The Future of AI Hardware is Commoditized

The AI hardware stack today looks like the best business any capital allocator has seen in a generation. Nvidia reported a 75.0% GAAP gross margin in its most recent quarter, on data-center revenue running past $75 billion in a single three-month stretch. Memory suppliers are posting similar numbers. Those are not operating-business margins. They are scarcity rents.

Scarcity rents end. They always have. Mainframes, disk drives, bandwidth, solar modules — every prior wave of hardware ran the same arc from premium pricing to utility economics, usually inside a decade. There is no serious reason to expect AI silicon to break that pattern, and quite a few concrete reasons to expect it to run faster.

For a long-term holder, that view matters more than the next quarter of chip pricing. It changes what a factory is worth, what a logistics footprint near a data center is worth, and what durable means in a world where the picks-and-shovels business turns into a toll road.

What AI Hardware Margins Actually Look Like Today

The margin stack in AI infrastructure is not evenly distributed. It is almost comically concentrated at two layers: the merchant GPU designer and the high-bandwidth memory stacked next to it. Everyone else works for wages.

Dell's AI server revenue grew 757% in a single quarter, and its non-GAAP gross margin fell to 18.1% while it was happening. Supermicro, the other large box builder, is operating in the 6% to 10% range. One teardown of a 24,000-chip Meta training cluster put the GPUs alone at 65.8% of the bill of materials. The box, the cabling, the integration labor, the rack-and-stack — all of it fights for the leftover 34%.

That shape is informative. It tells you where the price-maker sits and where the price-takers sit. It also tells you where competitive pressure will land first.

Every Prior Hardware Cycle Compressed Toward Utility

Mainframes carried IBM gross margins north of 70% through most of the 1970s. By the mid-1990s, servers were a thin-margin assembly business and the money had moved to software and services. Disk storage ran the same play: Seagate and Western Digital posted premium margins during each generational transition, then watched them collapse as capacity doubled and buyers consolidated. Long-haul bandwidth cost roughly $1,200 per Mbps per month in 1998 and under $1 within a decade. Solar modules fell from about $4 per watt in 2008 to under $0.20 in 2024.

The mechanism is always the same: fab capacity catches up, process nodes mature, buyers get sophisticated enough to design around proprietary stacks, and the second-source arrives. The spread between what a scarce chip can charge and what it costs to make narrows until it meets the cost of capital. The companies that survive the transition are the ones that already operate like utilities before they have to.

The point of history here is not nostalgia. It is that long-duration investors underwrite cycles, not peaks. A business whose unit economics only work at 2026 GPU rental prices is a business whose unit economics do not work. That has been true in enough prior cycles that treating it as the base case is the conservative read, not the contrarian one. Our general preference for non-cyclical cash flows applies here with unusual force.

Inference Prices Are the Leading Indicator

If GPU margins are the trailing indicator of hardware scarcity, inference prices are the leading indicator of hardware commoditization. The output of the chip, not the chip itself, is what customers actually buy. And that output is collapsing in price faster than almost any computing commodity in history.

Running a GPT-4-class model cost roughly $30 per million input tokens in March 2023, and the same capability now runs well under $0.50. Independent analysis from Stanford's AI Index found inference pricing for GPT-3.5-level performance fell from $20 per million tokens in November 2022 to about $0.07 by October 2024. a16z labelled this "LLMflation" and pegged the decline at roughly 10x per year for a fixed quality bar.

Not all of that is hardware. Algorithmic efficiency, distillation, better serving stacks, and open-weight competition each account for a share. But the direction is unambiguous: the end-customer price of the thing GPUs produce is falling at a rate that has to eventually show up in what GPUs themselves command. Hardware margins can lag end-product deflation for a while. They cannot do it forever.

Hyperscaler Silicon Removes the Pricing Umbrella

The second mechanism is already in motion. The four largest buyers of AI silicon have all built their own. Google's TPU v7, Amazon's Trainium 3, Microsoft's Maia 200, and Meta's MTIA are shipping or ramping, each on 3nm TSMC. Combined 2026 hyperscaler capex is tracking roughly $725 billion, with the custom-ASIC slice the fastest-growing allocation within it.

The TCO math is becoming hard to ignore. Industry analysis suggests custom ASICs are delivering a 40% to 65% total-cost-of-ownership advantage over merchant GPUs on captive workloads. Some analysts project Nvidia's share of the inference market could fall from above 90% today to the 20% to 30% range by 2028. Those are projections, not facts, and the moat around CUDA is real. But the direction holds: when your five largest customers each spend $60-plus billion a year and have in-house silicon programs, pricing power erodes from the top of the demand curve down.

Add the supply response. TSMC capacity at advanced nodes is expanding. HBM producers are adding fabs. RISC-V and open accelerator architectures are maturing. The second-source always arrives; the only question is which year it stops being optional.

What This Means for How We Underwrite

A commoditization thesis is only useful if it changes decisions. Four do, concretely.

First, manufacturing businesses adjacent to the AI buildout — precision machining for cooling systems, power distribution, enclosures, cable assemblies — are being underwritten today at multiples that assume the current demand curve continues. We prefer to buy them at multiples that assume it normalizes. The work is real and durable. The pricing power enjoyed during a capex supercycle is not. A business that earns a fair return at 2028 volumes is worth owning. One that only works at 2026 volumes is a trade.

Second, industrial services exposed to data-center construction — electrical contracting, mechanical systems, sitework — are structurally attractive, but the right lens is the cash-flow profile across a full cycle, not the current backlog. We discount backlog built during a scarcity moment. We pay up for recurring maintenance revenue, utility relationships, and the kind of regional density that holds pricing through a downturn.

Third, real estate near power. Land with firm interconnects, water rights, and permitted substation capacity will remain scarce long after the chips that run inside the buildings have commoditized. That asymmetry matters for how we think about sale-leaseback structures on industrial sites in the right geographies.

Fourth, the operating side. Falling inference prices are the single largest tailwind to the back-office work our portfolio companies are already doing with agentic systems, private LLMs, and automated customer-facing workflows. If a capability that cost $20 per million tokens three years ago now costs under $0.40, deploying it at scale across a holding company stops being an experiment and starts being an obvious operating decision. Our broader hardware thesis sits at hold.co/hardware; the applied work shows up quarter by quarter in the portfolio.

Decades, Not Quarters

None of this is a call against AI hardware as a sector. The buildout is real, the demand is real, and the total capital flowing into accelerators, memory, networking, and the facilities to house them is outpacing most prior categories of U.S. infrastructure on an inflation-adjusted basis. A great deal of economic value will be created. Some of it will accrue to shareholders of the chipmakers for years longer than the skeptics expect.

But a long-term holder is not paid to call the peak. A long-term holder is paid to own businesses whose economics survive the normalization. Hardware cycles compress. They have always compressed. The disciplined move is to buy the companies that sit downstream and alongside — the ones that benefit when compute gets cheap, not the ones whose multiples assume it never will.

Decades, not quarters. That has always been the frame, and the hardware cycle is a reminder of why.