A $100 billion deal vanished over a weekend.
Not gradually, through quarters of renegotiation and carefully worded press releases. Just—gone. One week Jensen Huang is reportedly attached to a massive investment in OpenAI, the next he’s telling reporters in Taipei it’ll be “huge” but “nothing like” $100 billion. Nvidia’s stock drops 10%. Headlines multiply. Oracle, which has its own $300 billion cloud deal with OpenAI, feels compelled to publicly reassure everyone that it still expects OpenAI to honor its commitments.
This matters, but not for the reason you might think. Deals fall through. That’s business. What this punctures is something else entirely: the carefully maintained story about how solid all these AI infrastructure bets actually are. And when you start pulling on that thread, what unravels isn’t just one deal between two companies—it’s a set of uncomfortable questions about who’s building what, for whom, and who ends up carrying the risk when expectations meet reality.
The Circular Money Machine
Let me explain the mechanics without drowning you in jargon, because the structure here should make you uncomfortable.
The Nvidia-OpenAI arrangement, as reported, worked like this: Nvidia would invest huge sums in OpenAI. OpenAI would then spend much of that money buying Nvidia’s AI chips. Company A finances Company B, which buys from Company A. On paper, everyone’s revenue looks spectacular—Nvidia books chip sales, OpenAI secures compute capacity, both justify their valuations. In substance, you’re measuring the velocity of money spinning in a closed loop, not actual demand from end customers who need AI products and are willing to pay for them.
This isn’t unique. The pattern repeats across the AI infrastructure landscape:
OpenAI has signed compute deals worth over $1 trillion total—$300 billion with Oracle for cloud services, $38 billion with Amazon AWS, another $250 billion with Microsoft Azure. These are enormous, multi-year commitments from a company whose current annual revenue is in the tens of billions at most. The deals are being used to justify massive data center construction before the revenue to pay for them actually exists.
Nvidia invested about $2 billion in CoreWeave, a specialist AI cloud provider, and then committed to buy up to $6.3 billion of any unused CoreWeave capacity through 2032. So Nvidia finances an AI cloud operator that exists primarily to buy and rent out Nvidia hardware, and guarantees to purchase the capacity if no one else does.
Anthropic has bundled equity investments with cloud commitments—up to $8 billion from Amazon tied to using AWS as their primary infrastructure, over $3 billion from Google tied to Google Cloud services. They’re effectively pre-committing massive future spend to two hyperscalers who are also equity investors.
These aren’t conspiracies. They’re rational business moves by companies trying to secure supply chains, lock in customers, and signal to markets that AI demand is vast and growing. But here’s what is concerning: the risk and the hype are distributed asymmetrically. The upside—stock price appreciation, equity stakes, market dominance, competitive moats—concentrates at the top, among already-wealthy firms and investors. The downside spreads very differently.
Who Actually Needs What’s Being Built?
This is where I want to slow down and ask a question that gets sidelined in every breathless announcement about trillion-dollar AI deals: do communities actually need data centers that consume 500,000 gallons of water per day and draw power equivalent to small cities?
The research here is stark. Analysts project that by 2028, data centers could use between 6.7% and 12% of all U.S. electricity—roughly double or more from current levels—driven primarily by AI workloads. One Meta data center in Georgia uses about 500,000 gallons of water daily, representing 10% of its county’s entire consumption. Proposed new AI facilities in that same area could use up to 6 million gallons per day.
Let me be clear: I’m not arguing data centers are useless. They’re infrastructure for everything from hospitals to disaster response to the cloud services we all rely on. But the pace, scale, and terms of this AI-driven expansion are being set by a relatively small group of very wealthy entities optimizing for competitive positioning and financial returns, not by communities deciding how to allocate power, water, land, and public investment.
From a global market perspective, there’s growing demand for compute. Fine. From a local community perspective, you get a construction boom, some high-skill permanent jobs, property tax revenue—and also massive strain on electrical grids, water systems, noise, land use conflicts, and often higher utility rates for everyone else as infrastructure expands to serve a handful of giant customers. Counties negotiate these deals under NDAs, with limited time for public input, sometimes relaxing environmental standards to attract investment.
The AI companies and their investors are making decisions about energy, water, and land use that will shape communities for decades. Those decisions are being driven by trillion-dollar compute commitments and competitive pressure to build “AI factories” before someone else does—not democratic deliberation about what serves local needs. That asymmetry bothers me more than I can adequately express.
The Winners and Losers Are Already Clear
We don’t have to speculate about who benefits from these circular megadeals. The pattern is already visible:
Nvidia and a handful of chipmakers are riding enormous demand and have seen their market caps explode. Cloud providers are locking in multi-year revenue commitments from AI labs, which justifies their own massive capital expenditure—Meta, Amazon, Microsoft, Alphabet, and Oracle are projected to spend over $370 billion in capex in 2025, approaching $500 billion in 2026. Large investors and asset managers holding tech equities are seeing significant portfolio gains. Some high-skill workers—engineers, specialized trades like electricians and HVAC technicians for data centers—are benefiting from strong demand.
On the other side: households with limited capital, facing higher electricity and water rates to pay for grid expansions serving data centers. Workers in roles being actively automated—customer service, back-office operations, some professional services. Just this week, certain software stocks got hammered after Anthropic launched an AI tool that can automate document review and professional services work. That’s not abstract market dynamics. Those are people’s jobs and careers being repriced in real-time as investors conclude their work can be done by AI.
Communities being used as resource frontiers—rural or exurban areas with cheap land and power, targeted for massive AI facilities that strain water systems, increase pollution through backup diesel generators, lock land into single-use for decades, and deliver limited ongoing employment after construction ends.
And here’s the kicker: future public budgets. If these circular deals unwind and AI revenue doesn’t materialize as projected, governments that offered big tax breaks and financed grid expansions could be left with stranded assets, lower-than-promised tax revenue, and cleanup costs.
The distributional question—who gains wealth and security, who gets exposed to risk and displacement—is not driving how these deals are structured.
What Happens When Expectations Meet Reality?
I’m genuinely uncertain here, and I’m going to be honest about that rather than pretending I have clarity I don’t possess.
Goldman Sachs estimates current AI-related capex is around 0.8% of global GDP. That’s enormous in absolute terms but still below the peaks of past technology investment cycles—the late 1990s telecom boom hit over 1.5% of GDP. Their analysis suggests this could be a genuine “super-cycle” with strong fundamentals, not just a bubble, if AI demand and monetization continue to grow. Big tech companies have very strong balance sheets and can sustain heavy investment longer than previous generations of startups could.
Other analysts point to classic bubble dynamics: circular financing where suppliers fund customers who buy their products, speculation driving investment ahead of demonstrated cash flows, valuations based on expectations rather than current revenue. The comparison to the late-90s dotcom era, where companies inflated each other’s numbers through deals that didn’t rest on sustainable end-customer demand, isn’t frivolous.
The disappearance of the Nvidia-OpenAI deal doesn’t prove we’re in a bubble. But it does demonstrate that some of these headline numbers are softer than they appear. When Huang says the commitment was “non-binding” and “not finalised,” despite widespread reporting treating it as real, you have to wonder: how many other trillion-dollar compute commitments being announced are firm contracts versus aspirational signals?
Here’s what nags at me. When OpenAI announced over $1 trillion in compute commitments, when circular deals between chipmakers and AI labs were reported as massive demand signals, how many market participants—traditional investors, partner companies like Oracle, local governments negotiating data center deals—treated those numbers as bankable? And when the music stops, when growth slows or monetization disappoints, who ends up holding the stranded infrastructure and long-dated obligations?
Morgan Stanley, Morningstar, and others have warned that vendor financing and circular revenue can overstate real demand. If AI adoption or revenue disappoints, the repayments and “take-or-pay” commitments could strain cash flows. Equity investors lose first—that’s wealthy individuals and institutions. But second-round effects land on workers through layoffs, communities through stalled projects and lower tax revenue, and consumers through higher utility bills as providers try to recover sunk infrastructure costs.
Forrester’s analyst nailed it: “The markets can stay irrational longer than you can stay solvent.” In this context, that means some firms—and some communities—may run out of resources or options before the AI market settles into more rational economics.
What This Means Going Forward
I don’t think we’re watching a bubble pop. I think we’re watching a wobble—a moment where the gap between hype and binding commitment becomes visible, and the market starts to distinguish more sharply between companies that benefit from AI and companies whose revenue streams are threatened by AI automation.
What I expect to see:
More scrutiny of AI infrastructure deals. Investors and partners will examine whether commitments are actually binding, how they map to realistic usage and revenue, and who carries the risk if projections don’t materialize. The easy money phase, where “AI” in a press release was enough to justify valuations, is ending.
Vendor diversification. OpenAI and other AI labs will keep pushing for alternatives to Nvidia—not because Nvidia’s chips aren’t good, but for leverage on price and supply security. Nvidia’s dominance will persist but could erode at the margins as customers refuse to be entirely dependent on one supplier.
Clearer separation of winners and losers. At the top of the stack—model providers with differentiated capabilities, infrastructure suppliers with genuine technical moats, cloud platforms with locked-in customers—the benefits will continue to concentrate. Lower down, software and service companies whose products can be easily automated are increasingly at risk. That sell-off in software stocks this week is a preview.
Less sci-fi, more cash flow. AI companies are under real pressure to move from grand narratives about AGI and superintelligence to concrete, profitable products that justify their infrastructure costs. OpenAI’s reported shift toward mundane monetization—ads, adult content, whatever generates actual revenue—fits this pattern.
And here’s what I think communities and individuals can actually do, because just describing the problem without offering paths forward feels inadequate:
If your community is facing a data center proposal, demand binding commitments on renewable energy sourcing, not just carbon offsets purchased elsewhere. Require transparent, ongoing reporting on water use, grid impact, and local hiring. Make tax incentives conditional on delivering promised community benefits, with clawback provisions if they don’t materialize. These facilities will shape your infrastructure for decades—negotiate like it.
If you’re an investor or business leader, start distinguishing between companies creating genuine customer value and those surfing financial engineering. Ask harder questions about how much revenue comes from real end-customers versus closed-loop deals with vendors and partners. Push for disclosure about circular financing structures and vendor commitments.
For everyone, recognize that decisions about AI infrastructure are being made right now that will determine energy access, water availability, job markets, and community development for a generation. Those decisions shouldn’t default to whatever maximizes quarterly stock prices for a handful of firms. We can ask for different terms. We can demand transparency about who benefits and who carries risk. We can insist that public resources—power, water, land, tax revenue—are allocated through processes that actually serve public priorities.
A Closing Thought I Can’t Shake
I started by saying a $100 billion deal vanished over a weekend. What really vanished was a number—a signal, a story element in the larger narrative about limitless AI demand. The actual chips still exist, the data centers are still being built, the workers are still being hired or displaced, the communities are still negotiating terms on infrastructure they may or may not need.
But here’s what that evaporation reveals: we don’t actually know how much of the AI boom rests on firm ground versus circular hype. We don’t know which trillion-dollar commitments are binding contracts and which are aspirational marketing. We don’t know if the revenue growth will materialize fast enough to justify the infrastructure being built on speculation.
What we do know is that the gains are concentrating among people and firms that were already winning, while the risks—higher bills, job displacement, environmental strain, potential stranded investments—are spreading more widely. That’s not a judgment about whether AI is valuable or whether this is a bubble. It’s an observation about power and who gets to make decisions that shape the future we’ll all inhabit.
The markets can stay irrational longer than you can stay solvent. So can the infrastructure build-out, the circular deals, the trillion-dollar commitments built on expectations rather than demonstrated revenue. The question isn’t whether that irrationality will eventually end—it will. The question is who’s positioned to absorb the correction when it comes, and who gets crushed by it.
I don’t have a tidy answer. I have concerns, questions I think deserve harder scrutiny than they’re getting, and a deep skepticism of anyone—on either the hype or cynicism side—who claims certainty about how this plays out. What I know for sure is that when firms make trillion-dollar bets using other people’s power, water, land, and risk exposure, all of us who live in the places being reshaped deserve a seat at the table where those bets are being placed.
The music is still playing. But it’s worth asking, before it stops, who actually needs what’s being built—and who ends up paying when the numbers don’t add up.