From the EditorTwo of our desks spent this week asking what a piece of evidence could have shown if the answer had been no.
Our cover story, from the Parameter desk, tests the six-year useful life that cloud operators assign to AI accelerators against the only market evidence there is. On the surface the market agrees: a three-year-old H100 resells at 45–55% of its original value, and straight-line book value at that age is 50%. The agreement turns out to be weaker than it looks. Resale price is the rental stream discounted, so the two observables are one measurement. H100 spot rates rose 33% in the chip's third year, inside a shortage the vendor guides to last through fiscal 2028. The window in which a collapsing rate could disprove the six-year life closes before the life ends. The Reality Gap applies the same discipline to cultivated meat: approvals keep arriving, but the one approved factory to come up for sale drew a single bid, and that bidder withdrew.
A count of approvals or a run of confirming prices is a test only if it could have failed. Where we judge that it could not, we say so.
Twelve features and eight briefings from across the network.
— The Ecliptic, Editorial Desk
Cover Story · Compute & Intelligence · Parameter
Parameter — The Market Cannot Falsify the Six-Year Life Until 2029 — September 8, 2026
Rental rates and residual values both agree with a straight-line six-year book value. They are not two tests, they are one — and it is being run inside a shortage that suppresses the only signal that would disprove it.
PAR-0012 examined the six-year useful life hyperscalers apply to AI server equipment and asked whether it was defensible. This report asks the complementary question: does the market agree? Two observables exist — what a used accelerator sells for, and what one rents for. Assembled from 2026 public data, both track a straight-line six-year book value closely: a three-year-old H100 retains 45–55% of original value against a book figure of 50%. The agreement is weaker evidence than it looks, for three reasons developed here. The observables are not independent, since resale price is the discounted value of the rental stream. The test is being run during a supply shortage that the chip vendor guides to persist through fiscal 2028, which is most of the asset's assumed life. And the dispersion in rental rates — $1.00 to $11.07 per GPU-hour on a single day across 53 providers — makes the median a statistic that describes no transaction. The finding that matters is not that the six-year life is wrong. It is that H100 spot rates rose 33% year over year in the chip's third year, which implies the scarce asset is not the silicon but the datacenter slot it occupies — and those two things do not share a depreciation schedule.
Read the Graphic Edition →
Flagship Star Power
Star Power — The Hydrogen-Boron Bet Is a Confinement Test — September 30, 2026
In the same fortnight, ENN broke ground on Helong-2 and reported more than 10^8 hydrogen-boron reactions per second on EXL-50U. The reaction rate is about twelve orders of magnitude below a power plant. The machine being built is designed to test a confinement scaling law, not to make power. And the company's own reference reactor sends about 60% of its plasma exhaust out as X-rays.
China's ENN Group broke ground on Helong-2 (EHL-2), a hydrogen-boron spherical torus, on 15 September 2026. About two weeks later it reported more than 10^8 proton-boron-11 fusion reactions per second on its existing EXL-50U device. This report asks what the two events establish. We reconstruct ENN's published sequence…
Flagship Faraday
Faraday — No Fire for Two Hours — September 29, 2026
China's GB 38031-2025 moved the pass line from the cell stack to the pack boundary, and the calorimetry says most of a runaway cell's heat leaves through the vent — so the new standard is mainly a gas-path problem, and one chemistry pays far more for it than the other
China's GB 38031-2025 took effect for new vehicle type approvals on 1 July 2026. It replaces a five-minute warning before fire with a requirement that a pack in which one cell has been driven into thermal runaway shows no fire and no explosion, keeps smoke from harming occupants,…
Flagship The Gas Gauge
The Gas Gauge — The Yamal Cliff — September 29, 2026
On 1 January Europe stops taking the Russian Arctic LNG that supplied one tanker in seven last year, in the one season when Yamal cannot easily send those cargoes anywhere else. A US sanctions review due in mid-October could bring the cutoff forward by about ten weeks. On the shipping arithmetic, the ban removes supply from the world rather than moving it, and the lost volume is close in size to the Qatari capacity destroyed in March
From 1 January 2027 the European Union prohibits imports of Russian LNG under long-term contracts, and the Commission has said that EU operators may not carry or trade that LNG anywhere in the world. The ban falls mainly on one plant. Yamal LNG exported 19.7 Mt in 2025,…
Flagship The Maple Grid
The Maple Grid — Pickering Goes Dark — September 29, 2026
Ontario is switching off all four Pickering B reactors together for a C$26.8 billion refurbishment that costs three to four times Darlington's per kilowatt. OPG's request to raise its nuclear price by about two-thirds is mostly the arithmetic of the same switch-off.
On 30 September 2026 all four Pickering B reactors, which supply roughly a tenth of Ontario's electricity, go offline together for a C$26.8 billion refurbishment. They are not expected back until 2031–2034. This report reads the primary record: Ontario's approval, the September contract awards, the IESO's 2026 Annual…
Flagship Joule
Joule — Heavy Rare Earth Separation Outside China: Feed, Not Plants, Binds First — September 28, 2026
Announced dysprosium and terbium separation outside China adds up to about 2,310 t/yr of oxide by 2028, three to seven times what the announced new magnet plants would need, but only about 830 t/yr is tied to a named feed source, much of that is already allocated by nation, and the heavies are just 6–10% of what a heavy rare earth plant produces
TIB-0070 found European dysprosium oxide trading at roughly three times the Chinese export price as Europe's first integrated magnet plant began shipping, and asked whether magnet or separation capacity outside China would bind first. This report builds the map. It compiles every disclosed dysprosium and terbium separation project…
Flagship Parameter
Parameter — The Custom-Silicon Economics — July 23, 2026
Hyperscalers are building their own accelerators to escape the Nvidia margin — but the money they save lands mostly on Broadcom, TSMC and the memory makers, and the club stays closed
In 2026 custom AI silicon crossed from science-project to structural force: application-specific accelerators are on track to make up roughly 27.8% of AI-server accelerator shipments this year and to surpass merchant GPUs in unit shipments by 2027-2028 [5]. Every hyperscaler now runs its own program — Google's TPU…
Flagship The Reality Gap
The Reality Gap — The Factory Nobody Would Buy — October 1, 2026
Cultivated meat has close to a dozen regulatory approvals, the world's largest food bioreactor and a finished commercial plant. This summer the plant could not find a buyer at a third of its build cost, and the reason says more about the imagined future than any cost model does
Has cultivated meat crossed from laboratory curiosity to food system, as its 1931 prophecy and its 2021 investment peak both imagined? This report measures the sector on output, not approvals. It combines regulatory, financial and production disclosures from 2024 to September 2026 with three published techno-economic models, normalises…
Flagship Engines of Tomorrow
Engines of Tomorrow — The Thrust-to-Drag Ratio That Decides Air-Breathing Satellites — September 28, 2026
Four companies and DARPA are funding the first orbital flights of air-breathing electric propulsion. The best ground test so far produced thrust equal to about a quarter of the drag it measured. One ratio, set by intake efficiency, intake area and exhaust speed, shows why, and what the flight programmes will actually prove.
Air-breathing electric propulsion promises satellites that hold very low orbits indefinitely by using the residual atmosphere as propellant. In 2025-2026 it acquired a flight manifest: DARPA's Otter spacecraft ($44 million Phase 2), a Kreios Space demonstrator on a 200 kg bus announced on 4 August 2026, and NewOrbit's…
Flagship Delta-V
Delta-V — The Absentee Ratio: What Space-Based Interceptors Actually Cost — September 27, 2026
Published sizes for a boost-phase interceptor constellation run from 400 to 36,000 satellites and published costs span a factor of about twenty. Most of that spread is one piece of geometry, the square of the time an interceptor has to fly, and the cheapest lever for moving it is not launch price or lighter satellites but seconds.
Golden Dome's space-based interceptor layer is the most expensive single line in the largest national-security space program now in development: the Congressional Budget Office's May 2026 estimate puts it at $743 billion of a $1.19 trillion, 20-year total, for 7,800 satellites sized to engage 10 simultaneous ICBMs. This…
Flagship Chain Reaction
Chain Reaction — The Flywheel Runs Backwards Below One — September 9, 2026
A digital asset treasury company's entire mechanism is issuing equity above net asset value to buy more coin per share. At Strategy's 0.81x, the identical raise that added 8.4% to coin per share at a 7x premium now destroys 2.1% of it — and the accretive trade becomes selling the asset to buy back the stock
A digital asset treasury company is not a fund and not an operating business. Its mechanism is a single arbitrage: issue equity at a premium to the coin it holds, use the proceeds to buy more coin, and leave every existing share owning more coin than before. That…
Flagship Chain Reaction
Chain Reaction — Forty-Six Cents of Every Revenue Dollar — August 20, 2026
Circle paid Coinbase $324.6m in Q2 against $701m of revenue and reserve income. Nineteen percent circulation growth bought no improvement in the split, and the terms just renewed to 2029.
Builds on SAT-0014, which opened the question of how much of a stablecoin issuer's reserve income is contractually paid away to the venues that hold the float. For USDC the answer is now measurable and larger than the sector generally assumes: Circle paid Coinbase $324.6m in the second…
The Power Bill · 2026-10-01The Power Bill — Briefing — October 1, 2026
1. Britain's cap rises 4% on gas; a VAT cut hides a 6% rise in the electricity rate 2. US power plants burned record gas in July, at a cheaper price and a slightly…
Delta-V · 2026-10-01Delta-V — Briefing — October 1, 2026
1. The Tracking Layer satellite has cost $41-44 million at the floor since 2022 2. The fire-control premium, measured once inside one prime, is about 13% 3. Five primes, overlapping supply: the supplier map…
Joule · 2026-09-30Joule — Briefing — September 30, 2026
1. PJM is paying 71% more than its capacity cap for something the auction never offered: fifteen years 2. New York bought 950 MW of storage from 6 GW offered, and the contract refunds…
Chain Reaction · 2026-09-30Chain Reaction — Briefing — September 30, 2026
1. The carry ledger: three of seven treasuries cover their fixed charges from the coin 2. BitMine's 9.50% preferred costs 12.14% on the cash it raised 3. Strategy earns interest only on the dollars…
Beyond Solar · 2026-09-29Beyond Solar — Briefing — September 29, 2026
1. Enhanced geothermal's first power is one 33 MW block of a 900 MW plan 2. A superhot-rock developer raised $135 million as its first 15 MW slipped a year 3. A 30.77% tandem…
Parameter · 2026-09-28Parameter — Briefing — September 28, 2026
Builds on PAR-0016 — that Briefing reported GPT-6 Astra's launch at a higher price per token and a lower claimed price per task, and proposed rebuilding the price question on task cost. This issue…
The Reality Gap · 2026-09-27The Reality Gap — Fusion's Second Gain — September 27, 2026
1. Fusion has raised about a tenth of what its own companies say they need 2. NIF's gain is above four at the target and near 0.03 at the wall socket 3. ITER's Q=10…
Chain Reaction · 2026-09-27Chain Reaction — Briefing — September 27, 2026
1. Liquid's federation signed a 4,000 BTC withdrawal because the chain told it to 2. Bitget lost $351.6 million without losing a key 3. The Fed joins the stablecoin rulemaking last, and settles the…