The Silicon Valley Consensus & AI Capex (Part 1)
On overbuilding AI infrastructure and its energy supply
Welcome back! First off, we hit 5,100 subscribers despite the episodic publishing schedule! Bear with me as I ramp it up, but thanks for joining and a special thanks to those of you who’ve become paid subscribers. To celebrate, there is a 25 percent sale for new paid subscriptions!
Today’s essay is (one half of) Part Three of my series on “artificial intelligence” and the second half will be out by Friday.
The first part looked at AI as a vector for revitalizing eugenics and shock therapy, which has become more apparent as the world’s richest man uses the pretense of algorithmic efficiency to finish the job Obama, Clinton, Reagan, and Carter couldn’t quite manage: burning away what little remains of the New Deal.
The second part examined the impact of the British Empire’s attempts to make factories resemble plantations on the birth of modern computation, the influence all this had on what we call artificial intelligence, and the weakness of popular frameworks for understanding Silicon Valley (surveillance capitalism, techno-feudalism, techno-authoritarianism, etc).
Part Three is about the Silicon Valley Consensus, a first stab at concretizing the constellation that links computational infrastructure, energy firms (from fossil fuel extractors to energy providers), and various financiers to help explain why various bullshit tech products are foisted upon us. My goal is to help explain how the scramble to build up infrastructure for artificial intelligence is part of a consistent but fragile pattern where new technologies and developments are backstopped by ongoing bids to bolster existing asset classes, synthesize new ones, and turn speculative gains into real wealth. SVC isn’t just about capital expenditures tacked onto the generative AI hype bubble, though this is certainly a major frontier of the froth—which is the subject of the first stab. SVC is ultimately about a now-familiar process that has come to dominate much of the economy and our daily lives: a bunch of independent profit-seeking actors have converged on sustaining a certain technology through a frenzy of overbuilding, overvaluing, and overinvesting in order to realize excessive gains that can be translated into political power aimed at restructuring society.
If you like my work and want to support me, consider becoming a paid subscriber for $7 a month ($5.25 a month with the 5K sale) or $70 a year ($52.50 a year with the 5K sale)! If you are a bit more well-off, take inspiration from Engels and become a Tech Bubble financier. And remember to share this newsletter with someone who you think might like it (or hate it)—and feel free to leave comments here or shoot me an email with further questions!
AI==computational infrastructure
AI Compute
Over the past two months, we’ve seen four companies announce they’ll spend a cumulative $320 billion on capital expenditures (capex) in 2025 as part of their push to massively expand AI infrastructure in anticipation of demand for generative AI products: Meta anticipates $65 billion, Alphabet promises $75 billion, Microsoft is shooting for $80 billion, and Amazon plans for $100 billion. It’s worth noting while this will not be exclusively on AI, each company insists most of this will be spent on it.
On the question of whether there is sufficient demand to justify such spending, I align with Ed Zitron’s repeated warnings that capital is being misallocated into a bubble that will only benefit firms and individuals motivated by profits, greed, power, and other forms of transparent self-interest. Still, I think it’s worth laying out why this overbuilding, overinvesting, and overvaluing frenzy is akin to lemmings running off a cliff—so we must begin with the actual stuff this sector is building.
Each company’s earnings calls hammer home the same constellation of talking points: we are investing in chips, custom silicon, servers, training clusters, data centers, network infrastructure, deep sea cables—in physical and digital resources that will let us build out the capacity necessary to meet anticipated demand for AI or pursue new developments in the field.
AI Capex
So, what is “AI capex” and why is there so much of it? At its crudest, it is simply the cost of building a physical data center: its land, its power, its cooling systems, its construction materials, its hardware and software components, and the supply chains that make all of it possible.
Let’s start with AI infrastructure itself—compute-intensive set of technologies that require eye-watering sums of capital to build, maintain, and replace. Generative AI requires specialized GPUs, which need to actively use a significant portion of the chip's processing power, which means they generate excessive heat that shortens the lifespan of the chips, racks, servers and clusters that rely on them. Nvidia’s new line of Blackwell GPUs—delayed because of overheating issues—will cost upwards of $3 million per 72-chip server (large data centers have tens of thousands of servers). On an earnings call last year, Nvidia CEO Jensen Huang laid out a rosy picture for the company amid the arms race to develop bigger, better foundation models:
You see now that at the tail end of the last generation of foundation models were at about 100,000 Hoppers. The next generation starts at 100,000 Blackwells. And so, that kind of gives you a sense of where the industry is moving with respect to pretraining scaling, post-training scaling, and then now very importantly, inference time scaling. And so, the demand is really great for all of those reasons.
The cost of a 100,000 Blackwell cluster is, however, only one piece of the puzzle: it leaves out the cost of buying the land, actually building the data center, powering the project, cooling the equipment, networking the equipment, recruiting talent, finding training data that isn’t tainted, the dominant path of model innovation (towards compute-intensive “reasoning” models), and replacing all of this over time at greater and greater scale. Nvidia intends on releasing new chips each year, so you’ll have to figure out how to balance burning out the chips to push ahead and burning out your wallet to buy next-gen chips to burn out to push ahead. And on Nvidia’s most recent earnings call, Huang claimed "Reasoning models can consume 100x more compute. Future reasoning can consume much more compute."
Translation: Save up a few billion dollars if you want to start playing with the majors, then another few billion to keep up! Then another few billion to make X amount of money off of products that don’t exist, work well, or see wide use—in part because of their ill-defined markets and applications, alongside unsustainable unit economics.
Over at venture capital firm Sequoia Capital (which has invested in Nvidia, OpenAI, and an AI firm started by OpenAI co-founder Ilya Sutskever), partner David Cahan has argued that the overbuilding caused by the AI capex arms race is actually rational from a game theory perspective:
In fact, the more you believe in AI, the more you might be concerned that AI model progress will outpace physical infrastructure, leaving the latter outdated. For example, once everyone has 100k clusters, big tech companies will need to figure out what to do with their 50k and 25k clusters. We’ve heard a few industry experts make comments along the lines of: No one will ever train a frontier model on the same data center twice—by the time the model has been trained, the GPUs will have become outdated, and frontier cluster sizes will have grown. There is also the issue of how much power you need for a given real estate footprint and how dense to pack your GPU racks, decisions that are dependent on GPU power efficiency—a moving target.
The key to understanding the pace of today’s infrastructure buildout is to recognize that while AI optimism is certainly a driver of AI CapEx, it is not the only one. The cloud players exist in a ruthless oligopoly with intense competition. This is no small prize to defend—the cloud business today is a $250B market, roughly the same size as the entire SaaS sector, combined. The cloud giants see AI as both a threat and an opportunity and do not have the luxury to wait and see how the technology evolves. They must act now.
The arms race between Microsoft, Amazon and Google is thus game theoretic. Every time Microsoft escalates, Amazon is motivated to escalate to keep up. And vice versa. We are now in a cycle of competitive escalation between three of the biggest companies in the history of the world, collectively worth more than $7T. At each cycle of the escalation, there is an easy justification—we have plenty of money to afford this. With more commitment comes more confidence, and this loop becomes self-reinforcing. Supply constraints turbocharge this dynamic: If you don’t acquire land, power and labor now, someone else will.
Regardless of whether or not we need this AI infrastructure, Cahan believes this is akin to a grand state infrastructure project—the risk is borne by the tech firms collectively spending hundreds of billions (instead of governments/taxpayers), and it’ll function as a subsidy to everyone who builds something with this new capacity. This metaphor, of course, is confused dogshit.
Infrastructure projects, when they’re not speculative time-bombs, tend to have some sense of revenue—they have some way to make good to investors and creditors. If we proposed a toll highway in a city where post-COVID there was massive migration out of the city, bringing traffic to an all-time low, it would still have a higher credit rating and a more concrete plan to make money (albeit declining ones) than the AI Capex as Infrastructure: spending hundreds of billions of dollars to make a few billion on disappointing and limited use products.
It’s worth looking at, however, complicating points Cahan has raised himself. In September 2023 and June 2024, he argued there was a huge gap between revenue expectations promised by the AI infrastructure overbuild and actual revenue growth—first a measly $200 billion, which then grew to $600 billion. Cahan is certain that a "huge amount of economic value is going to be created by AI,” that we are living through what may be a "generation-defining technology wave,” and that AI capex is akin to building railroads. Still he presses four points:
Lack of pricing power: Cahan believes that while railroads confer natural monopolistic advantages (there is only so much track space between two points), the same is not true of GPU data centers. GPU computing is becoming commoditized, AI compute can be offered through the cloud, and so prices are getting competed down to their marginal cost (airlines are offered as an example).
Investment incineration: Speculative investment frenzies are common! The underlying asset is not more impervious to zeroing out because large hoards of capital were poured into it; "It's hard to pick winners, but much easier to pick losers (canals, in the case of railroads).”
Depreciation: Compute differs from physical infrastructure in that it follows Moore’s Law. The continuous production of next-generation chips with lower costs and higher performance will accelerate the depreciation of the last-generation of chips. This goes both ways: markets will overestimate the value of today’s chips and under-appreciate the value of tomorrow’s chips. "Because the market under-appreciates the B100 and the rate at which next-gen chips will improve, it overestimates the extent to which H100s purchased today will hold their value in 3-4 years."
Winners vs losers: Long term, declining prices for GPU computing will be a boon for innovation/startups, but a bane for investors. "founders and company builders will continue to build in AI—and they will be more likely to succeed, because they will benefit both from lower costs and from learnings accrued during this period of experimentation."
If we entertain the idea that AI capex is like infrastructure, Cahan’s points complicate things more than he may realize:
Lack of pricing power: It’s not clear why GPU commodification will mean GPU data centers and AI cloud will be low barrier to entry. First, consider the differentiation that already exists when it comes to GPU software stacks, or the specialized infrastructure required to handle frontier models, or the geopolitics that may push countries to back national champions—why wouldn’t these forces drive markets towards monopolies or oligopolies? Second, why are we ignoring the tech industry’s history? Have large tech firms really never leveraged their scale into predatory pricing power? Never colluded on prices? Never carved up markets and ecosystems into walled gardens? Third, GPU data centers are capital intensive and GPU supply chains are strained by demand for Nvidia's newest chips swells. As hyperscalers build larger frontier clusters that demand a larger share of Nvidia’s next-gen chips, and energy/land/water constraints tether this all to physical assets with intrinsic value. Fourth, airlines aren’t a great example here! Not only do airlines employ revenue maximization schemes that bear no resemblance to marginal cost pricing, they were the some of the earliest adopters and innovators of algorithmic price-fixing, eventually setting up a clearinghouse for the industry to collude on pushing fares upwards—the DOJ filed a lawsuit but opted to settle, leaving much of the cartel in place.
Investment incineration: I agree. I would go a bit further and say firms burning AI capex are obviously incentivized to prevent incineration—but by inflating value instead of providing it. Venture capitalists, large tech firms, AI startup darlings, they all deploy thin marketing campaigns for unsustainable business models, but the argument seems to be they will eventually grow into profits on a long-enough time scale that also happens to gives every financier an out (exit, acquisition, inflation of other portfolio firms, political power, influence, etc.) even as they rack up inordinate externalities that will, of course, be foisted on the rest of us (ecological degradation, ascendance and empowerment of right-wing tech capitalists, enshittification of products and spheres of daily life onto which half-formed AI goods and services will be grafted onto, monopolization of innovation system in increasingly narrower hands, etc.)
Depreciation: If the assumption that Nvidia will keep producing better chips on a yearly schedule holds true, Cahan is right. But will it? Nvidia’s chief executive believes his firm’s progress in AI is surpassing Moore’s Law because they’ve discovered new scaling laws—Gary Marcus has written extensively about why these claims are likely bunk, but either way we will get a better picture this year of what’s what. The first phase of this accelerated rollout has already hit road bumps: Blackwell chips were delayed from 2024 to 2025, and the planned next-gen Rubin chips delayed from 2025 to 2026.
Winners vs losers: Here we come to the assumption that overbuilding will make it cheap enough for someone to build the next “killer app.” This is an unquestioned article of faith among most financiers and entrepreneurs in the space, but is it true?
Last year Goldman Sachs’ report on Generative AI “Too Much Spend, Too Little Benefit” got a little press but there wasn’t nearly enough appreciation for (or rebuttal of) Jim Covello, the head of Global Equity Research at Goldman Sachs. As Covello lays out, there’s little reason to believe Cahan’s framing about AI Capex driving down costs and spurring innovation:
The idea that technology typically starts out expensive before becoming cheaper is revisionist history. Ecommerce, as we just discussed, was cheaper from day one, not ten years down the road. But even beyond that misconception, the tech world is too complacent in its assumption that AI costs will decline substantially over time. Moore’s law in chips that enabled the smaller, faster, cheaper paradigm driving the history of technological innovation only proved true because competitors to Intel, like Advanced Micro Devices, forced Intel and others to reduce costs and innovate over time to remain competitive.
Today, Nvidia is the only company currently capable of producing the GPUs that power AI. Some people believe that competitors to Nvidia from within the semiconductor industry or from the hyperscalers—Google, Amazon, and Microsoft— themselves will emerge, which is possible. But that's a big leap from where we are today given that chip companies have tried and failed to dethrone Nvidia from its dominant GPU position for the last 10 years. Technology can be so difficult to replicate that no competitors are able to do so, allowing companies to maintain their monopoly and pricing power. For example, Advanced Semiconductor Materials Lithography (ASML) remains the only company in the world able to produce leading edge lithography tools and, as a result, the cost of their machines has increased from tens of millions of dollars twenty years ago to, in some cases, hundreds of millions of dollars today. Nvidia may not follow that pattern, and the scale in dollars is different, but the market is too complacent about the certainty of cost declines.
Covello occupies the position of a deeply pessimistic skeptic. I encourage you to read his interview in the report, but I’ll quote one more section where he's asked if AI is comparable to earlier transformative technologies in that their applications were difficult to imagine at first:
The idea that the transformative potential of the internet and smartphones wasn’t understood early on is false. I was a semiconductor analyst when smartphones were first introduced and sat through literally hundreds of presentations in the early 2000s about the future of the smartphone and its functionality, with much of it playing out just as the industry had expected. One example was the integration of GPS into smartphones, which wasn’t yet ready for prime time but was predicted to replace the clunky GPS systems commonly found in rental cars at the time. The roadmap on what other technologies would eventually be able to do also existed at their inception. No comparable roadmap exists today. AI bulls seem to just trust that use cases will proliferate as the technology evolves. But eighteen months after the introduction of generative AI to the world, not one truly transformative—let alone cost-effective—application has been found.
A recent Ed Zitron piece, “There Is No AI Revolution” hammers home this point and many more (gAI has few users, few clients, few use cases, but numerous costs and endless bulls), but I want to hammer in on several key points as we enter the final stretch:
Where is the money that this supposedly revolutionary, world-changing industry is making, and will make?
The answer is simple: I do not believe it exists. Generative AI lacks the basic unit economics, product-market fit, or market penetration associated with any meaningful software boom, and outside of OpenAI, the industry may be pathetically, hopelessly small, all while providing few meaningful business returns and constantly losing money.[sic]
Generative AI business models do not work. They require more training data than currently exists, hundreds of millions of dollars to hire, and burn money for companies offering their services even when customers are paying. In the case of OpenAI, as Zitron lays out, it spent $9 billion to make $4 billion in 2024 and spent all of its revenue + another billion on compute (running models and training them). It loses money on every single paid and free user, increases its burn rate with its paid subscriber base, and if it makes its projected $11.6 billion in 2025 will be "on course to burn over $26 billion in 2025 for a loss of $14.4 billion." All this despite its relationship with Microsoft:
It's also important to note that OpenAI's costs are partially subsidized by its relationship with Microsoft, which provides cloud compute credits for its Azure service, which is also offered to OpenAI at a discount. Or, put another way, it’s like OpenAI got paid with airmiles, but the airline lowered the redemption cost of booking a flight with those airmiles, allowing it to take more flights than another person with the equivalent amount of points. At this point, it isn’t clear if OpenAI is still paying out of the billions of credits it received from Microsoft in 2023 or whether it’s had to start using cold-hard cash.
As Zitron argues, convincingly I think, the few products OpenAI does offer (such as DeepResearch or Operator) are incredibly expensive and still lose the company money, being of such poor quality that users will engage them more than usual "to try and get the desired output, incurring further costs for OpenAI." It doesn’t matter how many users it has, it fundamentally loses money and burns more money the more users it garners because of its horrible unit economics and the low quality output of various products that make use even more costly, even when locked away at the highest tier.
Where does that leave us? Large tech firms are boosting the inevitability of an AI revolution and spending hundreds of billions of dollars on overbuilding capacity for AI.
Why are they doing that? They are locked in an arms race where billions, if not trillions, will be wasted by firms independently trying to create the first wheel (AGI).
Why is this arms race happening? Well, these firms believe creating this wheel will revolutionize human civilization and are pursuing products that require higher and higher levels of AI compute infrastructure, and because these firms argue they need even more compute to push the products further.
Do these products make money? No. There is a multi-hundred billion dollar revenue hole, pointed to by AI financiers and critics alike, that only grows because generative AI applications lose money on every user and have abysmally low conversion rates from free to paid users.
Are they widely used? No. As Zitron puts it, "the entire combined monthly active users of the Copilot, Claude, Gemini, DeepSeek, and Perplexity apps amount to 66 million or 19.47% of the entire monthly active users of ChatGPT's mobile app." Web traffic gets us to 161.6 million unique monthly visitors for those competitors—about 65.69 percent of ChatGPT.com's traffic. If we exclude Deepseek, however, its 39 million monthly active users of ChatGPT's competitor apps and 81.7 million unique monthly visitors to ChatGPT competitor sites.
Is there any evidence they will be substantially cheaper or more widely adopted in a time horizon that justifies the massive spend or overbuild beyond doing it for the sake of winning the arms race? No. Comparisons are often made to other technologies that have become foundational to the digital economy, but with revisionist histories that obscure how these digital technologies were always relatively cheap at the start and had early roadmaps to profitability or widespread adoption that are clear from the start.
So why is this arms race happening? Well, these firms believe creating this wheel will revolutionize human civilization and are pursuing products that require higher and higher levels of AI compute infrastructure, and these firms argue they need even more compute to push the products further.
Cloud Compute
Another place to start is with cloud computing infrastructure, of which AI infrastructure is a subset. The former allows customers to rent computing resources (databases, networking, servers, software, storage, etc.) from other companies that actually own or maintain prerequisite physical hardware. This means things like massive data centers, yes, but also CPUs that go into the servers (multiple CPUs) that go into the racks (dozens of servers) that go into clusters (dozens of racks) that go into the data centers (tens or hundreds of thousands of servers). It means the servers within that process and store data, networking equipment that ensure low-latency, virtual resources and machines, and so on. So capex for cloud looks like: buying land and buildings, constructing data centers, physical servers and networking equipment, renovations and upgrades, and maintenance and replacement of servers, among other things. Similar to AI infrastructure, but different when it boils down to the hardware and software.
If we look at who owns cloud compute infrastructure, we can broadly break it down to a few groups: hyperscalers (large tech firms that own and maintain massive data centers), colocation providers (firms that lease physical space to other firms that want to house hardware off-site), enterprises (large firms that need to own and operate their own data centers). We are going to focus on hyperscalers in this section, but the other groups will emerge in our discussion later.
Hyperscalers
Since 2018, three hyperscalers (Amazon Web Services, Microsoft Azure, and Google Cloud) have been the largest cloud service providers and currently control two-thirds of the global cloud market. Those “Big Three” cumulatively spent $68 billion that year on company-wide capex—not just on cloud infrastructure, but everything from land, office buildings, autonomous vehicles, satellites, warehouses. In 2019, they spent $73.5 billion, $97 billion in 2020, $124 billion in 2021, $127 billion in 2022, $127 billion in 2023, $212 billion in 2024 (a grand total of $828.5 billion). This year, they’ll spend about $255 billion and push the cumulative total to $1 trillion these past eight years.
Hyperscalers vary widely in the details they volunteer about their spending but, like every good ol’ corporation, try to err towards as little as possible, so we’ll focus on the company that offers a bit more than the rest: Amazon.
The basic story we are told is that enormous sums of capital invested in cloud compute has paved the way for AI, but it is still woefully insufficient as present and future demand for AI and non-AI services outstrip compute capacity. Cloud compute is important to think about because AI evangelicals often insist they’ll enjoy similar margins to this sector—a business unit which accounts a significant share of Amazon’s and Microsoft’s profits.
For the most of its history, Amazon’s capex has been directed towards building out the e-commerce/delivery/logistics operation. That’s meant hundreds of warehouses and delivery facilities over the years (from 2020-2022, it doubled the size of its entire network), an air cargo fleet of 97 planes, a hoard of last-mile delivery vans, sprawling physical locations (e.g. Whole Foods), and so on.
But Amazon has also invested heavily in AWS though: doubling its data center capacity as enterprises adopted cloud in 2018, taking advantage of the Covid-19 pandemic’s acceleration of e-commerce, telemedicine, remote work, and virtual education, and most recently in building out its infrastructure in response to the generative AI hype kicked off in 2022.
Across 2022, 2023, and 2024, AWS has brought in over $270 billion in revenue and logged just shy of $90 billion in operating income—profits that slightly edged out the parent company’s profits over the same period. And AWS profits have historically been a key part of Amazon’s anticompetitive strategy—the parent company would reinvest as much of its profits as possible into aggressive expansion, cutting costs and subsidizing prices to attract then trap consumers and clients and partners.
In 2023, when the Federal Trade Commission put out a request for information on cloud computing providers and their services, the Institute for Local Self-Reliance (ILSR) sent a comment letter laying out the consequence of hyperscalers (like AWS) dominating cloud compute. The letter is worth going through as it lays out well some key market dynamics of the cloud oligopoly:
Dominant Cloud Providers Have Durable Market Power and Persistently High Profits. In 2023, AWS controlled 40 percent of the market—larger than Microsoft Azure and Google Cloud combined. AWS commands a smaller lead (32 percent vs 23+10 percent) today, but the cloud market is still an oligopoly with these three controlling two-thirds. At the end of 2022, before Amazon ratcheted up its AWS capex as part of the AI arms race, ILSR notes AWS CEO Adam Selipsky said "more than 90 percent of all cloud-based startups use AWS, while 83 percent of the more than 1,000 technology "unicorns" rely on AWS infrastructure and software. It also dominated cloud services for federal, state, and local governments: over 7,500 local and national government agencies used AWS. Take this, throw in profit margins that are typically around 30 percent, add growth rates that outpaced worldwide growth of the market itself, and you get "durable market power warranting investigation and intervention by regulators."
Vertical Integration Across the Stack Impedes Competition. Hyperscalers are vertically integrated, meaning they own multiple parts of their own supply chain—here, the "infrastructure, platform, and software" that goes into cloud computing. The market's entry barriers may be relatively low, but hyperscalers can leverage their control over its capital-intensive infrastructure and the extent to which their products are integrated throughout the stack, to impose costs that should not exist. Some of the go-to methods for AWS are "customer lock-in and exploitative conduct."
Dominant Cloud Providers Erect Barriers to Switching and Multi-Cloud. AWS uses a variety of tactics to lock-in customers or exploit their reliance on its services. One reliable tool is to charge fees to customers who want to move data off its servers (known as "egress fees"), and make them exorbitant enough to discourage switching elsewhere or using multiple cloud providers. Another is to simply avoid interoperability for non-AWS products and services, creating "technical barriers to both switching cloud providers and integration operations across multiple providers, thus inhibiting competition." These tactics serve to ensure clients stay within the AWS product ecosystem, but to get users through the door in the first place it will offer discounts for bundled use of various tools as well as for long-term contracts that require "some threshold of money spent on services."
AWS Leverages Its Infrastructure Dominance to Preference Its Own Applications and Copy the Products Developed by Independent Software Providers. The sheer dominance of AWS, as with the sheer dominance of Amazon's retail operation, means that independent developers must offer their products on AWS's marketplace. This lets Amazon surveil and access their data and restrict their ability to promote those products elsewhere, but is a key part of Amazon and AWS’s long-documented practice of copying products offered by vendors on their platforms and aggressively self-preferencing the stolen designs.
Amazon Has Used Acquisitions to Cement Its Dominance in Cloud. This is a hallowed practice in our tech sector: acquire your way to a competitive advantage. ILSR's major example is Amazon's acquisition of Annapurna Labs in 2015, allowing for the development of Graviton microchips which attracted major clients to AWS, further entrenched its market power, and helped sustain stupendous growth for revenue and profits. As of 2021, Graviton servers generated over $5 billion in revenue each year.
Amazon Exploits AWS to Advantage Its Other Business Lines. AWS, as should be abundantly clear by now, provides "crucial" services to companies across the economy—including firms which are Amazon competitors. How does Amazon navigate these conflicts-of-interest? One way is by cross-leveraging AWS and other business lines to undermine its competition, as it did in 2020 when Amazon agreed to carry HBO Max on its Fire TV product when Warner Media agreed to renew its AWS contract. Other times, it will use "the market intelligence that Amazon gleans from providing cloud services" to "move into new industries with a built-in and unfair advantage." At the same time that AWS offered cloud-based tools for the health care sector, it pivoted towards the industry and acquired incumbent firms like One Medical to gain a foothold. If none of that sounds appealing, Amazon can also charge a non-cloud direct competitor for cloud services (while enjoying those same services itself at a discount for, say, its retail platform). When Amazon acquired Twitch in 2014, it acquired the tech behind what would become "AWS-based Amazon Interactive Video Service" (IVS). All and well, except that technology is the backbone of Kick, Twitch's main rival. "This means that Amazon competitor Kick must pay sizeable fees to Amazon for cloud services, including the pivotal IVS service, while Twitch likely pays much less for these same services."
Major Cloud Providers Control the Development of AI and Stand to Gain More Power Across Society as AI is Deployed in a Broad Array of Uses. AI infrastructure, built atop cloud compute infrastructure, will "further entrench the market power of the dominant cloud computing providers, as only they currently have the compute power, financial resources, and massive data sets needed to develop and train AI models. Because of these high thresholds, smaller AI companies must partner with one of the large cloud providers to viably develop and offer AI products, as Open AI did with Microsoft." Or as Amazon does with Anthropic. No one should entertain the idea of letting oligopolistic cloud providers like Amazon dominate the generative AI market, whether or not you believe it will create something of value, given how exploitative their business practices are in cloud.
AI infrastructure is a subset of cloud compute infrastructure, so what do we see when we examine the latter’s market dynamics? The vast majority of market is dominated by three firms who deploy deeply exploitative business practices that leverage their hyperscale infrastructure to extract rents, steal ideas, deter competitors, trap clients, and enter new markets with built-in advantages. One way to understand the sky-high levels of “AI capex” is that firms recognize there is an opportunity here: to create another industry that resembles cloud—not in its utility or ubiquity but in its rents, anti-competitive market intelligence, and persistently high margins. Or to put it another way, one reason why AI capex seems to only be going higher is that firms are eager to sustain levels of growth and profit that only come about by rigging a market’s infrastructure (its regulations, practices, prices, and participants). Spend enough money and you may get the first mover advantage, or be able to join the hallowed few who enjoy something like it.
AI==energy infrastructure
Our discussion of upward pressures on AI capex as it relates to compute infrastructure is only one part of the Silicon Valley Consensus. The next part boils down to how eager the most powerful and wealthy sectors of our society are to preserve their power and privilege with artificial intelligence. We will focus on energy supply and demand—specifically, the fossil fuel industry which wants to power AI infrastructure and be powered by it—but this can also be applied to the AI-mediated privatization of healthcare, ascent of defense tech startups, or coming deluge of fintech.
As Kate Aronoff writes for The New Republic, the few tech firms already firmly in control of key parts of global AI compute have made it no secret they felt constrained by net-zero pledges, sustainability commitments, and climate agreements—just like the fossil fuel industry:
On a recent quarterly earnings call, Amazon CEO Andy Jassy called AI the “biggest opportunity in business since the internet.” The main constraint on its growth, he added, is energy. “I think the world is still constrained on power from where I think we all believe we could serve customers if we were unconstrained,” he told investor analysts.
Fossil fuel companies are cashing in on Silicon Valley’s newfound thirst. “Data centers are learning a lot about the natural gas business and how critical it is to what they’re doing,” Marshall McCrea, CEO of the pipeline firm Energy Transfer, told investor analysts recently. “And man, we couldn’t be more pleased and excited about this new business.”
It’s also important to note that while Trump’s ascension is accelerating the shift here, it’s been in the works for a while. Back in September, Brian Merchant wrote a bit on how AI firms were making the climate crisis worse in at least three ways: selling tools to boost fossil fuel extraction, planning to exponentially build out energy-intensive compute, and giving fossil fuel companies alongside utilities excuses to overbuild infrastructure. Take Bloomberg’s reporting last year which revealed that the ongoing AI frenzy was already leading to a revival of US natural gas-fire power generation:
In the first six months of the year alone, companies have announced plans to build more new gas power capacity across the US than they did in all of 2020, data from Sierra Club show. And if the second half looks anything like the first, 2024 will mark the most new gas-power generation announced since at least 2017, when the environmental group started tracking the data.
Is this bad? Bloomberg notes that natural gas generation is likely to leak methane, which has "80 times the planet-warming impact of carbon dioxide in its first 20 years in the atmosphere" and which will come from new gas plants that will "run for 40 years or longer." Decarbonization goals are also being “quietly” revised downwards, renewable energy projects are being canceled, and a big thumbs up is given to building out data centers such that they’ll surge their domestic energy demand from 1 percent today to nine percent by 2030. Or consider Merchant’s observation: “If you think the web is overrun with AI content now, imagine a world where one tenth of all the electricity we generate is going into pumping out more of the stuff.”
Now, every single infrastructure project doesn’t get built simply because analysts project it or because the capital is out there. Bloomberg echoes the same point in their reporting:
Of course, not every announced plant makes it through construction or ultimately connects to the grid. Sierra Club says about 10 gigawatts of earlier plans were canceled last year compared to the nearly 45 gigawatts announced. Berkeley Lab estimates only about one-third of interconnection requests have historically resulted in active gas plants, though that’s higher than either wind (20%) or solar (13%). Some planned projects will may also get scrapped over concerns their cost will further inflate rising power prices.
Still, AI hysteria serves as a potent marketing strategy for actors looking to drive investment, hype products, inflate valuations, justifying the build out of new infrastructure (or entangling existing one). A recent report from the Institute for Energy Economics and Financial Analysis suggests ratepayers will subsidize the proposed overbuilding of electrical infrastructure that is largely being driven by "forecasted demand" for data centers:
“There is a serious risk of overbuilding electrical infrastructure to meet data center demand that may not materialize. Utilities already are financially incentivized to overbuild infrastructure, and this risk is exacerbated by the uncertainty surrounding data center demand, particularly as it relates to AI,” said Cathy Kunkel, IEEFA energy consultant and author of the report. “In the absence of proactive decisions by utilities or regulators, electric ratepayers will subsidize the building of new infrastructure that would not be needed in the absence of data centers and will be on the hook for overbuilt infrastructure.”
Tech has been bolstering fossil fuel extraction for some time now, as JS Tan wrote for Logic(s) nearly six years ago. Microsoft signed a 7 year contract with Chevron in 2017, announced major partnerships with oil giants in 2018, and signed a 2019 deal with ExxonMobil that was claimed to (then) be "the industry's largest [contract] in cloud computing." Why?
Oil companies like Chevron are the perfect customer for cloud providers. For years, they have been generating enormous amounts of data about their oil wells. Chevron alone has thousands of oil wells around the world, and each well is covered with sensors that generate more than a terabyte of data per day. (A terabyte is 1,000 gigabytes.)
At best, Chevron has only been able to use a fraction of that data. One problem is the scale of computation required. Many servers are needed to perform the complex workloads capable of analyzing all of this data. As a result, computational needs may skyrocket — but then abruptly subside when the analysis is complete. These sharp fluctuations can put significant pressure on a company like Chevron. During spikes, their data centers lack capacity. During troughs, they sit idly.
This is where the promise of the public cloud comes in. Oil companies can solve their computational woes by turning to the cloud’s renting model, which gives them as many servers as they need and allows them to pay only for what they use.
And as Karen Hao wrote last year, Microsoft has gone out of its way to pitch generative artificial intelligence as “a powerful tool for finding and developing new oil and gas reserves and maximizing their production—all while publicly committing to dramatically reduce emissions.” The sums involved are potentially massive: in 2024, the oil and gas sector spent a little over $3 billion on artificial intelligence, but Microsoft believes this will can grow to $75 billion soon. The defense Microsoft offers here is a bit unbelievable: yes, the products are being used to boost the productivity of fossil fuel extraction, but they are reducing the emissions generated by extraction itself. On top of this, it’s important to stop and think about what has to happen so that Microsoft can supposedly green Chevron’s fossil fuel extraction:
Lucas Joppa, Microsoft’s first chief environmental officer, who left the company in 2022, fears that the world will not be able to reverse the current trajectory of AI development even if the technology is shown to have a net-negative impact on sustainability. Companies are designing specialized chips and data centers just for advanced generative-AI models. Microsoft is reportedly planning a $100 billion supercomputer to support the next generations of OpenAI’s technologies; it could require as much energy annually as 4 million American homes. Abandoning all of this would be like the U.S. outlawing cars after designing its entire highway system around them.
Therein lies the crux of the problem: In this new generative-AI paradigm, uncertainty reigns over certainty, speculation dominates reality, science defers to faith. The hype around generative AI is accelerating fossil-fuel extraction while the technology consumes unprecedented amounts of energy. As Joppa told me: “This must be the most money we’ve ever spent in the least amount of time on something we fundamentally don’t understand.”
This sentiment is not limited to Microsoft, however—Silicon Valley’s ethos can be summed up as: why shouldn’t we? Or, as former Google chief executive Eric Schmidt put it: “we’re not going to hit the climate goals anyway” so why not gamble on artificial intelligence squaring this circle? Microsoft and Google (two of the largest hyperscalers) admit their data center expansion plans have ruined their sustainability goals: Microsoft’s emissions are up 29 percent since 2020 and Google’s are up 48 percent since 2019. A Guardian analysis found that from 2020 to 2022, real emissions from company-owned data centers of Google, Microsoft, Meta, and Apple where about 662 percent higher than officially reported thanks to "creative accounting.”
Similarly, the sentiment coming out of the fossil fuel sector is: why shouldn’t we? Or as one geoscientist turned financial analyst told writer Malcolm Harris at a Shell event: “We don’t plan on losing any money”—in fact we plan on making more money! As Harris lays out in the introduction of his forthcoming book What’s Left, oil companies have no reason to think they won’t make more: they're public firms with a responsibility to their investors to seek profits, and they’re eager to disregard studies that say the vast majority of oil and gas reserves are unextractable from an ecological perspective—from the market perspective, these reserves promise to heighten record levels of subsidies and profits.
If you want even greater levels of historic profits, this logic demands you look to the products Silicon Valley insists will revolutionize human civilization. And if you are a Silicon Valley firm looking for a proven use case for a technology sufficiently attractive to investors, why not look to the industry that is making more money than God destroying the world AND already planning out scenarios on how to best support your proliferation!
Going back to Aronoff's article, she cites one of the scenarios in Shell's latest Shell Energy Security Scenarios: "Surge," a future where our glorious private sector overinvests in and overbuilds compute infrastructure, fossil fuel infrastructure, and electrical infrastructure.
This AI-centric scenario sees “relative resilience for oil and gas production” owing to “more buoyant economic growth driving overall energy demand. Although new energy infrastructure appears rapidly, demand is such that existing energy production and systems are retained for longer.” Thanks to AI and surging economic growth, though, midcentury advancements in direct air capture and small modular nuclear reactors will keep global temperatures from rising by more than two degrees Celsius.
This is massively optimistic. In Surge, direct air carbon capture and storage, or DACCS, will remove more than 500 metric tons of carbon dioxide annually by 2030, and eight billion metric tons by 2100—a fifth of what the entire world currently emits each year. That technology is currently removing just 0.1 metric tons of carbon dioxide per year, per the International Energy Agency. The IEA’s own scenario for reaching net-zero emissions by 2050—which is relatively bullish about DACCS—forecasts DACCS capturing 65 metric tons by then, with the vast majority of reductions coming from phasing out fossil fuels and scaling up zero-carbon energy.
Some might take a moment to say “excuse me, hyperscalers are investing in nuclear power to fuel their AI infrastructure!” Are we supposed to just not believe tech firms insisting they will invest in and realize a new nuclear age to justify the ravenous growth of their questionable product? Amazon, Google, Microsoft, and Meta have all announced plans to invest in and develop nuclear power for their data centers—in fact, they’ve poured at least $1.5 billion this past year into small modular reactors!
It’s not clear why this shouldn’t be understood as anything more than greenwashing. This is the industry’s oldest trick—insisting the externalities of their profit-seeking will be mitigated by a moonshot project. Recall cryptocurrency? The speculative vehicle once spun as the future of money, cultural production, decentralized governance, and civilization itself? Just as crypto was struggling to find a use case beyond elaborate Ponzi schemes, financial deregulation, degenerate gambling, and electricity consumption, a new narrative was pitched: it was actually fighting climate change.
Bitcoin mining wasn’t straining energy grids worldwide, the industry said, but in fact optimizing them by consuming wasted energy. Emissions could be significantly curtailed by putting carbon credits on the blockchain, enmeshing them in more transparent, auditable, liquid, and accessible markets. And, more broadly, by fostering the growth of Web3 technologies (the umbrella term for decentralized blockchain-centric projects), we might even create new “regenerative finance” systems that promote nature-based and market-based solutions to climate change.
All of this turned out to be a crock of shit, of course. With each year that passed, crypto’s negative environmental impact grew alongside its persistent greenwashing campaigns. Combining crypto and carbon credits seems to have borne out the worst of both worlds, creating safe havens for speculators that resemble the ‘08 subprime mortgage market. When we turn to Web3, there is little substance that exists beyond self-referential activity but plenty of the same old nonsense—over $76 billion has been stolen since January 2021 within that ecosystem.
So if we return to reality, it’s obvious that compute infrastructure will substantially be scaled up with the help of fossil fuels—as we’ve been illustrating at length in this section. There is more momentum, more profit, and more political interest in doing this than not. Sure, peanuts will be thrown to the moonshot projects. They always are. But in the real world, on the ground, we will see projects like xAI’s Colossus which hopes to expand tenfold while already polluting South Memphis with unauthorized methane gas turbines.
This is the shape of what is to come. In the first section, we talked about the drive to overbuild compute infrastructure coming from firms that own it. In this section, we see another driver to overbuild compute infrastructure coming from fossil fuel extraction and energy production infrastructure that will power AI (or be powered by it). This dynamic can be extended more generally to other industries, but that’s something we’ll tackle in later iterations of the Silicon Valley Consensus thesis.
In the second half, I’ll talk about:
the role finance plays in overbuilding, overvaluing, and overinvesting AI infrastructure. This ranges from the monopolization of venture capital to the revenue models of data center landlords and the rise of asset manager data center portfolios.
the geopolitics driving the AI capex momentum, the ways close ties with the military-industrial complex would benefit speculators interested in overbuilding, overvaluing, and overinvesting AI infrastructure, and the reactionary political projects seeking to benefit from it.
the fragility of this drive to overbuild/overvalue/overinvest AI infrastructure and how many things have to go right to sustain the bubble thriving in the gap between reality (demand) and the predations of various sectors of our economy.
using SVC to map out frothy frontiers of Silicon Valley boosterism beyond AI capex: crypto, surveillance, defense tech, digitization, fintech, telemedicine, and more.