Power Transition: Nvidia and Control of AI Compute

The precedent: whoever owns the bottleneck dictates the outcome

In 1904 Standard Oil controlled roughly 90% of U.S. refining capacity. The mechanism was simple: own the chokepoint.

Anyone who wanted to bring oil to market went through its pipelines and its refineries. Competitors became customers, then dependents, and finally were absorbed. The structure mattered more than the finished product. The advantage was not in the price of the barrel. It was in control of the mandatory passage.

Today the same logic governs computing for artificial intelligence. The context changes, the structure stays identical. The bottleneck is the GPU, and its owner is called Nvidia. This is a power transition maturing before the markets' eyes. Anyone who reads only revenues does not see it. Anyone who reads capital flows does.

The current pattern: Groq re-enters Nvidia's orbit

Groq raised 350 million dollars to accelerate its transformation from chip maker to cloud infrastructure provider. The round, led by Disruptive with Nvidia expected to participate, values the company at 3.5 billion dollars.

The figure falls from 6.9 billion the previous September, as TechCrunch documents. In between, Nvidia had hired founder Jonathan Ross and part of the team, with a 20 billion licensing deal. The valuation nearly halved while the talent migrated toward the supplier. The two movements are not independent.

Groq was born to challenge Nvidia on inference with its LPUs. Today it runs Nvidia systems across 13 data centers and serves more than 6 million developers. The rival became a reseller. This is the turning point of the story: whoever wanted to replace the supplier now distributes it.

A spokesperson argues that this valuation establishes the worth of the company's version following the licensing deal. The reading is optimistic. The direction of the flows tells another story: talent and intellectual property migrated toward the center. A valuation drop presented as a new starting base remains a valuation drop.

The mechanism: supply the essential input and finance the customer

Nvidia supplies GPUs to CoreWeave, Lambda, Nebius and now the Groq ecosystem. At the same time it invests billions in those very operators.

The design is elegant. You sell the critical input, then you finance whoever consumes it, so you stabilize demand and bind the chain to yourself. Capital flows in a circle and returns to the center. Every dollar invested in the customer becomes a GPU order back to the supplier. Demand is no longer exogenous: it is engineered.

This replicates the Standard Oil pattern in a modern version: control of the bottleneck plus vertical integration of the customer. Three operators under the same dependency are enough to call it a pattern. The compute layer is concentrating, and concentration precedes pricing power. Pricing power precedes margin. Margin accumulates at the center.

My position: this is a structural transition

Competition in AI is decided at the compute layer. Nvidia is converting its own rivals into resellers: this is a structural power transition, distinct from a market cycle. A cycle reverses. A structural transition redefines who captures the value for a decade.

Consistent with my underlying thesis: whoever controls the fabs and the infrastructure controls the outcome. Models are replicable, fabs and GPU networks resist replication. Groq proves it in real time. It had the alternative chip. It was not enough.

What would change my reading? An inference silicon supplier that reaches industrial scale independent of Nvidia, or an antitrust break that separates GPU sales from investments in customers. Until then, the center strengthens. These are the only two scenarios that would reverse the direction of the flows.

Why risk models underestimate it

VAR models treat GPU supply as a diversifiable chain risk. This classification is wrong.

When a single supplier owns the input, finances the customers and holds the intellectual property via license, the risk becomes systemic and correlated. Diversification across CoreWeave, Lambda and Groq is illusory: they all point to the same node. Spreading capital across three resellers does not reduce exposure. It multiplies it on the same supplier.

A precedent illuminates the point. In the Wintel era of the 1990s, PC makers grew in volume while value migrated toward Intel and Microsoft. The margins of the assemblers stayed compressed for decades. Volume grew. Value stayed elsewhere.

The market priced the growth of inference. The market overlooked the concentration of its infrastructure. The open question remains the profitability of neoclouds, burdened by high capex, debt and hardware that depreciates fast. CoreWeave shows strongly growing revenues and, in parallel, the same cash fragility. Revenue growth does not offset the cost structure.

Three implications for capital

Concentration at the center imposes precise choices on whoever allocates today.

• Family offices and sovereign funds (36 months): exposure to the node (Nvidia and its chain) plus a hedge against regulatory break risk.
• CEOs and boards (18 months): dependence on a single compute supplier is a strategic risk absent from many plans.
• Chief risk officers (12 months): insert a neocloud correlation scenario into the models, today treated as independent.

For the CFO, the narrative to revise is the one of access to compute at falling costs. Pricing power migrates toward the center, and reseller margins compress. Whoever brings investors a story of abundant, cheap GPUs risks a rebuttal within eighteen months.

The reasonable reallocation targets the anchor of the chain and the intellectual property, reducing the weight of pure capacity resellers. Value accumulates where the bottleneck resides. This is the uncomfortable part for whoever manages risk: the hidden correlation emerges in moments of stress, when capacity and credit contract together. That is where models calibrated on independence fail.

The prediction

Groq aims to grow from 54 megawatts to more than 200 megawatts of capacity in 2027. I read this target as ambitious relative to the capital raised and to hardware depreciation.

The reasoning is mechanical. Scaling capacity requires high capex and debt, in a context of hardware that loses value rapidly. The 350 million raised, added to the 650 from June, sustain the expansion, yet the declared leap remains disproportionate to the endowment. Quadrupling capacity in a year requires more than the sum raised.

Prediction: Groq's operational capacity will remain below the 200 megawatt target, closing below 150 megawatts on 31 December 2027. Confidence: 62%. Horizon: 31 December 2027. Verification: the company's official operational capacity communication.

What to watch

Three indicators will confirm or refute the thesis in the coming months.

• New direct Nvidia investments in neoclouds that resell its GPUs.
• Operational capacity declared by Groq, measured in active megawatts.
• Antitrust signals on chip sales paired with equity stakes.

The concentration of compute advances through silent absorption. Watch the center, the margins matter less. The pattern is already readable. Related insights in our macro analysis.