Sign up to save your podcasts
Or
Share Audiobook: Chapter 3 Industrial AI Revolution
Share to email
Share to Facebook
Share to X
Share to email
Share to Facebook
Share to X
Or
Audiobook: Chapter 3 Industrial AI Revolution
CHAPTER 3: The Industrial AI Revolution
"That's exactly what I want!" My sister-in-law was thrilled to see a picture of a certain type of porch railing that she'd previously had only in her mind's eye. "Thank you so much. Now I can show the architect!"
My daughter grinned. She had just used AI to create a rendering of an architectural detail. She had translated her aunt's meandering description into a series of "prompts" typed into an AI tool which then generated a few possible images. The AI had been taught with the knowledge and practical skill of thousands of architects whose work was published in magazines, books, and on the web. My daughter was able to quickly and easily access and harness the architects' knowledge to accomplish a task without any prior experience or interest in porch architecture and design.
Before AI, knowledge transfer happened one-on-one, passed from an expert to an apprentice, one-to-many, passed from a teacher to a class of students, or one-to-more, passed from an author to readers of a book. Artificial intelligence offers a revolutionary path for the transfer of knowledge. When you teach the machine, it can transfer knowledge and skills from many experts to many, many more users of all backgrounds and skill levels across the globe.
We've been here before. Before the first Industrial Revolution, between roughly 1760 and 1860, an expert shoemaker had an exclusive on the knowledge and skill to make shoes that fit comfortably and worked well. During the Industrial Revolution, inventors transferred the shoemaker's expertise in leather-cutting and sole-stitching to automated machines that could do each task faster and just as, if not more, reliably. After the revolution, shoemakers were out of work, but more people could afford to buy shoes. We are now in the midst of the Industrial AI Revolution—and we can use the lessons of the past to give us an idea of where we are today.
Coal, Iron, and Steam
Coal, iron, the steam engine, and precision machining—that is, the use of machines to make other machines—were the interlocking factors that drove the Industrial Revolution. In 1760, a citizen of Great Britain or the United States likely worked on a farm, possibly growing wheat. Three generations later, their great grandson worked in a factory operating a bread-making machine powered by a steam engine heated by coal. The bread-making machine was made from iron parts cut on a lathe, which was itself powered by a steam engine heated by coal. People in country after country watched their economies industrialize, changing from primarily agricultural to primarily manufacturing. All except Japan, whose Shogunate government managed to keep out the modern world until 1863 when the Choshu Five snuck out of the country to learn the secrets of Western industry and power in London and returned to overthrow the Shogunate in 1868—proof that it is impossible to ignore or avoid the forward march of human progress.
An analogous change is happening today with AI: Countries with strong knowledge economies—focused on white-collar and professional services—will see their white-collar knowledge worker economies change from all human cognitive labor to a significant amount of AI cognitive labor. Or, if we demand it, primarily human-plus-AI. Artificial intelligence is already changing, and will continue to change, the work of doctors, lawyers, financial analysts, engineers, computer programmers, anyone who learns a lot about a subject and then uses that knowledge in their work. The more that person currently works alone with information on a computer, the more their work will be affected. People who work with their brains and hands will be less affected—plumbers, artists, musicians who perform live. People who empathize, relate, listen, negotiate, and communicate with other people will be less affected.
Factors of Production
Just as coal, iron, the steam engine, and precision machining drove the first Industrial Revolution, the AI Industrial Revolution is driven by its own factors: nuclear energy, data, graphical processing units, and foundation models.
I grew up on a farm in Pennsylvania near the Three Mile Island nuclear power plant, site of the worst nuclear accident in U.S. history when Unit 2 melted down in late March of 1979. Forty-five years later, in September of 2024, Microsoft ignored the negative association and announced that they had bought all of the electricity that Three Mile Island could produce for the next twenty years. Later in the fall, after Microsoft took the publicity hit, Amazon, Meta, and Google announced that they were all working on their own nuclear electricity generation projects—nuclear energy for computers that run AI.
So why do we need all this nuclear energy? In essence, it's about data, how we store it, and how AI companies process data to train AI. Artificial intelligence training happens in massive electricity-hungry cloud computing data centers sprouting up around the country. What's a cloud computing data center? Starting in the late 1990s, the new economics of the internet made it cheaper to locate your computers wherever you could get the best deal. Electricity and real estate prices became more important than the feel-good comfort of walking down the hall to look at your computer server. Enter the data center boom. A data center is kind of like a modular storage unit facility for your computer servers. Businesses rushed to move their computers from expensive floors in their headquarters to cheap leased space… wherever. Soon companies like Compaq realized they could rent computers in their own data center to other companies without having to box them up and ship them. This became the first generation of "cloud" The phenomenon was described first by young PhD candidate Ramnath Chellappa as a "computer paradigm where the boundaries of computing will be determined by economic rationale rather than technical limits alone" (Biswas, 2025). Marketing people who grew up watching Bob Ross paint happy little clouds decided the executives who signed the checks would swallow this concept if they called it "the cloud" and adopted Ross's imagery in their collateral. Not from the States, younger than me, or had cable when you were a kid? Look up Bob Ross on YouTube.
The smart-aleck phrase "There is no cloud; it's just someone else's computer" became my mantra when I had to negotiate high-stakes enterprise contracts for my employer with cloud service providers. The cloud service providers needed to be convinced that we weren't going to blindly go along with their convenient buzzwords. I put a sticker of the "There is no cloud…" phrase beneath a sad-faced cartoon cloud on my laptop for the well-intentioned cloud computing company representatives to contemplate while we negotiated terms that would protect our data and our institution.
Amazon Web Services was the first really big cloud computing service provider. Jeff Bezos hated that he had to pay for a mountain of computers to sit idle all year just so his website wouldn't crash when we all ordered books and DVDs at Christmas on Amazon.com. He told his team to figure out how to put them to work the other eleven months of the year, giving rise to even more buzzy terms like "elastic computing." This became the second generation of cloud.
The pendulum swings back as always in history, largely because AI promises to do its thing in real time. If you are a doctor relying on AI to help steer your scalpel during robotic surgery, you're going to want that AI on a computer close by so the camera data going to it and the scalpel-manipulation directions coming back don't take too long in transit. You want those systems connected on your own computer network, or at least on the edge of it. So now we have "edge computing," fancy words for somebody else's computer located down the street instead of somewhere in the "Eastern U.S. Region." A friend works for a company quietly buying up office buildings vacated during the pandemic so they can replace cubicles with computers to run real-time AI for the remaining businesses in the neighborhood where the edge of the AI computer network butts up against the edge of the customers' computer network.
Once you have a reliable supply of electricity for your data center, you need the next major factor: data. In addition to storing more than 17,000 of my emails going back to 2004, Google collects email data from more than 1.8 billion people worldwide, more than 130 million in the U.S. alone, where they are legally allowed to use all of it to train AI. Assuming I'm a typical user, in the U.S. that's 2,210,000,000,000 emails controlled by Google alone. And it doesn't stop with emails, or with Google. The data of the rest of our lives are captured variously by companies like Amazon (retail, publishing, pharmacy), Visa (retail financials), Apple (photos, videos, retail financials), Meta (photos, videos, communications, social activity), JP Morgan Chase (finance, banking), and Epic Systems Corporation (health), along with a host of other quiet but enormous data aggregation companies with benign names like MX, Mobius, and Plaid. All of this data is raw material being used to train AI. With electricity and data secured, it's time to process it.
"We got approval for you to use our graphical processing units!" Bittersweet news from thrilled sales reps of more than one of the cloud computing service providers my employer contracted with. Sweet because we were able to convince mega corporations to allow us to do important but financially dead-end pediatric AI research. Bitter because we'd "won" the chance to pay mid- to high-six figures for the privilege of renting their graphical processing units (GPU) for a few weeks. Graphical processing units are the computers used to teach AI. The "graphical" part of the name confusingly doesn't matter for AI; it's left over from their first major use in video games. At the same time we got the bittersweet "good news," the CEO of the leading maker of GPU engines that power AI used quarterly financials to proclaim, "The next industrial revolution has begun. Companies and countries are…using GPUs to…build AI factories to produce a new commodity: artificial intelligence" (NVIDIA, 2024a). He drove his analogy home by adding, "The age of AI is in full steam.…" The GPU is the new steam engine.
"Wow, it's like a third-year medical student." Early research showed this kind of sci-fi potential when we used large language models made available by big AI companies as a starting point—the foundation. Google, Amazon, Meta, and Microsoft all created massively powerful foundational AI—known as "foundation models"—based on the "Attention Is All You Need" paper, discussed in the Introduction to this book. These companies made them available to the marketplace in various ways we'll get into later. Researchers at my previous employer used foundation models to make new machines—previously unattainable AI that, by learning from medical data, could perform common clinical tasks like identifying a disease based on a description of symptoms. Use AI to build new AI. The foundation model is the new machine that makes machines.
The "Attention Is All You Need" authors who invented the Transformer kicked off a chain of events leading to OpenAI's GPT, the first foundation model AI to have a big impact. Nuclear energy, vast reserves of data, and GPUs all existed long before 2017, but foundation model AI—the machines to make machines, or in this case, other AI—did not. A foundation model is characterized not by what it can do directly, but by its potential to make other AI. Or rather, its potential to be "fine-tuned," or further taught to do something useful (see chapter 1). Before the invention of the Transformer described in the "Attention Is All You Need" paper, it didn't matter how many watts of energy, bytes of data, and GPUs you had, it was still prohibitively expensive in time to build a foundation model. Really powerful AI took exponentially more time to train with the tools available. Exponential growth means one plus one equals three. When you doubled the amount of data you used to teach AI, you tripled (or more) the amount of time it took to learn. Double your training data over and over so the AI "knows" enough? You're looking at decades or longer.
Ten years ago, at my former employer, we would stare wistfully at large volumes of data, money in the bank to pay the electricity bill, a healthy collection of GPUs, and the ability to buy more. And do nothing. With the tools we had at the time, we'd have to wait years to find out if the machine could learn anything useful. The same was true everywhere. Google, Microsoft, Amazon, Meta, and even the entire country of China had mind-boggling resources of data and dollars to invest. They were already buying GPUs and were ready to buy more. But the one thing they couldn't buy was time.
The Transformer broke the time barrier. The magic of the self-attention mechanism at the heart of the Transformer isn't that it's more accurate (it is). What made it revolutionary was that it didn't require exponentially more time. You could add more data, or teach the machine more lessons, and as long as you also added more GPUs, it would take the same amount of time. This meant experiments could happen in days instead of years. Sitting on the email data of a billion people with billions of dollars burning a hole in your pocket? Call up the GPU salesperson, restart a deactivated nuclear reactor, and with the Transformer, you could train really, really powerful AI in months instead of decades.
The Transformer-based AI was trained to translate between English and German using a well-known standard data set containing four and a half million English-German sentence pairs. A sentence pair is something like "my dog has fleas" matched to "mein Hund hat Flöhe." The Attention team figured out that they could teach their AI three hundred percent more lessons and it would only take fourteen percent longer without adding any more GPUs! And if they wanted an even more expert translation AI and used a data set of nine, eighteen, or thirty-six million sentence pairs but kept adding GPUs, it would take the same amount of time.
This meant the corporations and countries who had the data and could buy the GPUs and find enough electricity to run them could teach an entirely new class of AI: foundation model AI capable of making other AI. And that's what happened.
In 2017, Jen-Hsun (Jensen) Huang, the CEO of GPU-maker NVIDIA Corporation, reported annual revenue of $6.9 billion and stated, "We can now see that GPU-based deep learning will revolutionize major industries.…The era of AI is upon us." By 2023, all the big AI companies had trained foundational large language models. In fiscal year 2023, Mr. Huang reported revenue of $26.97 billion and said, "AI is at an inflection point, setting up for broad adoption reaching into every industry" (Choe & Parvini, 2023; NVIDIA, 2023).
NVIDIA, the company that makes GPUs that power AI, saw its revenue grow four hundred percent in the five years that included a global pandemic (Global Macro Monitor, 2024). Microsoft, Google, Amazon, and Meta all bought as many GPUs as possible in a race to build the machines that make machines for the coming industrialization of AI.
NVIDIA announced "partnerships" with the AI companies in this time period. Their biggest customers became something more. NVIDIA also tried and failed to buy their competitor chip designer Arm Holdings for $40 billion. The U.S. Federal Trade Commission squashed the deal with a lawsuit. Not to be deterred, NVIDIA went on the open market and quietly acquired hundreds of millions of dollars of Arm stock and announced new partnerships with Arm (Turner & Bloomberg, 2024).
Just as coal, iron, steam engines, and precision machine tools combined to industrialize human production work, nuclear energy, data, graphical processing units, and foundation models are combining to rapidly industrialize human knowledge work. Rapidly because instead of a century-long evolution driven by small entrepreneurs and inventors, AI industrialization is being driven by some of the wealthiest corporations on the planet.
Big AI
"Big Tech" companies are likely secretly happy with this moniker and tolerate its sometimes negative associations. They love it when a politician or an activist rails against "Big Tech" for controlling internet searches, computer operating systems, and social media. That means we're looking the other way while they quietly go about the business of harvesting our personal data and building up the capital to turn our data into profitable AI. The truth is that Big Tech has actually become Big AI.
The first Industrial Revolution played out over a hundred years, from 1760 to 1860. Change was driven by inventors, entrepreneurs, small, and then mid-sized, businesses in every community. The inventive shoemaker dreamed up a mechanized leather cutting machine. He partnered with the enterprising blacksmith next door to buy an early lathe to make his machine. They went in together on a shop next to the flour mill so they could run both their machines off the steam traction engine the innovative miller had bought to replace his water wheel. The shops sat next to a river full of barges shipping coal from independent mines, steel from small foundries. Now able to produce a larger volume of quality shoes, the shoemaker hired a farmer to run the leather-cutting machine in the afternoon after he had tended the fields. He sold shoes at a lower price, including a pair to the previously barefoot daughter of the farmer. The blacksmith trained the first machinist in the state to make leather-cutting machines which the partners then sold to shoemakers in nearby towns. There weren't any big companies, and the truly private corporation wasn't legal until politics and business conspired in the 1890s. Our cooperative of entrepreneurial small businesses could thrive. The farmer paid off a loan and hired a hand. His daughter went to school in the spring instead of planting corn. Life got better as the whole pie grew and more citizens got a slice. But there were limits to what could happen at this scale.
After this revolutionary period, large corporations began to emerge, at least in the U.S. Innovative, driven businessmen created new corporate and financial strategies with the goal of maximizing profit, often by beating the competition with a better product at a better price. Conglomerates in shoemaking equipment (United Shoe Machinery), oil (Rockefeller), steel (Carnegie), railroads (Stanford's Central Pacific), electricity (J.P. Morgan's General Electric), and tobacco (Duke's American Tobacco) rolled up their respective sectors. At their best, these conglomerates created economies of scale that allowed them to invest in innovation, efficiency, and the workforce, ultimately improving quality of life for more citizens. At their worst, the conglomerates put competitors out of business and did not invest in innovation, efficiency, and the workforce, ultimately leading to real and perceived economic injustice. Fifty years after the entrepreneurial frenzy of the Industrial Revolution, in some sectors, a very few businesses controlled all or most factors of production and sales in their industry. John D. Rockefeller's Standard Oil owned the oil fields, drilling companies, refineries, and effectively the railroads (through exclusive freight contracts), pipelines, heating oil distributors and gasoline service stations. Standard Oil's relentless pursuit of efficiency drove down the cost of oil nearly tenfold by the early 1900s. The first industrialists were also politically active, as they sought to influence regulatory and tax policy to their benefit, much as today's corporations hire lobbyists to advance their interests with whatever party is in power.
The political influence of Rockefeller and his contemporaries is best illustrated by the rapid evolution of corporate law. In the 1880s, Leland Stanford obtained "personhood" for his Southern Pacific Railroad, which established the precedent that a corporation has the same protection against discrimination as a human under the Fourteenth Amendment. In Stanford's case, the discrimination was a special tax on railroad property enacted by the State of California (Ballotpedia, n.d.). Not what the Fourteenth Amendment was meant to protect when it was enacted after the Civil War to protect the rights of freed slaves. Personhood gave corporations special protections, but they were still subject to strict controls on their power by the state legislatures that chartered them. J.P. Morgan created a permissive home for corporations by hiring a lawyer to rewrite New Jersey's corporate law to remove restrictions meant to prevent abuse of power. The changes were happily adopted by the governor and legislature in return for franchise taxes. Indeed, after Ohio broke up Rockefeller's Standard Oil Trust under state law, the oil conglomerate moved to New Jersey to take advantage of the state's enabling corporate laws. In the 1890s, public (and competitor) backlash against industrialists and their political influence, as well as real and perceived anti-competitive practices, brought new federal antitrust laws and ultimately a period of enforcement that resulted in the breakup of many monopolies, including Standard Oil in 1911 (Oller, 2019). But the industrialists hadn't gained power until the end of the first Industrial Revolution, after decades of improvement in standard of living, life expectancy, access to education, and other measures of life, liberty, and the pursuit of happiness.
Artificial intelligence industrialization went straight to conglomerate. Arguably, Big AI didn't even experience the speed bump of an extended or widespread entrepreneurial phase, given that it already had twenty years of momentum. Big AI rapidly invested enormous amounts of capital to consolidate control over facets of AI production, and wielded political influence to increase profits. By early 2025, Microsoft controlled the data of 400 million Microsoft 365 users (iron); had acquired a nuclear energy plant (coal); was in multiple financial relationships with a key GPU supplier, NVIDIA (steam engines); and had a $13 billion stake in OpenAI's GPT foundation models (machines to make machines). Microsoft, NVIDIA, BlackRock (the world's largest investment company), and MGX (the United Arab Emirates' sovereign wealth AI fund) are in a partnership to invest $80 to $100 billion in new data centers exclusively for AI controlled by Microsoft (BlackRock, 2025). Meanwhile, Google controls the data of 1.5 billion Gmail users (not to mention the data from YouTube and Google Photos), has built its own version of the GPU, is developing multiple nuclear energy projects and has its own foundation models plus a $2 billion stake in the foundation models of another wealthy AI company called Anthropic. Amazon controls data from more than 500 million Alexa devices plus the data of more than 350 million retail customers and the connected data of more than 200 million Prime users, has its own foundation models plus an $8 billion stake in Anthropic's models, buys as many of NVIDIA's GPUs as it's allowed to, while making its own GPUs, and is invested in multiple nuclear energy projects. Meta controls the data of the 3.9 billion people who use its apps every month, is in a partnership with AMD (an NVIDIA competitor) to design and build GPUs to install in its new $10 billion AI data center in Louisiana, and has its own foundation models.
As far as political influence goes, the inaugural U.S. National AI Advisory Committee (NAIAC), established by a law passed in 2020 and first convened in 2022, is made up of a who's who of delegates from NVIDIA, Google, Microsoft, Amazon, Anthropic, and others, along with academics and policymakers (U.S. Department of Commerce, 2022). Big AI is giving the advice. The CEOs of the Big AI companies exert more influence when they meet with presidents and presidents-elect of both parties in private, because AI is positioned as a national security issue. Mr. Huang from NVIDIA stated in 2024 that "Countries have awakened to the importance of developing their national AI and infrastructure." His use of the plural "countries" and the word "national" is in keeping with a sometimes arms-race response from elected officials around the world. Which certainly won't hurt business.
Industrial and corporate history is context for the emerging AI industry. AI will change our lives and the lives of our children and grandchildren just as the industrialization of manufacturing changed the lives of our great-great-grandparents. The AI produced by corporations is already changing the nature and value of human cognitive labor. Governments around the world are wrestling with novel policy questions, often with AI corporations in the room. Intertwined political and corporate interests are a nonpartisan phenomenon. Left, right, red, blue, or purple, it is our job as citizens, voters, and consumers to make informed choices that promote benefit and minimize harm of AI. My use of the term Big AI does not imply a negative, just as the use of the term Big Three to describe the major U.S. automakers is not a negative. Modern AI requires a scale of production that only very large corporations or governments can support (for now). Big is not necessarily bad for the individual citizen or society as a whole. Big is less likely to become bad when you have a sense of agency and are an active participant in the change brought by AI. You can be a driver of change. You can have an active role in this new, latest industrial revolution. And that role starts with your data. You teach the machines.
Follow The Data
The most important technology in human history since the printing press relies entirely on your data. When AI improves enough from learning from you and everyone around you, you will be more likely to pay for what it can do. Data is an asset that increases in value as it grows in size, and the big AI companies have been working toward this moment for decades. Ever hear the adage, "If it's free, there's a catch"? For years, I assumed that Gmail, Facebook, and Amazon's one-day delivery were free because "Big Tech" wanted to target me with ads. It was true. But only partially. What they really wanted was my data. Yes, with my data, they could learn of my interest in Iceland and sell a premium targeted ad to Icelandair. But the real money was in banking my data and the data of hundreds of millions of other people, because they knew value depended on volume when it came to training AI. Between 2010 and 2015, the data representing your digital life became a corporate asset with future value to train AI. It didn't cost much to store in the cloud, so collecting ever and ever larger volumes was a safe, long-term bet that, at a certain point, it would be feasible to use your data to industrialize AI. You teach the machines.
As discussed previously, when you read and agree to the Microsoft Service Agreement and Privacy Statement, it's clear that by clicking "I accept," you are allowing Microsoft to use your data to do anything permissible by law. An excerpt follows, accessed on April 22, 2025.
b. To the extent necessary to provide the Services to you and others, to protect you and the Services, and to improve Microsoft products and services, you grant to Microsoft a worldwide and royalty-free intellectual property license to use Your Content, for example, to make copies of, retain, transmit, reformat, display, and distribute via communication tools Your Content on the Services. If you publish Your Content in areas of the Service where it is available broadly online without restrictions, Your Content may appear in demonstrations or materials that promote the Service. Some of the Services are supported by advertising. Controls for how Microsoft personalizes advertising are available at https://choice.live.com (https://go.microsoft.com/fwlink/?LinkId=286759). We do not use what you say in email, chat, video calls or voice mail, or your documents, photos or other personal files, to target advertising to you. Our advertising policies are covered in detail in the Privacy Statement.
As you can see, per the Service Agreement, you "grant to Microsoft a worldwide and royalty-free intellectual property license to use Your Content." And if you follow the links to the Microsoft Privacy Statement, it's clear they have a royalty-free license to share your data with third parties like, say… OpenAI. That means the sentences and paragraphs I type into Microsoft 365 Word today could find their way into the dataset used tomorrow to train the next generation of ChatGPT. They are forthright enough to come right out and say that "As part of our efforts to improve and develop our products, we may use your data to develop and train our AI models."
To be clear, this degree of transparency means that Microsoft is behaving well under the U.S. laws to which I'm subject. And it doesn't just relate to content you type in Word. The "Services" referenced in this agreement are, at the time I'm writing this, an ever-expanding list of one hundred and thirty-eight Microsoft products from Excel (yes, your retirement spreadsheet) to Teams to your XBox.
OpenAI's GPT-3, the AI that blew all our minds, was trained on a large volume of unrestricted, publicly available data. Microsoft wrote OpenAI a check for $13 billion in 2020 possibly because they knew how much more "proprietary" data Microsoft had and were going to continue to acquire, and Microsoft was certain that they could return hundreds of billions of dollars more by combining your data with OpenAI's technology.
But it isn't just Big AI companies. I got an email from my bank, Wells Fargo, telling me that by continuing to use their banking services, I agreed to changes in their Online Access Agreement. The summary of changes includes a paragraph on their use of my information, accessed on April 22, 2025.
Updated Section 17(b) (Privacy and Use of Information – Acknowledgements and Agreements) to (i) clarify that your communications with us may be analyzed and processed (potentially through automated means), and may be shared with our service providers and other third parties, in accordance with our privacy policy and applicable law...
The "potentially through automated means" indicates AI. The "shared with our service providers and other third parties" means practically whoever they want, and I'll never know. Within the law, of course.
A quick word on privacy: We in the U.S. have a general sense that we have a good faith expectation, if not a right, to privacy, that our data will be kept private. That's why it's news (and a hit to share prices) when a company has a data breach. And if data is shared, we expect that our name and identifying details will be removed, making it anonymous or deidentified. There is no explicit protection of this expectation or right in the Constitution, even though the Supreme Court has inferred it from many of the ten amendments that make up the Bill of Rights. But U.S. laws and court decisions have trouble keeping up. Because there is no right protected by law, we choose to give our data up under contractual terms of use that almost nobody reads. And we are also surprised when we find out that a company is doing something… creepy with our data. Without going into the details, please accept for now that, for example, if a company has every email you've ever written, it doesn't matter if your name has been stripped from the email or not. The combination of unique information in those emails is easily linked to your identity. Claims that your data are stored without your identifying information, while technically true, are now irrelevant to practical privacy. When you see them, recognize that, for all practical purposes, they remain as a tonic to help you feel more comfortable with the idea that the book you're writing is being used wholesale to train AI, share with third parties, and make money under a royalty-free license. Microsoft actually does a pretty good job of not claiming to anonymize your data in the terms I've seen. At least they're more transparent!
It may seem like I'm picking on Microsoft, but it's all of the Big AI companies. And for now, everything these companies do with your data is completely legal. Is it right? Wrong? Creepy? That's for you to decide. But you know what else was thoroughly wrong and creepy, yet legal at the time it was done? Harvesting cells from a woman named Henrietta Lacks without her knowledge and then making billions of dollars from them.
The Story of Henrietta Lacks
On October 4,1951, at Johns Hopkins Hospital in Baltimore, a thirty-one-year-old woman named Henrietta Lacks died of complications from tumors growing all over her body. In her 2006 book The Immortal Life of Henrietta Lacks, author Rebecca Skloot writes about the injustice and institutional racism that led to a death that might have been avoided or at least made less painful. She also writes about how a researcher took cancer cells from Mrs. Lacks, without her knowledge or consent, that proved to be the first "immortal" cells to grow reliably in a test tube. The researcher and Johns Hopkins gave these cells away for free to scientists all over the world who used them to study cancer and many other diseases. Eventually, third parties like pharmaceutical and biotechnology companies acquired the cells and used them to make drugs and other biological products.
The book is about much more than what happened with these cells, and I encourage you to read it or watch the movie based on Mrs. Lacks's story. I also encourage you to consider possible parallels between Mrs. Lacks's cells and your own personal data. In her afterword, Ms. Skloot writes:
Beyond simply knowing their tissues are being used in research, some tissue-rights activists believe donors should have the right to say, for example, that they don't want their tissues used for research on nuclear weapons, abortion, racial differences, intelligence, or anything else that might run contrary to their beliefs. They also believe it's important for donors to be able to control who has access to their tissues, because they worry that information gathered from tissue samples might be used against them.
If we rewrite this paragraph, swapping the word tissue for data, we get:
Beyond simply knowing their data are being used in research, some data-rights activists believe donors should have the right to say, for example, that they don't want their data used for research on nuclear weapons, abortion, racial differences, intelligence, or anything else that might run contrary to their beliefs. They also believe it's important for donors to be able to control who has access to their data, because they worry that information gathered from data samples might be used against them.
This certainly seems reasonable to me. But it's wishful thinking at this point in the U.S. Even with sensitive data like your health records and financial information, healthcare providers are allowed to do whatever they want with your data so long as something called an Institutional Review Board determines that it would not be feasible to ask you for your permission and weighs in on the risks and benefits. Next time you go to the doctor, pay attention when you're asked to sign a digital pen pad acknowledging that you've read and understand the Notice of Privacy Practices. From a legal perspective, that signature means you have provided consent for your data to be used in any way allowable on the Notice and by law.
Does the government have anything to say about this? In short, in the U.S., as of May 2025: No. Maybe with some exceptions in California. You have to look elsewhere for another model. The European Union passed legislation called the General Data Protection Regulation, which requires more transparency and control of data for citizens of the EU. As a result, Amazon's terms governing your data (including the use of Alexa in your home) are 3,600 words in the United States and 8,700 in Germany.
Does this all seem like… a lot? It is. I work in the AI sector, and in some ways, I knew all of this before I sat down to write this chapter. But seeing it all laid out on the page in black and white is sobering. We all need to be savvy and protect ourselves.
Exercises: Try It Out
- If you have a Google account
- Log in and search "good present for my [best friend? sister? brother? spouse/partner?]
- Log out and do the same search. Do you see anything different?
- Look up and compare a data policy in the EU to a data policy in the U.S. for the same company.
- Next time you see your doctor, ask for a copy of the notice of privacy practices that you "acknowledge you've received" when you sign the little digital signature pad or complete the online check-in.
- Search a major national news website for "new nuclear reactor" or "proposed nuclear reactor." Look to see who the investors and backers are.
- Read NVIDIA's last couple of earnings reports, especially commentary from company leadership. Dig and find a transcript of the earnings call and see what questions analysts ask and how NVIDIA responds.
References
Ballotpedia, n.d. Santa Clara County v. Southern Pacific Railroad Company. (Retrieved on April 22, 2025, from Santa Clara County v. Southern Pacific Railroad Company - Ballotpedia)
Biswas, Sourya, 2025. History of Cloud Computing: A Journey of Innovation and Future Prospects. Cloud Tweaks. (Retrieved on May 8, 2025, from History Of Cloud Computing: A Journey Of Innovation And Future Prospects)
BlackRock, 2025. Press Releases: BlackRock, Global Infrastructure Partners, Microsoft, and MGX Welcome NVIDIA and xAI to the AI Infrastructure Partnership to Drive Investment in Data Centers and Enabling Infrastructure. (Retrieved on April 22, 2025, from AI Infrastructure Partnership | BlackRock)
Choe, Stan, & Sarah Parvini, 2023 (Updated January 28, 2025). Nvidia Faces a Reckoning as Chinese Upstart Raises Questions About Wall Street's Darling. AP. (Retrieved on April 22, 2025, from Nvidia faces a reckoning as Chinese upstart raises questions about Wall Street's darling | AP News)
Global Macro Monitor, 2024. NVIDIA's Stunning 400% Growth in Data Center Revenue. Macromon. (Retrieved on April 22, 2025, from NVIDIA's Stunning 400% Growth In Data Center Revenue | Global Macro Monitor)
Hindle, Brooke, & Steven Lubar, 1986. Engines of Change: The American Industrial Revolution, 1790‒1860. Smithsonian Institution Press.
Knipprath, Joerg, 2017. Justice Stephen J. Field (1816‒1899). Constituting America. (Retrieved on April 22, 2025, from Justice Stephen J. Field (1816-1899) – Constituting America)
McCarthy, Eugene, 2023. Corporate Law, Business Schools, and White-Collar Crime. St. Louis University Law Journal, 67. (Retrieved on April 22, 2025, from https://scholarship.law.slu.edu/cgi/viewcontent.cgi?article=2330&context=lj)
Microsoft, 2024. Microsoft Services Agreement. (Retrieved on April 22, 2025, from Microsoft Services Agreement)
NVIDIA Corporation, 2017. Newsroom: NVIDIA Announces Financial Results for Fourth Quarter and Fiscal 2017. (Retrieved on April 22, 2025, from NVIDIA Announces Financial Results for Fourth Quarter and Fiscal 2017 | NVIDIA Newsroom)
NVIDIA Corporation, 2023. Press Release Details: NVIDIA Announces Financial Results for Fourth Quarter and Fiscal 2023. (Retrieved on April 22, 2025, from NVIDIA Corporation - NVIDIA Announces Financial Results for Fourth Quarter and Fiscal 2023)
NVIDIA Corporation, 2024a. Computer Industry Joins NVIDIA to Build AI Factories and Data Centers for the Next Industrial Revolution. (Retrieved on May 8, 2025, from Computer Industry Joins NVIDIA to Build AI Factories and Data Centers for the Next Industrial Revolution | NVIDIA Newsroom)
NVIDIA Corporation, 2024b. NVIDIA Announces Financial Results for Third Quarter Fiscal 2025. (Retrieved on April 22, 2025, from NVIDIA Announces Financial Results for Third Quarter Fiscal)
Oller, John, 2019. White Shoe: How a New Breed of Wall Street Lawyers Changed Big Business and the American Century. Dutton.
Skloot, Rebecca, 2011. The Immortal Life of Henrietta Lacks. Crown.
Turner, Nick, & Bloomberg, 2024. Nvidia Tried and Failed to Buy Arm for $40 billion in 2020, but It Just Reported a Stake Worth $147.3 Million. Fortune. (Retrieved on April 22, 2025, from Nvidia discloses $147.3 million stake in Arm | Fortune)
U.S. Department of Commerce, 2022. U.S. Department of Commerce Appoints 27 Members to National AI Advisory Committee. (Retrieved on April 22, 2025, from U.S. Department of Commerce Appoints 27 Members to National AI Advisory Committee | U.S. Department of Commerce)
Wells Fargo, n.d. We Are Updating Our Online Access Agreement. (Retrieved on April 22, 2025, from We Are Updating Our Online Access Agreement)
View all episodes
By Jeff Pennington and MJ PenningtonCHAPTER 3: The Industrial AI Revolution
"That's exactly what I want!" My sister-in-law was thrilled to see a picture of a certain type of porch railing that she'd previously had only in her mind's eye. "Thank you so much. Now I can show the architect!"
My daughter grinned. She had just used AI to create a rendering of an architectural detail. She had translated her aunt's meandering description into a series of "prompts" typed into an AI tool which then generated a few possible images. The AI had been taught with the knowledge and practical skill of thousands of architects whose work was published in magazines, books, and on the web. My daughter was able to quickly and easily access and harness the architects' knowledge to accomplish a task without any prior experience or interest in porch architecture and design.
Before AI, knowledge transfer happened one-on-one, passed from an expert to an apprentice, one-to-many, passed from a teacher to a class of students, or one-to-more, passed from an author to readers of a book. Artificial intelligence offers a revolutionary path for the transfer of knowledge. When you teach the machine, it can transfer knowledge and skills from many experts to many, many more users of all backgrounds and skill levels across the globe.
We've been here before. Before the first Industrial Revolution, between roughly 1760 and 1860, an expert shoemaker had an exclusive on the knowledge and skill to make shoes that fit comfortably and worked well. During the Industrial Revolution, inventors transferred the shoemaker's expertise in leather-cutting and sole-stitching to automated machines that could do each task faster and just as, if not more, reliably. After the revolution, shoemakers were out of work, but more people could afford to buy shoes. We are now in the midst of the Industrial AI Revolution—and we can use the lessons of the past to give us an idea of where we are today.
Coal, Iron, and Steam
Coal, iron, the steam engine, and precision machining—that is, the use of machines to make other machines—were the interlocking factors that drove the Industrial Revolution. In 1760, a citizen of Great Britain or the United States likely worked on a farm, possibly growing wheat. Three generations later, their great grandson worked in a factory operating a bread-making machine powered by a steam engine heated by coal. The bread-making machine was made from iron parts cut on a lathe, which was itself powered by a steam engine heated by coal. People in country after country watched their economies industrialize, changing from primarily agricultural to primarily manufacturing. All except Japan, whose Shogunate government managed to keep out the modern world until 1863 when the Choshu Five snuck out of the country to learn the secrets of Western industry and power in London and returned to overthrow the Shogunate in 1868—proof that it is impossible to ignore or avoid the forward march of human progress.
An analogous change is happening today with AI: Countries with strong knowledge economies—focused on white-collar and professional services—will see their white-collar knowledge worker economies change from all human cognitive labor to a significant amount of AI cognitive labor. Or, if we demand it, primarily human-plus-AI. Artificial intelligence is already changing, and will continue to change, the work of doctors, lawyers, financial analysts, engineers, computer programmers, anyone who learns a lot about a subject and then uses that knowledge in their work. The more that person currently works alone with information on a computer, the more their work will be affected. People who work with their brains and hands will be less affected—plumbers, artists, musicians who perform live. People who empathize, relate, listen, negotiate, and communicate with other people will be less affected.
Factors of Production
Just as coal, iron, the steam engine, and precision machining drove the first Industrial Revolution, the AI Industrial Revolution is driven by its own factors: nuclear energy, data, graphical processing units, and foundation models.
I grew up on a farm in Pennsylvania near the Three Mile Island nuclear power plant, site of the worst nuclear accident in U.S. history when Unit 2 melted down in late March of 1979. Forty-five years later, in September of 2024, Microsoft ignored the negative association and announced that they had bought all of the electricity that Three Mile Island could produce for the next twenty years. Later in the fall, after Microsoft took the publicity hit, Amazon, Meta, and Google announced that they were all working on their own nuclear electricity generation projects—nuclear energy for computers that run AI.
So why do we need all this nuclear energy? In essence, it's about data, how we store it, and how AI companies process data to train AI. Artificial intelligence training happens in massive electricity-hungry cloud computing data centers sprouting up around the country. What's a cloud computing data center? Starting in the late 1990s, the new economics of the internet made it cheaper to locate your computers wherever you could get the best deal. Electricity and real estate prices became more important than the feel-good comfort of walking down the hall to look at your computer server. Enter the data center boom. A data center is kind of like a modular storage unit facility for your computer servers. Businesses rushed to move their computers from expensive floors in their headquarters to cheap leased space… wherever. Soon companies like Compaq realized they could rent computers in their own data center to other companies without having to box them up and ship them. This became the first generation of "cloud" The phenomenon was described first by young PhD candidate Ramnath Chellappa as a "computer paradigm where the boundaries of computing will be determined by economic rationale rather than technical limits alone" (Biswas, 2025). Marketing people who grew up watching Bob Ross paint happy little clouds decided the executives who signed the checks would swallow this concept if they called it "the cloud" and adopted Ross's imagery in their collateral. Not from the States, younger than me, or had cable when you were a kid? Look up Bob Ross on YouTube.
The smart-aleck phrase "There is no cloud; it's just someone else's computer" became my mantra when I had to negotiate high-stakes enterprise contracts for my employer with cloud service providers. The cloud service providers needed to be convinced that we weren't going to blindly go along with their convenient buzzwords. I put a sticker of the "There is no cloud…" phrase beneath a sad-faced cartoon cloud on my laptop for the well-intentioned cloud computing company representatives to contemplate while we negotiated terms that would protect our data and our institution.
Amazon Web Services was the first really big cloud computing service provider. Jeff Bezos hated that he had to pay for a mountain of computers to sit idle all year just so his website wouldn't crash when we all ordered books and DVDs at Christmas on Amazon.com. He told his team to figure out how to put them to work the other eleven months of the year, giving rise to even more buzzy terms like "elastic computing." This became the second generation of cloud.
The pendulum swings back as always in history, largely because AI promises to do its thing in real time. If you are a doctor relying on AI to help steer your scalpel during robotic surgery, you're going to want that AI on a computer close by so the camera data going to it and the scalpel-manipulation directions coming back don't take too long in transit. You want those systems connected on your own computer network, or at least on the edge of it. So now we have "edge computing," fancy words for somebody else's computer located down the street instead of somewhere in the "Eastern U.S. Region." A friend works for a company quietly buying up office buildings vacated during the pandemic so they can replace cubicles with computers to run real-time AI for the remaining businesses in the neighborhood where the edge of the AI computer network butts up against the edge of the customers' computer network.
Once you have a reliable supply of electricity for your data center, you need the next major factor: data. In addition to storing more than 17,000 of my emails going back to 2004, Google collects email data from more than 1.8 billion people worldwide, more than 130 million in the U.S. alone, where they are legally allowed to use all of it to train AI. Assuming I'm a typical user, in the U.S. that's 2,210,000,000,000 emails controlled by Google alone. And it doesn't stop with emails, or with Google. The data of the rest of our lives are captured variously by companies like Amazon (retail, publishing, pharmacy), Visa (retail financials), Apple (photos, videos, retail financials), Meta (photos, videos, communications, social activity), JP Morgan Chase (finance, banking), and Epic Systems Corporation (health), along with a host of other quiet but enormous data aggregation companies with benign names like MX, Mobius, and Plaid. All of this data is raw material being used to train AI. With electricity and data secured, it's time to process it.
"We got approval for you to use our graphical processing units!" Bittersweet news from thrilled sales reps of more than one of the cloud computing service providers my employer contracted with. Sweet because we were able to convince mega corporations to allow us to do important but financially dead-end pediatric AI research. Bitter because we'd "won" the chance to pay mid- to high-six figures for the privilege of renting their graphical processing units (GPU) for a few weeks. Graphical processing units are the computers used to teach AI. The "graphical" part of the name confusingly doesn't matter for AI; it's left over from their first major use in video games. At the same time we got the bittersweet "good news," the CEO of the leading maker of GPU engines that power AI used quarterly financials to proclaim, "The next industrial revolution has begun. Companies and countries are…using GPUs to…build AI factories to produce a new commodity: artificial intelligence" (NVIDIA, 2024a). He drove his analogy home by adding, "The age of AI is in full steam.…" The GPU is the new steam engine.
"Wow, it's like a third-year medical student." Early research showed this kind of sci-fi potential when we used large language models made available by big AI companies as a starting point—the foundation. Google, Amazon, Meta, and Microsoft all created massively powerful foundational AI—known as "foundation models"—based on the "Attention Is All You Need" paper, discussed in the Introduction to this book. These companies made them available to the marketplace in various ways we'll get into later. Researchers at my previous employer used foundation models to make new machines—previously unattainable AI that, by learning from medical data, could perform common clinical tasks like identifying a disease based on a description of symptoms. Use AI to build new AI. The foundation model is the new machine that makes machines.
The "Attention Is All You Need" authors who invented the Transformer kicked off a chain of events leading to OpenAI's GPT, the first foundation model AI to have a big impact. Nuclear energy, vast reserves of data, and GPUs all existed long before 2017, but foundation model AI—the machines to make machines, or in this case, other AI—did not. A foundation model is characterized not by what it can do directly, but by its potential to make other AI. Or rather, its potential to be "fine-tuned," or further taught to do something useful (see chapter 1). Before the invention of the Transformer described in the "Attention Is All You Need" paper, it didn't matter how many watts of energy, bytes of data, and GPUs you had, it was still prohibitively expensive in time to build a foundation model. Really powerful AI took exponentially more time to train with the tools available. Exponential growth means one plus one equals three. When you doubled the amount of data you used to teach AI, you tripled (or more) the amount of time it took to learn. Double your training data over and over so the AI "knows" enough? You're looking at decades or longer.
Ten years ago, at my former employer, we would stare wistfully at large volumes of data, money in the bank to pay the electricity bill, a healthy collection of GPUs, and the ability to buy more. And do nothing. With the tools we had at the time, we'd have to wait years to find out if the machine could learn anything useful. The same was true everywhere. Google, Microsoft, Amazon, Meta, and even the entire country of China had mind-boggling resources of data and dollars to invest. They were already buying GPUs and were ready to buy more. But the one thing they couldn't buy was time.
The Transformer broke the time barrier. The magic of the self-attention mechanism at the heart of the Transformer isn't that it's more accurate (it is). What made it revolutionary was that it didn't require exponentially more time. You could add more data, or teach the machine more lessons, and as long as you also added more GPUs, it would take the same amount of time. This meant experiments could happen in days instead of years. Sitting on the email data of a billion people with billions of dollars burning a hole in your pocket? Call up the GPU salesperson, restart a deactivated nuclear reactor, and with the Transformer, you could train really, really powerful AI in months instead of decades.
The Transformer-based AI was trained to translate between English and German using a well-known standard data set containing four and a half million English-German sentence pairs. A sentence pair is something like "my dog has fleas" matched to "mein Hund hat Flöhe." The Attention team figured out that they could teach their AI three hundred percent more lessons and it would only take fourteen percent longer without adding any more GPUs! And if they wanted an even more expert translation AI and used a data set of nine, eighteen, or thirty-six million sentence pairs but kept adding GPUs, it would take the same amount of time.
This meant the corporations and countries who had the data and could buy the GPUs and find enough electricity to run them could teach an entirely new class of AI: foundation model AI capable of making other AI. And that's what happened.
In 2017, Jen-Hsun (Jensen) Huang, the CEO of GPU-maker NVIDIA Corporation, reported annual revenue of $6.9 billion and stated, "We can now see that GPU-based deep learning will revolutionize major industries.…The era of AI is upon us." By 2023, all the big AI companies had trained foundational large language models. In fiscal year 2023, Mr. Huang reported revenue of $26.97 billion and said, "AI is at an inflection point, setting up for broad adoption reaching into every industry" (Choe & Parvini, 2023; NVIDIA, 2023).
NVIDIA, the company that makes GPUs that power AI, saw its revenue grow four hundred percent in the five years that included a global pandemic (Global Macro Monitor, 2024). Microsoft, Google, Amazon, and Meta all bought as many GPUs as possible in a race to build the machines that make machines for the coming industrialization of AI.
NVIDIA announced "partnerships" with the AI companies in this time period. Their biggest customers became something more. NVIDIA also tried and failed to buy their competitor chip designer Arm Holdings for $40 billion. The U.S. Federal Trade Commission squashed the deal with a lawsuit. Not to be deterred, NVIDIA went on the open market and quietly acquired hundreds of millions of dollars of Arm stock and announced new partnerships with Arm (Turner & Bloomberg, 2024).
Just as coal, iron, steam engines, and precision machine tools combined to industrialize human production work, nuclear energy, data, graphical processing units, and foundation models are combining to rapidly industrialize human knowledge work. Rapidly because instead of a century-long evolution driven by small entrepreneurs and inventors, AI industrialization is being driven by some of the wealthiest corporations on the planet.
Big AI
"Big Tech" companies are likely secretly happy with this moniker and tolerate its sometimes negative associations. They love it when a politician or an activist rails against "Big Tech" for controlling internet searches, computer operating systems, and social media. That means we're looking the other way while they quietly go about the business of harvesting our personal data and building up the capital to turn our data into profitable AI. The truth is that Big Tech has actually become Big AI.
The first Industrial Revolution played out over a hundred years, from 1760 to 1860. Change was driven by inventors, entrepreneurs, small, and then mid-sized, businesses in every community. The inventive shoemaker dreamed up a mechanized leather cutting machine. He partnered with the enterprising blacksmith next door to buy an early lathe to make his machine. They went in together on a shop next to the flour mill so they could run both their machines off the steam traction engine the innovative miller had bought to replace his water wheel. The shops sat next to a river full of barges shipping coal from independent mines, steel from small foundries. Now able to produce a larger volume of quality shoes, the shoemaker hired a farmer to run the leather-cutting machine in the afternoon after he had tended the fields. He sold shoes at a lower price, including a pair to the previously barefoot daughter of the farmer. The blacksmith trained the first machinist in the state to make leather-cutting machines which the partners then sold to shoemakers in nearby towns. There weren't any big companies, and the truly private corporation wasn't legal until politics and business conspired in the 1890s. Our cooperative of entrepreneurial small businesses could thrive. The farmer paid off a loan and hired a hand. His daughter went to school in the spring instead of planting corn. Life got better as the whole pie grew and more citizens got a slice. But there were limits to what could happen at this scale.
After this revolutionary period, large corporations began to emerge, at least in the U.S. Innovative, driven businessmen created new corporate and financial strategies with the goal of maximizing profit, often by beating the competition with a better product at a better price. Conglomerates in shoemaking equipment (United Shoe Machinery), oil (Rockefeller), steel (Carnegie), railroads (Stanford's Central Pacific), electricity (J.P. Morgan's General Electric), and tobacco (Duke's American Tobacco) rolled up their respective sectors. At their best, these conglomerates created economies of scale that allowed them to invest in innovation, efficiency, and the workforce, ultimately improving quality of life for more citizens. At their worst, the conglomerates put competitors out of business and did not invest in innovation, efficiency, and the workforce, ultimately leading to real and perceived economic injustice. Fifty years after the entrepreneurial frenzy of the Industrial Revolution, in some sectors, a very few businesses controlled all or most factors of production and sales in their industry. John D. Rockefeller's Standard Oil owned the oil fields, drilling companies, refineries, and effectively the railroads (through exclusive freight contracts), pipelines, heating oil distributors and gasoline service stations. Standard Oil's relentless pursuit of efficiency drove down the cost of oil nearly tenfold by the early 1900s. The first industrialists were also politically active, as they sought to influence regulatory and tax policy to their benefit, much as today's corporations hire lobbyists to advance their interests with whatever party is in power.
The political influence of Rockefeller and his contemporaries is best illustrated by the rapid evolution of corporate law. In the 1880s, Leland Stanford obtained "personhood" for his Southern Pacific Railroad, which established the precedent that a corporation has the same protection against discrimination as a human under the Fourteenth Amendment. In Stanford's case, the discrimination was a special tax on railroad property enacted by the State of California (Ballotpedia, n.d.). Not what the Fourteenth Amendment was meant to protect when it was enacted after the Civil War to protect the rights of freed slaves. Personhood gave corporations special protections, but they were still subject to strict controls on their power by the state legislatures that chartered them. J.P. Morgan created a permissive home for corporations by hiring a lawyer to rewrite New Jersey's corporate law to remove restrictions meant to prevent abuse of power. The changes were happily adopted by the governor and legislature in return for franchise taxes. Indeed, after Ohio broke up Rockefeller's Standard Oil Trust under state law, the oil conglomerate moved to New Jersey to take advantage of the state's enabling corporate laws. In the 1890s, public (and competitor) backlash against industrialists and their political influence, as well as real and perceived anti-competitive practices, brought new federal antitrust laws and ultimately a period of enforcement that resulted in the breakup of many monopolies, including Standard Oil in 1911 (Oller, 2019). But the industrialists hadn't gained power until the end of the first Industrial Revolution, after decades of improvement in standard of living, life expectancy, access to education, and other measures of life, liberty, and the pursuit of happiness.
Artificial intelligence industrialization went straight to conglomerate. Arguably, Big AI didn't even experience the speed bump of an extended or widespread entrepreneurial phase, given that it already had twenty years of momentum. Big AI rapidly invested enormous amounts of capital to consolidate control over facets of AI production, and wielded political influence to increase profits. By early 2025, Microsoft controlled the data of 400 million Microsoft 365 users (iron); had acquired a nuclear energy plant (coal); was in multiple financial relationships with a key GPU supplier, NVIDIA (steam engines); and had a $13 billion stake in OpenAI's GPT foundation models (machines to make machines). Microsoft, NVIDIA, BlackRock (the world's largest investment company), and MGX (the United Arab Emirates' sovereign wealth AI fund) are in a partnership to invest $80 to $100 billion in new data centers exclusively for AI controlled by Microsoft (BlackRock, 2025). Meanwhile, Google controls the data of 1.5 billion Gmail users (not to mention the data from YouTube and Google Photos), has built its own version of the GPU, is developing multiple nuclear energy projects and has its own foundation models plus a $2 billion stake in the foundation models of another wealthy AI company called Anthropic. Amazon controls data from more than 500 million Alexa devices plus the data of more than 350 million retail customers and the connected data of more than 200 million Prime users, has its own foundation models plus an $8 billion stake in Anthropic's models, buys as many of NVIDIA's GPUs as it's allowed to, while making its own GPUs, and is invested in multiple nuclear energy projects. Meta controls the data of the 3.9 billion people who use its apps every month, is in a partnership with AMD (an NVIDIA competitor) to design and build GPUs to install in its new $10 billion AI data center in Louisiana, and has its own foundation models.
As far as political influence goes, the inaugural U.S. National AI Advisory Committee (NAIAC), established by a law passed in 2020 and first convened in 2022, is made up of a who's who of delegates from NVIDIA, Google, Microsoft, Amazon, Anthropic, and others, along with academics and policymakers (U.S. Department of Commerce, 2022). Big AI is giving the advice. The CEOs of the Big AI companies exert more influence when they meet with presidents and presidents-elect of both parties in private, because AI is positioned as a national security issue. Mr. Huang from NVIDIA stated in 2024 that "Countries have awakened to the importance of developing their national AI and infrastructure." His use of the plural "countries" and the word "national" is in keeping with a sometimes arms-race response from elected officials around the world. Which certainly won't hurt business.
Industrial and corporate history is context for the emerging AI industry. AI will change our lives and the lives of our children and grandchildren just as the industrialization of manufacturing changed the lives of our great-great-grandparents. The AI produced by corporations is already changing the nature and value of human cognitive labor. Governments around the world are wrestling with novel policy questions, often with AI corporations in the room. Intertwined political and corporate interests are a nonpartisan phenomenon. Left, right, red, blue, or purple, it is our job as citizens, voters, and consumers to make informed choices that promote benefit and minimize harm of AI. My use of the term Big AI does not imply a negative, just as the use of the term Big Three to describe the major U.S. automakers is not a negative. Modern AI requires a scale of production that only very large corporations or governments can support (for now). Big is not necessarily bad for the individual citizen or society as a whole. Big is less likely to become bad when you have a sense of agency and are an active participant in the change brought by AI. You can be a driver of change. You can have an active role in this new, latest industrial revolution. And that role starts with your data. You teach the machines.
Follow The Data
The most important technology in human history since the printing press relies entirely on your data. When AI improves enough from learning from you and everyone around you, you will be more likely to pay for what it can do. Data is an asset that increases in value as it grows in size, and the big AI companies have been working toward this moment for decades. Ever hear the adage, "If it's free, there's a catch"? For years, I assumed that Gmail, Facebook, and Amazon's one-day delivery were free because "Big Tech" wanted to target me with ads. It was true. But only partially. What they really wanted was my data. Yes, with my data, they could learn of my interest in Iceland and sell a premium targeted ad to Icelandair. But the real money was in banking my data and the data of hundreds of millions of other people, because they knew value depended on volume when it came to training AI. Between 2010 and 2015, the data representing your digital life became a corporate asset with future value to train AI. It didn't cost much to store in the cloud, so collecting ever and ever larger volumes was a safe, long-term bet that, at a certain point, it would be feasible to use your data to industrialize AI. You teach the machines.
As discussed previously, when you read and agree to the Microsoft Service Agreement and Privacy Statement, it's clear that by clicking "I accept," you are allowing Microsoft to use your data to do anything permissible by law. An excerpt follows, accessed on April 22, 2025.
b. To the extent necessary to provide the Services to you and others, to protect you and the Services, and to improve Microsoft products and services, you grant to Microsoft a worldwide and royalty-free intellectual property license to use Your Content, for example, to make copies of, retain, transmit, reformat, display, and distribute via communication tools Your Content on the Services. If you publish Your Content in areas of the Service where it is available broadly online without restrictions, Your Content may appear in demonstrations or materials that promote the Service. Some of the Services are supported by advertising. Controls for how Microsoft personalizes advertising are available at https://choice.live.com (https://go.microsoft.com/fwlink/?LinkId=286759). We do not use what you say in email, chat, video calls or voice mail, or your documents, photos or other personal files, to target advertising to you. Our advertising policies are covered in detail in the Privacy Statement.
As you can see, per the Service Agreement, you "grant to Microsoft a worldwide and royalty-free intellectual property license to use Your Content." And if you follow the links to the Microsoft Privacy Statement, it's clear they have a royalty-free license to share your data with third parties like, say… OpenAI. That means the sentences and paragraphs I type into Microsoft 365 Word today could find their way into the dataset used tomorrow to train the next generation of ChatGPT. They are forthright enough to come right out and say that "As part of our efforts to improve and develop our products, we may use your data to develop and train our AI models."
To be clear, this degree of transparency means that Microsoft is behaving well under the U.S. laws to which I'm subject. And it doesn't just relate to content you type in Word. The "Services" referenced in this agreement are, at the time I'm writing this, an ever-expanding list of one hundred and thirty-eight Microsoft products from Excel (yes, your retirement spreadsheet) to Teams to your XBox.
OpenAI's GPT-3, the AI that blew all our minds, was trained on a large volume of unrestricted, publicly available data. Microsoft wrote OpenAI a check for $13 billion in 2020 possibly because they knew how much more "proprietary" data Microsoft had and were going to continue to acquire, and Microsoft was certain that they could return hundreds of billions of dollars more by combining your data with OpenAI's technology.
But it isn't just Big AI companies. I got an email from my bank, Wells Fargo, telling me that by continuing to use their banking services, I agreed to changes in their Online Access Agreement. The summary of changes includes a paragraph on their use of my information, accessed on April 22, 2025.
Updated Section 17(b) (Privacy and Use of Information – Acknowledgements and Agreements) to (i) clarify that your communications with us may be analyzed and processed (potentially through automated means), and may be shared with our service providers and other third parties, in accordance with our privacy policy and applicable law...
The "potentially through automated means" indicates AI. The "shared with our service providers and other third parties" means practically whoever they want, and I'll never know. Within the law, of course.
A quick word on privacy: We in the U.S. have a general sense that we have a good faith expectation, if not a right, to privacy, that our data will be kept private. That's why it's news (and a hit to share prices) when a company has a data breach. And if data is shared, we expect that our name and identifying details will be removed, making it anonymous or deidentified. There is no explicit protection of this expectation or right in the Constitution, even though the Supreme Court has inferred it from many of the ten amendments that make up the Bill of Rights. But U.S. laws and court decisions have trouble keeping up. Because there is no right protected by law, we choose to give our data up under contractual terms of use that almost nobody reads. And we are also surprised when we find out that a company is doing something… creepy with our data. Without going into the details, please accept for now that, for example, if a company has every email you've ever written, it doesn't matter if your name has been stripped from the email or not. The combination of unique information in those emails is easily linked to your identity. Claims that your data are stored without your identifying information, while technically true, are now irrelevant to practical privacy. When you see them, recognize that, for all practical purposes, they remain as a tonic to help you feel more comfortable with the idea that the book you're writing is being used wholesale to train AI, share with third parties, and make money under a royalty-free license. Microsoft actually does a pretty good job of not claiming to anonymize your data in the terms I've seen. At least they're more transparent!
It may seem like I'm picking on Microsoft, but it's all of the Big AI companies. And for now, everything these companies do with your data is completely legal. Is it right? Wrong? Creepy? That's for you to decide. But you know what else was thoroughly wrong and creepy, yet legal at the time it was done? Harvesting cells from a woman named Henrietta Lacks without her knowledge and then making billions of dollars from them.
The Story of Henrietta Lacks
On October 4,1951, at Johns Hopkins Hospital in Baltimore, a thirty-one-year-old woman named Henrietta Lacks died of complications from tumors growing all over her body. In her 2006 book The Immortal Life of Henrietta Lacks, author Rebecca Skloot writes about the injustice and institutional racism that led to a death that might have been avoided or at least made less painful. She also writes about how a researcher took cancer cells from Mrs. Lacks, without her knowledge or consent, that proved to be the first "immortal" cells to grow reliably in a test tube. The researcher and Johns Hopkins gave these cells away for free to scientists all over the world who used them to study cancer and many other diseases. Eventually, third parties like pharmaceutical and biotechnology companies acquired the cells and used them to make drugs and other biological products.
The book is about much more than what happened with these cells, and I encourage you to read it or watch the movie based on Mrs. Lacks's story. I also encourage you to consider possible parallels between Mrs. Lacks's cells and your own personal data. In her afterword, Ms. Skloot writes:
Beyond simply knowing their tissues are being used in research, some tissue-rights activists believe donors should have the right to say, for example, that they don't want their tissues used for research on nuclear weapons, abortion, racial differences, intelligence, or anything else that might run contrary to their beliefs. They also believe it's important for donors to be able to control who has access to their tissues, because they worry that information gathered from tissue samples might be used against them.
If we rewrite this paragraph, swapping the word tissue for data, we get:
Beyond simply knowing their data are being used in research, some data-rights activists believe donors should have the right to say, for example, that they don't want their data used for research on nuclear weapons, abortion, racial differences, intelligence, or anything else that might run contrary to their beliefs. They also believe it's important for donors to be able to control who has access to their data, because they worry that information gathered from data samples might be used against them.
This certainly seems reasonable to me. But it's wishful thinking at this point in the U.S. Even with sensitive data like your health records and financial information, healthcare providers are allowed to do whatever they want with your data so long as something called an Institutional Review Board determines that it would not be feasible to ask you for your permission and weighs in on the risks and benefits. Next time you go to the doctor, pay attention when you're asked to sign a digital pen pad acknowledging that you've read and understand the Notice of Privacy Practices. From a legal perspective, that signature means you have provided consent for your data to be used in any way allowable on the Notice and by law.
Does the government have anything to say about this? In short, in the U.S., as of May 2025: No. Maybe with some exceptions in California. You have to look elsewhere for another model. The European Union passed legislation called the General Data Protection Regulation, which requires more transparency and control of data for citizens of the EU. As a result, Amazon's terms governing your data (including the use of Alexa in your home) are 3,600 words in the United States and 8,700 in Germany.
Does this all seem like… a lot? It is. I work in the AI sector, and in some ways, I knew all of this before I sat down to write this chapter. But seeing it all laid out on the page in black and white is sobering. We all need to be savvy and protect ourselves.
Exercises: Try It Out
- If you have a Google account
- Log in and search "good present for my [best friend? sister? brother? spouse/partner?]
- Log out and do the same search. Do you see anything different?
- Look up and compare a data policy in the EU to a data policy in the U.S. for the same company.
- Next time you see your doctor, ask for a copy of the notice of privacy practices that you "acknowledge you've received" when you sign the little digital signature pad or complete the online check-in.
- Search a major national news website for "new nuclear reactor" or "proposed nuclear reactor." Look to see who the investors and backers are.
- Read NVIDIA's last couple of earnings reports, especially commentary from company leadership. Dig and find a transcript of the earnings call and see what questions analysts ask and how NVIDIA responds.
References
Ballotpedia, n.d. Santa Clara County v. Southern Pacific Railroad Company. (Retrieved on April 22, 2025, from Santa Clara County v. Southern Pacific Railroad Company - Ballotpedia)
Biswas, Sourya, 2025. History of Cloud Computing: A Journey of Innovation and Future Prospects. Cloud Tweaks. (Retrieved on May 8, 2025, from History Of Cloud Computing: A Journey Of Innovation And Future Prospects)
BlackRock, 2025. Press Releases: BlackRock, Global Infrastructure Partners, Microsoft, and MGX Welcome NVIDIA and xAI to the AI Infrastructure Partnership to Drive Investment in Data Centers and Enabling Infrastructure. (Retrieved on April 22, 2025, from AI Infrastructure Partnership | BlackRock)
Choe, Stan, & Sarah Parvini, 2023 (Updated January 28, 2025). Nvidia Faces a Reckoning as Chinese Upstart Raises Questions About Wall Street's Darling. AP. (Retrieved on April 22, 2025, from Nvidia faces a reckoning as Chinese upstart raises questions about Wall Street's darling | AP News)
Global Macro Monitor, 2024. NVIDIA's Stunning 400% Growth in Data Center Revenue. Macromon. (Retrieved on April 22, 2025, from NVIDIA's Stunning 400% Growth In Data Center Revenue | Global Macro Monitor)
Hindle, Brooke, & Steven Lubar, 1986. Engines of Change: The American Industrial Revolution, 1790‒1860. Smithsonian Institution Press.
Knipprath, Joerg, 2017. Justice Stephen J. Field (1816‒1899). Constituting America. (Retrieved on April 22, 2025, from Justice Stephen J. Field (1816-1899) – Constituting America)
McCarthy, Eugene, 2023. Corporate Law, Business Schools, and White-Collar Crime. St. Louis University Law Journal, 67. (Retrieved on April 22, 2025, from https://scholarship.law.slu.edu/cgi/viewcontent.cgi?article=2330&context=lj)
Microsoft, 2024. Microsoft Services Agreement. (Retrieved on April 22, 2025, from Microsoft Services Agreement)
NVIDIA Corporation, 2017. Newsroom: NVIDIA Announces Financial Results for Fourth Quarter and Fiscal 2017. (Retrieved on April 22, 2025, from NVIDIA Announces Financial Results for Fourth Quarter and Fiscal 2017 | NVIDIA Newsroom)
NVIDIA Corporation, 2023. Press Release Details: NVIDIA Announces Financial Results for Fourth Quarter and Fiscal 2023. (Retrieved on April 22, 2025, from NVIDIA Corporation - NVIDIA Announces Financial Results for Fourth Quarter and Fiscal 2023)
NVIDIA Corporation, 2024a. Computer Industry Joins NVIDIA to Build AI Factories and Data Centers for the Next Industrial Revolution. (Retrieved on May 8, 2025, from Computer Industry Joins NVIDIA to Build AI Factories and Data Centers for the Next Industrial Revolution | NVIDIA Newsroom)
NVIDIA Corporation, 2024b. NVIDIA Announces Financial Results for Third Quarter Fiscal 2025. (Retrieved on April 22, 2025, from NVIDIA Announces Financial Results for Third Quarter Fiscal)
Oller, John, 2019. White Shoe: How a New Breed of Wall Street Lawyers Changed Big Business and the American Century. Dutton.
Skloot, Rebecca, 2011. The Immortal Life of Henrietta Lacks. Crown.
Turner, Nick, & Bloomberg, 2024. Nvidia Tried and Failed to Buy Arm for $40 billion in 2020, but It Just Reported a Stake Worth $147.3 Million. Fortune. (Retrieved on April 22, 2025, from Nvidia discloses $147.3 million stake in Arm | Fortune)
U.S. Department of Commerce, 2022. U.S. Department of Commerce Appoints 27 Members to National AI Advisory Committee. (Retrieved on April 22, 2025, from U.S. Department of Commerce Appoints 27 Members to National AI Advisory Committee | U.S. Department of Commerce)
Wells Fargo, n.d. We Are Updating Our Online Access Agreement. (Retrieved on April 22, 2025, from We Are Updating Our Online Access Agreement)