Big Tech's AI Spending Is Finally Deploying Its Idle Cash
For the last year, the dominant narrative has been simple: AI is a bubble. Too much hype, too much capital, too many promises, and eventually it pops.
The warnings are everywhere. Nobel laureate Daron Acemoglu calls AI models “hyped up” and says “we’re investing more than we should.” Sam Altman admits investors are “overexcited” about AI. One analyst called it “the biggest and most dangerous bubble the world has ever seen.”
I don’t buy that framing, or at least not the version that predicts a 2008-style collapse. Big tech is finally deploying tens of billions of dollars in capital that’s been sitting idle for years, because until now there was nowhere sensible to put it. AI didn’t create the money. It gave it a destination.
The cash problem
If you’ve never worked inside big tech, it’s hard to grasp the scale of the cash problem. I mean cash, not revenue or valuation. I’ve been at both Meta and AWS. At neither company did I ever feel like there was a lack of capital. You hear about it at All Hands, and there’s documentation around it. When leadership talks about spending, it’s a source of pride: look how much capital we’re able to deploy to this cause. Employees would like more justification for it, honestly, but that information lives at the executive level and it doesn’t come down.
Between roughly 2015 and 2022, companies like Meta, Google, Apple, and Microsoft accumulated enormous cash reserves. By the end of 2021, Big Tech held a combined $476.7 billion in cash, cash equivalents, and marketable securities. Alphabet consistently sat north of $100B. Apple peaked near $167B before aggressive buybacks. Microsoft hovered in the $80-130B range. Meta held $50-60B. Amazon reached $89B by Q2 2024.
The size matters less than the fact that there were very few places to deploy that cash at scale without destroying returns or triggering regulatory scrutiny. You can’t buy meaningful companies without antitrust issues. You can’t launch moonshots at $50B scale without burning shareholder trust. You can’t dump tens of billions into financial markets without distorting them. So the cash accumulated, slowly, relentlessly. By the late 2010s, the constraint on big tech was deployability, not capital.
I’ve lived this one. Inside these companies the tension is between CapEx and OpEx. I’m on a high CapEx team at AWS. We could get budget for infrastructure tomorrow, but if I could trade a million dollars of our CapEx budget for headcount, I would do it in a heartbeat, and we can’t. The money is there, but it’s locked into infrastructure spending rather than people. That’s the system working as designed. Infrastructure is a depreciating asset companies can write off. Headcount is an ongoing operational expense that hits margins differently.
That’s why AI infrastructure spending feels so natural for these companies. The money was already earmarked for infrastructure. AI just gave it a more compelling label and a clearer return profile. AWS gets returns on a server rack quickly, and it’s almost always profitable to build data centers or add new buildings to existing ones because the demand is already there. Internally, we’re pushed to avoid certain regions because they’re already in a capacity crunch. I’ve seen this before: at Meta during the pandemic I was on the infrastructure teams kicking projects off of production and demoting them because we didn’t have enough capacity. Monthly active users jumped from 1.2 billion in 2020 to 2 billion in 2021. That increase in traffic is insane, and the demand has always been real.
Why this isn’t 2008
The 2008 bubble was built on leverage, fragility, and financial engineering. In 2008, capital was borrowed, risk was hidden, and the system collapsed when liquidity disappeared. This moment is built on cash already on hand. When Meta spends $65-72B on capex in 2025, it isn’t borrowing against the future. It’s spending money that already exists on its balance sheet. The combined numbers are staggering:
That’s over $320 billion in a single year, deployed into data centers, power infrastructure, and compute capacity. And people miss the timeline: when Amazon announces $200B in 2026 capital expenditure, that order was placed in 2024. Procurement cycles for server hardware go out months or years in advance. That’s a supply chain that moves in years. Hype spending doesn’t work that way.
Even skeptics acknowledge the structural difference. IMF Chief Economist Pierre-Olivier Gourinchas stated: “An AI investment boom may lead to bust, but not likely systemic crisis… Unlike the 2000 dot-com bust or 2008 financial crisis, AI investments are generally not debt-financed.” The money is already there.
Amazon and the Alexa fumble
There’s a narrative that AWS and Amazon were slow movers on AI. I think that’s true, but not for the reasons everyone assumes. Amazon built Alexa and sold over 500 million units. They had an AI-ready piece of hardware in people’s houses before anyone else, and fumbled it because they didn’t turn it agentic, didn’t invest in LLMs early enough, and didn’t see the interface shift coming. The hardware was there. The software vision wasn’t.
But calling Amazon a slow mover ignores the infrastructure reality: AWS has the most data centers and the most market share. Its year-over-year growth is lower than other companies’, which is what you’d expect at its size. AWS and Meta know how to build data centers; it’s part of the core business. The slow part is permits and the hoops you go through to break ground. The capex is an infrastructure play on a depreciating asset that people will need regardless. Whether we keep building more efficient models and hardware to justify the capacity is the open question, and so far the demand curve says yes.
AI didn’t arrive as a miracle product. It arrived as an infrastructure excuse. Training clusters, inference fleets, power contracts, data centers, networking, cooling, specialized hardware, land: boring, real, industrial investments that absorb enormous capital without promising overnight returns, which is exactly what big tech needed. Multi-year data center builds, power generation and grid commitments, long-term hardware supply chains, fiber and interconnects, custom silicon programs: these are industrial-scale projects with decade-long horizons. A bubble implies fragile demand and speculative excess. What we’re seeing is closer to a re-industrialization of the digital economy, funded by cash that’s been sitting on balance sheets for years, waiting for somewhere to go.
Where I could be wrong
I should be honest about where this framing is weakest. If the productivity gains from AI infrastructure don’t materialize, big tech is left with expensive data centers that don’t generate sufficient returns. The spending is real, but the value capture isn’t guaranteed. Even the skeptical scenarios suggest compute demand keeps growing, though, so the question is rate of return rather than total write-off.
And yes, every speculative excess looks like rational investment to someone at the time. The difference I’d point to is that this capital is owned rather than borrowed, the spending is on physical infrastructure with residual value, and the companies deploying it are profitable on their core businesses regardless of AI outcomes. That argument applies to big tech with fortress balance sheets, not to the AI startup ecosystem, where plenty of companies are burning cash on the assumption that scaling laws continue indefinitely. Some of those bets will fail.
On the API economics: if you do the math on AI subscriptions, the compute costs mean they don’t make money off the subscriptions themselves. The real play is loyalty to a model ecosystem, with the actual money in API calls at scale. That’s a bet on adoption curves, not current profitability.
None of this makes AI a sure thing. It means the structure of this moment is different from 2008, and that matters for how we think about systemic risk.
For founders, the companies spending $320B need software, services, and tooling, and that’s a procurement cycle you can sell into. But build products, not features. Steve Jobs famously told Drew Houston that Dropbox was a feature, not a product. Before you build, decide which one you have: something that gets rolled into a bigger company’s platform, or a standalone product that can live on its own. Google Drive and Dropbox are both still around, and they’re very different businesses because of that choice.
For engineers, the layoffs are real. I’ve been through three rounds at Amazon. From what I’ve seen, AI isn’t the cause. They’re timed to earnings reports, and if you watch the pattern, each round waits for the previous severance period to end before announcing the next. There are plenty of AI initiatives inside these companies and teams being built around them, but they aren’t replacing engineers with AI. The message from leadership is “ruthless prioritization,” which gets old when you’ve just laid off 15,000 people and then tell the rest to be more efficient.
For investors, watch capex efficiency, not hype cycles. The question I’d ask is what the return on this infrastructure looks like once it’s fully built out. The winners might be the picks-and-shovels players rather than the model makers.
If you’re waiting for a clean pop, a moment where AI collapses under its own hype, you’re probably waiting for the wrong thing. The capital is real, it’s owned, and it’s being deployed into physical infrastructure that has value regardless of whether the current generation of AI models lives up to the hype. The conversation I keep hearing, that AI is going to take our jobs, is aimed at the wrong target. I think AI is a tool that shifts what work looks like, the same way every major technology has before it. The bigger disruption is happening in warehouses, on highways, and in logistics networks: physical automation at industrial scale. Almost nobody is talking about it.
That’s Part II.