If US data centers need around 68GW of power between 2026 and 2028, and roughly 38GW is still missing, what is the most valuable AI asset now: the chip, or the ability to finance and energize the infrastructure around it?
A Scarce Megawatt Is Only the First Layer
Roughly 15GW of the projected requirement is under construction, with another 15GW available or contracted. That leaves an estimated 38GW shortfall before alternative power sources.
Traditional colocation providers reportedly generate roughly $3 million to $4 million in annualized revenue per active megawatt. CoreWeave and Nebius reportedly generate roughly $9 million to $10 million by combining power with Nvidia GPUs, networking, and software. Nebius management reportedly said during Q2 earnings that new short term deals were being priced at $40 million to $50 million per megawatt.
That progression suggests power is only the first layer. Integrated compute, networking, and software can produce more revenue from each scarce megawatt.
But revenue per megawatt is not return on capital. Capacity creates value only if utilization, operating costs, financing terms, and durable cash flow support the investment.
Nvidia Is Turning Its Balance Sheet Into Distribution
Nvidia may provide up to $105 billion of financing for an OpenAI data center at the PORTS-Pike Technology Campus in Pike County, Ohio. The project would support an initial 4.25GW of computing capacity, with an option for another 3.75GW.
Nvidia recorded $48.5 billion of quarterly free cash flow, 18 times the level three years earlier, after 12 consecutive quarters of revenue growth above 55%. It also signed a memorandum of understanding with Goldman Sachs, Apollo Global Management, Blackstone, and BlackRock for GPU financing, including an option to backstop 25% of each loan.
Capital is becoming part of Nvidia’s competitive position. The company may help customers finance the infrastructure required to deploy its GPUs at enormous scale.
That creates circularity risk. If Nvidia finances data centers that purchase Nvidia GPUs, it may be financing its own revenue. Supporters, including Cantor analysts, see evidence of a durable AI investment cycle. Critics see demand being pulled forward into eventual overcapacity.
Both interpretations can be true for a period. The decisive question is whether utilization and cash generation arrive before financing burdens expose weak economics.
Nebius presents the same tension at a smaller scale. It proposed $4.5 billion of convertible senior notes after earlier raises of $2.75 billion and $4.3 billion. The funding could expand capacity while demand exceeds construction capacity, but repeated issuance creates leverage and dilution before adequate returns are demonstrated.
Google’s Marvell Warrant Pays for Performance
Google received a warrant to buy up to 58,970,907 Marvell shares at $206.58 per share, exercisable through August 2033. Only a limited portion vests on a fixed schedule.
The remaining 57,610,040 shares are divided into 240 equal tranches. One tranche vests for every $500 million of Custom Products revenue that Google generates for Marvell through fiscal 2033. Full vesting corresponds to a $120 billion revenue ladder.
That is not a $120 billion revenue commitment.
It is a structure that ties nearly all potential dilution to realized commercial performance. Marvell’s upside becomes measurable, while Google gains long duration alignment with a supplier without handing it a blank check.
Supplier Alignment Does Not Prove TPU Displacement
The warrant supports the view that Google is broadening its custom silicon supplier ecosystem. It does not establish that Marvell is replacing Broadcom in Google’s core TPUs.
One interpretation is that the agreement concerns products attached to the TPU ecosystem rather than the core TPUs themselves. Broadcom had also announced a long term Google agreement running through 2031.
The warrant creates a transparent way for Marvell to participate if actual Custom Products revenue arrives. It does not settle how the opportunity will be divided.
Scarce power, financing capacity, and custom silicon relationships can support attractive economics. But scarcity is a moat only when capital becomes utilized infrastructure and durable cash flow. Until then, financing commitments and warrants reveal opportunity, not intrinsic value.