Tracking the Stack: A Trail of Receipts
- Rich Washburn
- 2 hours ago
- 8 min read
The grid became the constraint. Wall Street built the teams. NVIDIA started financing the mine. Now the world's largest capital providers are building the rails.
I keep coming back to the same rule:
Follow the constraint. Then follow the money trying to get around it.
Individual headlines are noisy.
Sequences are not.
And over the last several months, the AI infrastructure story has produced a sequence that is becoming increasingly difficult to dismiss as coincidence.
This isn't an "I told you so" article.
It's a receipt check.
Because if we're going to talk seriously about where this market is going, we should occasionally go back and compare what the stack appeared to be telling us with what actually happened next.
So let's do that.
Receipt One: The Money Was Already There
On April 3, I published $710 Billion and Nowhere to Plug It In.
At the time, Microsoft, Google, Amazon and Meta were staring at roughly $710 billion in combined capital expenditure across 2025 and 2026.
The remarkable part wasn't the number.
It was what the number couldn't buy fast enough.
I wrote:
"The land is available. The capital is available. The power is not."
That was the point.
AI had already escaped the software economy.
Training clusters were becoming industrial facilities. Hundreds of megawatts. Eventually gigawatts. Transformers, substations, transmission, cooling, fiber, generators and land.
And those systems run on physical timelines.
You can authorize another $20 billion before lunch.
You cannot authorize a transmission line into existence.
Grid interconnection queues were already stretching years. Transformer lead times were extending. Northern Virginia was tightening. Other major markets were beginning to feel the same pressure.
Once that happens, the definition of the valuable asset changes.
A piece of land with power isn't merely real estate anymore.
It's time.
If one site can deploy 200 megawatts years sooner than another, that compression can be worth billions.
The first receipt was simple:
capital was no longer the primary constraint. Power was.
Receipt Two: Wall Street Started Building the Team
The same day, I published When Wall Street Builds a Team, Follow the Money.
Goldman Sachs, Morgan Stanley and Jefferies had begun building formal investment-banking capabilities around AI infrastructure.
Not AI applications.
Not SaaS.
Not another chatbot company.
Power generation. Cooling. Fiber. Data centers. Debt. Equity. M&A.
That mattered.
Banks don't build senior coverage teams because something is trending on LinkedIn.
They build them because they expect transactions.
At the time, I wrote:
"The asset class is still being defined."
And that was exactly what was happening.
The banks were positioning ahead of an infrastructure transaction cycle involving data centers, power assets, fiber, private generation, credit and increasingly complex ownership structures.
The machines hadn't just gotten bigger.
The financial ecosystem around the machines was beginning to mature.
Another receipt:
Wall Street could see the transaction layer forming before the asset class had even finished acquiring a name.
Receipt Three: The Canary Was the Grid
Then came The Canary Is Dead.
That piece pulled the signals together.
Compute was being secured like inventory.
Infrastructure spending was outrunning application spending.
Power had become a gating factor.
The banks were staffing.
The queues were extending.
The physical assets required to turn compute into useful output were becoming scarce.
I wrote:
"The moment power became the constraint, everything changed."
And this was the larger point:
Constraints don't disappear when technology advances.
They move.
We solved one class of problem with accelerated compute and created another farther down the stack.
The constraint moved from software capability to chips.
Then from chips to facilities.
Then from facilities to power.
And every time the bottleneck moves, value migrates with it.
That's why I ended that piece with the question I still think matters most:
Can you see where the constraints are moving before they become obvious?
Because once everyone sees the bottleneck, the assets sitting on the right side of it tend to already have owners.
Receipt Four: NVIDIA Started Financing the Mine
Then things got considerably more interesting.
Five days ago I published I Owe My GPUs to the Company Store.
The subject was NVIDIA's increasingly unusual position inside the AI economy.
NVIDIA sells the accelerators.
Its customers need enormous facilities to deploy those accelerators.
Those facilities require enormous amounts of financing.
And NVIDIA had increasingly begun helping remove the financing barriers preventing customers from building infrastructure filled with NVIDIA equipment.
The simplified loop looked something like this:
NVIDIA financial strength → AI customer → data center → NVIDIA GPUs → NVIDIA revenue.
That doesn't make the demand fake.
The chips still exist.
The buildings still exist.
The debt still exists.
And if somebody eventually can't pay, the losses still land somewhere very real.
But the development was important enough that I wrote:
"They're not just selling the shovel anymore. They're financing the mine."
Then I pushed the thought one step further:
"That's what vertically integrated domination looks like when it reaches the capital layer."
That sentence didn't have much time to age.
Receipt Five: Welcome to the Capital Layer
On August 10, NVIDIA announced memorandums of understanding with:
Apollo.
BlackRock.
Blackstone.
Brookfield.
Goldman Sachs.
KKR.
The stated objective is to create independent financing platforms capable of mobilizing more than $500 billion of third-party capital over time for AI infrastructure.
Important distinction:
This is not a $500 billion fund.
It is not $500 billion already committed.
It is potentially something more structurally significant.
It is financial plumbing.
Repeatable platforms designed to independently underwrite AI infrastructure projects and connect enormous pools of institutional capital with NVIDIA's ecosystem.
And NVIDIA didn't bury the thesis.
Its headline literally says the platforms are intended to turn NVIDIA compute and full-stack AI infrastructure into:
"an investable asset class for global capital."
Remember April?
"The asset class is still being defined."
Four months later, NVIDIA and six of the largest names in global finance are explicitly building mechanisms to finance it.
Receipt.
The Company Store Just Invited Wall Street Inside
This is the part that matters.
A few days ago, the question was whether NVIDIA was becoming unusually involved in financing the customers buying NVIDIA infrastructure.
Now the model is expanding.
The company doesn't have to become JPMorgan.
It can bring JPMorgan's world to the company store.
Apollo brings private credit.
BlackRock brings enormous pools of long-duration institutional capital.
Blackstone and Brookfield bring deep infrastructure expertise.
Goldman brings origination, structuring and distribution.
KKR brings another massive infrastructure and private-capital engine.
NVIDIA brings the compute ecosystem.
And suddenly the capital layer stops being an improvised collection of one-off deals.
It begins to look like a market.
Goldman CEO David Solomon said the partnership creates an opportunity to develop a market for credit backed by NVIDIA compute.
That phrase deserves attention.
Because once lenders become comfortable treating productive compute as financeable collateral, an entire financial architecture can begin forming around it.
Compute leases.
Infrastructure debt.
Capacity agreements.
Usage-linked financing.
Residual-value models.
Securitized compute exposure.
AI factories start looking less like rooms full of computers and more like another category of productive infrastructure.
That is a fundamental shift.
But Capital Isn't the End of the Stack
Here's where this gets interesting.
The capital constraint is being attacked aggressively.
That does not mean capital becomes infinite.
Projects can still be bad.
Customers can still fail.
GPUs can depreciate faster than underwriting models expect.
Token economics can change.
Hardware generations can make residual-value assumptions look stupid very quickly.
There is plenty of financial risk here.
But capital has a remarkable property:
When returns are attractive, humans are very good at manufacturing financial structures.
We create funds.
Credit vehicles.
Leases.
Guarantees.
Project companies.
Special-purpose vehicles.
Securitizations.
Private-credit facilities.
Capital adapts quickly.
Physics does not.
And Then NVIDIA Published the Power Receipt
The timing here is almost too perfect.
On August 11 — one day after announcing the new financing platforms — NVIDIA published another piece:
"Why Scaling AI Compute Performance Requires a New Power Architecture."
The opening problem?
Every generation of accelerated computing increases demand on the infrastructure beneath it.
Higher rack density.
More efficient power distribution.
More sophisticated electrical architecture.
NVIDIA summarizes the issue simply:
The bottleneck isn't just wattage.
And earlier this year, NVIDIA's Vera Rubin DSX announcement described energy as the biggest bottleneck for AI infrastructure buildouts, citing more than $300 billion in equipment backlogs and more than 200 gigawatts of projects waiting in U.S. interconnection queues.
So look at what happened in the span of roughly 24 hours.
NVIDIA:
Here is how we intend to help unlock hundreds of billions of dollars of capital.
Also NVIDIA:
We need an entirely new power architecture underneath the compute.
That's Tracking the Stack in two press releases.
Follow the Constraint
Now lay the receipts out in order.
$710 billion in hyperscaler CAPEX.
↓
The grid becomes the limiting factor.
↓
Permitted, energized infrastructure becomes disproportionately valuable.
↓
Wall Street builds dedicated AI infrastructure teams.
↓
The asset class starts forming.
↓
NVIDIA begins helping finance the ecosystem purchasing NVIDIA equipment.
↓
NVIDIA brings Apollo, BlackRock, Blackstone, Brookfield, Goldman and KKR into dedicated compute-financing platforms.
↓
Compute begins being discussed as collateral.
↓
Capital gets easier to mobilize.
↓
Power remains difficult to manufacture.
That's the stack.
And the farther we move into the industrial phase of AI, the less useful it becomes to think of this as a software boom.
This is compute infrastructure.
Electrical infrastructure.
Financial infrastructure.
They are becoming one system.
The Important Part Isn't the $500 Billion
The giant number will get the headlines.
That's understandable.
But $500 billion is not the most interesting part.
The interesting part is that the machinery now exists to potentially raise the next $500 billion.
And the next.
Assuming the projects generate returns.
That's what infrastructure finance does.
It takes something prohibitively capital intensive and builds repeatable structures allowing other people's money to own pieces of the productive asset.
Railroads did it.
Telecom did it.
Energy did it.
Commercial aviation did it.
Now compute is moving in that direction.
Which means the question changes again.
It isn't simply:
Who can afford the AI factory?
Increasingly, the answer may be:
Who can finance one?
And if financing becomes increasingly available, we circle right back to the question from April:
Where are you going to plug it in?
Another Receipt in the File
No single one of these events proves where AI infrastructure ends up.
That's not the point of this series.
The point is to maintain the data set.
Watch the spend.
Watch the financing.
Watch the deals.
Watch the physical bottlenecks.
Watch where the capital redirects when it hits resistance.
Most importantly: watch the constraint.
Because four months ago the asset class was being defined.
Five days ago NVIDIA looked like it was financing the mine.
Today NVIDIA is standing next to some of the largest financial institutions on Earth, telling global capital that compute itself is investable infrastructure.
That's not the end of the story.
It's another receipt.
And the stack is getting easier to read.
Rich Washburn is a technologist and strategist working at the intersection of AI, infrastructure, and capital. He is Managing Partner and Chief AI Officer at Eliakim Capital.
Source Links
NVIDIA — AI Compute Infrastructure Financing Platforms With Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR
NVIDIA — News Archive: AI Compute as an Investable Asset Class / New Power Architecture
NVIDIA — Vera Rubin DSX AI Factory Reference Design
Rich Washburn — $710 Billion and Nowhere to Plug It In
https://www.richwashburn.com/post/710-billion-and-nowhere-to-plug-it-in
Rich Washburn — When Wall Street Builds a Team, Follow the Money
https://www.richwashburn.com/post/when-wall-street-builds-a-team-follow-the-money-1
Rich Washburn — The Canary Is Dead
Rich Washburn — I Owe My GPUs to the Company Store
https://www.richwashburn.com/post/i-owe-my-gpus-to-the-company-store
Comments