The Carcinisation of Compute

Wang Dongsheng's 2010 account of the display-panel industry described a recurring pressure: standard product prices fall by roughly half over a three-year interval while performance and effective technical capability must continue to improve. The curve is not a promise that profits always fall and then recover. It is a survival relationship between price pressure and value creation.

In compute, preserve the survival relationship — not the panel industry's exact cadence or outcome.

Stanford AI Index 2025 recorded the price of model calls at GPT-3.5-level MMLU performance falling from approximately $20 per million tokens in November 2022 to approximately $0.07 in October 2024 — a decline of more than 280×. This is a cross-model comparison at a defined capability threshold. It is not evidence that a specific GPU rental rate, purchase price, or operating cost fell by 280×.

In a simple explanatory model, if the selling-price index falls from 100 to 50 while cost only falls from 70 to 60, unit profit moves from 30 to negative 10. If engineering lowers cost to 30, the same price pressure can still leave 20 of unit profit. Procurement is only one lever: effective throughput per watt, communication wait time, failure recovery, and sellable capacity utilization matter just as much.

CapEx is a one-time asset purchase; OpEx is incurred throughout operations. Comparing equipment purchase totals with one year of electricity cannot show which cost dominates. This analysis annualizes equipment and facility investment over their economic lives before comparing them with electricity and maintenance.

The direction of the mining analogy is useful: if equipment power grows faster than its purchase price, hardware dollars per watt decline and electricity gains relative weight. The financial result still depends on delivered power price, load factor, facility losses, and economic life — not on watts per machine alone.

To isolate hardware deflation, the model holds a 1 MW rated IT facility constant: 90% average IT load, PUE 1.15, delivered electricity at $0.06/kWh, a 4-year hardware life, $7 million/MW of facility investment amortized over 15 years, and $200,000 of annual other operations and maintenance. The scenarios are stress tests, not forecasts.

Electricity becomes the largest single line item in the strongest stress case, but remains below half of total annualized cost. Carcinisation does not require electricity to exceed 50% of TCO. A rising share, thinner margins, and constrained supply are already sufficient to make energy engineering more important.

Under the model assumptions, a 1 MW facility purchases approximately 9.0666 million kWh per year. Moving delivered power from $0.06 to $0.03/kWh saves approximately $272,000 per year: 4.4% of baseline total cost, but 18.6% of the 95%-hardware-deflation scenario. The same energy advantage matters much more after hardware deflation.

The checkmark curve describes the competitive survival requirement. CapEx/OpEx analysis describes how that pressure enters the cost structure. Carcinisation describes the engineering response. These mechanisms are connected, but they do not require a single fixed sequence.

Reduce redundancy only after comparing the annualized cost of additional redundancy with the expected loss from interruption. Checkpointing can reduce training rollback loss; it cannot keep cooling pumps, switches, storage, or compute operating during a power failure.

Hyperion is a starting point in the sense that it makes compute carcinisation visible at gigawatt scale, with long-life infrastructure and institutional capital. It is not the first industrial compute campus, nor proof that the final commercial model has been validated.

Meta announced in July 2026 that the Richland Parish, Louisiana campus would expand to 5 GW, alongside more than $50 billion of regional investment. Those figures describe planned scale and a regional investment disclosure; they do not mean 5 GW is already operating or establish a uniform facility cost per MW.

Public materials support Hyperion's direction toward closed-loop liquid cooling, full grid connection, and next-generation networking. They do not establish a whole-campus tent deployment, a universal UPS configuration, a fixed PUE, or second-by-second fault recovery.

Industrialization changes the reliability budget. It does not remove reliability requirements.

A prior scenario used 5,000 MW and hypothetical facility unit costs of $9.0 million/MW versus $5.25 million/MW. The arithmetic is valid: $45.0 billion versus $26.25 billion, a $18.75 billion difference and a 41.7% reduction. The unit costs lack a verifiable project budget basis, so this is an illustrative model, not a claim that Hyperion achieved those savings.

The correct optimization target is the discounted cost of effective output. Lower facility investment can reduce annualized burden, but equipment idle time, slow recovery, maintenance, or inadequate networking can raise the final cost per delivered task.

If hardware deflation continues, investment analysis should place more weight on deliverable power, complete delivered-energy cost, cross-generation facility adaptability, and customer contract quality. Hardware supply, network capacity, software efficiency, and customer acquisition remain material; the system does not reduce to electricity alone.

The 1 MW model uses 1 MW rated IT power × 90% average IT load × 8,760 hours × 1.15 PUE = 9,066.6 MWh/year. At $60/MWh, annual electricity is $543,996. Facility annualization is $7,000,000 ÷ 15 = $466,667; other O&M is $200,000; hardware annualization is $5,000,000, $1,000,000, or $250,000 across the three scenarios.

The revision removes unsupported universal thresholds: the claim that HBM is always 63% of component cost, that checkpointing replaces UPS, that Hyperion uses tent structures campus-wide, and that a modeled $18.75 billion difference represents realized Hyperion savings. These remain research questions or illustrative arithmetic, not established facts.