At the Gates of Compute

The Machines May Not Take Control. We May Hand It Over—and Make Them Our Universal Excuse

A response to Jensen Huang’s proposal to finance AI compute as infrastructure

Chain of Prompts: Generate Your Own Insights First

Before reading the argument below, try running this chain of prompts through the GenAI system of your choice. Do not ask it to agree with the premise. Ask it to challenge each step.

1. What exactly did Jensen Huang announce about financing AI compute? Separate the actual financing mechanism from the narrative around it.

2. Assume the proposal works exactly as intended. What first-order economic problems does it solve, and who benefits?

3. If AI compute becomes a financeable infrastructure asset, what behaviors will that financing mechanism incentivize among labs, enterprises, governments, investors and users?

4. What second-order effects emerge if the supply of compute receives increasingly reliable financial returns before the downstream value produced by that compute is fully demonstrated?

5. What third-order effects emerge if society begins treating compute not merely as infrastructure, but as something resembling currency, collateral or a claim on future economic productivity?

6. Who actually controls the major capability chokepoints in this system, and which consequential decisions remain human decisions?

7. Where could responsibility for human decisions later be attributed to “AI,” “the algorithm,” automation or technological inevitability?

8. How is AI compute materially different from electricity, telecommunications and other infrastructure layers? Specifically examine whether AI requires continuing inputs from collective human knowledge, behavior, feedback and participation.

9. Who contributes to the resulting AI capability, who owns the chokepoints, who captures the economic returns, and who absorbs negative externalities?

10. Compare two extremes: compute as a broadly guaranteed resource versus compute as a tightly gated strategic resource. What fails at either extreme?

11. Design intermediate architectures that preserve broad productive access while increasing accountability as compute becomes more consequential.

12. How could AI itself be used to model the second- and third-order consequences of compute allocation without transferring final decision authority or accountability from humans to machines?

Now compare what you discover with the argument that follows.

Jensen’s Proposal Makes Sense

On August 10, NVIDIA announced partnerships with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR intended to establish independent compute-financing platforms capable of mobilizing more than $500 billion in third-party capital for AI infrastructure. NVIDIA’s proposition is to make its compute and full-stack AI infrastructure an investable asset class and broaden customers’ access to large-scale AI factories. Jensen Huang compressed the economic thesis into four words:

“In AI, compute is revenue.”

There is considerable logic behind it.

AI infrastructure is extraordinarily capital intensive. Demand and capital do not always meet efficiently. If long-duration investors can finance compute infrastructure much as they finance factories, telecommunications networks, aircraft, power plants and other productive assets, more organizations can obtain the capacity required to create AI products and services.

The announcement also moves beyond ordinary equipment financing. NVIDIA describes compute as fungible and transferable across customers and workloads, while Goldman Sachs explicitly points toward the possibility of creating credit backed by NVIDIA compute. The announced partnerships remain subject to final agreements.

This could be an important financial innovation.

The question is not whether we should stop it.

The question is whether we understand what else we may be creating with it.

From Compute Is Revenue to Compute Is Currency

There is a tempting extension of Jensen’s proposition:

Compute is currency.

As metaphor, it is powerful.

If compute is scarce, transferable, productive and capable of generating recurring revenue, financing its provision becomes increasingly rational. Providing compute could eventually resemble providing productive capital itself.

But language can alter incentives.

There is a significant difference between:

Compute enables revenue

and:

Compute is revenue.

There is an even larger jump to:

Compute is currency.

The further we move along that chain, the easier it becomes to financially recognize the value of supplying compute before proving the independent economic value created downstream.

That creates a possible progression:

Capital → Compute → Capability → Expected Value → Financial Claim

Notice where the financial claim can appear.

It can precede the final proof of value.

That is not automatically bad. Much of modern finance exists precisely because capital must precede production.

But AI introduces an unusual complication.

AI Compute Is Not Just Electricity

The electricity analogy gets us only so far.

Electricity is consumed by factories, offices, homes and machines. A power plant does not generally require factories to continually contribute their engineering knowledge, conversations, decisions, corrections, creative work and behavioral patterns so electricity can remain economically useful.

AI does.

Not in the simplistic sense that every prompt automatically retrains a model. It does not.

But the broader AI ecosystem recursively depends upon collective participation.

Humans provide accumulated knowledge.

Humans provide current context.

Humans discover new things.

Humans evaluate outputs.

Humans correct errors.

Humans reveal preferences.

Humans reorganize workflows.

Humans generate demand.

Humans create the new reality against which tomorrow’s AI must again be calibrated.

The loop increasingly resembles:

Collective human activity → AI capability → AI output → changed human activity → new collective inputs

Compute therefore does more than transform electricity into tokens.

It provides leverage over collectively supplied intelligence.

And whoever controls the scarce compute layer can occupy a disproportionately powerful position within that loop.

Scarcity can justify a premium.

It does not automatically establish sole authorship of the resulting value.

The Capability Chokepoint

This is where financing becomes a governance question.

Suppose compute becomes increasingly available because financiers can earn attractive and relatively dependable returns from supplying it.

Excellent.

But somebody still decides:

which capabilities receive compute,

which applications get subsidized,

which markets receive capacity,

what gets automated,

what models get developed,

what gets restricted,

what gets priced below cost to accelerate adoption,

and what kinds of human participation are amplified, replaced or rendered economically marginal.

Those decisions are not being made by silicon.

They are being made by people and institutions controlling the levers surrounding silicon.

That suggests a simple principle:

The greater your control over a consequential capability chokepoint, the greater your traceable human accountability should become.

Yet complex technological systems can produce the opposite effect.

Capability increases.

Complexity increases.

Causal chains become longer.

Consequences arrive later.

Attribution becomes harder.

And eventually we start hearing:

“AI eliminated the jobs.”

“The algorithm decided.”

“Automation made it inevitable.”

“The market demanded AI.”

Perhaps.

But somewhere upstream, humans allocated the capital, compute, authority and incentives that made those outcomes possible.

The Hinton Inversion

This creates a strange path toward one version of Geoff Hinton’s warning about losing control to increasingly capable AI.

It might not require machines deciding to dominate us.

We could construct the outcome ourselves.

Humans could finance increasingly capable systems.

Humans could reorganize institutions around their expected productivity.

Humans could progressively surrender decisions to those systems because doing so appears economically rational.

Humans could allow the resulting architecture to become too deeply embedded to unwind easily.

And when negative consequences eventually arrive, humans could blame the machines.

Hinton could therefore become directionally right through a mechanism quite different from the one most people imagine.

The machines need not seize human agency.

Humans can mortgage their agency to machines, institutionalize the arrangement, and later describe the resulting constraints as technological inevitability.

That would be one of history’s stranger accountability inversions.

Appropriation Can Move Faster Than Accountability

There is another asymmetry worth examining.

Financial claims can propagate very efficiently.

A lender has a contract.

An infrastructure investor has a claim.

A GPU provider receives revenue.

A cloud provider charges for consumption.

A laboratory charges for inference.

An enterprise books productivity improvements.

The economic architecture knows remarkably well how to record these claims.

But what about the distributed contribution underlying AI capability?

What about displaced human capability?

Foregone careers?

Knowledge absorbed from institutions?

Public infrastructure?

Behavioral feedback?

Social disruption?

Changes to bargaining power?

Future dependency?

The economic ledger becomes much less precise.

This produces an uncomfortable asymmetry:

Appropriation can be automated long before accountability can be reconstructed.

And there can be a sophisticated lag between decision and consequence.

By the time society discovers that some allocation of AI capability produced undesirable systemic effects, the infrastructure may already be financed, contracted, depreciating, politically protected, operationally indispensable and employing people whose livelihoods now depend upon its continuation.

The question quietly changes from:

Should we do this?

to:

How could we possibly stop doing this?

That is lock-in.

Guaranteed Compute or Gated Compute?

None of this argues for making compute artificially scarce.

That produces another dangerous concentration of power.

If access to advanced intelligence becomes tightly controlled by a handful of laboratories, hyperscalers, governments or financial institutions, whoever manages the compute gate acquires extraordinary influence over who gets to experiment, compete, learn and build.

So we face two unattractive extremes.

At one end:

Compute as guaranteed resource.

Cheap, abundant access could unleash experimentation and democratize capability. It could also subsidize enormous quantities of low-value inference, extraction, manipulation, automated activity and capability escalation whose externalities others must absorb.

At the other:

Compute as gated resource.

Scarcity could reduce waste and constrain dangerous capability. But excessive gating could entrench incumbents and transform compute owners into governors of economic participation.

The sweet spot is probably neither.

It is likely a gradient.

Basic and civic compute could approach guaranteed access.

Ordinary productive compute could be broadly available and financeable.

Higher-scale allocations could require stronger evidence of productive use and resource efficiency.

Frontier capability could require substantially stronger accountability.

Exceptionally consequential capability could require explicit, attributable human authorization.

In other words:

Guaranteed → Accessible → Governed → Gated

with boundaries capable of moving as evidence changes.

Put More Adults in the Room

“Adults in the room” should not become code for a small committee of regulators.

Nor should it mean laboratories regulating themselves.

Nor bankers.

Nor politicians.

Nor technologists.

The point is plural accountability.

If one institution controls the resource, finances its deployment, measures its success and defines its social benefit, there is almost no meaningful independent reality check.

The important chokepoints therefore need actors exposed to different consequences.

Capital providers.

Compute providers.

Model developers.

Enterprises.

Governments.

Domain specialists.

Workers.

Communities.

Users.

Independent evaluators.

Not everyone deciding everything.

But enough independent perspectives crossing the consequential perimeter that no single party can simultaneously exercise the leverage and define whether exercising it was successful.

Use AI to Govern AI Compute—But Do Not Let It Become the Governor

Here we reach an interesting opportunity.

If the second- and third-order effects of AI allocation are becoming too complex for unaided human reasoning, perhaps one of the most valuable uses of advanced AI is understanding what happens when we deploy advanced AI.

Before committing massive compute to a consequential use case, we can ask AI systems to simulate competing causal pathways.

Who gains?

Who becomes dependent?

Who loses bargaining power?

What happens if competitors imitate the strategy?

What resources become bottlenecks?

What activity gets cannibalized?

What new jobs emerge?

Which disappear?

What becomes difficult to reverse?

Where could value merely circulate between AI suppliers rather than reaching independent customers?

What happens if the expected productivity never arrives?

Which party bears the downside?

We can run competing models against one another.

We can deliberately generate counterarguments.

We can search historical analogues.

We can stress-test assumptions.

We can continually compare predictions with reality.

This may become a legitimate new category of AI work:

consequence inference.

But AI should expand the room, not become the adult in it.

The system can tell us:

Here are the probable consequences.

A human institution must still say:

We authorize this.

And its name should remain attached to that decision when the consequences arrive.

Finance the Expansion—but Keep the Human Hand on the Lever

Jensen Huang may be right that AI compute has become a productive, investable infrastructure asset.

And financing it at global scale may be exactly what the next stage of AI development requires.

But if compute becomes financial leverage, societal leverage follows.

If it becomes collateral, somebody acquires claims against future productivity.

If it becomes currency, we must ask what that currency is purchasing.

And if it becomes a strategic resource layer, we must know who controls the gate.

So the response to NVIDIA’s announcement should not be fear.

It should be maturity.

Finance the compute.

Expand access.

Build the infrastructure.

Experiment aggressively.

But simultaneously use the intelligence we are building to understand the second- and third-order consequences of allocating it.

Keep multiple adults in the room.

Keep accountability attached to the humans controlling the capability chokepoints.

Keep access broad enough that governance does not become monopoly.

Keep consequential capability bounded enough that abundance does not become irresponsibility.

And above all, resist one very convenient future explanation:

“The machines did it.”

They did not finance themselves.

They did not allocate themselves.

They did not decide who should control the chokepoints.

They did not choose what economic system should form around them.

Those remain human decisions.

If AI eventually becomes consequential enough that humanity feels it has lost control, our first question should therefore not be:

When did the machines take over?

It should be:

At which human-controlled chokepoint did we decide to hand the lever away?

Perhaps the real opportunity in Jensen’s announcement is not merely to invent a new way to finance compute.

It is to recognize, early enough, that we also need to invent a better way to remain accountable for what financed compute allows us to do.