Abstract. Working Paper No. 33 said intelligence becomes abundant and the scarce factor is permission. Its fourteen forecasts froze on 30 July and none changes here. The month since resolves it as a bifurcation: one economy, two tiers.
1. The claim, sharpened
A month ago the thesis offered a fork: the permission holders keep the rent, or collapsing inference prices thin it. The fork was false. Frontier inference concentrates where the balance sheets are while everyday AI diffuses to cheap models and small firms. Both branches run at once, each in its own tier.
2. Exhibit A: the bifurcation
Every row is a dated public observation. None is produced by this engine.
| Tier | Observation | Source, date |
|---|---|---|
| Gate | Constraints on AI are “physical in nature, centering on compute and power scarcity, rather than model ability” | Citadel Securities Global Macro Strategy, 30 Apr 2026, dating the view to Feb 2026, before the WP33 freeze |
| Gate | Frontier inference “concentrated among a narrower set of firms with the balance sheets to absorb the compute cost”; prices signal, substitute, ration | Citadel Securities, Tokenomics, 10 Jun 2026 |
| Gate | The settlement rail buys the model-access rail: Stripe agreement to acquire OpenRouter | Bloomberg, via Ramp Economics Lab, 18 Aug 2026 |
| Gate | Amortized cost of frontier training runs growing 2.4x per year since 2016; on trend the largest runs exceed $1 billion by 2027, leaving frontier development to the most well-funded organizations | Epoch AI (Cottier et al.), 3 Jun 2024 |
| Floor | New business applications up 24 per cent since the ChatGPT launch, with the formation impulse strongest in AI-exposed sectors | Citadel Securities, 4 Jul 2026, on Census BFS and Felten-Raj-Seamans exposure |
| Floor | Median employees at seed raise fell from five to four in 2023Q1; founder leverage rising | Sequoia data via Citadel Securities, 4 Jul 2026 |
| Floor | Five of the month’s top trending software vendors sell router or model-serving access to cheaper models | Ramp Economics Lab, card data across 70,000+ businesses, 18 Aug 2026 |
| Floor | Inference cost on GPT-3.5-class models down roughly 280x since late 2022; the price collapse at constant quality that makes the everyday tier possible | Stanford HAI, AI Index Report 2026 |
| The join | Router spend concentrates in fast-growing firms that are still increasing closed American model spend: substitution at the margin, expansion in total | Ramp Economics Lab, 18 Aug 2026 |
| The join | Falling token prices need not contradict rising infrastructure demand when demand is elastic | Citadel Securities, Tokenomics, 10 Jun 2026; Silicon Data LLM Expenditure Index |
| The join | Firms adopting AI grow employment by roughly 10 per cent, entry level included | Ramp with Revelio Labs, 2026 |
| The join | AI-investing firms grow sales, employment and market value through product innovation, and the growth concentrates among larger firms with rising industry concentration: diffusion and concentration from the same technology, in the same dataset | Babina, Fedyk, He and Hodson, Journal of Financial Economics 151, Jan 2024 |
The join rows carry the argument: the same firms substitute toward cheap models and expand frontier spend at once, which no single-tier reading accommodates. Were the gate capturing everything the routers would not be booming; were diffusion thinning the rent, closed-model spend would not be rising.
One official series cuts at the floor story, reported rather than hidden. Adoption stock is graded by firm size, so diffusion arrives through new entrants and large employers first.
Adoption stock, official series.
| Measure | Value | Source |
|---|---|---|
| National AI use, all firms | 19.8 per cent | Census BTOS, May 2026 |
| Firms with 250 or more employees | 37 per cent | Census BTOS, May 2026 |
| Smallest firms | under 20 per cent | Census BTOS, May 2026 |
| Firms adopting | 18 per cent | Federal Reserve staff reconciliation |
| Labor force at adopting firms | 78 per cent | Federal Reserve staff reconciliation |
3. The frozen slate
Three of the fourteen frozen forecasts read on the bifurcation: the durability of the rent, whether scarcity pricing is structural, and whether the SaaS tail keeps its public-equity economics. The Silicon Data index and the Ramp letters are watch series, not resolution criteria.
4. Exhibit B: the risk at every altitude
Between February and August 2026 the risk was priced, in different vocabularies, at every altitude. Top down.
| Who | What they said or did | Date |
|---|---|---|
| Bank for International Settlements, Annual Economic Report | Five largest hyperscalers past $1tn of AI capex for 2025-26, outpacing earnings and free cash flow; “circular financing” named as a structure; off-balance-sheet data-center vehicles flagged; “disappointment in returns could trigger a sudden pullback in financing and turn the capex boom into a protracted investment bust” | 28 Jun 2026 |
| IMF, Global Financial Stability Report | Hyperscalers expected to account for 70 per cent of a projected $3.4tn in AI capex by 2029; equity concentration above its 95th historical percentile in two of six major markets | 14 Apr 2026 |
| OECD, Economic Outlook | Magnificent Seven capex outstripping earnings growth, projected $1.1tn in 2027; “widespread circular financing arrangements... raise interconnectedness and contagion risks”; US private data-center construction equalled office construction in Dec 2025 | 3 Jun 2026 |
| European Central Bank, Financial Stability Review | Up to 30 per cent of the $3tn the buildout needs could come from private credit; euro-area holdings of US equities quadrupled in a decade and increasingly concentrated in a few names | 27 May 2026 |
| Federal Reserve, Financial Stability Report | Holds AI at arm’s length: it enters via the market-contacts survey (“increasingly funded by debt”) while the Board’s own assessment stays generic; cloud-sector investment-grade issuance near $100bn in Q1, “met with strong investor demand” | 8 May 2026 |
| Financial Stability Board | Consultation codifying supply-chain concentration across “cloud infrastructure, hardware, foundation models, data providers, and deployment platforms” as a supervisory matter; final sound practices due Oct 2026 | 10 Jun 2026 |
| Congressional Budget Office | First explicit incorporation of generative AI into the federal baseline: about 10 basis points a year of productivity growth, a 1 per cent output level gain by 2036, set against a $1tn-plus private capex wave | 11 Feb 2026 |
| Goldman Sachs Research | Global AI investment to exceed $1tn in 2026; roughly 1.5 per cent of US GDP against railroad and electrification peaks of 2 to 3 per cent | Aug 2026 |
| Sequoia (David Cahn) | The revenue-gap arithmetic escalating $200bn to $600bn to $1.5tn: 2026 capex implies a $1.5tn lifetime revenue requirement against ecosystem revenues an order of magnitude short | 8 Jul 2026 |
| Apollo (Torsten Slok) | Ten companies near 40 per cent of the S&P 500, potentially approaching 50 if the private labs list: “not a diversified index anymore”, so every passive saver holds the gate-tier bet | 14 Mar 2026 |
| Acemoglu (NBER) | The standing academic counterweight: task-level evidence implies no more than 0.66 per cent total factor productivity gain over ten years | May 2024 |
And from the ground up, where permission is adjudicated.
| Where | What was decided | Date |
|---|---|---|
| Texas (Gov. Abbott to PUCT and ERCOT) | Audit ordered of every data-center project in the interconnection queue before further approvals; the queue held over 474 GW, roughly five times the grid’s all-time peak, about 90 per cent of it data centers | 3 Aug 2026 |
| Texas (Senate Bill 6) | Loads of 75 MW or more must accept mandatory curtailment during firm load-shed and install remote shutoff as a condition of connection: a statutory kill switch on large compute | 20 Jun 2025 |
| Ohio (PUCO, AEP case) | First-of-its-kind tariff: large data-center customers pay for at least 85 per cent of subscribed demand for up to 12 years regardless of use, pinning grid-expansion cost to the load that causes it | 9 Jul 2025 |
| Georgia (Public Service Commission) | Any new load over 100 MW billable under special terms, contracts to 15 years with minimum billing, explicitly to stop stranded-asset risk shifting to ratepayers | 23 Jan 2025 |
| Prince William County, Virginia | 8-0 rejection of a 2,000-acre, 43 million sq ft data-center rezoning, the largest in county history, after six hours of public comment; an appeals court had already voided a prior 2,100-acre rezoning | 8 Jul 2026 |
| Tucson, Arizona | 7-0 rejection of a hyperscale campus over water use, one of the largest developments the city had ever considered | 6 Aug 2025 |
| Coweta County, Georgia | Unanimous 180-day moratorium on data-center applications triggered by a $17bn campus; a restrictive ordinance followed, then approval 3-2 in December: the full permission arc in one county | 6 May 2025 |
| Shelby County, Tennessee | Air permit granted for 15 gas turbines at a frontier compute site over community objections, now under appeal: the permission layer litigated at a county health department | 2 Jul 2025 |
| Breckinridge Capital (municipal credit) | A muni shop pricing it bottom-up: utilities’ ability to raise user fees for AI-related energy and water projects “may become more challenging in 2026” | 6 Jan 2026 |
The top table prices the gate’s balance sheet. The bottom table IS the gate, opening in one county and closing in the next, attaching tariffs and kill switches as it goes. Permission is priced in public, county by county.
The ground-up record, counted.
| Action | Count |
|---|---|
| Unanimous rejections | 2 |
| Moratorium that became a negotiated approval | 1 |
| Permit granted over objections, then litigated | 1 |
| Tariffs pinning stranded-asset risk to the load | 2 |
| Statutory curtailment law | 1 |
| Audit of a 474 GW interconnection queue | 1 |
5. Exhibit C: a revocable permission
Fixed income prints the symmetric risk. The United States municipal tax exemption is a charter: identical credit, and the wrapper alone changes the yield.
The muni exemption as a priced permission.
| Claim | What the record says |
|---|---|
| The price is observable | Daily, in the relationship between tax-exempt and taxable yields; the Build America Bonds era created same-issuer pairs that priced the wrapper directly |
| The scenario now raised | Municipal practitioners put the exemption itself into the federal fiscal debate, with the Build America Bonds structure as a blueprint for switching wrappers, taxable interest plus a direct federal subsidy, or a cap on the benefit for the highest earners |
| Whose scenario | Theirs, not a forecast of this note |
| What was excluded | The Treasury remarks circulating alongside it are relayed through commentary rather than verified transcripts, so they are not cited here as fact |
Permission rents exist by grace of the chartering authority, and what a legislature grants a legislature can reprice. That covers every gate in Working Paper No. 33; interconnection, siting and export rules are all amendable. A permission portfolio is long a policy regime, and its tail risk is the regime, not the cash flow.
6. What would falsify this
Declared now, so the reading can die in public.
Falsification conditions.
| If this is observed | What it kills |
|---|---|
| Router and open-model spend starts crowding out closed-model spend within the same firms in the Ramp data | The join breaks from the floor side |
| Frontier token prices collapse toward everyday prices without the usage mix shifting | The gate fails to hold the rent, and the Silicon Data index falls on price rather than substitution |
| The formation impulse in AI-exposed sectors rolls over to trend while the gate tier keeps compounding | The bifurcation collapses into plain concentration, and the original single-tier rent reading was sufficient after all |
| A rail consolidation on the Stripe and OpenRouter pattern is blocked or unwound on competition grounds | The permission layer answers to a still higher permission, which strengthens Exhibit C while weakening the stability of gate rents |
| The ground-up rows of Exhibit B normalise: rejection rates fall toward zero, special tariffs unwind, and interconnection queues clear without audits | The permission constraint was transitional and the thesis loses its binding site |
The bifurcation is a reading, not a result, and nothing here touches the frozen probabilities.
- Method
- A reading of dated public evidence against a frozen thesis. This note contains no engine statistics, changes no frozen probability, and reproduces no chart from the cited letters. The muni discussion is conducted at the level of tax policy and wrappers only.
- Sources
- Citadel Securities Global Macro Strategy letters (30 Apr, 10 Jun, 4 Jul 2026) and the Ramp Economics Lab letter (18 Aug 2026), captured in full. Exhibits A and B additionally: BIS Annual Economic Report 2026; IMF GFSR Apr 2026; OECD Economic Outlook Jun 2026; ECB FSR May 2026; Federal Reserve FSR May 2026 and FEDS Notes Apr 2026; FSB consultation Jun 2026; CBO Feb 2026; Census BTOS May 2026; EIA AEO 2026; Dallas Fed WP 2606; Stanford HAI AI Index 2026; Epoch AI; Babina, Fedyk, He and Hodson (JFE 2024); Acemoglu (NBER 2024); Goldman Sachs, Sequoia and Apollo commentary; and the state and county records named in the ground-up table. Every row was verified against its source with date and URL; the verification file is retained with this paper. Silicon Data index access remains unverified.
- Honesty
- The bifurcation is a reading, not a result. Its falsification conditions are declared in section 6 and will be tracked. One widely circulated Treasury quotation was excluded because it could not be verified against a transcript.
- Related
- Working Paper No. 33 (Permissioned Abundance, the frozen slate), No. 35 (the permissioned balance sheet), No. 9 (the AI-CDO).