Speaking to reporters at an artificial intelligence summit in Scotland, Nvidia chief executive Jensen Huang offered an aggressive forecast for the hardware supply chain. He stated that the company expects to sell twice as many chips next year as it does this year. The projected volume covers not only flagship data center graphics processing units across the Blackwell and upcoming Rubin architectures, but also central processing units, NVLink switches, optical networking silicon, and robotics modules. Around the same time, Nvidia outlined long-term expectations pointing toward fiscal 2028 revenue of roughly 673 billion dollars, an increase of approximately 70 percent over the prior year.
The numerical spread between a 100 percent increase in physical shipments and a 70 percent increase in revenue is the first detail worth pricing. A gap of thirty percentage points between unit expansion and top-line dollar growth reflects either an intentional broadening of the product mix into lower-priced networking and edge silicon, or the early stages of average selling price compression. For three years, accelerator pricing operated under conditions of near-total inelasticity, where buyers paid whatever premium was asked simply to secure allocation. As packaging capacity catches up, unit volumes can surge even as the average realization per die begins to normalize.
The more significant economic friction sits outside the semiconductor fabrication plant. Wafer fabrication and advanced packaging are flexible over a twelve-month horizon. With contract manufacturers like TSMC expanding CoWoS lines and tooling up advanced packaging nodes, fab capacity can double within several quarters. Power delivery infrastructure, by contrast, operates on an entirely different physical schedule.
High-voltage substation transformers carry delivery lead times between 36 and 48 months. Interconnection study queues across regional transmission organizations such as PJM, ERCOT, and MISO currently average four to seven years. A semiconductor vendor can fabricate, package, and crate twice as many server racks in a calendar year, but the electrical grid cannot double the volume of energized megawatts delivered to data center substations over that same twelve-month window.
When the physical output of compute hardware expands faster than the power infrastructure required to plug it in, the industry faces an energization gap. The excess hardware must follow one of two paths, and both carry specific financial penalties.
The first path is physical delay: hardware sits uncrated in logistics facilities or unpowered data center shells, waiting for substation construction and utility interconnect approvals. Unlike commercial real estate or traditional factory equipment, high-performance computing accelerators carry a severe time decay. Hardware architectures advance rapidly; memory bandwidth and compute density improvements typically halve the economic competitiveness of a flagship chip within three to four years. Every month an accelerator sits on a pallet without drawing power, its remaining economic half-life burns away while generating zero cash flow. For leveraged infrastructure operators holding floating-rate debt facilities, that carrying cost creates immediate negative drag.
The second path is cannibalization: installing new accelerators into existing, already-energized data centers by removing older hardware. From an engineering standpoint, this is a clean solution. A rack of Blackwell or Rubin accelerators delivers substantially higher floating-point operations per megawatt than an older Hopper installation. By replacing older clusters with denser silicon, an operator increases overall compute capacity while staying strictly within an existing substation power envelope.
The financial cost of that replacement cycle shows up on the corporate balance sheet. Over the past three fiscal years, major cloud providers extended the accounting depreciation schedules on their server infrastructure from three or four years out to five or six years. That accounting change deferred billions of dollars in annual depreciation expense, flattering reported operating income across the sector.
If two-year-old H100 clusters are pulled from racks early to make room for incoming hardware, those assets have not reached the end of their accounting lives. The unamortized book value cannot remain capitalized indefinitely if the hardware is decommissioned or relegated to secondary storage. Cloud operators would face accelerated depreciation charges or outright non-cash asset impairments. You cannot maintain the position that server assets have a six-year useful life for accounting purposes while simultaneously cycling hardware every twenty-four months to navigate a local power cap.
There is also a downstream effect on compute rental pricing. In capital-intensive industries with high fixed equipment costs, marginal production costs are low. Once an accelerator cluster is racked, cooled, and powered, the incremental cost of generating an additional token is largely the price of wholesale electricity. If total energized compute capacity expands rapidly while software application demand grows at a linear enterprise adoption pace, spot market pricing per GPU-hour must adjust downward to clear available capacity. Lower spot compute rates benefit developers and end users, but they compress the cash flows required to amortize expensive hardware tranches.
Steelman the hardware expansion. Generative reasoning architectures, recursive agent loops, and multi-token search algorithms consume orders of magnitude more compute per query than single-turn completions. If token demand grows along an exponential curve, every newly minted chip will find a workload waiting for it, and the higher efficiency of newer silicon will justify writing off earlier hardware investments.
I want the distribution, not the point estimate. When hardware supply doubles into a power grid with multi-year transmission queues, the outcome is rarely uniform absorption. The return distribution splits. Operators who secured multi-gigawatt interconnection rights three years ago capture strong operational rents, while buyers acquiring hardware ahead of energization rights absorb the carrying cost of depreciating silicon. Tracking the AI capital cycle requires watching the substation queues as closely as the wafer starts.
Originally published at https://deanlee.info/essays/nvidia-chip-volume-and-the-energization-gap/.
Top comments (2)
Strong piece, and I think the core call holds. I spent the afternoon trying to break it and couldn't — modelling the pipeline in TLA+ and running the numbers over sourced intervals puts the stranding-forced regime at about 80%, with a median ~20% of shipments landing ahead of the megawatts to run them.
What surprised me is that the best argument for your thesis isn't one you make. Enterprise racks draw 8–10 kW; a GB300 NVL72 draws 132–142. Converting a rack slot therefore recovers only ~6% of what the new hardware needs — and conventional server energy is flat in LBNL's model, so the "10–13% annual fleet turnover" that looks like a reuse pool mostly isn't one. New hardware can't hide in old power. That's a harder constraint than the queue arithmetic.
Two things worth checking, though. The $673B FY2028 figure isn't Nvidia's — it's CNBC's arithmetic on an LSEG consensus ($396B × 1.70). "673" appears nowhere in the 8-K, the 10-Q, or the corrected transcript; Kress guided the rate, and attributed the gap to supply rather than price on that same call. And the 100%-vs-70% spread composes multiplicatively: 1.70/2.00 − 1 = −15%, not −30.
One genuine disagreement: GPU rental prices bottomed in Dec 2025 and are up ~32% since — Nvidia disclosed A100 rental up ~15% y/y. Which your own mechanism predicts, I think: if energized megawatts are the binding constraint, hardwarein an energized slot gets priced off the scarcity of the slot.
Full workings, data and spec: copyleftdev.github.io/does-the-gri...
Happy to be wrong on any of it. The weakest link is mine, not yours — scaling two utility meter readings to a national figure, which no source actually provides.
The rack density arithmetic is sharper than the queue math. If an enterprise cabinet delivers 10 kW and an NVL72 demands 140 kW, you can clear out fourteen legacy server bays and still lack the power infrastructure to light up a single AI rack. Legacy data center retrofits run into feeder and substation limits almost immediately.
Your point on GPU rental pricing also hits the central economic mechanic. When energized power is the binding constraint, the economic rent migrates from the chip fabricator to the entity holding the energized interconnection queue. The 30 percent rebound in compute lease rates is simply the shadow price of megawatts asserting itself over unenergized silicon.
Fair points on the CNBC consensus math and the multiplicative compounding rate. I will take a close look at your TLA+ spec.