Your Favorite Tech Giant Is Quietly Turning Into a Utility
High capex crushes market valuation
Welcome to the latest edition of Buy the Rumor; Sell the News. In today’s post, I take a look at the hyperscalers’ capex craze, and what it means for their market caps. Look out below!
If you like what you read here, consider subscribing, and if you’re already subscribed, consider upgrading to a paid subscription. If you want to connect with me directly, my contact information is at the end of this post.
A few weeks ago, Stripe’s John Collison put Meta’s CFO Susan Li on the spot. He asked whether the soaring capital expenditures at Microsoft, Google, and Meta were ultimately bottlenecked by power, and whether they could keep growing at this rate.
She didn’t duck. She admitted these are massive bets, hard to repurpose if demand stumbles, and entirely dependent on breakthroughs in distributed training and energy efficiency.
That should have been a bigger headline.
Because it reveals how these companies, once poster children for asset-light software models, are morphing into something closer to regulated utilities. And it’s happening fast.
Jevons in a hoodie
Optimists like to point out that model performance roughly scales with the logarithm of compute. You’ll hear:
We can get 90% of GPT-4 performance with just 30% of the compute.
Sparse training and clever distillation will keep costs under control.
And they’re right, locally. But they’re missing the systemic truth.
Enter Jevons paradox. In the 1800s, steam engines became dramatically more efficient. Did that cut coal demand? No. It exploded. Lower marginal costs unlocked thousands of new applications, and aggregate consumption soared.
Same with GPUs. As training gets cheaper per token, we don’t save money. We train ten times more models and serve a hundred times more inference calls. The total compute, and therefore power demand, still goes vertical.
Yes, Jevons isn’t some immutable law. In consumer mobile, for instance, more efficient chips have actually reduced total power draw because you’re constrained by battery size and how often you’ll recharge. But AI isn’t bounded by your willingness to plug in. It’s bounded only by new economic use cases. Which means the paradox probably holds.
The CapEx cliff
This is why Big Tech is pivoting from pure software to heavy industry.
Just look at their FY25 guidance. Microsoft plans over $55 billion in capital expenditures. Google, around $47 billion. Meta, roughly $40 billion.
That puts them on par with, or even above, top-tier regulated utilities. Duke Energy, for comparison, plans about $65 billion over a five-year horizon. The scale is no longer that different.
And it’s not just servers. The real kicker is power. Hyperscalers are locking in 10- to 20-year contracts for renewables, peaker plants, even microgrids. They have to. AI power demand is doubling roughly every 18 months. You can’t run trillion-parameter models on goodwill and ESG decks.
The new balance sheet reality
Most people assume these power deals stay off the books. Like the old days of operating leases.
But under today’s accounting rules (ASC 842 in the US, IFRS 16 globally), long-term dedicated-use contracts often get treated like owned assets. That means new right-of-use assets on the balance sheet with matching lease liabilities.
Sure, some contracts slip by if the hyperscaler doesn’t control how a plant operates. But the trend is clear: regulators want these obligations front and center.
So even if power is only 15-20% of lifetime data center costs, locking in decades of exclusive capacity transforms these companies. They start to resemble utilities, just with flashier logos and keynote presentations.
Geopolitics piles on
Now layer on national security and sovereignty.
The U.S. doesn’t have to seize data centers at gunpoint. Export controls on GPUs and foundation models, CFIUS reviews, and Commerce Department rules already slow global optimization. Meanwhile, Europe, India, and countless others insist on local data clusters for privacy and latency reasons.
All this forces duplicative infrastructure. If today’s redundancy factor is maybe 1.3 (meaning 30% more capacity than strict demand requires), tomorrow it could be 1.6 with barely any incremental revenue.
This compounds the Jevons effect. It locks hyperscalers into a capital treadmill where the only way out is even more buildout.
Why it means valuation compression
For decades, markets valued Microsoft, Google, Meta like software firms.
Low fixed costs. Near-infinite scalability. Minimal capital intensity.
Now? They’re drifting toward industrial infrastructure profiles. Huge upfront investments, long payback periods, enormous fixed costs rolling into uncertain demand cycles.
Yes, today Microsoft still posts around 24% ROIC. That’s extraordinary. They have unmatched pricing power.
But as permanent CapEx piles up, the margin structure inevitably shifts. Depreciation and financing start eating into free cash flow. Meanwhile, every 100 basis point rise in WACC cuts terminal value by roughly 15-20%.
So even if top-line revenues keep climbing, multiples compress. Maybe not overnight. As Keynes (perhaps apocryphally) warned, markets can stay irrational longer than you can stay solvent. But the structural economics always catch up.
This bleeds into startups and VC
Late-stage venture valuations don’t live in a vacuum. They’re tethered, explicitly or implicitly, to public comps.
If Microsoft and Meta start clearing 8-12x EBITDA instead of 25-30x, the whole late-stage multiple stack reprices in lockstep. It doesn’t matter if your seed deck says AI fifty times.
When the foundational businesses get treated like utilities, the exit math for everything built on top gets crushed.
Index funds won’t bail you out
Some argue index funds hold too much Big Tech to let prices really fall.
It’s true: S&P 500, Nasdaq 100, global ETFs, pensions, sovereign wealth: all mechanically buy these names.
But passive holders don’t set prices. The marginal active investor does. Passive flows add volume, reduce volatility, soften immediate drops. But the valuation multiple itself gets set by discretionary capital: hedge funds, long-only growth, crossover funds shifting between tech, infrastructure, and staples.
So even if 40% of Meta is held in index funds, if active investors decide it now looks like a utility, they’ll pay 12x EBITDA, not 30x. The passive funds will just keep holding at a lower mark.
In fact, passive indexing absorbs price moves. It doesn’t resist them. So you’re more likely to see a slow erosion of multiples, as allocators quietly rebalance into private infra or hard assets to chase returns.
What could change this?
There are wildcards.
Maybe cheap small modular reactors (SMRs) slash energy costs.
Maybe smarter, distributed, quantized models drastically cut training loads.
Maybe sovereign AI clusters protect hyperscaler margins by breaking up workloads.
But each of these only dampens the shift. They don’t reverse it. And the heavy fixed investments are already happening, locked into contracts and depreciation schedules. Efficiency gains will mostly just fuel more demand. The Jevons engine runs on.
So what should you do?
Allocators: Stop underwriting Big Tech like it’s 2015. Stress-test your models with infra-like cash flow yields, explicit WACC sensitivity, and geopolitical redundancy costs.
VCs: The days of paying nosebleed forward-revenue multiples on faith in software-style exits are ending. Get sharper on capital structure exposure, customer concentration, and hyperscaler dependence. Many AI wrappers simply won’t clear the cost-of-capital bar.
Founders: Your moat isn’t just clever UI on someone else’s foundation. It might be compliance, local market capture, or your own integrated energy footprint. Your cost structure increasingly inherits your upstream providers’ capital intensity.
Bottom line
If you’re still underwriting tech like it’s 2015, where you’re hoping for asset-light hypergrowth and effortless multiple expansion, you’re setting yourself up to be crushed.
Better to see it clearly now. There’s still time to reposition. But not much.
Coda
If you enjoy this newsletter, consider sharing it with a colleague.
Most posts are public. Some are paywalled.
I’m always happy to receive comments, questions, and pushback. If you want to connect with me directly, you can:
But unlike common energy utilities, there are many fewer of these behemoths. This concentration gives them extraordinary market pricing power and rent capture ability. This can sustain the high multiples for the winners.
Another excellent post - thank you