11 Apex
In October 2022, OpenAI held a company-wide off-site in Monterey, California, a beautiful coastal city two hours south of San Francisco. By then the company had grown to roughly three hundred employees, from less than two hundred the year before. For a photo that weekend, they all posed outside Monterey’s tiny airport, grinning and wearing their OpenAI-branded gear. Altman looked relaxed, sitting on the ground in the front of the pack, knees against his chest, arms loosely crossed, feet pointed up.
Over two days, the executive team presented updates on their vision and implementation: Altman on the massive data centers that Microsoft was scaling up for OpenAI; Steve Dowling and his deputy Hannah Wong on the top-tier publications lining up to cover the company; Anna Makanju on its expanding footprint in Washington, DC. Then there were demos, one after the other, of the research and product teams’ latest projects. The sheer impressiveness of it all was palpable. “Everything together was so mind-boggling,” remembers a former employee who was present.
Brockman got onstage to discuss the latest plans for GPT-4 and began to tell a story about his wife, Anna, who started having abdominal pains one day that wouldn’t go away. Anna was a frequent fixture at the office. She had a desk next to his and often came with him to meetings even though she didn’t officially work at the company. Besides Altman and Sutskever, she was, Greg had once told me, the person he relied on most for support, his best friend, his confidant. They were often attached at the hip, going everywhere together. Greg went with Anna to multiple doctors, he explained to employees, and none could figure out the problem. But when he asked GPT-4, it suggested she might have a condition that they hadn’t considered. “And then she did!” he exclaimed. It was a retelling of the same story he’d shared with me in 2019, about the promise of AGI solving health care for people like his friend who had gone to myriad specialists to diagnose a problem—this time with a different character.
Greg would repeat the story again on X weeks after the board’s attempt to fire Altman, with a new variation. He would recount new medical challenges that Anna had faced and her struggle over five years “seeing more doctors and specialists than in her whole life prior” to finally get a diagnosis. It was her allergist who finally put all the pieces together and realized she had hypermobile Ehlers-Danlos Syndrome, a genetic mobility disorder. There was still a long way for AGI to work “in high-stakes areas like medicine,” he wrote, keeping up his drumbeat on behalf of OpenAI’s rallying ambition, “but the promise is getting increasingly clear.”
Within weeks of returning from the off-site, rumors began to spread that Anthropic was testing—and would soon release—a new chatbot. The Superassistant team was midway through designing its chat interface. If it didn’t launch first, OpenAI risked losing its leading position, which could deliver a big hit to morale for employees who had worked long and tough hours to retain that dominance. Worse still for some leaders, OpenAI would lose to Anthropic.
Anthropic had not in fact been planning any imminent releases. It was also in the midst of its own problems. With the sudden collapse of FTX in the early days of November, it was getting swept up in the fallout. Just months before, it had raised $580 million, $500 million of which an FTX press release had said was from SBF and other senior leaders. Financial documents released during the trial would later find that that money had been sourced from the billions that SBF had embezzled from FTX customer deposits, turning the Anthropic investment into a central issue during the trial over whether the significant returns generated from it could be used to pay back customers. (A judge would rule that it could, and FTX would sell off its Anthropic shares in batches through 2024 for a total of $1.3 billion.)
But for OpenAI executives, the rumors were enough to trigger a decision: The company wouldn’t wait to ready GPT-4 into a chatbot; it would release Schulman’s chat-enabled GPT-3.5 model with the Superassistant team’s brand-new chat interface in two weeks, right after Thanksgiving. The Superassistant team instantly pivoted, pulling in several other members as they sprinted to integrate everything and build out the remaining features. To the rest of the company, leadership framed the effort carefully. ChatGPT—the name they settled on—would not in fact be a product launch but a “low-key research preview,” just like DALL-E 2. In the same way, it wouldn’t be monetized but “get the data flywheel going”—in other words, amass more data from people using it—which would help improve GPT-4 and the Superassistant product.
Outside of the Superassistant team, everyone took the executives literally. A low-key research preview didn’t require their attention; they needed to stay focused on the GPT-4 launch for early 2023. The trust and safety team felt they barely had enough time to build out its monitoring infrastructure in time for that launch. By comparison, ChatGPT seemed like a nonissue. GPT-3.5 had already been refined with RLHF; it was inherently safer than the version of GPT-3, which had not been, still available on the API. OpenAI had also posted a version of 3.5 without chat features on its developer platform for developers to test out the model’s capabilities. Did adding a chat interface really make a difference? People in the Safety clan, occupied with testing and tuning GPT-4, agreed. For the first time, a model release flew through the checks with little resistance.
Even within the Superassistant team, no one truly fathomed the societal phase shift they were about to unleash. They expected the chatbot to be a flash in the pan. Much like DALL-E 2, it would generate a lot of fanfare on social media and then quiet down after a few weeks. The night before the release, things felt remarkably calm after such an intense sprint to the finish. They placed bets on how many users might try the tool by the end of the weekend. Some people guessed a few thousand. Others guessed tens of thousands. To be safe, the infrastructure team provisioned enough server capacity for one hundred thousand users.
The following day, on Wednesday, November 30, most other employees didn’t even realize that the launch had happened. Like OpenAI’s own debut, ChatGPT’s release coincided with the annual NeurIPS proceedings, that year being held in New Orleans, Louisiana. The conference had earlier announced its Test of Time Award, an honor bestowed each year to a paper published ten years earlier that had had a critical impact on the field. The award went to Hinton, Sutskever, and Krizhevsky’s 2012 ImageNet paper that introduced the world to the power of deep learning.
That evening a small group of employees hosted an OpenAI party near the conference convention center to represent the company and recruit interested candidates among the nearly ten thousand in-person attendees. DeepMind, Meta, and Google were holding competing recruitment parties at the exact same time throughout the city. As the party went on, a recruiter at the event noticed an OpenAI engineer working nonstop on his computer.
He finally went over to talk to the engineer: “Bro, have a drink. We’re all here. Be social.”
The engineer didn’t move. “No, all the GPUs are melting. Everything is crashing.”
That night, Japan had been first to wake up and to deliver a massive and unexpected swell in the traffic. The following day, the number of users continued to surge, as time zone by time zone the rest of the world came online. Musk played no small part in boosting the climb. “Lot of people stuck in a damn-that’s-crazy ChatGPT loop “,” he tweeted, receiving some seventy-five thousand likes.
The instant runaway success of ChatGPT was beyond what anyone at OpenAI had dreamed of. It would leave the company’s engineers and researchers completely miffed even years later. GPT-3.5 hadn’t been that much of a capability improvement over GPT-3, which had already been out for two years. And GPT-3.5 had already been available to developers. The interface and format had made the model more accessible, certainly, but it wasn’t the fundamental step change that employees had felt with GPT-4. Altman later said that he’d believed ChatGPT would be popular but by something like “one order of magnitude less.” “It was shocking that people liked it,” a former employee remembers. “To all of us, they’d downgraded the thing we’d been using internally and launched it.”
Within five days, Brockman tweeted that ChatGPT had crossed one million users. Within two months, it had reached one hundred million, becoming what was then the fastest-growing consumer app in history. (Meta’s X rival, Threads, later claimed the title by reaching the same user count in less than five days; pundits argued that it didn’t count because Meta was primarily tapping into an existing base of users.)
ChatGPT catapulted OpenAI from a hot startup well-known within the tech industry into a household name overnight. Indeed, at an AI research conference several months later in Kigali, Rwanda, over nine thousand miles away from San Francisco, a researcher based in the country would gush to me that post-ChatGPT, his parents finally understood what he did for work. “You know a technology is accessible to anyone when your mother tells you about it,” he’d say.
At the same time, it was this very blockbuster success that would place extraordinary strain on the company. Over the course of a year, it would polarize its factions further and wind up the stress and tension within the organization to an explosive level.
In the immediate aftermath, the whole company was firefighting. OpenAI’s servers crashed repeatedly as the infrastructure team struggled to scale up its capacity as fast as possible, in the most compressed timeline in the history of Silicon Valley. The team cannibalized some of the Research division’s compute to support ChatGPT’s growth and still didn’t have enough to keep the app up and running. The trust and safety team, numbering just over a dozen people, scrambled to understand and catch bad behavior among the floods of new users, heavily handicapped by spotty monitoring. It had struggled to hire the engineers needed to implement its limited reactive enforcement plan and was still in the middle of building the necessary systems. ChatGPT derailed the project. All efforts to finish new tooling halted as engineering resources were redirected to stabilize what already existed. When the servers crashed, so did the platform for monitoring traffic, grinding the ability to do any scaled enforcement to a complete halt.
The severe shortage of GPUs also derailed another effort. In an attempt to leverage the company’s own technology, the trust and safety team had prototyped a plan internally called Fact Factory, which OpenAI publicly touted, for using GPT-4 to content moderate its own outputs and that of other OpenAI models. The implementation didn’t exactly scale; it required giving GPT-4 extremely long prompts to capture enough nuance. Even when the servers were working, it would cost too many computational resources. And the servers were not consistently up.
To many in the Safety clan, ChatGPT was the most alarming example yet of the limitations of OpenAI’s foresight. One Safety person raised the question in an all-hands meeting: How could the company have failed to predict user behavior and ChatGPT’s popularity so badly? What did that say about the company’s ability to calibrate and forecast the future impacts of its technologies?
To much of the rest of the company, the crashing servers, while an extraordinary source of stress, were even more so an extraordinary mark of triumph. OpenAI had built a technology so profound, in such wild demand, that it had lit up the world and transformed it overnight. They had set their sights on all of humanity and had really done it. Everyone, all eight billion people, was now living in OpenAI’s world.
Altman didn’t indulge the moment. He reminded employees that the company ultimately had a mission to achieve something far bigger than building “the biggest product in the history of Silicon Valley.” He urged every team to stay the course and press forward. As he’d expected, OpenAI had woken up all of its competitors: Anthropic was on its way to releasing its chatbot, Claude; Google had sounded a “code red” alarm internally and would soon consolidate its AI divisions into Google DeepMind to throw its full weight behind launching a similar product. Though OpenAI had hit the market first with its 10x better offering, it needed to keep running to stay number one.
With every team stretched dangerously thin, managers begged Altman for more head count. There was no shortage of candidates. After ChatGPT, the number of job applicants clamoring to join the rocket ship had rapidly multiplied. But Altman worried about what would happen to company culture and mission alignment if the company scaled up its staff too quickly. He believed firmly in maintaining a small staff and high talent density. “We are now in a position where it’s tempting to let the organization grow extremely large,” he had written in his 2020 vision memo, in reference to Microsoft’s investment. “We should try very hard to resist this—what has worked for us so far is being small, focused, high-trust, low-bullshit, and intense.
“The overhead of too many people and too much bureaucracy can easily kill great ideas or result in sclerosis. Unlike the big-iron engineering projects of the past, we could fulfill our mission with a surprisingly small number of great people.”
He was now repeating this to executives in late 2022, emphasizing repeatedly during head count discussions the need to keep the company lean and the talent bar high, and add no more than one hundred or so hires. Other executives balked. At the rate that their teams were burning out, many saw the need for something closer to around five hundred or even more new people.
Over several weeks, as the discussions continued, the executive team finally compromised on a number somewhere in the middle, between two hundred fifty and three hundred. The cap didn’t hold. By summer, there were as many as thirty, even fifty, people joining OpenAI each week, including more recruiters to scale up hiring even faster. By fall, the company had blown well past its own self-imposed quota.
The sudden growth spurt indeed changed company culture. A recruiter wrote a manifesto about how the pressure to hire so quickly was forcing his team to lower the quality bar for talent. “If you want to build Meta, you’re doing a great job,” he said in a pointed jab at Altman, alluding to the very fears that the CEO had warned about of the company rapidly diluting its talent density and mission orientation, while increasing its bureaucracy. The rapid expansion was also leading to an uptick in firings. During his onboarding, one manager was told to swiftly document and report any underperforming members of his team, only to be let go himself sometime later. Terminations were rarely communicated to the rest of the company. People routinely discovered that colleagues had been fired only by noticing when a Slack account grayed out from being deactivated. They began calling it “getting disappeared.”
To new hires, fully bought into the idea that they were joining a fast-moving, money-making startup, the tumultuousness felt like a particularly chaotic, at times brutal, manifestation of standard corporate problems: poor management, confusing priorities, the coldhearted ruthlessness of a capitalistic company willing to treat its employees as disposable. “There was a huge lack of psychological safety,” says a former employee who joined during this era. “It is like the opposite of ‘a company as a family’—which is fair, you know, it is a company.” Many people coming aboard were simply holding on for dear life until their one-year mark to get access to the first share of their equity. One significant upside: They still felt their colleagues were among the highest caliber in the tech industry, which, combined with the seemingly boundless resources and unparalleled global impact, could spark a feeling of magic difficult to find in the rest of the industry when things actually aligned. “I would say OpenAI is one of the best places I’ve ever worked but also probably one of the worst,” the former employee says.
For some employees who remembered the scrappy early days of OpenAI as a tight-knit, mission-driven nonprofit, its dramatic transformation into a big, faceless corporation was far more shocking and emotional. Gone was the organization as they’d known it; in its place was something unrecognizable. “OpenAI is Burning Man,” Rob Mallery, a former recruiter, says, referring to how the desert art festival scaled to the point that it lost touch with its original spirit. “I know it meant a lot more to the people who were there at the beginning than it does to everyone now.”
In those early years, the team had set up a Slack channel called #explainlikeimfive that allowed employees to submit anonymous questions about technical topics. With the company pushing six hundred people, the channel also turned into a place for airing anonymous grievances. In mid-2023, an employee posted that the company was hiring too many people not aligned with the mission or passionate about building AGI.
Another person responded: They knew OpenAI was going downhill once it started hiring people who could look you in the eye.
ChatGPT also surprised Microsoft. OpenAI leaders had told its partner, as they’d told their own employees, that the chatbot would be a “low-key research preview.” It was clearly anything but.
The mismatch initially peeved the tech giant’s executives. ChatGPT had completely stolen the thunder of Microsoft’s chatbot for Bing. When Microsoft pushed out Bing AI the following February, the product would also take a PR hit with an article by New York Times columnist Kevin Roose about it pushing him to divorce his wife. It was far from the reception Microsoft had hoped for and, by comparison, had made OpenAI look even better.
But the crossed wires weren’t nearly enough to dampen Microsoft’s enthusiasm for OpenAI. The enormously positive reception to ChatGPT was contagious, and the continuously improving capabilities of OpenAI’s models made the giant’s executives even more excited. Microsoft was now readying a whole new slate of Copilots for the tech giant’s products based on GPT-3.5 and GPT-4, which it planned to release one after the other in a steady drumbeat of announcements. After Bing in February, March was for Microsoft 365 Copilot, bringing an AI-powered chat-based interface to every Office product from Word to Outlook to Teams.
The way in which Microsoft executives talked internally about the OpenAI partnership was also rapidly shifting. Before, Microsoft felt like it had power over OpenAI; now some of the giant’s executives felt like OpenAI had power over Microsoft. There was a creeping sense of inadequacy within parts of Microsoft that its own AI research efforts had failed to achieve what OpenAI had pulled off. If Microsoft walked away as OpenAI’s main investor, the startup could find other investors, a former Microsoft employee remembers of some of the executives’ thinking. But if OpenAI walked away from Microsoft, would the tech giant find another OpenAI?
At the same time, many executives were no longer talking merely about beating Google. Where they had once responded to OpenAI’s strange talk about AGI and highfalutin language about its power to invent the future with polite skepticism, they were now believers. AI, AGI, generative AI—whatever you wanted to call it—this technology was the future, and Microsoft was shepherding it hand in hand with OpenAI. The more Microsoft believed, the more the company reoriented its rhetoric and strategy. “I saw the incentives at Microsoft push more and more toward a narrow conception of the future,” the former employee says. “I saw the technology become narrowed into something propped up by narrative rather than reality.”
Nadella implemented a new strategy for distributing Microsoft’s computing resources. He shifted GPUs away from Microsoft’s research teams to support OpenAI. The company also consolidated all of its GPUs into one pool for better supporting generative AI workloads. “The typical Microsoft employee had no fucking clue what OpenAI was before January last year,” one Microsoft employee remembers. Now they were receiving urgent directives from their superiors about finding ways to intersect their work with OpenAI technologies.
The tech giant would experience a rapid proliferation of over one hundred new generative AI projects within just a few months as employees experimented with various ways of using GPT-4 and ChatGPT. In an ironic twist, the aggressive adoption would force Microsoft to grapple with many of the same challenges that other companies would face as they raced to adopt generative AI without fully understanding it. That included causing headaches for the risk and compliance teams. Not everyone was using Microsoft’s internal versions of the technologies; some were opting to use the free version of ChatGPT straight from OpenAI, which trained on user data, raising concerns over whether that could leak Microsoft customer information or interfere with regulatory compliance. While some employees found the tools a big productivity boost, many also found them exhausting. “There is this culture of ‘Use AI, use AI, use AI,” says one. But “it’s like, okay, this doesn’t help us. We don’t want to use it. And it feels like it’s everywhere and we can’t escape it.”
In switching from Microsoft’s own models to OpenAI’s, many employees also lost control and visibility into a core infrastructure layer of their work. Within the tech giant, access to the startup’s underlying models was tightly guarded, even though they were trained and stored on Microsoft’s servers. Most Microsoft employees could no longer examine the training data or tweak the weights of the models they were using. OpenAI’s models were instead delivered through an API, as they were to other OpenAI customers.
But in exchange for these trade-offs, Microsoft was being richly rewarded. The company was seeing a massive surge in inbound customers for its Azure AI platform as the only cloud provider able to offer the typical benefits of the cloud, including simpler data storage and management, alongside the ability to process that data with OpenAI’s capabilities. “Azure OpenAI Service is getting us in the door with many new customers these days,” Eric Boyd, the corporate vice president of the AI platform, wrote in an email to his division in May 2023. In August, he enthused once more. “Every now and then it’s great to take a step back and marvel at just how far we’ve come in just one year,” he wrote to his division, adding that the platform that year had seen a “21x increase in customers.” The following month, Boyd celebrated a new milestone. After centralizing Microsoft’s fractured AI efforts onto the platform, and with the thousands of new customers who had joined Azure OpenAI Service, traffic on the platform had grown tenfold in just nine months. In January 2023, it had been receiving one hundred billion monthly inference requests; now in September, it was receiving one trillion.
That summer, Microsoft CTO Kevin Scott was effusive with his praise during an OpenAI all-hands meeting. “We have stopped, like, our AI machine learning investments in a bunch of places to the point that, like, people are like, ‘Hey, you know, fuck you, Microsoft,’ ” Scott said, referring to the shifting away of GPUs from some of Microsoft’s own internal research. “And we’ve taken the bet because we believe that you all are doing the absolute best work in the industry.”
ChatGPT firmly codified OpenAI’s turn away from nonprofit and toward commercialization. Altman and other executives pushed to build on the momentum of the chatbot’s success by launching a slew of paid products. In February 2023, it released a paid version of ChatGPT; in March, one after the other, it released an API version, the Whisper API, and finally GPT-4. “After ChatGPT, there was a clear path to revenue and profit,” a former employee says. “You could no longer make a case for being an idealistic research lab. There were customers looking to be served here and now.”
The burst of new products overwhelmed the trust and safety team anew. For a while, OpenAI had enticed users to join its API by giving them an initial twenty dollars’ worth of free usage credits. With the mega-popularity of ChatGPT also sparking a dramatic surge in API usage, this sign-up incentive now posed a problem: Many users were creating new accounts at scale to cash in repeatedly on the bonus. In some cases, users were also spinning up new accounts to evade bans and suspensions on their old ones. The mass fraud was leading OpenAI to lose huge amounts of revenue as costs climbed with its need for more and more servers. Still numbering fewer than twenty people and with its reactive enforcement efforts severely hampered, the trust and safety team redirected its personnel once again to whack-a-mole the new vector of abuse.
Soon enough, the constant whiplash would push Willner to severe burnout. Within months he and several of his staffers would depart the company. By the end of that year, the team would dissolve and some of its remaining members be folded under a broader safety systems operation, headed by a longtime OpenAI researcher Lilian Weng. Some of the trust and safety people would come to feel that the persistent clashing between Applied and the Safety clan, with their overemphasis on Doomerism, had cultivated a culture among many of the company leadership to heavily discount any kinds of “safety” concerns, leading to an environment that made their function, already disempowered at most tech companies, even more so at OpenAI.
Indeed, with every new launch, the clashing continued, including over the release of GPT-4. While many in Applied felt the six-month delay in launching the model to be abundantly cautious, some in Safety felt it still hadn’t given them enough time to finish their comprehensive testing and alignment.
The model’s high rate of hallucinations, for example, had continued to prove particularly difficult to get under control, even with a concerted RLHF effort to address the problem. In November 2022, as users latched on to ChatGPT as if it were a search tool, spawning widespread speculation that it could unseat Google, an internal document noted that OpenAI’s model had hallucinated during an internal test on roughly 30 percent of so-called closed-domain questions.
Closed-domain questions are meant to be the easiest category of questions: when users ask the model only about the information they give it—for example, uploading a pdf and asking for a summary, or providing bullet points and asking for a rewrite to complete sentences. This is in contrast to open-domain questions, when a user asks the model a question without reference material—pop culture, ancient history, high school biology—the way you would a typical search engine.
Meanwhile, GPUs became an ever-present constraint on OpenAI’s research and expansion. New research and product or feature launches had to be repeatedly delayed or shelved due to a lack of chip capacity. After ChatGPT went viral, SemiAnalysis, a trade newsletter focused on the semiconductor industry, estimated that the company was spending some $700,000 a day on compute costs alone. After executives reallocated chips from the Research division, Applied made commitments to return them by a certain date. That date came and went, but Applied couldn’t return them. With the continued rapid growth in users, the division needed more chips, not fewer.
The pressure accelerated OpenAI’s research into more-efficient models. During its work to improve the company’s compute efficiency, the Research division had figured out a new method for developing Transformer-based models that were cheaper to serve to users. As they used that method, which they named DUST, they assigned code names to the resulting models to follow the desert-based theme. The first one, an optimized version of GPT-3.5, they called Sahara, which they released in February 2023 under the public name GPT-3.5 Turbo. Another they called Gobi, which would be an optimized version of one of its text-and-image models.
A third one, meant to be an optimized version of GPT-4, they called Arrakis, the desert planet from the science fiction epic Dune. But after months of work, the team was still struggling to make Arrakis more efficient while maintaining the same performance. The project ate up significant computational resources. Shortly thereafter, leadership scrapped it to free up GPUs for other projects.
As Microsoft worried about whether OpenAI could just leave the relationship, OpenAI felt its own vulnerabilities about whether Microsoft would stop cooperating if the startup didn’t work hard to please its partner. Arrakis felt like a particular setback in this regard. In the hopes of impressing the giant, OpenAI had reworked its road map to prioritize delivering the model over its own more strategically aligned projects, including an effort to apply GPT-4 to a search engine product. Instead, the failed effort left some senior Microsoft executives disappointed.
There was also a new awkward reality: OpenAI and Microsoft were beginning to compete for contracts. Codex and DALL-E 2 had convinced OpenAI to retain control of delivering its technologies directly to users. ChatGPT and GPT-4 were showing that OpenAI could also make its own money. That meant directly pitching to customers the very same technology that it was handing over to Microsoft, which was then pitching to the exact same customers.
Handing off the technology had its own challenges. As OpenAI’s release schedule picked up, so did Microsoft’s. But Microsoft completely dwarfed OpenAI, leading to a dynamic where a single OpenAI employee could get pinged by dozens of Microsoft counterparts across various departments with all sorts of questions about technical or logistical details with every new product. It was growing increasingly frustrating and overwhelming for OpenAI staff to support Microsoft releases while focusing on their own road map.
Much of the smoothing over of the relationship was left to Murati. Murati sought to work closely with Scott and other Microsoft executives on coordinating the timing of product releases, strategizing how OpenAI and Microsoft would differentiate their offerings and finding more productive ways for the two organizations to work together. In the summer of 2023, in an attempt to cut back the communication burden, a team of Microsoft engineers began embedding inside OpenAI with full access to everything to streamline transfers of technology.
With the deeper integration, both Altman and Nadella were growing more involved than ever before, especially with the management of compute resources, making tough calls on how to redistribute chips and money for yet more chips to OpenAI’s ever-compute-hungry operations. Nadella would tell The New York Times that OpenAI’s demands would grow so fast and so high that Altman would start calling him every day saying, “I need more, I need more, I need more.”
OpenAI didn’t just need more data centers to serve ChatGPT. It still needed far-more-powerful supercomputers to train its future generations of models. To fulfill that aggressive and escalating demand, the two companies were sketching out a new unprecedented project called Stargate to OpenAI and Mercury to Microsoft: a single supercomputer that, for its construction alone, would cost an estimated $100 billion. The empire of AI was returning to the exact same form of expansion as the empires of old: To fuel its growth, it needed more material resources and, crucially, more land.