Gary Levi analyzes the nature of speculative investment bubbles in the tech industry and conjectures on the likely direction that labor struggles will take in response to the coming crash of the AI frenzy.
The 2001 tech bust was the worst job market for programmers in at least 20 years,1 and probably still a worse job market than today. There were over two million layoffs that year; many left the industry and never came back. Almost 2 trillion in market value vanished from the stock market. We knew there had been a bubble, and we knew there had been a bust; only recently did I begin to understand some of the underlying dynamics behind it, and how they relate to what has happened in the industry where we work since.
Especially in the first internet bubble in the early 90s, there was a lot of talk of the “information economy” which was unlike the physical economy, and that justified why growth would be unbounded. But like so much tech hype that we see with each cycle, this was false. What we called the “dotcom” bubble was only secondarily a dotcom bubble. Underlying it was a telecom bubble; this was not a bubble in domain names and valuations of websites. It was a bubble in the underlying network cables and hardware on which these websites ran, the physical economy of information technology. The evaporation of market capitalization and loss of jobs were not concentrated only in the dotcom startups, but overwhelmingly in telecoms like WorldCom, Global Crossing, and Lucent – companies whose names are almost forgotten today.2 To understand where these companies came from, we need to go back further.
The Birth of the Internet
Lots of folks know the internet was invented by DARPA – the Defense Advanced Research Projects Agency. But it was built out by the National Science Foundation as a civilian government project to connect major universities – known as NSFNET. Built during the height of 1980s Reaganite neoliberalism, it was constructed with public funds but destined for privatization from the outset. Having constructed the backbone, the government gave it away to private firms for pennies on the dollar, in the early 1990s. Where these firms come from goes to the AT&T breakup. It was the liberal answer to the AT&T monopoly in the form of “competition” rather than rational public management, which set the stage for the goldrush.3
In 1995, NSFNET was essentially given to a group of private companies which the government had first contracted to build it – MCI, Cogent, and Time-Warner, among others. There was no congressional vote – this happened purely behind closed doors. In 1996, under Clinton, the Telecom Act was signed: furthering deregulation, ostensibly by opening up competition, but in actuality by giving away huge portions of public airwaves and bandwidth to the private sector, allowing more mergers and amalgamations and acquisitions. The creation of new monopoly behemoths. The privatization of NSFNET, the deregulation of the telcos and the conversion of bandwidth into a private commodity were the real acts of creation of the modern internet.4
With these giveaways came a goldrush, and that goldrush was fundamentally about building, owning, and acquiring bandwidth.
19th Century Analogs of the Tech Economy’s Boom and Bust
Marx’s Capital, Volume II is often considered the least interesting of the volumes, but it is very rich, particularly in providing some surprising insights about the tech economy. In Part 2 of the volume, Marx analyzes “turnover,” which is the duration from an investment to its payoff.5 Making a t-shirt has fast turnover. Making a cruise ship has slow turnover. In Chapter 16, he examines turnover in the railroad industry–which tied up capital for especially long periods of time. He describes that in capitalist society, where reason only asserts itself after the event and “great disturbances may and must constantly occur,” such large scale investments require enormous loans or bond investments. These, in turn, are fostered by environments where capital is easily procured–i.e. where interest rates are low and borrowing is cheap. This creates a great deal of paper wealth–productive capital withdrawn from the market and only “an equivalent in money” thrown on the market in their place–in turn generating rises in prices of productive materials and the cost of subsistence goods. Such vast credit outlays create speculation and grift, land booms where rails might run though, negotiations for sweetheart construction contracts and so forth. In the course of this, latent workers are drawn into the market and production is overstimulated to meet speculative demand, leading to local bubbles of overproduction. As Marx notes, “This lasts until the inevitable crash again releases the reserve army of labour and wages are once more depressed to their minimum, and lower.”6
What Marx was describing in railroads was exemplified in the Credit Mobilier scandal of the 1860s, where, in a single grift, roughly $44 million–in today’s money almost a trillion dollars–was embezzled, much of which turned into bribes to government officials. This history and understanding translates in many ways to the telecom bubble.
Underneath the dotcom asset bubble in stock valuation was a classic crisis of overproduction: that specific form described by Marx – the capitalist free market, chaotic and uncontrolled, with many firms at once trying to build capacity, and hence producing overcapacity, and in so doing flooding the market with cheap cash, invested in speculation on use of that capacity. This is one sense in which the analogies between figures such as Elon Musk and Peter Thiel and the robber barons of the 1800s are much more than surface-level. There’s a very important paradox of capitalism here. Long term infrastructure investment, eminently rational, gives rise in capitalist free markets to its inverse, short term speculative frenzy.
Where Are We Today?
From 2001 to 2007, the economy was left in a period of chronic overcapacity, of excess network bandwidth and core hardware investment. What was lacking was enough consumer computing devices to make use of that capacity. A shift came with the introduction of smartphones that put the internet in the hands of everybody, and finally opened a genuine mass consumer market that could allow a new wave of growth – just in time to receive all the excess capital that sought returns in the wake of the great financial crash. This was the basis of the growth of the second internet bubble – fueled on smartphones, continued low interest rates, and especially a glut of speculative capital withdrawn from financial products after the bank crashes.
This new bubble came with and developed a new material underpinning in physical infrastructure – no longer network cables, but increasingly datacenters.7 At first, these were parasitic on the slowly resurgent startup world–the “cloud” was a rent one could pay to bootstrap a startup without needing your own machine infrastructure, which then captured profit by vendor lock-in. A small company would pay a cloud vendor a monthly fee for use of its hardware rather than needing to purchase its own–but they would then become dependent on the services tied to that hardware, and unable to shift as fees piled up. But a feedback loop emerged. The bubble in the cloud led to excess datacenter capacity. In turn, that meant the capacity would be sold cheaply, because you can’t “store” compute cycles – a processor idled today can’t compute for tasks it will be given tomorrow–that potential profit is lost. Cheap capacity would then lead to new applications for cheap computation. Most applications wouldn’t go anywhere, but occasionally one would take off. And when one of those applications struck gold, so to speak, there would be a new wave of demand for datacenters, and so on – a constant churn of hype-driven miniature business-cycles while capacity continues to grow.
One of the first of these cycles was “big data”–which really meant making use of the consumer data to target marketing. That later became “machine learning”–more sophisticated advertising tech with a fancy name. In the advertising mini-bubble it was claimed there was more valuation on companies driven by advertising dollars than the total amount of ad spends, i.e. that the total amount possible to spend on advertisements was dwarfed by the speculative value of companies that could only derive revenue from such advertisements.8 But the cycles were at such a rapid pace that as one bubble began to deflate, the next had already arrived–bouncing workers between startups or bouncing teams at large companies from one half-completed project chasing one buzzword to the next. Rather than a crash, the combination of cheap credit and new “innovative” uses for capacity kept things going. Instead, there was just an industry littered with abandoned half-built toys, along with “swag” and merch from legions of forgotten startups. The products hardly mattered, just the next funding cycle. This is what they called “innovation.”
In a certain philosophical sense, cryptocurrency mining is the culmination of this process, or at least its most conceptually perfect expression. The computation of microprocessors no longer needs to be directed towards calculating things requested by people for some ostensible satisfaction of some human desire. Mining a bitcoin directly transmutes compute cycles into money–or a facsimile thereof. No longer is there the claim that computation is at the service of a product, a producer, or a consumer. Instead, there is just a dream of machines making money themselves – that a processor on its own, solving little puzzles towards no purpose, is producing wealth. When Amazon introduced the cloud, raw computation began to be seen as a commodity. Bitcoin made this manifest – a “real abstraction” so to speak.
After cryptocurrency came the crypto bust – with the collapse of FTX, Three Arrows, Terra, the mass souring on NFTs, and so on. That was precisely when venture capital began to, right on schedule, promote generative AI – now fueling a new and even greater speculative boom in datacenter construction. Take two examples: Microsoft and Amazon. Microsoft has announced spending plans for billion dollar “hyper-scale” datacenters in various cities – massive behemoths that soak up electric and water capacities for entire regions. This includes a gargantuan $100 billion datacenter as a joint effort with OpenAI. Microsoft also announced an initial $30 billion fund with BlackRock for AI infrastructure. Amazon similarly announced a projected $100 billion in datacenter investment over the next decade.
Much has already been written on how these companies are overselling AI, as well as the harms of AI–to workers and consumers alike.9 Still, some may argue: aren’t these tech companies in the business of making money? Clearly, they must see some profitable applications in the future. After all, executives may be greedy and cynical and callous, but they’re not fools.
But the long historic view, and some political economy, show that hype cycles tend to be disconnected from underlying applications, and why that is so. The very dynamics of the market are such that those who rise to the top are liars, grifters, thieves, and con-men. Indeed, the less they care about making things that work well, and matter, the better equipped they are to make money.
Another paradox is that there is a current situation of massive corporate valuations and expenditures for tech companies, while their workers face a wave of layoffs and a brutal job market. This is unlike 2001, where the depressed job market was the evident result of the economic bust–here the layoffs are preceding the bubble popping, not following it.
What happened in 2001 was an uncontrolled crash that was resolved by flooding the market with cheap credit, combined with an infusion into the economy via the Iraq war. Today’s market, on the other hand, is the result of a controlled demolition. In the normal business cycle, roughly every 10-20 years there is a significant crash. From the standpoint of the capitalist economy, this is like a forest fire – it clears out the underbrush; less profitable companies fail and are absorbed into bigger ones, less profitable manufacture is shuttered, and, as old-machinery is end-of-lifed, that destruction of capital means that production resumes with a higher profit rate–a better ratio of exploited labor to fixed capital investment. The covid crisis was not such a typical crisis, but rather a sharp, fast “external shock.” It accomplished much of that consolidation and clearing out for the capitalists very effectively. However, the “rebound” was unexpectedly rapid, leading to a demand shock for labor, and a situation where workers have been understanding their worth, organizing, demanding higher wages, improved conditions and respect.
Very explicitly and consciously to discipline labor, the government raised interest rates to cool demand for workers. In most sectors this had only a modest impact. In tech, because all our work is on the speculative edge of new production, it has been devastating. Interest rates were raised to levels not seen since 1999–where Greenspan’s brief rate hike to that point was part of what immediately induced the telecom crash. Before that moment, the last time rates exceeded this was 1990.
High interest rates mean that the “future value” of money is higher – i.e. a 2-year Treasury bill will pay around 4%, so any investment with a 2-year timespan had better, on average, be more profitable than that. What the capitalists then desire are investments that pay off sooner rather than later, and with certainty. Investing in workers to build new products is speculative investment that will pay off, maybe, sometime in the future. Investing in physical infrastructure in the midst of a bubble-fueled demand surge will pay off surely, and immediately.
Marx writes about the special delusion that arises in a financial bubble–the capitalist forgets that value comes from the creative labor of humans. They begin to believe that it is money that reproduces itself, that it is capital and not labor that begets more capital. And in this speculative bubble with AI there is a similar delusion–they believe and want us to believe that it is raw compute itself that possesses value and begets profit, not the creative human labor that can put that power to meaningful use in service of human needs.
The Nature of Tech Work
In the Part 3 of Capital, Volume II, Marx introduces his famous two-department model of production,10with companies classified into production of goods for consumption, and production of means of production. Some tech work falls into the latter–for example, workers building and maintaining data centers and their network equipment, laying cable or manufacturing chips. Some fall into the former–for example, site reliability engineers maintaining user-facing applications–“producing” so to speak the social media feeds we consume.
But a lot of tech work isn’t in direct production at all. Rather it is supplementary to these classifications, effectively in research and development. Producing the ability to produce new types of goods, or produce more efficiently. This includes not just what we think of as R&D, but even mundane tasks like fixing bugs, implementing new dropdown menus, A/B testing form layouts–anything that does not just reproduce or deliver the product, but somehow changes it. These jobs can require a lot of specific skill and knowledge, and have a multiplicative effect on what is produced and how efficiently it is produced. When demand is high, labor may be scarce and there is plenty of room for compensation to scale. But this job is never really necessary for day-to-day profit. They can pause or downsize at any moment. When core profitability sneezes, R&D catches the flu. Tech workers in this category may be skilled and well compensated, but they are also intensely subject to the fluctuations of market forces, and, depending on circumstances, may have significantly less structural power than the less compensated workers who actually keep systems running. Solidarity across these distinctions is essential to all workers, not just the least well treated.
Right now, the AI bubble is doomed, like prior bubbles. In a sense, that is good, because it is very stupid. In another sense, it will be terrible for the workers who will be the victims of the crash. We are going to continue to be in a period where struggle and solidarity are possible. However, I think these will largely not be for vast improvements in pay and conditions. Rather, they will be defensive struggles, based on preserving conditions, reducing layoffs or at least increasing severance pay–minimizing the pain of the crisis, not fixing it. But these are also conditions where modest but real demands can give rise to mass organizing, the growth of solidarity, and the recognition of our place as workers, those whose labor is extracted.
One more point on Capital, Volume II. In the part of the book, Marx expands on his discussion about the circuit of capital. From money to commodity then production to new commodity and back to money again. The name of the conference, “Circuit Breakers” can be read as a pun about Luddites and machine smashing. But there’s another meaning, with reference to that circuit of capital. Workers are at the key point in the circuit of capitalist production, and there we can be literal circuit breakers, able to interrupt accumulation. Capital doesn’t build things, workers do. CEOs don’t build things, workers do. Private equity doesn’t build things, workers do.
Most tech workers want to build things that matter, and that will stick around. We take pride in our work. We want to think in years, not sprint cycles. But that runs up against how capitalist markets operate. And it runs into an especially deep irony, or contradiction–which is that in capitalist markets it is precisely long-term investments in the future that give rise to rampant speculation, bubbles, hype cycles, and fraud, because that is how long-term investment is given present monetization. One aspect of our organizing should be that we, the workers, care more about building things correctly–in ways that are maintainable, stable, and useful to humans–than our bosses, structurally, ever can. It is not simply that we are in a struggle over our wages and conditions, but that our class position is such that we want to build things that are worthwhile and durable, and their class position is such that they, often, do not. This simple truth should spur and guide our organizing.
- This presentation was given on October 12 at Circuit Breakers – a conference for labor organizing in tech organized by Tech Workers Coalition and Collective Action in Tech. It has been edited for publication.
- Paul Starr, “The Great Telecom Implosion,” The American Prospect, September 8, 2002, https://www.princeton.edu/~starr/articles/articles02/Starr-TelecomImplosion-9-02.htm;
Doug O’Laughlin, “Lessons from History: The Rise and Fall of the Telecom Bubble,” Fabricated Knowledge, October 16, 2023, https://www.fabricatedknowledge.com/p/lessons-from-history-the-rise-and;
Roger Lowenstein, “How Lecent Lost It,” MIT Technology Review, February 1, 2004, https://www.technologyreview.com/2005/02/01/231676/how-lucent-lost-it/. - Jake Kobrick, “The Break Up of ‘Ma Bell’,” Federal Judicial Center, https://www.fjc.gov/history/spotlight-judicial-history/breakup-ma-bell.
- Yasha Levine, Surveillance Valley: The Secret Military History of the Internet. PublicAffairs, 2018.
- https://www.marxists.org/archive/marx/works/1885-c2/ch07.htm.
- https://www.marxists.org/archive/marx/works/1885-c2/ch16.htm#3.
- The Tech Won’t Save Us podcast has coincidentally just launched an excellent four part miniseries, “Data Vampires” about the explosive growth of these: https://techwontsave.us/episode/241_data_vampires_going_hyperscale_episode_1.
- “The Advertising Bubble,” Idle Words, November 11, 2015, https://idlewords.com/2015/11/the_advertising_bubble.htm.
- See e.g. Edward Zitron, “The Subprime AI Crisis,” Where’s Your Ed At, September 16, 2024, https://www.wheresyoured.at/subprimeai/.
- https://www.marxists.org/archive/marx/works/1885-c2/ch20_01.htm#2.