THE LAST ENCLOSURE

Shanaka Anslem Perera

July 31, 2026

On June 12, 2026, Anthropic received an immediate export-control instruction from the United States government and suspended Claude Fable 5 and Claude Mythos 5 for every user worldwide. Fable had been generally available for three days. Mythos had been available only to a limited partner group.

Nothing in the company’s published account indicates a breach, outage, model defect or infrastructure failure. Access stopped because the new controls required restriction by nationality and Anthropic said it could not verify nationality across its global user base in real time. The controls were lifted on June 30. Fable returned globally on July 1. Mythos access had already resumed for a set of US organisations after separate approval, with broader access still under discussion.

What customers did during the interruption is not public. Migration, contract loss, fallback volume and recapture remain unidentified. The record establishes something narrower and more useful. A technically functioning service became globally unavailable through a compliance decision taken by a reachable provider. The intervention required no new Act of Congress, public rulemaking cycle or shareholder vote.

The event does not prove permanent economic change. It is a clean demonstration of service-discontinuity risk, and the argument that follows is about the clocks through which that risk reaches a balance sheet.

Each institution in this market measures its domain. Laboratories measure capability. Controllers estimate useful lives and depreciation. Lenders test coverage, collateral and maturity. Regulators measure compliance. Insurers estimate frequency, severity and accumulation. Each discipline can be competent and still leave the system incompletely observed, because each is reading a different clock.

Selected model capabilities and price-performance can reprice within months. That statement does not describe every task and does not convert unlike benchmarks into one universal frontier lag. It describes the speed with which adequate substitutes can appear for particular workloads.

Physical capacity moves on several slower schedules. Existing accelerators can be delivered or redeployed within months. Greenfield fabrication, data-centre construction, transmission and grid interconnection can require years. The clock varies, but capital cannot compress its slowest components.

Capital matures across years to decades. Servers, buildings, leases, power contracts, debt, purchase commitments and depreciation estimates have different lives, yet each can outlast several model generations and several rounds of price-performance repricing.

Evidence may decay whenever a material part of the deployed system changes. A model update, new tool, altered workflow, shifted data distribution or emerging threat can invalidate a previously sound claim while the certificate remains in the file.

Liability is allocated and priced before loss, but its practical incidence becomes observable after an incident, claim, enforcement action, settlement or judgment. The contract may allocate risk in advance. The real allocation is learned only when somebody invokes it.

Accounting standards already require useful-life, residual-value and impairment judgments. Credit analysts already examine obsolescence. Insurers already model concentration. Benchmark laboratories already measure capability. The missing object is not awareness. It is a standard public crosswalk that links the repricing clock of a comparable workload to the remaining duration of the assets and commitments financing it.

The exposure does not hide because nobody sees obsolescence. It hides because each profession sees a different part of the same balance sheet.

On Vercel’s AI Gateway in June, open-weight models handled roughly 29% of routed tokens while accounting for under 4% of spend imputed from published list prices. Anthropic represented about 61% of that imputed spend on roughly 32% of tokens, and at least 72% in every use case Vercel classifies as high-stakes.

Those figures describe Vercel’s gateway. Spend imputed from list prices is not provider revenue, realised invoices, margin or willingness to pay. Token volume is not work completed, value created or loss avoided. The high-stakes taxonomy is Vercel’s own, with no independent replication located here. Within those boundaries, the divergence is real. Volume moved toward cheaper models while imputed spend remained concentrated elsewhere.

Kimi K3 is a current test of what that divergence does and does not mean. Moonshot released the model in July and published its weights on July 27. It is a 2.8 trillion parameter mixture-of-experts system with 104 billion active parameters and a one-million-token context window. It is distributed under a custom licence granting broad use and modification rights, with conditions for certain model-service operators and very large products. The precise description is open-weight under a custom licence, not open source under every definition.

Capability remains task-conditional. The preliminary joint assessment by the UK AI Security Institute and the US Center for AI Standards and Innovation found K3 behind leading American systems on the tested cyber tasks. It reached an average of step 17 in a 32-step intrusion range against roughly 28.5 for the strongest comparison systems, and succeeded on 0 of 41 arbitrary-code-execution samples against roughly 20. The evaluators disclosed that American closed systems were tested with safeguards disabled to reduce refusals and measure maximum capability. The results support a narrow task comparison. They do not establish a general law about refusal behaviour, production safety or open versus closed deployment.

The inference is not that open models have won or the frontier stopped. Anthropic released Opus 5 on July 24, calling it frontier intelligence at half Fable’s cost. Capability reprices within the frontier, although adequacy depends on workload, configuration and institutional requirements.

Micron’s fiscal third-quarter materials describe a memory market in which demand exceeds supply and tight conditions extend beyond calendar 2027. The company reported that DRAM selling prices rose in the low-60% range sequentially and said HBM’s trade ratio keeps increasing. Its March 2024 industrywide estimate explicitly said that HBM3E consumed roughly three times the wafer supply of DDR5 to produce an equal number of bits in the same technology node.

That trade matters because capacity allocated to one product is not free capacity for another. Greenfield fabrication has long lead times. Data-centre projects unfold in phases. Grid interconnection can take several years, while brownfield retrofits, equipment deployment and software efficiency can arrive much faster. Capacity is not one clock, but the slow portions govern at the moment they bind.

Operators describe the constraint directly. Microsoft said on July 29 that Azure demand continued to exceed available capacity and that efficiency gains across its CPU and GPU fleet were monetised quickly. The same call supplied the strongest public counterargument. Microsoft described its infrastructure as “pretty fungible” across model families and customer choices. Scarcity and fungibility can coexist. One supports current utilisation. The other determines how much value survives when a specific workload is repriced.

Two prices have moved in opposite directions. Adequate machine cognition became far cheaper for many value-tier tasks while memory, energised capacity and advanced infrastructure remained constrained. The mechanism is conditional. If lower model prices expand usage faster than software and hardware efficiency reduce physical demand, abundance at the model layer becomes inflationary at the substrate layer. If efficiency or demand saturation dominates, it does not.

What makes capacity a clock is not that every component takes years. It is that the slowest indispensable complement sets the delivery date for the governed outcome.

Alphabet’s second-quarter filing is one of the most instructive documents in the sector because its operating result, investment mark and cash statement tell three different truths.

The company reported a pre-tax gain on equity securities of $99.031bn. Its disclosure states that the gain increased the tax provision by $21.9bn and net income by $77.1bn, adding $6.26 to diluted earnings per share. Net income available to common stockholders was approximately $112.1bn against income from operations of roughly $40.8bn. Comparing the pre-tax gain with after-tax net income is a category error. The filing says the gain arose primarily from SpaceX and one unnamed private company. It provides no issuer split and does not identify the second company.

Quarterly free cash flow was negative $5.9bn against capital expenditure of $44.9bn, while trailing twelve-month free cash flow remained positive at roughly $53.3bn. Full-year capex guidance rose to $195bn to $205bn. Operating strength, a net-income figure dominated by non-cash marks, and negative quarterly cash conversion under an exceptional investment load are separate facts. None invalidates the others.

The accounting clock moved again on July 29. Microsoft extended the estimated useful lives of datacentres and office buildings from 15 to 25 years. It said the change would have minimal benefit to fiscal 2027 operating income, but would cause more future datacentre leases to be classified as operating rather than finance leases, reducing the capex presentation to approximately $175bn while leaving calendar-year physical investment expectations unchanged. The change concerns buildings and datacentres, not a disclosed extension of GPU life. Its significance is mechanical. A change in estimated life can alter depreciation timing and lease classification without changing the underlying physical investment expectations that management publicly stated.