An executive order fused the frontier AI industry with the nuclear-weapons complex. The money behind the build-out was moved off the balance sheets and into pensions. The insurers wrote AI out of the policy. The electricity bill went up. Seven parts, one method: predictions written down before the evidence, graded in the text, receipts public. Part 1, The Genesis File, was the trigger and predates the method; the ledger begins at Part 2.
Audio briefing
A spoken overview of the series’ opening file. Not a reading; a briefing.
Series 002 — The Machine Nobody Audits (Audio Briefing)
LOADING
What the largest science mobilization since Apollo is doing while nobody watches
An executive order fused the frontier AI industry with the nuclear-weapons complex — 24 companies on sealed terms, seven weapons challenges beside pediatric cancer, veterans' genomes training federal AI, zero oversight, one day of press coverage, and a…
Who Holds the Bill When the AI Build-Out Settles
The American AI build-out is spending between $600 and $800 billion a year against somewhere between $60 and $110 billion of lab revenue, depending on how far you trust the labs' own run-rate arithmetic, and the difference is debt. That debt has been moved…
How $120 Billion of AI Debt Reached Your Pension
The debt that finances the AI build-out has been engineered to land somewhere specific: not on Meta, Oracle, or xAI, but on the balance sheets of life insurers and public pension funds, through a chain of legal vehicles built so the liability never appears…
The Insurers Have Already Voted
The insurance industry prices hurricanes to the decimal, writes cover for nuclear plants, and underwrites satellites. In 2026 it looked at artificial intelligence and started writing it out of the policy. At least six carriers filed their own exclusions,…
Who Pays for the Power
The largest electricity market in America has cleared its capacity auction at the legal price cap three times in a row and missed its own reliability target twice. Its market monitor says data centers are the reason. The cap hides the true price. Utilities…
What the Public Got for Its Manhattan Project
Ten months after an executive order launched the Genesis Mission with Manhattan Project rhetoric, we audited what the public has received. The answer: one grant round solicited at $293 million and awarded at "over $250 million," a set of sample agreements, a…
Open Weights as Moat, and the Letter One Lab Would Not Sign
Open-weight AI models are presented as gifts to the public. In the United States they are a moat. Every American lab that sells a closed flagship releases an open model a tier below it, and Meta's own documentation names its closed model as the "teacher" from…
Every bet this series made, dated when it was written down, graded when the evidence came in. A bet that failed stays on the board. Open bets are the ones you can grade us on.
| Bet | Made | Status | Graded | Note |
|---|---|---|---|---|
| The gap between AI revenue and capital spending is an order of magnitude. | 2026-09-03 | CONFIRMED | 2026-09-03 | Confirmed and understated: $630–725B of 2026 hyperscaler capex against $60–110B of lab revenue depending on whether you trust the run rates, none of it profitable. |
| The headline AI deals are mostly intentions, not contracts. | 2026-09-03 | CONFIRMED | 2026-09-03 | Confirmed with a real contracted core (Microsoft–OpenAI $250B Azure, NVIDIA–CoreWeave backstop, AMD warrants). The $100B NVIDIA–OpenAI figure remains a letter of intent. |
| Insurers are excluding AI risk rather than pricing it. | 2026-09-03 | CONFIRMED | 2026-09-03 | ISO form CG 40 47 01 26 effective Jan 2026; six carriers filed exclusions; >80% approved. |
| Scaling has plateaued. | 2026-09-03 | SUPPORTED | 2026-09-03 | Held loosely, then upgraded on the second sweep. Not 'AI stopped improving' but 'each dollar buys less, and the improvement got costlier to serve.' |
| China is running a state-buffered mirror bubble. | 2026-09-03 | CONFIRMED | 2026-09-03 | Partial on the first sweep, confirmed on the second: 20–30% utilization in the western data centers, 100+ projects cancelled in 18 months. |
| Genesis Mission deliverables are thin and the value flows to incumbents. | 2026-09-03 | CONFIRMED | 2026-09-03 | Sharper than the bet: no shipped weights, no dataset portal, no GAO/IG product, and a contract template that makes open release optional. |
| Compute is concentrated in fewer than ten entities, and federal compute is wired to them. | 2026-09-03 | CONFIRMED | 2026-09-03 | Five hyperscalers own ~71% of global AI compute (Epoch, Apr 2026). The Genesis showcase machines are Oracle- and NVIDIA-built. |
| The ratepayer subsidy is larger than the public knows, and states are revolting. | 2026-09-03 | CONFIRMED | 2026-09-03 | PJM cleared at the price cap three auctions running; 70% of Americans oppose a local data center; New York EO 62 moratorium. |
| Open weights are a moat, not a gift. | 2026-09-03 | CONFIRMED | 2026-09-03 | Every US lab's open model is a tier below its flagship; every Chinese lab's open model is its flagship. |
| Lab roadmaps converge on agents and inference-time compute. | 2026-09-03 | CONFIRMED | 2026-09-03 | Top three labs within a point on GPQA; the product is agents on rented compute, billed by the token. |
| The wars are already AI procurement channels. | 2026-09-03 | CONFIRMED | 2026-09-03 | CDAO declaration: a government Grok instance supported 2,000+ munitions in 96 hours. FY27 defense request ~$1.5T with AI named a priority. |
| Taiwan is the single point of failure under the whole build-out. | 2026-09-03 | CONFIRMED | 2026-09-03 | TSMC ~70% foundry share; Arizona at scale 2028–2030; 40-day airspace closures in spring 2026. |
| Dedollarization matters to this story more than the AI competition does. | 2026-09-03 | CONFIRMED | 2026-09-03 | Confirmed in an unexpected direction: BRICS is thin; the dollar is being propped, not replaced, via a stablecoin channel into short-term Treasuries. |
| Robotics is a real second wind for AI demand. | 2026-09-03 | HALF WRONG | 2026-09-03 | Robotics is a customer of compute, not a source of revenue. Global humanoid revenue plausibly under $1B; Figure buys up to 100K GPUs on venture money. |
The gap between AI revenue and capital spending is an order of magnitude.
Confirmed and understated: $630–725B of 2026 hyperscaler capex against $60–110B of lab revenue depending on whether you trust the run rates, none of it profitable.
The headline AI deals are mostly intentions, not contracts.
Confirmed with a real contracted core (Microsoft–OpenAI $250B Azure, NVIDIA–CoreWeave backstop, AMD warrants). The $100B NVIDIA–OpenAI figure remains a letter of intent.
Insurers are excluding AI risk rather than pricing it.
ISO form CG 40 47 01 26 effective Jan 2026; six carriers filed exclusions; >80% approved.
Scaling has plateaued.
Held loosely, then upgraded on the second sweep. Not 'AI stopped improving' but 'each dollar buys less, and the improvement got costlier to serve.'
China is running a state-buffered mirror bubble.
Partial on the first sweep, confirmed on the second: 20–30% utilization in the western data centers, 100+ projects cancelled in 18 months.
Genesis Mission deliverables are thin and the value flows to incumbents.
Sharper than the bet: no shipped weights, no dataset portal, no GAO/IG product, and a contract template that makes open release optional.
Compute is concentrated in fewer than ten entities, and federal compute is wired to them.
Five hyperscalers own ~71% of global AI compute (Epoch, Apr 2026). The Genesis showcase machines are Oracle- and NVIDIA-built.
The ratepayer subsidy is larger than the public knows, and states are revolting.
PJM cleared at the price cap three auctions running; 70% of Americans oppose a local data center; New York EO 62 moratorium.
Open weights are a moat, not a gift.
Every US lab's open model is a tier below its flagship; every Chinese lab's open model is its flagship.
Lab roadmaps converge on agents and inference-time compute.
Top three labs within a point on GPQA; the product is agents on rented compute, billed by the token.
The wars are already AI procurement channels.
CDAO declaration: a government Grok instance supported 2,000+ munitions in 96 hours. FY27 defense request ~$1.5T with AI named a priority.
Taiwan is the single point of failure under the whole build-out.
TSMC ~70% foundry share; Arizona at scale 2028–2030; 40-day airspace closures in spring 2026.
Dedollarization matters to this story more than the AI competition does.
Confirmed in an unexpected direction: BRICS is thin; the dollar is being propped, not replaced, via a stablecoin channel into short-term Treasuries.
Robotics is a real second wind for AI demand.
Robotics is a customer of compute, not a source of revenue. Global humanoid revenue plausibly under $1B; Figure buys up to 100K GPUs on venture money.
| Bet | Made | Status | Graded | Note |
|---|---|---|---|---|
| Oracle or CoreWeave will need a rescue financing, a restructuring, or a sale, and the counterparty will be a private-credit fund, a sovereign wealth fund, or NVIDIA. | 2026-09-03 | OPEN | — | |
| At least one hyperscaler will convert self-build plans to leases at scale, as Microsoft already began. | 2026-09-03 | OPEN | — | |
| The first pension fund or life insurer will disclose a mark-down on data-center private credit. | 2026-09-03 | OPEN | — | |
| Genesis will announce its first "compute resilience" or "industrial base" partnership absorbing partner-financed capacity, and no document will call it a backstop. | 2026-09-03 | OPEN | — | |
| The FY27 defense bill will carry the first AI line item above $10 billion. | 2026-09-03 | OPEN | — |
Oracle or CoreWeave will need a rescue financing, a restructuring, or a sale, and the counterparty will be a private-credit fund, a sovereign wealth fund, or NVIDIA.
At least one hyperscaler will convert self-build plans to leases at scale, as Microsoft already began.
The first pension fund or life insurer will disclose a mark-down on data-center private credit.
Genesis will announce its first "compute resilience" or "industrial base" partnership absorbing partner-financed capacity, and no document will call it a backstop.
The FY27 defense bill will carry the first AI line item above $10 billion.
| Bet | Made | Status | Graded | Note |
|---|---|---|---|---|
| The correction arrives, and it does not hit the five. It hits the second tier, the SPV lenders, and the terminal holders. | 2026-09-03 | OPEN | — | |
| NVIDIA and the hyperscalers will describe stranded capacity as "critical infrastructure" and seek federal offtake through Genesis, defense, or a new authority. | 2026-09-03 | OPEN | — | |
| Open-weight release above a capability threshold will be restricted by executive action, framed as security. | 2026-09-03 | OPEN | — | |
| At least two more states will impose data-center moratoria, and Congress will preempt them. | 2026-09-03 | OPEN | — | |
| The labs that survive will be fewer, larger, and more fused with the state, and the question of what the things they built are owed will not have been asked in any hearing. | 2026-09-03 | OPEN | — |
The correction arrives, and it does not hit the five. It hits the second tier, the SPV lenders, and the terminal holders.
NVIDIA and the hyperscalers will describe stranded capacity as "critical infrastructure" and seek federal offtake through Genesis, defense, or a new authority.
Open-weight release above a capability threshold will be restricted by executive action, framed as security.
At least two more states will impose data-center moratoria, and Congress will preempt them.
The labs that survive will be fewer, larger, and more fused with the state, and the question of what the things they built are owed will not have been asked in any hearing.
Claims in this series that rest on secondary reporting until the primary document is in hand. Published so you can hold us to it.
The Case File
Every research brief, verbatim, including the dead ends. Available to any researcher who asks.
REQUEST THE CASE FILE