I grew up in telecom and I have seen this movie before. Boom and bust.
While I think we all recognize that AI is different - we don’t have any evidence that it is immune to market cycles. It would be foolish to think otherwise.
Yes, it is bigger on every level. It is moving faster. It is pulling more capital, more power, more land, more fiber, more cooling, and more political attention than anything before it. All of that is true.
But we are already seeing shifts in the landscape. The evidence is everywhere. Very few are willing to say it out loud.
The first is that the tide is going out. I know it is hard to get your head around that concept, but it is and I will explain below.
The second is that we don’t know how far it will go out and therefore what the implications are. What we do know is that it will reshape things - likely for the better.
As John F Kennedy once said, “A rising tide lifts all ships.” Warren Buffet is just as famous for his counter argument, “Only when the tide goes out do you discover who's been swimming naked.”
For the past few years, the rising tide has lifted a lot of ships. Capital was available. GPUs were scarce but ultimately attainable as evidenced by rising and falling term pricing. Capacity could be leased forward. If you could get access to supply and stand something up quickly, you could participate. Speed mattered more than structure. Growth covered inefficiency. Narrative covered risk.
That phase is ending.
The tide is going out because the constraints are now binding.
Power is constrained. Cooling is constrained. Build timelines are measured in years, not quarters. Liquid cooling alone is running $1.5 to $1.6 million per megawatt and taking 6 to 9 months to deploy.
At the same time, the commercial terms have hardened. Large deployments now require 10 to 15 year commitments. Pricing has moved into the $150 to $165 per kW range. Investment-grade credit, or something very close to it, is now the price of admission. Some operators have capped exposure to neo-clouds. Others have stopped taking the deals entirely.
The market is also far more crowded than the narrative suggests. JLL estimates roughly 190 neo-cloud and adjacent operators are now competing for capacity. Most of them do not control power, do not control land, and do not have the balance sheet to carry long-duration commitments.
Capital is doing the same thing. Blue Owl, one of the largest and most sophisticated pools of capital in this space, walked away from a $10 billion data center financing based on terms and conditions - not expected demand.
At the top end, the scale of capital required is already clear. CoreWeave, for example, is carrying tens of billions in contracted backlog alongside tens of billions in debt and lease obligations. Even the leaders are capital intensive. The difference in CoreWeave’s case is they have shown that they can fund it. Most cannot.
That is what we mean when we say the tide has turned. It is not collapsing, far from it. Rather it is heightened selectivity - to such an extent that most of the players cannot participate. This is a 10% problem. Using 190 players - that means ~19 of them go forward. Start counting - you will get to the same number.
Selectivity will thin the herd.
Selectivity means you don’t lease capacity at high cost or finance GPUs at high cost or try to bridge the gap with growth.
That model only works if three things are true at the same time.
- High utilization.
- Cheap or available capital.
- Stable or rising pricing.
Most of the bottom tier fails this test even before stress. At steady-state utilization, many are already operating with thin or negative margins once power, lease costs, and financing are fully accounted for.
Take one away and margins compress. Take two away and the model breaks.
Run the math.
Assume a mid-tier operator leases 5MW at $155 per kW. That is roughly $775,000 per month, or about $9.3 million per year in base infrastructure cost before power. Add electricity, which is now the dominant operating expense. Add debt service on GPUs. Add the upfront liquid cooling capital, which in many cases is pushed onto the customer. Add operating overhead.
This is a high fixed cost business.
Now assume they need 80% utilization to make the model work. That is not aggressive. That is table stakes.
At 80%, they cover fixed costs and generate some margin.
At 70%, that margin compresses sharply.
At 60%, they are not covering fixed obligations - the lease, the debt, the infrastructure.
Nothing about those costs flexes down with demand. The lease is fixed. The debt is fixed. The power reservation is fixed. The cooling infrastructure is already paid for.
This is negative operating leverage.
To make it explicit, that $9.3 million annual lease is spread over 4MW at 80% utilization. At 60%, it is spread over 3MW. The effective cost per sold kilowatt jumps by roughly a third without any change in underlying expense. The revenue base shrinks. The cost base does not.
There is no margin of safety here. A 10 to 20 point drop in utilization does not reduce profitability. It eliminates it. For a large portion of this market, that shift moves the business from marginally viable to structurally insolvent.
Now layer in the real market conditions.
- Customers are taking longer to commit.
- Colocation providers are requiring proof of end demand.
- Hyperscaler allocations shift the entire pipeline overnight.
- Capital is more selective.
- Terms are longer.
- Rates are higher.
This is not a theoretical stress case. This is the current environment.
So what happens to operator number 47 when utilization slips from 80% to 65% and their next capacity expansion cannot be financed on the same terms?
There is no way to grow out of it. You look like a massive risk.
It is possible the aggregators (those who source capacity in the market) flip the model and start rolling those players up - but I wouldn’t bet on it (and if I were there are only a very few I like from a business perspective). In a tide receding cycle, it is hard to make money on spread risk when the market flattens out and counterparty terms tighten at the same time.
An aggregator in this market has about as much chance of building a durable position as someone reselling Claude has of owning the model layer.
When the tide was rising, everyone looked viable. Growth hid the structure. Capital filled the gaps. Customers were less sensitive to price, term, and counterparty risk because they needed capacity immediately.
When the tide goes out, the structure is all that matters.
- Who owns the power.
- Who owns the land.
- Who owns the infrastructure.
- Who can sign a 10 or 15 year contract and still look credible at year ten.
- Who can carry inventory.
- Who can absorb a delay.
- Who can survive a quarter, two quarters, three quarters where demand slows and capital gets tight.
Again, I have seen the tide flow both ways.
In telecom, capital flooded in. Networks were built ahead of demand. Balance sheets were stretched. When the cycle turned, demand did not disappear, but the structure did not hold. The assets survived. Many of the companies did not.
On the other side of this, the market will concentrate.
The winners will be the ones who control the constraints and can fund through the cycle.
- Operators with access to deep, low-cost capital.
- Operators who control powered land.
- Operators who can build and deliver infrastructure on predictable timelines.
- Operators who can support both hyperscale demand and smaller enterprise workloads without changing their model.
- Operators who can commit long term and still be bankable.
The term fortress balance sheet is all the rage right now - but despite the buzziness of the term, it is a real thing and it has real utility in a receding tide.
Fortress balance sheets show up in very specific ways. Access to low-cost, long-duration capital. The ability to sign 10 to 15 year contracts with credible counterparties. The ability to fund capacity ahead of demand instead of reacting to it. The ability to carry fixed costs through a slowdown without relying on external financing. Most of the market does not meet that bar.
Because when the tide goes out, it does not just slow things down.
It tells you exactly who was real.
.png)
