From the EditorA pattern runs through almost everything our desks filed this week, and it is worth naming: the number that settles the argument is the number nobody publishes.
Our cover story finds the six-year useful life that now carries hundreds of billions of dollars of AI infrastructure resting not on how long a chip works, but on what a five-year-old chip will still be worth doing — a residual value its own owners are spending heavily to compete away. The same shape appears three more times in this issue. Constellation operators disclose how often they dodge, never how hard, so the propellant cost of collision avoidance spans a factor of ten and nobody can close it. K-12 software is priced per enrolled student in states that have just made enrolment and activation different numbers, and no district has yet published the consent rate that decides the category's economics. Graduate programmes cross-subsidise the universities around them at a scale that appears in no filing, which is why the institutions most exposed to this summer's lending change are absent from every closure-risk list.
We would rather report a ten-fold range honestly than a false point estimate. Twelve features and eight briefings from across the Orrery network.
— The Ecliptic, Editorial Desk
Cover Story · Compute & Intelligence · Parameter
Parameter - The Six-Year Assumption - August 27, 2026
The depreciation schedule that carries AI capex rests on a residual-value tail its own owners are spending billions to compete away
Major cloud operators moved server and accelerator useful life from three or four years to six, cutting collective annual depreciation by roughly USD 18 billion and, on one 2024 estimate, from USD 39 billion to USD 21 billion — a 46% reduction in reported expense on an unchanged asset base. The extension is defended by a value cascade in which years one and two serve training, three and four real-time inference, and five and six batch inference and analytics. This report examines the terminal stage of that cascade and finds it is the stage most directly targeted by the operators' own custom inference silicon, which is deployed precisely because it beats older training GPUs on inference total cost of ownership. The six-year assumption therefore depends on a residual-value tail that its owners are spending capital to eliminate. We size the exposure, show why the 2025 divergence between operators is the most informative disclosure in the debate, and argue the estimate is an input to capital approval rather than a consequence of it.
Read the Graphic Edition →
Flagship Joule
Joule — The Factory Is Not Where the Money Is — August 24, 2026
Running every electrolyser factory in the world at gigawatt throughput would remove about USD 350/kW from an installed system priced near USD 2,000/kW. That is the entire prize from fixing utilization — roughly one-fifth of the capex, and less than a fifth of what a production tax credit is worth.
TIB-0051 found the world's electrolyser factories running near 5% utilization and argued a feedback loop: low throughput raises unit cost, which raises capex, which suppresses final investment decisions, which lowers throughput. It left the size of the effect unquantified. This report quantifies it. Using NREL's bottom-up manufacturing cost…
Flagship Beyond Solar
Beyond Solar — The Crossover Depth — August 22, 2026
Floating offshore wind's cost curve is nearly flat in water depth and fixed-bottom's is steep, so a crossover must exist. The expected crossover moved from 80 metres to 60 metres in six years, and that 20-metre shift is the only number in the sector that measures real progress
Does floating offshore wind have a definable point at which it becomes the cheaper foundation, and where is it? We build the question from the structural economics rather than from LCOE headlines: fixed-bottom foundation cost rises steeply with water depth because the structure must resist hydrostatic and wave…
Flagship Beyond Solar
Beyond Solar — The Tandem Haircut: What Survives the Trip From a 35.5% Cell to a Shipped Module
A perovskite-on-silicon cell certified at 35.5% in July. The best tandem module you can actually buy is 24.5%. The 11-point gap is not a laboratory curiosity — it is the entire investment case.
If a perovskite-on-silicon tandem cell is certified at 35.5% and the best silicon module in mass production is 25.0%, how much of that 10.5-point advantage survives the trip to a shipped, warranted product — and what is the surviving fraction actually worth in dollars? This report rebuilds the…
Flagship Joule
Joule — The Transformer Shortage Is a Steel Shortage — August 7, 2026
Grain-oriented electrical steel grew 14.3% last year; power-transformer demand has compounded at 14.0% since 2019. On those two rates the 30% deficit closes in roughly a century.
Why has a four-year transformer lead time persisted through a capital-investment cycle that should have relieved it? This report separates the transformer bottleneck into its assembly layer and its materials layer, and finds that the binding constraint has moved upstream to grain-oriented electrical steel. Global GOES output rose…
Flagship The Gas Gauge
The Marginal Bcf: Power Burn Versus LNG Feedgas Through 2027
The EIA's own July forecast has every cubic foot of 2027 supply growth already claimed — and the claim that cannot flex is the one that shows up for ninety days a summer
Does the United States have enough incremental gas to serve both the LNG buildout and a record power burn through 2027? This report rebuilds the balance directly from the EIA's July 2026 Short-Term Energy Outlook and normalises every claim against the same supply increment. Dry production is forecast…
Flagship Engines of Tomorrow
Engines of Tomorrow — The Power Ceiling Breaks: 20 kW in Orbit and the Industrialisation of Electric Propulsion
A 20 kW Hall thruster fired in orbit in March, a 120 kW lithium plasma thruster fired on the ground in February, and a 20 kWe reactor is manifested for Mars in 2028. Every one of those numbers is a power-supply number, not a thruster number.
Electric propulsion spent thirty years stuck below about 5 kW per thruster in flight. In five months of 2026 that ceiling broke three times: K2 Space fired a 20 kW Hall thruster in orbit in late March, JPL fired a lithium-fed magnetoplasmadynamic thruster at 120 kW on the…
Flagship The Reality Gap
The Reality Gap — The Data Centre Went to Space and Sent Back a Bill
Orbital compute is real, flying, and training models. The two most careful cost analyses ever published on it agree the whole idea turns on one number — and it is not a compute number.
TAC-0009 found that beaming solar power from orbit to Earth fails on economics and named one escape route: consume the power in orbit and skip the beam. Nine months on, that route has hardware on it — an NVIDIA H100 trained a language model in orbit, Google radiation-tested…
Flagship Delta-V
The Cost-to-Orbit Collapse: Reusable Heavy Lift
How reusability is dragging launch from $54,500/kg to a sub-$1,000/kg world — and what that unlocks
How far, how fast, and how durably is the marginal cost of reaching orbit falling — and what does a sub-$1,000/kg world unlock? This deep-dive reconstructs the $/kg trajectory from the Space Shuttle's ~$54,500/kg [1] to a reused Falcon 9 list price near $2,700–$3,250/kg [2][3] and an internal…
Flagship Chain Reaction
Chain Reaction — The Rent Is $1.8 Million a Megawatt-Year — August 7, 2026
Four of the largest miner-to-AI leases price within 5% of each other on a per-megawatt-year basis. The one investment-grade direct lease prices 32% below them — and that discount is the whole story.
Public bitcoin miners have announced more than $70 billion of AI and HPC contract value, and the sector is being valued on those headline totals. This report converts each disclosed lease into a single comparable metric — contracted revenue per critical megawatt per year — and finds a…
Flagship Chain Reaction
Power Landlords of AI: The Bitcoin-Miner Pivot and the $50 Billion Execution Gap
Public bitcoin miners have signed $110B+ of AI/HPC leases in two years — reclassifying grid access as the asset. Only about a quarter of it has been delivered. This report models the capital-allocation math, the second-order counterparty concentration, and who is exposed.
Can a bitcoin miner become an AI-infrastructure company, and what is that transition actually worth? Public miners have announced more than $110 billion of AI and high- performance-computing (HPC) leases across roughly 17 deals and ~6 GW in two years, with listed miners projected to derive up to…
Delta-V · 2026-08-27Delta-V - The Tank Is the Constraint - August 27, 2026
Builds on DLV-0014 — *The Maneuver Burden* (August 10, 2026), whose second thread asked whether avoidance is now competing with station-keeping and deorbit reserve for the same tank, and proposed deriving end-of-life mass margins…
The Reality Gap · 2026-08-25The Reality Gap — Briefing — August 25, 2026
1. A five-year slip delivered in one line of a quarterly report 2. Zero certificated electric aircraft, and an independent charging estate of about thirteen airports 3. The robotaxi gap is no longer capability…
The Gas Gauge · 2026-08-25The Gas Gauge — Briefing — August 25, 2026
1. The 2026 price deck was cut more than 6% into a demand forecast that is still rising 2. A record end-October inventory that is 1.5 days of production 3. Production sets an annual…
Engines of Tomorrow · 2026-08-25Engines of Tomorrow — The Certification Clock — August 25, 2026
1. The number does not exist, and the regulator declines to estimate it 2. Electric propulsion is 0.137% of the FAA's certification workload 3. The existing-standards path produces a certification basis 2.8× as often…
Beyond Solar · 2026-08-25Beyond Solar — The Offtake-First Screen — August 25, 2026
1. A backlog that went from roughly 20 GWh to 80 GWh 2. The $750 million arrived after the contracts, not before them 3. Three buyer species in a single announcement, and only one…
The Power Bill · 2026-08-24The Power Bill — Briefing — August 24, 2026
Builds on CAL-0021 — that report ended by asking whether PJM's demand curve can reproduce MISO's result, noting that MISO cleared 9.1 GW of demand response at a 100% clearance rate against three consecutive…
Star Power · 2026-08-24Star Power — Briefing — August 24, 2026
Builds on TRP-0013 — this briefing takes up that report's first thread, the concentration ratio as an industry-health metric: tracking one developer's share of cumulative fusion capital, on the reasoning that a falling share…
Chain Reaction · 2026-08-24Chain Reaction — Briefing — August 24, 2026
The week's move was large, fast and almost entirely exogenous. Both named catalysts came from outside the asset class — one from the Treasury's debt management desk, one from the White House — and…