Are the Hyperscalers Spending Too Much on AI, or Not Nearly Enough?

The Numbers Are Staggering

Microsoft Google Amazon and Meta have collectively forecast more than $300 billion in combined capital spending between 2025 and 2026. Say that number out loud once. Three hundred billion dollars from four companies were spent primarily on data centers chips and the electricity to run them

The downstream effects are manifesting far beyond the tech sector itself. NVIDIA's data center revenues now exceed $35 billion quarterly. Energy infrastructure is scarce enough for hyperscalers to directly negotiate supply with nuclear plants. Data center construction is reshaping real estate markets in Northern Virginia Phoenix and Dallas places that used to be known for entirely different industries

Then in late June 2026 Alphabet shares fell 10% in a single session due to concerns about the departure of AI talent. Amazon Oracle and Meta fell 2% to 4% that same day due to broader capital spending concerns

The Bull Case: Every GPU Is Already Working

The strongest argument for spending is simple: It looks like demand-driven capacity not speculative capacity. When Cisco and Nortel built fiber optic networks in 1999 and 2000 they were laying cables years ahead of any proven demand for that amount of bandwidth. Building AI looks different at that specific point. Every GPU that a hyperscaler stacks up starts generating revenue almost immediately running inference industry shorthand for using a model already trained to answer a real question for ChatGPT Gemini Copilot or an advertising algorithm that decides what to show you next

Usage numbers back this up. Microsoft's Copilot products have more than 300 million users. Google's AI overviews now appear in the majority of search queries in the US. Meta's AI-powered recommendation system drives more than half the time people spend on the platform. That's not projected future demand. That's today's product running today on chips purchased this year

There is a second piece to the bull argument and it has less to do with revenue than with fear: competitive necessity. If you fall behind in AI infrastructure not only will you lose advantages but you will begin to lose share in the cloud in advertising in enterprise software compared to those who did not blink. The investment bill is enormous. The bull case says that the bill for not paying it would be worse

The Bear Case: The Math Doesn't Close at These Prices

The bearish case does not question whether GPUs are being used. It questions whether their use actually allows them to pay for them. You cannot run a long-lasting business spending $75 billion a year on infrastructure to offer a product that most customers use for free or at a subsidized price well below its true cost

And the price is falling rapidly. The cost of running a single query through a leading AI model has dropped about 90% over the past two years as models become more efficient and competition drives down prices. If unit economics are deflating at that rate much more volume significantly higher prices or both are needed simply to keep revenue growing at the pace of the infrastructure bill. None of that is guaranteed

Then there's the Alphabet talent story from late June. It's a different kind of bearish signal one that has nothing to do with capital spending. If the people who make an AI investment valuable start leaving the capital stored in the data center doesn't depreciate any faster but the case for having spent it becomes shakier. I think the market's reaction was fair. Infrastructure without the talent to make it valuable is a cost center wearing the clothes of a growth story

Training Versus Inference: Why the Mix Matters

Not all AI calculations are from the same animal and the distinction is important when reading this entire debate. Training A frontier model requires massive concentrated bursts of computing: thousands of NVIDIA H100 GPUs running for weeks or months at a time all targeting a single job. Inference running that trained model to actually answer a user's question needs a completely different smaller consistent and scalable form of computing no matter how many people are using the product at the moment

Those two workloads entail different economics and different competitive dynamics which is where custom silicon comes into play.Google's TPU chips Amazon's Trainium and Microsoft's Maia are increasingly competitive for inference workloads because once you know the exact architecture of the model you're running you can build a chip optimized specifically for that job. NVIDIA remains dominant in the training space largely because the CUDA software ecosystem is based on NVIDIA hardware and the switching costs are real

This is the part that I think is undervalued in most of the coverage: As the overall AI computing mix moves from training-heavy to inference-heavy which is the natural direction as products mature and user bases grow the hyperscalers' own custom silicon bets become relatively more valuable than NVIDIA's. That's a second capex bet within the first a bet that owning its own inference chips is better.than renting NVIDIA ones at scale for years

The Real Decision: Overbuild or Underbuild

Let's put aside the bull and bear cases for a second and ask what a hyperscaler's financial team really looks at when approving an investment budget. It's not a forecast. There are two ways to be wrong and both ways don't cost the same

If you build insufficiently demand shows up anyway and you turn away paying customers or worse they sign up with a competitor that has spare capacity and never come back. In a market that grows this fast that's not a quarter of a mistake. Enterprise AI and cloud customers are notoriously sticky once they've built a platform so losing one to a rival's spare capacity can be a decade-long loss disguised as a single lost sale

Overbuilding and demand don't show up on time and you're sitting on idle GPUs burning energy and losing value with nothing to show for it. That's a real cost. However it's usually not a business end for a business with the balance sheet of Amazon or Microsoft because some of what was built - the power contracts the land the cooling infrastructure the fiber connecting the site - has value even when the specific chip generation inside the building doesn't

That asymmetry a painful but survivable mistake on the one hand versus a potentially permanent and aggravating loss of market position on the other is why I think every hyperscaler CFO is inclined to spend more than a simple discounted cash flow model would recommend.measuring risk

A Worked Example: What Return Does This Capex Need to Clear?

Here's one way to make the abstract debate concrete using illustrative numbers not a single company's actual numbers. Let's say a hyperscaler commits $50 billion to a set of AI infrastructure: GPUs servers data center structures power contracts. Calculate the useful life of that bucket five years before it needs a major upgrade and call cost of capital what investors expect to earn given the risk of the business 10%

The question raised by a return on invested capital perspective is: what does that $50 billion need to generate every year for five years just to cover its cost of capital before counting anything above that figure as profit?

The math uses a standard annuity formula. annuity factor for a 10% discount rate for five years is one minus 1.1 to the negative fifth power divided by 0.10.1.1 to the fifth power is 1.61051. One divided by that is about 0.6209. One minus 0.6209 is 0.3791. Divide that by 0.10 and the annuity factor comes out to about 3.79

Divide the $50 billion of capital spending by that factor of 3.79 and you get the level of annual pre-tax cash flow this investment needs to produce each year for five years to clear exactly the 10% hurdle: about $13.2 billion a year

But cash flow is not revenue. Let's say this infrastructure runs at an illustrative 40% incremental operating margin a reasonable substitute for a mature AI or cloud services business rather than a subsidized product that is given away for free. Divide the $13.2 billion of required cash flow by that 40% margin and the required annual revenue works out to about $33 billion each year for five years before the investmenthas earned a penny above its cost of capital

Sit with that number next to the actual guidance numbers cited above. A company that spends $50 billion to $80 billion on AI infrastructure in a single year needs tens of billions of dollars of durable healthy-margin annual revenue tied to that specific crop of spending indefinitely just to break even on a risk-adjusted basis. That's the bar the bull case must clear. It's not impossible justMicrosoft's Copilot has 300 million users but it's a real verifiable bar and it's much higher than simply pointing out that GPUs are being used

An honest warning: This model is a simplification. It ignores taxes whatever the value of the infrastructure is at the end of five years the fact that real capital spending comes in continuously and not as a lump sum and the reality that margins rarely stay constant as a product grows. Treat it as a way of thinking about order of magnitude not as a valuation model that you would hand in to get a rating

Case Study: The Fiber Glut That Wasn't Wasted

If you want a precedent for the idea that we might be overbuilding and that capacity might still matter later the clearest one is the deployment of fiber optics in the late 1990s. Telecommunications companies including Global Crossing and WorldCom spent huge sums of money laying fiber optic cables across the country and under the oceans betting that Internet traffic would grow fast enough to eventually fill them. Much of that capacity was dark fiber already buried cable that wasn't even lit up for use because there was simply more than the Internet needed in 1999

The bet for the moment was wrong very wrong. When the dot-com bubble burst and supplier financing dried up companies that had gone into deep debt to build that capacity were wiped out. Global Crossing filed for bankruptcy in 2002. WorldCom collapsed the same year in what was then the largest bankruptcy in U.S. history a collapse entangled with accounting fraud that had been covering up how much expenses had exceeded revenue. Those companies' shareholders virtually lost it.everything.Nortel the equipment manufacturer that had financed much of that construction through supplier loans also ultimately failed

Here's the twist that makes this worth counting as more than just a warning: Fiber itself went nowhere. Over the next decade and beyond as adoption of broadband then video streaming and then cloud computing grew that dark fiber was lit up and put to work. Part of the same physical cable that bankrupted its original owners in 2002 ended up carrying Netflix traffic then cloud workloads and quite possiblytoday some of the data traffic behind the very construction of AI that this article is about. The capacity was overbuilt relative to the demand in 1999 and was still needed just on a longer term than the people who funded it could survive

That's the uncomfortable lesson for anyone trying to label the hyperscaler capex debate as simply right or wrong. Overbuilt and eventually essential are not opposites. Both can be true about the exact same asset measured over different time horizons. People who get burned aren't necessarily wrong about the long term. They're wrong about how long they can wait and who else is willing to keep funding the wait

Where the Bull Case Breaks

I've laid out reasons to think this capex cycle is different than 1999 and I believe most of them. The honest version of this article has to say where that comparison stops being valid because it stops being valid in at least one important place

Fiber optic cable once buried barely degrades. Glass buried in 1999 still shines today and is physically almost as good as new. A GPU is nothing like that. AI chips improve fast enough that a data center full of cutting-edge hardware can go from cutting-edge to energy inefficient and commercially uncompetitive in a few years well beyond the lifespan assumed by a payback calculation like the above. If demand for AI computing eventually reachescurrent growth in the same way that internet traffic reached fiber in 1999 there is a real risk that it will reach a newer generation of chips instead of the ones that are idle today. Capacity does not wait patiently like dark fiber did. It ages

There's also a financing gap and it cuts the other way. Much of the telecom development in 1999 was debt-financed much of it supplier financing from equipment makers desperate to book sales exactly the kind of leverage that turns a bad forecast into a bankruptcy filing. Today's largest hyperscalers are financing much of this capital spending with their own operating cash flow and investment-grade balance sheets meaning a demand shortfall manifests itself asPressure on margins and stock price rather than necessarily a bankruptcy filing. That makes a Global Crossing-style wipeout less likely this time around. It doesn't make the spending right. It just changes who absorbs the error and how visibly

And there's a simpler failure mode that doesn't need any historical analogy: Efficiency could continue to improve faster than usage grows. If the cost of running a query continues to fall near the roughly 90% two-year pace already underway and usage grows more slowly than the falling cost curve it's possible to have real growing profitable usage of AI and still not enough revenue growth to justify the size of the infrastructure bill. That scenario doesn't require anyone to be wrong about theimportance of AI. It only requires the economics to serve that importance to remain cheaper faster than the market grows

How I Actually Think About This

My read and I want to be clear that this is a read not a prediction is that the headline capex numbers are almost useless on their own. Three hundred billion dollars means nothing without knowing what you're buying and how quickly it's being used. So the way I would really follow this story is not through the capex guidance headlines at all

You'd look at revenue per compute unit over time even a rough version because that's the metric that tells you whether utilization is catching up or falling behind. You'd look at the custom silicon numbers TPU Trainium Maia as a proportion of total compute because a hyperscaler that shifts more of its own workloads to its own chips is a company voting with its engineering resources that inference economics is important enough to fight for. AndI would look at stock reactions to hard data such as Alphabet's late June crash less as verdicts on the entire thesis and more as evidence of how nervous the market already is. A 10% move in a talent story not even a loss of revenue indicates that the market is prepared to punish this group with little evidence. This is useful information about sentiment separate from what the underlying economy eventually turns out to be

I admit that I go back and forth on which side of this I really believe. The bullish argument that this is demand-driven and that the usage numbers are real is the strongest argument today. But "today" is working hard on that phrase because the example above shows that the bar these companies have to clear is really high and I don't think it's been cleared yet as much as it's been guided toward. If I had to pick the most important number to watch over the next year it wouldn't be the capital spending guidance.next quarter. It would be whether revenue per computing dollar increases or decreases. Everything else in this debate is behind that line

The Bottom Line

Microsoft Alphabet Amazon and Meta are spending more than $300 billion combined on AI infrastructure between 2025 and 2026 and the market has begun to openly question whether the spending curve is outpacing the revenue curve. The bull case says that each GPU generates revenue from day one unlike the speculative fiber developments of the late 1990s. The bear case says that unit economics of inference are collapsing.deflating too quickly for the expense to pay for itself at these prices. Both are defensible today and that is exactly why the question in the title of this article still has no clear answer

The decision itself is not really a forecasting problem it is a problem of asymmetry of results: the substructure risks a lasting loss of market share while excessive construction risks a costly but viable mismatch for such strong balance sheets. The worked example puts a real figure on how high the revenue bar is approximately $33 billion a year on an illustrative investment of $50 billion at a 10% cost of capital one more verifiable hurdle thanvague. The expansion of fiber optics in the late 1990s shows that overbuilding and eventually overbuilding can be both for the same asset just at different clocks although the faster obsolescence of chips versus glass wire is a real reason why this cycle might not resolve in the same way. My own approach is to stop looking at the capex headlines and start looking at revenue per computing dollar because that's the number that will eventually tell the whole story that the guidance figuresthey can't