Nadella's Paradox. The Microsoft's CEO on his AI balancing act. Nervous customers and a volatile partnership with open Ai are complicating things for the world's most valuable company. By Austin Carr and Dina Bass Read aloud by Mark Leidorf. Satya Nadella arrived at the World Economic Forum in January ready to talk up his triumphs in artificial intelligence when a dangerous threat emerged. A little known Chinese start up named deep Seek had just released an AI model that quickly became the talk of Davos, Switzerland. Nadella, the chief executive officer of Microsoft, gathered his lieutenants to assess the out of nowhere competition. They set up a virtual war room on where Else Microsoft teams to coordinate a response. The new model. Deep Seek R I could deliver results roughly on par with those of Open Ai at a fraction of the price. Computer processing that would cost one thousand dollars through open AI ran for just thirty six dollars through R one. Even crazier, Deep Seek made R one open source, meaning anyone could install versions of it for free if they had a powerful enough computer. Open AI has been so far ahead that no one's really come close, Nadella tells Bloomberg BusinessWeek, Deep Seek and R one in particular, was the first model I've seen post some points to the schmoozers in Davos. This seemed like a huge problem for Microsoft. The company had invested thirteen point seventy five billion dollars in open AI by that point, and had already committed to spend eighty billion dollars on AI data centers in this fiscal year alone, all under an assumption that better AI required more computing resources. Nadella immediately ordered his team to conduct a security review of deep Seek R I. They scrutinized a research paper Deep Seek published detailing its work, and contacted the startups engineers, ring them with questions about the model. Few others have even been able to get the opaqu Chinese company to respond to emails. Soon, roughly one hundred Microsoft employees were coming in and out of the team's video conference rooms, testing the security of Deepsek's code base and exchanging notes. People didn't sleep, says Asha Sharma, the company's AI platform head who spearheaded the effort. It was forty eight hours of going through every single thing, R one appeared to be legit, But instead of trying to stomp out this new rival, Nadella chose to embrace it. He instructed his team to install R one on Microsoft's cloud and sell access to it to customers, alongside products from Open Ai and Microsoft itself. Get it out, Nadella recalls telling his staff, Nadella's primary allegiance now isn't to Open AI's very expensive skunk works. His ultimate objective is to sell whatever AI his customers might want through Microsoft's platforms. Nadella had spent three years running various parts of the company's cloud business called Azure before he became CEO in twenty fourteen, and that's now central to his AI strategy. Customers can choose from over nineteen hundred different models on Azure, including ones made by Meta and Open Ai, and upstarts such as coher, Mistral, Stability AI, and now deep Seek. Some, though, such as Google's Gemini, aren't available to Microsoft for competitive reasons. Whether a model's usage costs a customer ten dollars through open Ai or ninety cents via deep Seek, Microsoft gets paid for the cloud computing cybersecurity protections data storage and other upsolt services. The Deep Seek episode highlights another arguably more revealing part of Nadella's thinking. AI is rapidly commoditizing, and this is a good thing for Microsoft. While everyone in Davos was focused on AI consumption, Nadella was contemplating the history of coal production. One of his favorite economic theories is the Jevins paradox, which posits that as a resource becomes more accessible and its usage more efficient, consumption increases. This happened with coal during the eighteenth and nineteenth centuries, and more recently with plane travel, when plummeting operational costs and airfares helped create frequent flyers, new flight destinations, and booming sales for airlines. Nadella believes a similar phenomenon will play out with AI. This econ mindset has been driving the company toward creating its own AI architectures, including some tiny models with capabilities similar to those of deep sek r I. Over the past year, it's also been training a series of large language models called MAI two, the latest iteration of Microsoft's in house alternatives to open AIS models, which it had been developing in secret. The goal is to build AI that requires less computing power than Chat GPT and bring down the cost to operate Microsoft's equivalent service Copilot. The company will still regular tap open AI's bleeding edge technology, but Nadella is convinced Microsoft can deliver near chat GPT quality for a lot less. What this all means for Microsoft's arrangement with OpenAI is complicated. Six years on, what began as a nurturing kinship has turned into an intense sibling rivalry. A power struggle with open Ai in twenty twenty three was a turning point in the relationship between the companies, and in various ways, Microsoft is asserting its dominance. The MAI two project, for example, offers protection from Microsoft to avoid being left exposed in case anything catastrophic happens to open Ai, says Mustapha Sullimann, the head of Microsoft's consumer co pilot efforts. The relationship with OpenAI has to date been pretty amazing, he says, but this is a fifty year old company that needs to be in an amazing place. In twenty thirty, twenty thirty five, and twenty forty, open Ai CEO Sam Altman has repeatedly signaled that his company's mission is to create an artificial general intelligence, a final frontier for the current AI era. With MAI two, Nadella is instead pursuing maximum cost efficiency, even though it won't be as intelligent as open AI's most advanced models. He already feels a sense of ownership over open AI's intellectual property. MAAI is not a clone, he says. When I have a contract with open ai that says, oh, I'm essentially funding it and have IP rights, it would be stupid to sort of do it twice, so we avoided that stupidity. If Nadella sounds defensive, that's because he's facing a concentrated version of the trial. Practically every CEO is undergoing now. Even with Microsoft's considerable resources. Nadella must decide how AI will reshape his business by making a set of extremely difficult trade offs between embracing newfangled technology and shielding his employees and business partners from disruptive systems, and between marching in line with what's worked in the past and blindly leaping into the future. Up to this point, shareholders have lauded Nadella's performance, making Microsoft the most valuable company on Earth. If he falls flat, though his may be one of the first jobs threatened by AI. The alliance between Microsoft and open Ai, and between Nadella and Altman, was the foundation of the present AI boom. To realize their ambitions, Altman's brainy engineers needed Nadella's money and data centers. Nadella's one billion dollar investment in open Ai in mid twenty nineteen was precient, giving Microsoft early and exclusive access to the research labs technology. Two years later, Microsoft rolled out gethub Copilot, a coding assistant that quickly became a must have for programmers. In twenty twenty two, OpenAI held demonstrations for senior Microsoft executives of a groundbreaking new model. Over the next several months, groups inside Micro tinkered with other ways they might adapt open ais technology to the daily needs of the modern office worker. One promising prototype, called intelligent Recap, would summarize conference calls in teams. Before Microsoft could release any of it, though, OpenAI built a deceptively simple chat interface around its language engine, and people went nuts. After the release of Chat GPT in November twenty twenty two, Microsoft found itself in a position similar to that of every other tech company playing catch up. It was especially bruising for Microsoft, almost as if the company had been holding the components for the original iPhone and didn't know what to do with them. Nadella took a firmer hand, directing the AI projects and setting aggressive deadlines. Within a month or so of chat GPT's debut, a software team held a demonstration for Nadella showing some open AI powered functions built into office programs. The idea was for users to select an option in outlook to have AI composed an email reply, or hit a button in word to convert an outline into a rough draft. Satya said, Look, I just feel like we're missing something, recalls Jared Spataro, who oversees marketing for Microsoft's workplace apps. This feels like just a set of features. Instead of bearing these new tools in different menus. Nadella pushed the teams to bundle all the open AI enhancements into a single assistant later branded Copilot, pinned to the side of Microsoft's apps. Some later grumbled that the vision wasn't totally clear, due in part to a lack of coordination between teams and a reticence to mess too much with what a former Copilot product manager calls the Crown Jewels, Windows Word Excel. The resulting AI assistants were inconsistent. The PowerPoint copilot behaved much differently than the GitHub or team's co pilot. It felt like there were thirteen different copilots, says a former Microsoft design leader, who, like some other ex employees interviewed by BusinessWeek, asked not to be an to avoid career repercussions. In March twenty twenty three, the company unveiled its business oriented copilot for office programs, which are part of its Microsoft three to sixty five subscription. By that point, chat gpt boasted more than one hundred million active users and was selling a twenty dollars a month subscription for its most advanced AI. To get early access to copilot, businesses were soon asked to purchase a minimum of three hundred licenses at thirty dollars a month each on top of their existing Microsoft three sixty five plans. Many customers who tried this copilot founded an inferior product, and an alarming pattern emerged. Users often preferred to go to chat gpt for questions and paste its responses into word. A notable exception. People loved the intelligent recap feature in teams. When it arrived in May twenty twenty three, chat GPT didn't have that skill. Yet. Part of what slowed down Microsoft was its preoccupation with software security compliance, data protections, and legal reviews left its technology lagging, but eventually proved to be an excellent selling point to corporations. It's been great for us to go to customers and say, hey, you like chat GPT. This is that, but better, more powerful user experience, better security, says Jeff Tepper, who runs Teams. Yet it was clear something else was holding the company back from taking bigger risks with its products. The former copilot manager describes it as a kind of post traumatic stress disorder from embarrassments that stretch as far back as the proto chatbot Clippy, the animated paper clip that became an enduring punchline. Chat GPT could perform way more actions with fewer bureaucratic limitations, and Microsoft's engineers knew it. This person says he estimates that of the hundreds of people in his organization, three quarters of them paid out of their own pocket for chat GPT because Microsoft wouldn't let them expense it. Tepper says it's common for employees to tech other company services to better understand the market, Microsoft's customers were gravitating to chat GPT two. Maderna found Open AI's options to customize chat GPT made it far more versatile than Copilot, says Brad Miller, the biotech company's former chief information officer. Microsoft was like, it's embedded in office, just use it in office, and we were like, there's so much more that we want to do. He says that fall, as Microsoft was ready and Copilot for a general release, more than eighteen thousand businesses were accessing open ai models through Azure, whereas Copilot's corporate subscribers numbered in the hundreds. Microsoft still made oodles of money from renting access to open AI's models and a revenue share agreement on chat GPT subscriptions, but generally not as much as it would have if clients paid for Copilot instead. Nadella tried to build momentum for Microsoft's AI services at the company's annual Developer conference in November twenty two, twenty three, showcasing a tool to create custom copilots similar to what businesses were already doing with chat GPT. The effort was overshadowed by News on the last day of the event that open AI's board had ousted Altman as CEO. Nadella, despite having increased his financial commitment to open Ai by ten million dollars earlier that year, only learned of Altman's firing after it transpired. Microsoft management now refers to the infamous four day period when Altman lost his job, accepted an executive role at Microsoft, and then wrested back control of open Ai from his board using an array of euphemisms that particular incident, or the November open Ai sort of situation. In truth, that particular incident irrevocably damaged the partnership between the companies. Just before, Nadella had been considering another massive cash infusion in the tens of billions of dollars to keep open Ai hurtling forward ahead of Amazon and Google, according to a person with direct knowledge of the deliberations, no more. Nadella figured Microsoft had risked enough on the immaturely run start up. Nadella had been bonding with Suliman, an influential figure in the ai world who'd co founded the lab that became Google deep Mind and was at the time overseeing a startup called Inflection AI. Inflection had a chatbot geared more for EQ than IQ, positioning the product as a sort of extremely knowledgeable friend. Inflection's chatbot never reached anywhere near the prominence of chat GPT, but Nadella loved its crisp, playful responses. It was Nadella's turn to surprise Open Ai. In March twenty twenty four, Microsoft spent six hundred and fifty million dollars to license Inflection software and hire most of its team. Microsoft informed open Ai a day before the deal was announced. Nadella put Suliman, a forty year old Britain, in charge of Microsoft's Consumer AI, a portfolio that incl It's bing the copilot app, the Edge web browser, and MSN Web portal. The slate was in weak shape, especially among young people who aren't exactly in MSN dot COM's core demographic. Generation Z is growing up with iPhones and MacBooks, Gmail and Google Docs, Instagram and TikTok and now chat GPT. The way that Saya put it to me is that we're going through this generational shift where younger people today who are going to become the CEOs and industry leaders in a decade's time and make purchasing decisions in the enterprise, will have not really experienced Microsoft. Sulliman says that's a big structural challenge. Working out of offices in London and Mountain View, California, Sulliman's employees set about creating a version of Copilot designed for life away from the office. As they'd done at Inflection. They taught this Copilot empathy, humor, and kindness. Do people want an emotional connection with the company that makes excel Suliman thinks so, or at least he believes it'll be harder to switch to a competitor if the user sees Copilot as a friend or therapist. But Suliman grappled with a bigger challenge. His work essentially consisted of modifying open Ai technology, and he says Microsoft wasn't getting access to the startup's new models as early as it was entitled to. OpenAI would roll out a model to chat GPT customers with a splash before he had enough time to integrate it into Microsoft's software. Soliman, who has a reputation for having a hot temper despite his EQ values, could be indelicate on conference calls. Tensions escalated as he pushed open Ai to share more and earlier access to its models. Brad Lightcap, OpenAI's chief operating officer, says that technology transfer is an incredibly complex operational challenge and that OpenAI is one of Microsoft's biggest Azure customers and expects to remain a large client after their current agreement expires in twenty third. Suliman says he's not surprised the fast growing startup occasionally forgets some of its obligations to Microsoft. I'm quite empathetic for them, he says. Empathy is a big theme for Suliman. For a demonstration of what he's been working on in Copilot, he taps the voice agent on his iPhone app and tells it he's in a difficult life state and struggling with anxieties. A soothing male voice with a British accent responds tough situations can fog up clear thinking and gently asks if he can share more about what's stirring up his emotions. Copilot is programmed to sense a user's comfort boundaries and carefully diagnose issues before suggesting solutions. Sulliman says, did you hear pauses? Did you hear the volume go down? He asks with pride. That's extremely hard to do. We'll be right back with Nadella's paradox. Welcome back to Nadella's Paradox. On a frigid February morning, at Microsoft's headquarters in Redmond, Washington, more than a dozen managers from Estay Lauder are receiving a crash course in Copilot. The cosmetics company has been exploring ways for Microsoft's AI to assist in a corporate restructuring intended to improve management of ingredients, product lines, and supplies. Although Estay Lauder is a longtime Copilot customer with fifteen thousand licenses, employees are only just learning about some of the ways it can be used. When a few of them express interest in a tool to automatically manage the menuche around hiring, a Microsoft sales rep informs them they've had access to that Copilot feature for a while. For Microsoft, these moments when a customer realizes Copilot can do something genuinely useful for them are essential. Roberto canavari, Esta, Latter's executive vice president who oversees its global supply chain, says he experienced his own just weeks earlier after spending an entire weekend studying and annotating a thirty page operations report. He asked Copilot on a lark for its summary. The resulting one pager was almost as good as the one he wrote. It made me feel very old and not very smart, he says jokingly. While Sulliman worked on a more empathic Copilot for consumers, his cohorts in the Enterprise Group spent the past year focused on automating routine and monotonous workflows, such as reconciling bank invoices or reviewing warehouse inventory. Microsoft found that office workers were most impressed when Copilot could do the really boring parts of their job for them. Instead of talking in the abstract about AI's world changing potential, Microsoft sales reps now pitch Copilot's ability to add a slide about budget variances to a PowerPoint. The approach is resonating. In a quarterly financial disclosure in January, Microsoft said it's on track to generate more than thirteen billion dollars a year in AI revenue. It earns money from Copilot subscriptions, usage fees from AI models running on Edge, and a revenue share with open Ai on Chat, GPT memberships, GitHub Copilots are past fifteen million users as of March, and the corporate version of Copilot now boasts hundreds of thousands of businesses as customers, with their engagement tripling over the past year. Nesley has found Copilot wildly useful for comparing contracts and for language translations for customer service teams, says Christopher Wright, CEIO at the Swiss food conglomerate. Nesley pays for more than twenty five thousand Microsoft three sixty five licenses with Copilot, and Wright estimates that ten minutes of time savings per day covers the cost of each employee's subscription. The company is also starting to pay extra to use a product called Copilot Studio to craft custom AI agents with subject matter expertise. Wright describes it as shockingly cheap for what the company gets out of it. Don't tell Microsoft that, please, He says. A lot of these AI capabilities still depend on models from Open Ai, a fact Microsoft is less keen to advertise lately. Copilot will automatically deliver results using whatever Microsoft or OpenAI model is best for a particular prompt, rather than ask customers to choose. Meanwhile, Chat GPT gives users the ability to select from a jumbled list that includes four to oh four point five, three four mini and four mini high designations that are probably more meaningful to robots than to people. Spataro, the Microsoft marketing leader, scoffs at the design choice man. The interface is getting very loaded, he says. Altman has said OpenAI is working on simplifying its offerings. Judson Aldhoff, Microsoft's sales chief, says few car buyers care what engine is under the hood, so long as the vehicle can get them from point A to B. Is it a four cylinder turbo an eight cylinder? He asks, how much the whole package costs is what they care about. If Chat GPT is a top of the line range rover, MAAI two will be like a prias, not as fast or powerful, but much more efficient for your daily commute. And Microsoft has been training even smaller and more efficient models that specialize in math or speech recognition, part of a portfolio called FI. These AI efforts could someday allow you to summarize a lengthy document or process voice commands using Copilot on a standard laptop without an Internet connection. The key to getting there, according to Suliman, can be found in a mathematical concept called the Pareto frontier. The Wikipedia article helpfully describes it as the set of all Pareto efficient solutions, based on an efficiency theory developed by a nineteenth century Italian economist. The term refers to the sweet spot between a set of desirable attributes. It's thrown around in meetings by software engineers seeking to find an equilibrium between how fast and high quality an AI model is versus what it costs to operate. Google CEO Sunder PITCHAI boasted on a recent earnings call about its own advances in the Pareto frontier so the company can deliver more AI features on Android phones, its search engine, and YouTube without breaking the bank. Apple, too, has cited Pareto in its research around building language models that can run on an iPhone. Nadella offers various reasons for Microsoft to strengthen its in house AI units after spending billions on Open AI. Partly, it's about exercising in the gym, he says, to build up the company's AI model training muscle so it's not entirely dependent on Open AI. He says he won't necessarily favor one Lab's technology over the others. Sometimes the priority will be Sulliman's cheaper models, sometimes Altman's more proficient ones. The rush to embrace deep Seek in January wasn't a one off either. Microsoft added another deep Seek model to its cloud catalog in March, renting it for half the price of an equivalent from OpenAI. Someday Copilot could run certain queries through Deepseek or another model available in Microsoft's cloud, something the consumer and enterprise teams have been testing. These moves are not so much about hedging Nadella's bet on OpenAI. It's almost like he's tossing chips every which way on the roulette table. Of course, Nadella isn't always winning on the day. In January, when deep Seek first dethroned chat GPT atop the Apple App Store in the US, Copilot wasn't anywhere in the top two hundred, according to research firm Sensor Tower. Although Deepsek's open source models are available on Azure, the startup runs its own mobile chatbot app, delivering no benefit to Microsoft. Other latecomers, such as Google's Gemini and Elon Musk's Grock, are also consistently ahead. Nadella points out that Microsoft was once less relevant in video conferencing too, everybody would say hey, Zoom, zoom zoom, whom Nadella says, we won that in the enterprise. A Zoom spokesperson declined to comment. In AI, the company's products are scattered and not in a good way, says Ethan Mollick, a professor at the Wharton School of the University of Pennsylvania who studies generative AI and wrote the book co Intelligence, Living and Working with AI. But Microsoft has an edge on every other company out there. Their advantage is they own the workplace, he says. Corporate America runs on Microsoft. Competitors have accused Microsoft of using anti competitive tactics in regard to its alliance with open AI. Donald Trump's Federal Trade Commission is pursuing a broad antitrust probe of the company that was opened by the Biden administration. Microsoft has said its cooperating with the agency. Many incentives inside Microsoft are aligned to push Copilot. On May nineteenth, the company will hold its biggest annual developer conference, where Copilot is expect to be the star of the show. Already, the company boasts that about seventy percent of Fortune five hundred companies use it, and that Blackrock, Price, Waterhouse Coopers and many others have signed massive contracts for the AI license. Arman Waworski, an engineer overseeing AI solutions for Austria's Revice and Bank International, says Microsoft's sales team is consistently pitching his staff to use Copilot. Use Copilot, use Copilot. Wiwarski says. The bank's employees do love to use Copilot in teams for meeting summaries, but way more of them over twenty five thousand rely on a customized corporate version of Chat gpt RBI built it using Microsoft's cloud tools for compliance reasons. Maybe they don't deliver everything from the very beginning, or sometimes for me they are too slow, Wworski says, but in the end Microsoft completely understands how to bring certain technologies into our environment. Copilot consumes Nadella's life outside the office's well. He likes podcasts, but instead of listening to them, he loads transcripts into the Copilot app on his iPhone so he can chat with the voice assistant about the content of an episode in the car on his commute to Redmond at the office. He relies on Copilot to deliver summaries of messages he receives in Outlook and teams and toggles. Among at least ten custom agents from co Pilot Studio, he views them as his AI chiefs of staff, delegating meeting prep, research and other tasks to the bots. I'm an email typist, Nadella jokes of his job, noting that Copilot is thankfully very good at triaging his messages. Nadella, who grew up in Hyderabad, India, the son of a Marxist economist, was seven when Bill Gates and Paul Allen started Microsoft in nineteen seventy five. He spent more than half his life at the company. During Nadella's tenure, he's seen the emergence of the world Wide Web, the creation of Internet Explorer, the dot com crash, the USC government's landmark antitrust lawsuit against Microsoft, and the disastrous losses that followed to Google in search and Apple in phones. AI could top them all. At an event in April to celebrate the company's fiftieth anniversary, Gates quipped that Copilot would eventually take over as CEO. For now, Nadella is still trying to earn his keep. He sought to rally the troops by telling them to forget about the successes of the past and be ready to abandon what's worked up to now. Says Charles Lamana, a vice president who oversees Copilot Studio and some other products. The last five years we spent building it doesn't matter. It's not worth anything anymore, he recalls, Nadella, telling them burn the ships. One ship recently set ablaze. In May, Nadella shut down Skype, which Microsoft once spent eight point five billion dollars to acquire, only to later replace it with teams. With AI promising or threatening such existential change, there remain in weighty questions about who will win and who will lose from this next technological transformation. In public, Nadella often sounds like a macroeconomist. He draws parallels between this moment in AI and the industrial Revolution and the convergence of the global North and South, and cites economic and labour theorists from David Ottor to Friedrich Hayek to Herbert Simon. In particular, the Jevins paradox, first described in an eighteen sixty five book on British coal production by William Stanley Jevins has greatly informed Nadella's understanding of Microsoft's place in the AI revolution. Jevons found that the invention of a more fuel efficient steam engine made coal a more desirable fuel source, igniting a surge in coal consumption. That's why Nadella appears unphased by the perception of deep seek undermining the value of his investment in open AI or the buildout of Microsoft data centers. More efficient AI will eventually mean a greater need for AI services and infrastructure to power them. In his view. That's not to suggest there won't be a reckoning for excessive spending on AI model training and in Nvidia's graphics processing units GPUs. Open AI is working with Oracle and soft Bank on a massive data center endeavor called the Stargate Project, promising to spend a whopping five hundred billion dollars over the next four years. Nadella chose not to be part of the financing, given the tens of billions he's already investing annually in Microsoft's own AI infrastructure. I wouldn't hold it against anyone, certainly not a CEO of a public company, to realize when there are limits of what their company can bear says Lightcap, the OpenAI COO. At the same time, Microsoft is pulling back on some of its own data center projects around the world. These moves are an acknowledgment that it doesn't make sense to keep piling money into what is effectively a real estate asset, says Lyla Tredikoff, a former deputy chief technology officer at Microsoft. On the other hand, the reckoning could be far more consequential for the broader society. Microsoft has been very careful with how it talks about AI, says a person with direct knowledge of its AI marketing who has not to be identified. Copilot is supposed to help accomplish tasks, not automate jobs. Interviews with Microsoft customers and salespeople find that the latter is already occurring. They describe contractors, outsourced programmers, interns, and summer analysts all losing work to AI. In his book The Coming Wave, published before he joined Microsoft, Suliman laid out the risks more bluntly. These tools will only temporarily augment human intelligence. They will make us smarter and more efficient for a time, and will unlock enormous amounts of economic growth, but they are fundamentally labor replacing, he wrote, Suliman advocated for regulatory oversight and other government interventions, such as new taxes on autonomous systems and a universal basic income to prevent a socioeconomic collapse. Nadella is more optimistic than his new deputy. In an interview at Microsoft headquarters in conference room number thirty four thirty five sixty three, sitting next to his human chief of staff, Nadella makes the case that his co pilot assistants won't replace her as she stoically types notes of our conversation on her tablet. He acknowledges that AI will cause hard displacement and changes in labor pools, including for Microsoft. Altoff, the sales chief, says Nadella is pressuring his team to find ways to use AI to increase revenue from its operation instead of boosting headcount. Meanwhile, Microsoft said on May thirteenth that it's cutting six thousand employees, about three percent of its workforce. But Nadella contends that AI could end up delivering more societal benefits than the Industrial Revolution did. When you create abundance, he says, then the question is what one does with that abundant to create more surplus. He argues again that it comes down to economics. If Microsoft spends hundreds of billions of dollars on AI projects, he estimates that will likely generate trillions of dollars in economic activity. More data centers means more construction, energy use, manufacturing equipment, and so forth. More AI use means more opportunities to create new kinds of businesses, especially as AI bridges the skills gap between an entry level employee, expert engineer, and executive. There may be a new startup with one hundred founders, Nadella says, or if the AI ever gets good enough, perhaps one hundred startups with just one founder. It will take time for the jobs of the AI era to arrive, he says, and then the huge economic and productivity boosts will follow, just as they did after the advent of the steam engine, electricity, the computer, and the internet. When that happens, then everything computes at a societal level, Nadallas says. Otherwise, all you have is a bunch of GPUs

Nadella's Paradox. The Microsoft's CEO on his AI balancing act. Nervous customers and a volatile partnership with open Ai are complicating things for the world's most valuable company. By Austin Carr and Dina Bass Read aloud by Mark Leidorf. Satya Nadella arrived at the World Economic Forum in January ready to talk up his triumphs in artificial intelligence when a dangerous threat emerged. A little known Chinese start up named deep Seek had just released an AI model that quickly became the talk of Davos, Switzerland. Nadella, the chief executive officer of Microsoft, gathered his lieutenants to assess the out of nowhere competition. They set up a virtual war room on where Else Microsoft teams to coordinate a response. The new model. Deep Seek R I could deliver results roughly on par with those of Open Ai at a fraction of the price. Computer processing that would cost one thousand dollars through open AI ran for just thirty six dollars through R one. Even crazier, Deep Seek made R one open source, meaning anyone could install versions of it for free if they had a powerful enough computer. Open AI has been so far ahead that no one's really come close, Nadella tells Bloomberg BusinessWeek, Deep Seek and R one in particular, was the first model I've seen post some points to the schmoozers in Davos. This seemed like a huge problem for Microsoft. The company had invested thirteen point seventy five billion dollars in open AI by that point, and had already committed to spend eighty billion dollars on AI data centers in this fiscal year alone, all under an assumption that better AI required more computing resources. Nadella immediately ordered his team to conduct a security review of deep Seek R I. They scrutinized a research paper Deep Seek published detailing its work, and contacted the startups engineers, ring them with questions about the model. Few others have even been able to get the opaqu Chinese company to respond to emails. Soon, roughly one hundred Microsoft employees were coming in and out of the team's video conference rooms, testing the security of Deepsek's code base and exchanging notes. People didn't sleep, says Asha Sharma, the company's AI platform head who spearheaded the effort. It was forty eight hours of going through every single thing, R one appeared to be legit, But instead of trying to stomp out this new rival, Nadella chose to embrace it. He instructed his team to install R one on Microsoft's cloud and sell access to it to customers, alongside products from Open Ai and Microsoft itself. Get it out, Nadella recalls telling his staff, Nadella's primary allegiance now isn't to Open AI's very expensive skunk works. His ultimate objective is to sell whatever AI his customers might want through Microsoft's platforms. Nadella had spent three years running various parts of the company's cloud business called Azure before he became CEO in twenty fourteen, and that's now central to his AI strategy. Customers can choose from over nineteen hundred different models on Azure, including ones made by Meta and Open Ai, and upstarts such as coher, Mistral, Stability AI, and now deep Seek. Some, though, such as Google's Gemini, aren't available to Microsoft for competitive reasons. Whether a model's usage costs a customer ten dollars through open Ai or ninety cents via deep Seek, Microsoft gets paid for the cloud computing cybersecurity protections data storage and other upsolt services. The Deep Seek episode highlights another arguably more revealing part of Nadella's thinking. AI is rapidly commoditizing, and this is a good thing for Microsoft. While everyone in Davos was focused on AI consumption, Nadella was contemplating the history of coal production. One of his favorite economic theories is the Jevins paradox, which posits that as a resource becomes more accessible and its usage more efficient, consumption increases. This happened with coal during the eighteenth and nineteenth centuries, and more recently with plane travel, when plummeting operational costs and airfares helped create frequent flyers, new flight destinations, and booming sales for airlines. Nadella believes a similar phenomenon will play out with AI. This econ mindset has been driving the company toward creating its own AI architectures, including some tiny models with capabilities similar to those of deep sek r I. Over the past year, it's also been training a series of large language models called MAI two, the latest iteration of Microsoft's in house alternatives to open AIS models, which it had been developing in secret. The goal is to build AI that requires less computing power than Chat GPT and bring down the cost to operate Microsoft's equivalent service Copilot. The company will still regular tap open AI's bleeding edge technology, but Nadella is convinced Microsoft can deliver near chat GPT quality for a lot less. What this all means for Microsoft's arrangement with OpenAI is complicated. Six years on, what began as a nurturing kinship has turned into an intense sibling rivalry. A power struggle with open Ai in twenty twenty three was a turning point in the relationship between the companies, and in various ways, Microsoft is asserting its dominance. The MAI two project, for example, offers protection from Microsoft to avoid being left exposed in case anything catastrophic happens to open Ai, says Mustapha Sullimann, the head of Microsoft's consumer co pilot efforts. The relationship with OpenAI has to date been pretty amazing, he says, but this is a fifty year old company that needs to be in an amazing place. In twenty thirty, twenty thirty five, and twenty forty, open Ai CEO Sam Altman has repeatedly signaled that his company's mission is to create an artificial general intelligence, a final frontier for the current AI era. With MAI two, Nadella is instead pursuing maximum cost efficiency, even though it won't be as intelligent as open AI's most advanced models. He already feels a sense of ownership over open AI's intellectual property. MAAI is not a clone, he says. When I have a contract with open ai that says, oh, I'm essentially funding it and have IP rights, it would be stupid to sort of do it twice, so we avoided that stupidity. If Nadella sounds defensive, that's because he's facing a concentrated version of the trial. Practically every CEO is undergoing now. Even with Microsoft's considerable resources. Nadella must decide how AI will reshape his business by making a set of extremely difficult trade offs between embracing newfangled technology and shielding his employees and business partners from disruptive systems, and between marching in line with what's worked in the past and blindly leaping into the future. Up to this point, shareholders have lauded Nadella's performance, making Microsoft the most valuable company on Earth. If he falls flat, though his may be one of the first jobs threatened by AI. The alliance between Microsoft and open Ai, and between Nadella and Altman, was the foundation of the present AI boom. To realize their ambitions, Altman's brainy engineers needed Nadella's money and data centers. Nadella's one billion dollar investment in open Ai in mid twenty nineteen was precient, giving Microsoft early and exclusive access to the research labs technology. Two years later, Microsoft rolled out gethub Copilot, a coding assistant that quickly became a must have for programmers. In twenty twenty two, OpenAI held demonstrations for senior Microsoft executives of a groundbreaking new model. Over the next several months, groups inside Micro tinkered with other ways they might adapt open ais technology to the daily needs of the modern office worker. One promising prototype, called intelligent Recap, would summarize conference calls in teams. Before Microsoft could release any of it, though, OpenAI built a deceptively simple chat interface around its language engine, and people went nuts. After the release of Chat GPT in November twenty twenty two, Microsoft found itself in a position similar to that of every other tech company playing catch up. It was especially bruising for Microsoft, almost as if the company had been holding the components for the original iPhone and didn't know what to do with them. Nadella took a firmer hand, directing the AI projects and setting aggressive deadlines. Within a month or so of chat GPT's debut, a software team held a demonstration for Nadella showing some open AI powered functions built into office programs. The idea was for users to select an option in outlook to have AI composed an email reply, or hit a button in word to convert an outline into a rough draft. Satya said, Look, I just feel like we're missing something, recalls Jared Spataro, who oversees marketing for Microsoft's workplace apps. This feels like just a set of features. Instead of bearing these new tools in different menus. Nadella pushed the teams to bundle all the open AI enhancements into a single assistant later branded Copilot, pinned to the side of Microsoft's apps. Some later grumbled that the vision wasn't totally clear, due in part to a lack of coordination between teams and a reticence to mess too much with what a former Copilot product manager calls the Crown Jewels, Windows Word Excel. The resulting AI assistants were inconsistent. The PowerPoint copilot behaved much differently than the GitHub or team's co pilot. It felt like there were thirteen different copilots, says a former Microsoft design leader, who, like some other ex employees interviewed by BusinessWeek, asked not to be an to avoid career repercussions. In March twenty twenty three, the company unveiled its business oriented copilot for office programs, which are part of its Microsoft three to sixty five subscription. By that point, chat gpt boasted more than one hundred million active users and was selling a twenty dollars a month subscription for its most advanced AI. To get early access to copilot, businesses were soon asked to purchase a minimum of three hundred licenses at thirty dollars a month each on top of their existing Microsoft three sixty five plans. Many customers who tried this copilot founded an inferior product, and an alarming pattern emerged. Users often preferred to go to chat gpt for questions and paste its responses into word. A notable exception. People loved the intelligent recap feature in teams. When it arrived in May twenty twenty three, chat GPT didn't have that skill. Yet. Part of what slowed down Microsoft was its preoccupation with software security compliance, data protections, and legal reviews left its technology lagging, but eventually proved to be an excellent selling point to corporations. It's been great for us to go to customers and say, hey, you like chat GPT. This is that, but better, more powerful user experience, better security, says Jeff Tepper, who runs Teams. Yet it was clear something else was holding the company back from taking bigger risks with its products. The former copilot manager describes it as a kind of post traumatic stress disorder from embarrassments that stretch as far back as the proto chatbot Clippy, the animated paper clip that became an enduring punchline. Chat GPT could perform way more actions with fewer bureaucratic limitations, and Microsoft's engineers knew it. This person says he estimates that of the hundreds of people in his organization, three quarters of them paid out of their own pocket for chat GPT because Microsoft wouldn't let them expense it. Tepper says it's common for employees to tech other company services to better understand the market, Microsoft's customers were gravitating to chat GPT two. Maderna found Open AI's options to customize chat GPT made it far more versatile than Copilot, says Brad Miller, the biotech company's former chief information officer. Microsoft was like, it's embedded in office, just use it in office, and we were like, there's so much more that we want to do. He says that fall, as Microsoft was ready and Copilot for a general release, more than eighteen thousand businesses were accessing open ai models through Azure, whereas Copilot's corporate subscribers numbered in the hundreds. Microsoft still made oodles of money from renting access to open AI's models and a revenue share agreement on chat GPT subscriptions, but generally not as much as it would have if clients paid for Copilot instead. Nadella tried to build momentum for Microsoft's AI services at the company's annual Developer conference in November twenty two, twenty three, showcasing a tool to create custom copilots similar to what businesses were already doing with chat GPT. The effort was overshadowed by News on the last day of the event that open AI's board had ousted Altman as CEO. Nadella, despite having increased his financial commitment to open Ai by ten million dollars earlier that year, only learned of Altman's firing after it transpired. Microsoft management now refers to the infamous four day period when Altman lost his job, accepted an executive role at Microsoft, and then wrested back control of open Ai from his board using an array of euphemisms that particular incident, or the November open Ai sort of situation. In truth, that particular incident irrevocably damaged the partnership between the companies. Just before, Nadella had been considering another massive cash infusion in the tens of billions of dollars to keep open Ai hurtling forward ahead of Amazon and Google, according to a person with direct knowledge of the deliberations, no more. Nadella figured Microsoft had risked enough on the immaturely run start up. Nadella had been bonding with Suliman, an influential figure in the ai world who'd co founded the lab that became Google deep Mind and was at the time overseeing a startup called Inflection AI. Inflection had a chatbot geared more for EQ than IQ, positioning the product as a sort of extremely knowledgeable friend. Inflection's chatbot never reached anywhere near the prominence of chat GPT, but Nadella loved its crisp, playful responses. It was Nadella's turn to surprise Open Ai. In March twenty twenty four, Microsoft spent six hundred and fifty million dollars to license Inflection software and hire most of its team. Microsoft informed open Ai a day before the deal was announced. Nadella put Suliman, a forty year old Britain, in charge of Microsoft's Consumer AI, a portfolio that incl It's bing the copilot app, the Edge web browser, and MSN Web portal. The slate was in weak shape, especially among young people who aren't exactly in MSN dot COM's core demographic. Generation Z is growing up with iPhones and MacBooks, Gmail and Google Docs, Instagram and TikTok and now chat GPT. The way that Saya put it to me is that we're going through this generational shift where younger people today who are going to become the CEOs and industry leaders in a decade's time and make purchasing decisions in the enterprise, will have not really experienced Microsoft. Sulliman says that's a big structural challenge. Working out of offices in London and Mountain View, California, Sulliman's employees set about creating a version of Copilot designed for life away from the office. As they'd done at Inflection. They taught this Copilot empathy, humor, and kindness. Do people want an emotional connection with the company that makes excel Suliman thinks so, or at least he believes it'll be harder to switch to a competitor if the user sees Copilot as a friend or therapist. But Suliman grappled with a bigger challenge. His work essentially consisted of modifying open Ai technology, and he says Microsoft wasn't getting access to the startup's new models as early as it was entitled to. OpenAI would roll out a model to chat GPT customers with a splash before he had enough time to integrate it into Microsoft's software. Soliman, who has a reputation for having a hot temper despite his EQ values, could be indelicate on conference calls. Tensions escalated as he pushed open Ai to share more and earlier access to its models. Brad Lightcap, OpenAI's chief operating officer, says that technology transfer is an incredibly complex operational challenge and that OpenAI is one of Microsoft's biggest Azure customers and expects to remain a large client after their current agreement expires in twenty third. Suliman says he's not surprised the fast growing startup occasionally forgets some of its obligations to Microsoft. I'm quite empathetic for them, he says. Empathy is a big theme for Suliman. For a demonstration of what he's been working on in Copilot, he taps the voice agent on his iPhone app and tells it he's in a difficult life state and struggling with anxieties. A soothing male voice with a British accent responds tough situations can fog up clear thinking and gently asks if he can share more about what's stirring up his emotions. Copilot is programmed to sense a user's comfort boundaries and carefully diagnose issues before suggesting solutions. Sulliman says, did you hear pauses? Did you hear the volume go down? He asks with pride. That's extremely hard to do. We'll be right back with Nadella's paradox. Welcome back to Nadella's Paradox. On a frigid February morning, at Microsoft's headquarters in Redmond, Washington, more than a dozen managers from Estay Lauder are receiving a crash course in Copilot. The cosmetics company has been exploring ways for Microsoft's AI to assist in a corporate restructuring intended to improve management of ingredients, product lines, and supplies. Although Estay Lauder is a longtime Copilot customer with fifteen thousand licenses, employees are only just learning about some of the ways it can be used. When a few of them express interest in a tool to automatically manage the menuche around hiring, a Microsoft sales rep informs them they've had access to that Copilot feature for a while. For Microsoft, these moments when a customer realizes Copilot can do something genuinely useful for them are essential. Roberto canavari, Esta, Latter's executive vice president who oversees its global supply chain, says he experienced his own just weeks earlier after spending an entire weekend studying and annotating a thirty page operations report. He asked Copilot on a lark for its summary. The resulting one pager was almost as good as the one he wrote. It made me feel very old and not very smart, he says jokingly. While Sulliman worked on a more empathic Copilot for consumers, his cohorts in the Enterprise Group spent the past year focused on automating routine and monotonous workflows, such as reconciling bank invoices or reviewing warehouse inventory. Microsoft found that office workers were most impressed when Copilot could do the really boring parts of their job for them. Instead of talking in the abstract about AI's world changing potential, Microsoft sales reps now pitch Copilot's ability to add a slide about budget variances to a PowerPoint. The approach is resonating. In a quarterly financial disclosure in January, Microsoft said it's on track to generate more than thirteen billion dollars a year in AI revenue. It earns money from Copilot subscriptions, usage fees from AI models running on Edge, and a revenue share with open Ai on Chat, GPT memberships, GitHub Copilots are past fifteen million users as of March, and the corporate version of Copilot now boasts hundreds of thousands of businesses as customers, with their engagement tripling over the past year. Nesley has found Copilot wildly useful for comparing contracts and for language translations for customer service teams, says Christopher Wright, CEIO at the Swiss food conglomerate. Nesley pays for more than twenty five thousand Microsoft three sixty five licenses with Copilot, and Wright estimates that ten minutes of time savings per day covers the cost of each employee's subscription. The company is also starting to pay extra to use a product called Copilot Studio to craft custom AI agents with subject matter expertise. Wright describes it as shockingly cheap for what the company gets out of it. Don't tell Microsoft that, please, He says. A lot of these AI capabilities still depend on models from Open Ai, a fact Microsoft is less keen to advertise lately. Copilot will automatically deliver results using whatever Microsoft or OpenAI model is best for a particular prompt, rather than ask customers to choose. Meanwhile, Chat GPT gives users the ability to select from a jumbled list that includes four to oh four point five, three four mini and four mini high designations that are probably more meaningful to robots than to people. Spataro, the Microsoft marketing leader, scoffs at the design choice man. The interface is getting very loaded, he says. Altman has said OpenAI is working on simplifying its offerings. Judson Aldhoff, Microsoft's sales chief, says few car buyers care what engine is under the hood, so long as the vehicle can get them from point A to B. Is it a four cylinder turbo an eight cylinder? He asks, how much the whole package costs is what they care about. If Chat GPT is a top of the line range rover, MAAI two will be like a prias, not as fast or powerful, but much more efficient for your daily commute. And Microsoft has been training even smaller and more efficient models that specialize in math or speech recognition, part of a portfolio called FI. These AI efforts could someday allow you to summarize a lengthy document or process voice commands using Copilot on a standard laptop without an Internet connection. The key to getting there, according to Suliman, can be found in a mathematical concept called the Pareto frontier. The Wikipedia article helpfully describes it as the set of all Pareto efficient solutions, based on an efficiency theory developed by a nineteenth century Italian economist. The term refers to the sweet spot between a set of desirable attributes. It's thrown around in meetings by software engineers seeking to find an equilibrium between how fast and high quality an AI model is versus what it costs to operate. Google CEO Sunder PITCHAI boasted on a recent earnings call about its own advances in the Pareto frontier so the company can deliver more AI features on Android phones, its search engine, and YouTube without breaking the bank. Apple, too, has cited Pareto in its research around building language models that can run on an iPhone. Nadella offers various reasons for Microsoft to strengthen its in house AI units after spending billions on Open AI. Partly, it's about exercising in the gym, he says, to build up the company's AI model training muscle so it's not entirely dependent on Open AI. He says he won't necessarily favor one Lab's technology over the others. Sometimes the priority will be Sulliman's cheaper models, sometimes Altman's more proficient ones. The rush to embrace deep Seek in January wasn't a one off either. Microsoft added another deep Seek model to its cloud catalog in March, renting it for half the price of an equivalent from OpenAI. Someday Copilot could run certain queries through Deepseek or another model available in Microsoft's cloud, something the consumer and enterprise teams have been testing. These moves are not so much about hedging Nadella's bet on OpenAI. It's almost like he's tossing chips every which way on the roulette table. Of course, Nadella isn't always winning on the day. In January, when deep Seek first dethroned chat GPT atop the Apple App Store in the US, Copilot wasn't anywhere in the top two hundred, according to research firm Sensor Tower. Although Deepsek's open source models are available on Azure, the startup runs its own mobile chatbot app, delivering no benefit to Microsoft. Other latecomers, such as Google's Gemini and Elon Musk's Grock, are also consistently ahead. Nadella points out that Microsoft was once less relevant in video conferencing too, everybody would say hey, Zoom, zoom zoom, whom Nadella says, we won that in the enterprise. A Zoom spokesperson declined to comment. In AI, the company's products are scattered and not in a good way, says Ethan Mollick, a professor at the Wharton School of the University of Pennsylvania who studies generative AI and wrote the book co Intelligence, Living and Working with AI. But Microsoft has an edge on every other company out there. Their advantage is they own the workplace, he says. Corporate America runs on Microsoft. Competitors have accused Microsoft of using anti competitive tactics in regard to its alliance with open AI. Donald Trump's Federal Trade Commission is pursuing a broad antitrust probe of the company that was opened by the Biden administration. Microsoft has said its cooperating with the agency. Many incentives inside Microsoft are aligned to push Copilot. On May nineteenth, the company will hold its biggest annual developer conference, where Copilot is expect to be the star of the show. Already, the company boasts that about seventy percent of Fortune five hundred companies use it, and that Blackrock, Price, Waterhouse Coopers and many others have signed massive contracts for the AI license. Arman Waworski, an engineer overseeing AI solutions for Austria's Revice and Bank International, says Microsoft's sales team is consistently pitching his staff to use Copilot. Use Copilot, use Copilot. Wiwarski says. The bank's employees do love to use Copilot in teams for meeting summaries, but way more of them over twenty five thousand rely on a customized corporate version of Chat gpt RBI built it using Microsoft's cloud tools for compliance reasons. Maybe they don't deliver everything from the very beginning, or sometimes for me they are too slow, Wworski says, but in the end Microsoft completely understands how to bring certain technologies into our environment. Copilot consumes Nadella's life outside the office's well. He likes podcasts, but instead of listening to them, he loads transcripts into the Copilot app on his iPhone so he can chat with the voice assistant about the content of an episode in the car on his commute to Redmond at the office. He relies on Copilot to deliver summaries of messages he receives in Outlook and teams and toggles. Among at least ten custom agents from co Pilot Studio, he views them as his AI chiefs of staff, delegating meeting prep, research and other tasks to the bots. I'm an email typist, Nadella jokes of his job, noting that Copilot is thankfully very good at triaging his messages. Nadella, who grew up in Hyderabad, India, the son of a Marxist economist, was seven when Bill Gates and Paul Allen started Microsoft in nineteen seventy five. He spent more than half his life at the company. During Nadella's tenure, he's seen the emergence of the world Wide Web, the creation of Internet Explorer, the dot com crash, the USC government's landmark antitrust lawsuit against Microsoft, and the disastrous losses that followed to Google in search and Apple in phones. AI could top them all. At an event in April to celebrate the company's fiftieth anniversary, Gates quipped that Copilot would eventually take over as CEO. For now, Nadella is still trying to earn his keep. He sought to rally the troops by telling them to forget about the successes of the past and be ready to abandon what's worked up to now. Says Charles Lamana, a vice president who oversees Copilot Studio and some other products. The last five years we spent building it doesn't matter. It's not worth anything anymore, he recalls, Nadella, telling them burn the ships. One ship recently set ablaze. In May, Nadella shut down Skype, which Microsoft once spent eight point five billion dollars to acquire, only to later replace it with teams. With AI promising or threatening such existential change, there remain in weighty questions about who will win and who will lose from this next technological transformation. In public, Nadella often sounds like a macroeconomist. He draws parallels between this moment in AI and the industrial Revolution and the convergence of the global North and South, and cites economic and labour theorists from David Ottor to Friedrich Hayek to Herbert Simon. In particular, the Jevins paradox, first described in an eighteen sixty five book on British coal production by William Stanley Jevins has greatly informed Nadella's understanding of Microsoft's place in the AI revolution. Jevons found that the invention of a more fuel efficient steam engine made coal a more desirable fuel source, igniting a surge in coal consumption. That's why Nadella appears unphased by the perception of deep seek undermining the value of his investment in open AI or the buildout of Microsoft data centers. More efficient AI will eventually mean a greater need for AI services and infrastructure to power them. In his view. That's not to suggest there won't be a reckoning for excessive spending on AI model training and in Nvidia's graphics processing units GPUs. Open AI is working with Oracle and soft Bank on a massive data center endeavor called the Stargate Project, promising to spend a whopping five hundred billion dollars over the next four years. Nadella chose not to be part of the financing, given the tens of billions he's already investing annually in Microsoft's own AI infrastructure. I wouldn't hold it against anyone, certainly not a CEO of a public company, to realize when there are limits of what their company can bear says Lightcap, the OpenAI COO. At the same time, Microsoft is pulling back on some of its own data center projects around the world. These moves are an acknowledgment that it doesn't make sense to keep piling money into what is effectively a real estate asset, says Lyla Tredikoff, a former deputy chief technology officer at Microsoft. On the other hand, the reckoning could be far more consequential for the broader society. Microsoft has been very careful with how it talks about AI, says a person with direct knowledge of its AI marketing who has not to be identified. Copilot is supposed to help accomplish tasks, not automate jobs. Interviews with Microsoft customers and salespeople find that the latter is already occurring. They describe contractors, outsourced programmers, interns, and summer analysts all losing work to AI. In his book The Coming Wave, published before he joined Microsoft, Suliman laid out the risks more bluntly. These tools will only temporarily augment human intelligence. They will make us smarter and more efficient for a time, and will unlock enormous amounts of economic growth, but they are fundamentally labor replacing, he wrote, Suliman advocated for regulatory oversight and other government interventions, such as new taxes on autonomous systems and a universal basic income to prevent a socioeconomic collapse. Nadella is more optimistic than his new deputy. In an interview at Microsoft headquarters in conference room number thirty four thirty five sixty three, sitting next to his human chief of staff, Nadella makes the case that his co pilot assistants won't replace her as she stoically types notes of our conversation on her tablet. He acknowledges that AI will cause hard displacement and changes in labor pools, including for Microsoft. Altoff, the sales chief, says Nadella is pressuring his team to find ways to use AI to increase revenue from its operation instead of boosting headcount. Meanwhile, Microsoft said on May thirteenth that it's cutting six thousand employees, about three percent of its workforce. But Nadella contends that AI could end up delivering more societal benefits than the Industrial Revolution did. When you create abundance, he says, then the question is what one does with that abundant to create more surplus. He argues again that it comes down to economics. If Microsoft spends hundreds of billions of dollars on AI projects, he estimates that will likely generate trillions of dollars in economic activity. More data centers means more construction, energy use, manufacturing equipment, and so forth. More AI use means more opportunities to create new kinds of businesses, especially as AI bridges the skills gap between an entry level employee, expert engineer, and executive. There may be a new startup with one hundred founders, Nadella says, or if the AI ever gets good enough, perhaps one hundred startups with just one founder. It will take time for the jobs of the AI era to arrive, he says, and then the huge economic and productivity boosts will follow, just as they did after the advent of the steam engine, electricity, the computer, and the internet. When that happens, then everything computes at a societal level, Nadallas says. Otherwise, all you have is a bunch of GPUs