressure to demonstrate progress toward AGI — to release models that were perceived as advancing toward AGI — became a significant factor in the company’s product strategy. The reasoning models, released in late 2024 and 2025, were partly a response to this pressure — a demonstration that OpenAI was making progress toward the goal of AGI.

The third effect was to intensify the concerns of the AI safety community. The combination of escalating AGI rhetoric, the for-profit conversion, the safety researcher departures, and the deepening Microsoft relationship created a perception that OpenAI was prioritising the race to AGI over the safety considerations that the AI safety community regarded as essential. The perception was not entirely fair — OpenAI continued to invest in safety research, and the company’s safety teams continued to produce valuable work — but it was widespread, and it shaped the broader conversation about OpenAI’s role in the AI industry.

The Apple Partnership and the ChatGPT-as-OS Strategy

The Apple partnership, announced at WWDC in June 2024, was the most consequential commercial deal of the post-crisis period. Apple, which had been perceived as lagging in AI, integrated ChatGPT into Siri and across its operating systems — making OpenAI’s technology the default AI layer for hundreds of millions of Apple devices. The deal did not involve direct payment to OpenAI — the value was in distribution and brand association — but it established OpenAI as the AI partner of choice for the world’s most valuable consumer technology company.

The partnership was strategically significant for several reasons. The first was the distribution it provided. Apple’s user base — over a billion active devices — was a distribution channel that no other AI company could match. The integration of ChatGPT into Siri and across Apple’s operating systems would expose OpenAI’s technology to a vast audience that had not previously used it. The potential for user acquisition was enormous.

The second was the legitimisation it provided. Apple’s decision to partner with OpenAI — rather than build its own AI capabilities or partner with a competitor — was a powerful endorsement of OpenAI’s technology. Apple is known for the quality of its partnerships and the selectivity of its choices. The partnership signalled to the broader market that OpenAI’s technology was the best available — the technology that Apple, with its resources and standards, had chosen to integrate into its products.

The third was the strategic positioning it provided. The partnership was part of a broader strategy to position ChatGPT as the operating system for the AI era — the layer that would sit between users and their devices, applications, and services, providing AI capabilities across all of them. The Apple partnership was a significant step in this strategy — it established ChatGPT as the AI layer for Apple’s devices, complementing the existing integrations with Microsoft’s products.

The ChatGPT-as-OS strategy was an ambitious one. It positioned OpenAI not just as an AI model provider but as a platform company — the company that would provide the AI infrastructure for the next generation of computing. The strategy required significant investment in product development — in the ChatGPT application itself, in the developer tools that would allow third parties to build on the platform, in the enterprise tools that would allow businesses to integrate ChatGPT into their workflows. The investment was substantial, but the potential payoff was enormous — if ChatGPT became the default AI layer for consumer and enterprise computing, OpenAI would occupy a position in the AI era comparable to Microsoft’s position in the personal computing era.

The strategy also raised questions about competition and lock-in. If ChatGPT became the default AI layer, the switching costs for users and businesses would become significant — the integration of ChatGPT into workflows, applications, and devices would make it difficult to switch to a competitor. The competitive implications of this lock-in were significant — and were beginning to attract the attention of antitrust regulators who were concerned about the concentration of power in the AI industry.

The Competitors: Anthropic, Google, Meta Closing In

The post-crisis period was marked by intensifying competition in the frontier AI race. OpenAI’s perceived lead — which had been substantial in the immediate aftermath of ChatGPT’s launch — was narrowing as competitors closed the gap. Anthropic, Google, and Meta were all investing heavily in frontier AI, and each was making significant progress.

Anthropic was the most direct competitor. Founded in 2021 by former OpenAI researchers, Anthropic had positioned itself as the safety-focused alternative to OpenAI — a company that would pursue frontier AI capabilities while maintaining a stronger commitment to safety research and a more cautious approach to deployment. The Claude models — Claude 2, Claude 3, Claude 3.5, Claude 4 — were competitive with OpenAI’s models on most benchmarks, and in some cases superior. Anthropic’s revenue was growing rapidly, and the company had raised significant capital at valuations that, while lower than OpenAI’s, were substantial. The flow of safety researchers from OpenAI to Anthropic was strengthening Anthropic’s research capabilities and reinforcing its safety-focused positioning.

Google was the most resourced competitor. Google’s AI research organisation — Google DeepMind, after the merger of Google Brain and DeepMind in 2023 — was the largest and most experienced AI research organisation in the world. Google’s Gemini models were competitive with OpenAI’s models, and Google’s infrastructure — its custom TPUs, its data centres, its distribution through Google Search, Android, and Workspace — gave it capabilities that OpenAI could not match. Google’s integration of AI into its products was accelerating, and the company was beginning to close the gap that OpenAI had opened with ChatGPT.

Meta was the most strategically distinctive competitor. Meta’s decision to release its Llama models as open weights — making them freely available for anyone to use, modify, and deploy — was a fundamentally different strategy from OpenAI’s closed, API-based approach. The Llama models were not as capable as OpenAI’s frontier models, but they were close enough for many use cases, and their open availability made them the default choice for many developers, researchers, and businesses. Meta’s strategy was not to compete with OpenAI directly in the frontier model market but to commoditise the model layer — to make the models themselves so widely available that the value would shift to the applications built on top of them. The strategy was having significant effects on the competitive landscape, and was the subject of the companion piece E26 — The Open Source Wars.

The intensifying competition had several effects on OpenAI. The first was to increase the pressure to maintain the perceived lead. OpenAI’s competitive position depended in part on the perception that it was ahead — that its models were the best, that its products were the most advanced, that its research was the most cutting-edge. The perception was increasingly difficult to maintain as competitors closed the gap. The product cadence — the regular release of new models and products — was partly a response to this pressure, a way of demonstrating continued progress.

The second effect was to increase the pressure on revenue growth. The competitors were investing heavily, and OpenAI needed to invest heavily as well to maintain its position. The investment required significant revenue, which required significant commercial growth. The pressure on the commercial side of the business — to find new revenue sources, to expand existing ones, to develop products that could generate significant income — was intensifying.

The third effect was to increase the strategic importance of the Microsoft relationship. As competitors with their own infrastructure — Google, Meta — closed the gap, OpenAI’s dependence on Microsoft’s infrastructure became more consequential. The ability to train frontier models depended on access to large-scale compute, and Microsoft’s Azure was the primary source of that compute for OpenAI. The relationship was both a competitive advantage — Microsoft’s infrastructure was among the best in the world — and a strategic dependency — OpenAI’s ability to train frontier models depended on Microsoft’s continued support.

The Lawsuits: NYT, Authors Guild, the IP Reckoning

The post-crisis period was marked by the escalation of legal challenges to OpenAI’s training data practices. The most significant was the lawsuit filed by the New York Times in December 2023, alleging that OpenAI had used millions of NYT articles to train its models without permission and in violation of the Times’s copyright. The lawsuit was followed by similar suits from the Authors Guild, from individual authors, from music publishers, and from other content creators.

The lawsuits were not unique to OpenAI — similar suits were filed against other AI companies — but they were particularly consequential for OpenAI because of the company’s prominence and because of the volume and quality of the training data it was alleged to have used. The NYT suit, in particular, was significant — the Times is one of the most respected news organisations in the world, and its articles represent a significant portion of the high-quality text that is valuable for training language models.

The legal question — whether training AI models on copyrighted works constitutes fair use or infringement — was being actively litigated, and the outcomes of the cases would have significant implications for OpenAI and for the broader AI industry. If the courts ruled that training was fair use, OpenAI and other AI companies would be free to continue training on copyrighted works without compensation. If the courts ruled that training was infringement, the AI companies would need to develop licensing schemes, opt-out mechanisms, or other ways of compensating creators — and the cost of training models would increase significantly.

OpenAI’s response to the lawsuits was twofold. The first was legal — defending the cases on the grounds that training on publicly available works constituted fair use. The second was commercial — developing licensing agreements with content creators that would allow the creators to be compensated for the use of their work in training. The licensing agreements were partly a commercial strategy — they provided access to high-quality training data that could improve model performance — and partly a legal strategy — they demonstrated that OpenAI was willing to compensate creators, which could support a fair use defence by showing that the company was acting in good faith.

The lawsuits were a significant strategic challenge for OpenAI. The potential liability — if the company were found to have infringed copyright on a large scale — was substantial. The potential impact on the company’s training data practices — if the courts required licensing for all copyrighted works — was significant. And the reputational impact — the perception that OpenAI was profiting from the work of creators without compensating them — was damaging, particularly in the creative communities whose work was at issue.

The resolution of the lawsuits would take years, and the outcomes were uncertain. But the lawsuits were already having an effect on the AI industry — they were increasing the attention being paid to training data practices, they were encouraging the development of licensing schemes and opt-out mechanisms, and they were contributing to a broader conversation about the relationship between AI companies and the creators whose work the AI systems were built on.

The Mission-vs-Commerce Question: What the Return Revealed

The deepest question raised by the post-crisis period was the question of OpenAI’s mission — whether the organisation was still pursuing the mission it had been founded to pursue, or whether the mission had been subordinated to commercial considerations. The question was not new — it had been raised since the company’s founding, and the November 2023 crisis had been, in part, a manifestation of it. But the post-crisis period made the question more urgent, because the changes that followed the crisis — the for-profit conversion, the safety researcher departures, the AGI rhetoric escalation, the deepening Microsoft relationship — all pointed in the same direction: toward a more commercial, more competitive, more conventionally corporate OpenAI.

The case that the mission had been subordinated was straightforward. The for-profit conversion weakened the structural commitment to the mission. The safety researcher departures suggested that the safety research that the mission required was being deprioritised. The AGI rhetoric escalation suggested that the company was more interested in being perceived as building AGI than in ensuring that AGI, when built, would benefit all of humanity. The deepening Microsoft relationship suggested that the company’s strategic priorities were increasingly aligned with Microsoft’s commercial interests rather than with the broader public interest.

The case that the mission had not been subordinated was also straightforward. The nonprofit remained a significant shareholder and a voice in the company’s governance. The safety research continued, even after the departures — the company’s safety teams continued to produce valuable work, and the company continued to invest in safety research at a level that few other AI companies matched. The AGI rhetoric, while escalated, was consistently framed in terms of the benefits that AGI would bring to humanity. The Microsoft relationship, while deepening, was a partnership between two independent companies, not an absorption of OpenAI into Microsoft.

The reality was probably more complex than either case suggested. The mission was not abandoned, but it was being interpreted in a particular way — a way that emphasised the importance of building AGI quickly, of maintaining the commercial momentum that would allow the company to invest in the research needed to build AGI, and of partnering with Microsoft to access the resources needed to compete in the frontier AI race. This interpretation was not indefensible — one could argue that building AGI was the most important thing, that commercial success was necessary to fund the research, that the Microsoft partnership was the best available vehicle for competing in the race. But the interpretation was not the only possible one, and it was not the one that the organisation’s founders had explicitly endorsed.

The question of whether the mission had been subordinated was, ultimately, a question about Altman’s own commitments and beliefs. Altman had been the architect of the post-crisis strategy — the for-profit conversion, the product cadence, the Microsoft deepening, the AGI rhetoric escalation. If Altman genuinely believed that these choices served the mission, then the mission was being pursued, even if in a way that critics disputed. If Altman was using the mission as cover for commercial ambitions, then the mission had been subordinated.

The evidence was ambiguous. Altman’s public statements consistently framed the company’s strategy in terms of the mission. His private actions — the deal-making, the valuation negotiations, the competitive manoeuvring — were consistent with the strategy he described publicly. The internal culture of OpenAI, while more commercially focused than before, retained a significant mission orientation — the researchers and engineers who worked there did so in part because they believed in the mission, and their belief was a check on the company’s drift toward pure commercialism.

The question would not be resolved by analysis. It would be resolved by the company’s actions over the coming years — by the choices it made about safety research, about deployment, about the balance between commercial and mission considerations. If the company continued to invest in safety, to deploy its technology responsibly, to use its commercial success to fund mission-oriented research, then the mission would be shown to have been preserved. If the company deprioritised safety, deployed its technology recklessly, or used its commercial success to enrich its investors and employees at the expense of the broader public interest, then the mission would be shown to have been subordinated. The answer was, and remains, in the future.

The Sora, o1, and the New Product Cadence

The post-crisis period saw the establishment of a new product cadence at OpenAI — a rhythm of releases that demonstrated the company’s continued progress and maintained its competitive position. The cadence included Sora (video generation, previewed February 2024), GPT-4o (multimodal, May 2024), o1 (reasoning model, September 2024), the o1 full release (December 2024), o1-mini and o3-mini (early 2025), GPT-4.5 (early 2025), o3 (early 2025), and the continuous improvement of ChatGPT’s features and capabilities.

The cadence was strategic. Each release demonstrated progress — in capabilities, in modalities, in performance — that supported the company’s positioning as the leader in the frontier AI race. The releases were timed to maintain visibility — to keep OpenAI in the news, to keep the perception of progress alive, to give customers and investors reasons to continue their engagement with the company. The cadence was, in part, a marketing strategy — a way of managing the perception of the company’s position in the competitive landscape.

But the cadence was also a genuine reflection of technical progress. The models were improving — they were becoming more capable, more efficient, more useful. The pace of improvement was rapid, and the releases were the visible manifestation of that improvement. The cadence was not just marketing — it was the outward sign of a research programme that was making real progress on the technical challenges of building more capable AI systems.

The cadence had effects on the broader AI industry. The regular releases set a pace that competitors felt compelled to match — and the matching was difficult, because the research progress that underlay the releases was not easily replicated. The cadence also set expectations — among users, among developers, among investors — about the pace at which AI capabilities would improve. The expectations, once set, became a constraint — the company needed to continue releasing new models and products at a pace that met the expectations, or risk losing the perception of leadership.

The cadence also had effects on the company’s internal culture. The pace of releases required a pace of work that was demanding — the engineers and researchers who produced the releases worked under significant pressure, and the pressure contributed to the burnout and turnover that the company experienced. The cadence was a source of pride — the employees were proud of what they were producing — but it was also a source of stress, and the long-term sustainability of the pace was a question that the company’s leadership needed to address.

The 2025 Reorganisation: OpenAI as For-Profit Empire

The culmination of the post-crisis transformation was the 2025 reorganisation of OpenAI’s corporate structure. The reorganisation, which was implemented through 2025 and finalised in late 2025, converted OpenAI from the capped-profit structure that had defined it since 2019 to a more conventional for-profit structure, with the nonprofit parent retaining a significant but non-controlling stake.

The reorganisation was the logical conclusion of the changes that had followed the November 2023 crisis. The for-profit conversion, the valuation escalation, the deepening Microsoft relationship, the AGI rhetoric escalation, the product cadence — all of these had been pushing OpenAI toward a more conventional corporate structure. The reorganisation was the formalisation of what had already happened in practice.

The reorganisation had several implications. The first was the simplification of the company’s governance. The complex capped-profit structure, with its tensions between the nonprofit board and the for-profit subsidiary, was replaced by a more conventional structure in which the for-profit company was governed by its board of directors, with the nonprofit as a significant shareholder. The governance was simpler, clearer, and more familiar to investors and regulators.

The second was the alignment of incentives. Under the capped-profit structure, the incentives of the for-profit subsidiary were complicated by the caps on investor returns — investors had limited upside, which affected their willingness to invest and their expectations of returns. Under the new structure, the incentives were more conventional — investors could participate fully in the upside, which made the company more attractive to capital and allowed it to raise money more efficiently.

The third was the weakening of the structural commitment to the mission. The nonprofit, as a significant shareholder, would continue to benefit from the company’s commercial success — and could use its resources to pursue the mission. But the nonprofit’s ability to control the for-profit’s decisions was reduced. The mission, going forward, would depend less on structural guarantees and more on the culture and values of the company’s leadership and employees.

The reorganisation was controversial. Critics argued that it represented the final abandonment of OpenAI’s founding mission — that the company that had been created to ensure AGI benefited all of humanity had become, in effect, just another for-profit AI company. Supporters argued that the capped-profit model had been a failed experiment, that the November 2023 crisis had demonstrated its unworkability, and that the reorganisation was a necessary adaptation to the realities of building frontier AI at scale.

Altman defended the reorganisation as a refinement rather than an abandonment of the mission. The nonprofit, he argued, would remain a significant shareholder and would continue to play a role in the organisation’s governance. The mission — to build AGI that benefits all of humanity — remained unchanged. What was changing was the structure through which the mission would be pursued. The new structure, he argued, would allow OpenAI to raise the capital needed to build AGI while preserving the mission orientation through the nonprofit’s ongoing role.

Whether the reorganisation would actually preserve the mission was an open question — the same open question that the for-profit conversion had raised, now sharpened by the formalisation of the new structure. The answer would depend on the company’s actions over the coming years, on the choices made by its leadership, on the culture that developed within it, and on the broader competitive and regulatory environment in which it operated.

What the Year Made: Altman’s Position in History

The eighteen months following the November 2023 crisis transformed Sam Altman’s position in the AI industry and in the broader technology landscape. Before the crisis, Altman was the CEO of OpenAI — a powerful position, but one in which his authority was constrained by the board, by the capped-profit structure, and by the need to balance the commercial and research sides of the organisation. After the crisis and the transformation that followed, Altman was the CEO of OpenAI without those constraints — the leader of the most valuable AI company in the world, with a board aligned to his vision, a for-profit structure that allowed him to raise capital efficiently, a deepened Microsoft partnership, an Apple partnership, and a competitive position that, while challenged, was still perceived as leading.

The transformation had made Altman more powerful, but it had also made him more controversial. The decisions that had increased his power — the for-profit conversion, the safety researcher departures, the AGI rhetoric escalation — were the same decisions that had intensified the concerns of the AI safety community, the creative community, the regulatory community, and the broader public. Altman was both more powerful and more distrusted than he had been before the crisis. The combination was an uncomfortable one — for Altman, for OpenAI, and for the broader AI ecosystem that depended on the company’s decisions.

The question of how to evaluate Altman’s post-crisis tenure depended on the perspective from which it was evaluated. From the perspective of OpenAI’s commercial success, the evaluation was positive — the company’s revenue, valuation, and market position had all improved dramatically. From the perspective of OpenAI’s research progress, the evaluation was also positive — the company had continued to produce cutting-edge models and to push the frontier of AI capabilities. From the perspective of OpenAI’s safety research, the evaluation was mixed — the company had continued to invest in safety, but the departures of key safety researchers and the perceived deprioritisation of safety relative to capability were concerning. From the perspective of OpenAI’s mission, the evaluation was contested — the company’s defenders argued that the mission was being pursued effectively, while the company’s critics argued that the mission had been subordinated to commercial considerations.

Altman’s position in history — the place he would occupy in the broader story of AI’s development — would depend on the outcome of the project he was leading. If OpenAI achieved AGI and the achievement benefited humanity, Altman would be remembered as the leader who navigated the most important technology transition in human history. If OpenAI achieved AGI and the achievement harmed humanity — through misalignment, through concentration of power, through unintended consequences — Altman would be remembered as the leader who failed to navigate the transition safely. If OpenAI did not achieve AGI, and the race to AGI was won by a competitor, Altman would be remembered as the leader who built the most influential AI company of the era but did not achieve the ultimate goal.

The outcome was not yet known. The race was ongoing. The technology was developing. The competitive landscape was shifting. The regulatory environment was evolving. The public conversation was changing. The next several years would determine the outcome — and would determine, in the process, how Altman’s post-crisis tenure would be evaluated. The year that made OpenAI had also made Altman’s position in the AI industry more powerful and more consequential than it had ever been. What he would do with that power, and how history would judge it, was the question the future would answer.

Further Reading

  • “The Man Who Would Build AGI: Sam Altman and the OpenAI Project” by Steven Levy (Wired, 2024) — A long-form profile of Altman in the post-crisis period, covering the for-profit conversion, the Microsoft deepening, and the AGI rhetoric escalation.
  • “The OpenAI Board Crisis: A Reconstruction” by Karen Hao (MIT Technology Review, 2024) — The most detailed journalistic reconstruction of the November 2023 crisis, based on interviews with the participants.
  • “The For-Profit Conversion: What It Means for OpenAI’s Mission” by the AI Safety Community (Various, 2024–2025) — A collection of analyses from the AI safety community on the implications of the for-profit conversion.
  • “OpenAI and Microsoft: The Alliance That Defined the AI Era” by Brad Smith (Microsoft, 2024) — Microsoft’s president’s account of the Microsoft-OpenAI relationship, providing insight into the strategic logic of the partnership.
  • “The Departure of Ilya Sutskever: What It Means for OpenAI” by Will Knight (Wired, 2024) — Coverage of Sutskever’s departure and the broader safety researcher brain drain.
  • “GPT-4o and the Multimodal Moment” by OpenAI (2024) — OpenAI’s own technical report on GPT-4o, providing insight into the model’s capabilities and the strategic vision behind its development.
  • “The Apple-OpenAI Partnership: What It Means for the AI Industry” by Mark Gurman (Bloomberg, 2024) — The most detailed reporting on the Apple-OpenAI partnership and its strategic implications.
  • “The NYT v. OpenAI Lawsuit: The Legal Question That Will Shape the AI Industry” by Rebecca Tushnet (Harvard Law Review, 2024) — A legal analysis of the NYT lawsuit and its implications for AI training data practices.

Next in the Profiles series: this is the final profile. The Minds & Machines series concludes with the companion pieces A26 — The Future of Creativity: Art, Music, and Writing in the AI Age and E26 — The Open Source Wars: When AI Went Free. The conversation continues in Minds & Machines: Beyond the Series — standalone essays extending the themes, profiles, and events explored in these 75 articles, without the chronological-act structure.

Minds & Machines: The Story of AI is published weekly. If the story of Sam Altman’s return and the year that made OpenAI illuminates something about the contradictions and consequences of building AI at the frontier of commercial and mission considerations, share it with someone who would find the illumination valuable.

Comments

Reply on Bluesky → (opens in a new tab)