Near-zero code-generation cost will translate into near-zero prices for complete enterprise products.

Code is only one input. Data migration, reliability, permissions, compliance, integration, support, distribution, and accountability remain costly and can sustain vendor pricing even when implementation gets cheaper.

Most large enterprises will prefer bespoke AI-built workflows over broad ERP, CRM, and productivity suites.

AI will expand internal development and replace some point tools, but core systems benefit from standardization, ecosystems, auditability, and someone else carrying operational risk. Adoption will vary sharply by workflow.

The incumbent software market will shrink to roughly one-tenth of its current size.

The figure is unsupported, and spending data through 2026 points the other way. A shift in market composition and pricing is plausible; a precise 90% collapse needs far more evidence.

AI agents can move spending from labor budgets into a larger software addressable market.

Agent products increasingly automate service workflows, so some labor and outsourcing spend can become software revenue. The magnitude and net employment effect remain uncertain.

Customers will keep buying packaged software when its value exceeds the hassle and risk of building internally.

This matches standard buy-versus-build economics and the continued growth of commercial software despite increasingly capable development tools.

Non-developers will be able to specify, test, deploy, and integrate production software through chat interfaces.

Capability is moving in this direction, but dependable production operation—especially across permissions, legacy systems, and failure recovery—is a much higher bar than generating an interface or prototype.

AI can make bespoke internal software the better solution for many company-specific workflows.

Lower development costs improve the case for custom tools where packaged products force awkward processes. The advantage weakens for standardized or highly regulated core systems.

Most companies will continue buying core systems because maintaining bespoke replacements is not their business.

Core systems carry switching costs, integration networks, compliance obligations, and operational risk that do not disappear when code generation gets cheap.

Cheaper software creation expands the total market by converting services and previously uneconomic workflows into products.

This is consistent with rapid agent adoption and rising software spending, though expansion will not protect every incumbent or pricing model.

Regulated organizations will require human QA and accountable controls for probabilistically generated software for years.

Clinical, financial, and other regulated systems impose validation and audit obligations. AI can assist those controls, but it does not remove the need for accountable evidence and sign-off.

The nighttime drone reports were plausibly part of a search for the missing radioactive source.

The source was recovered at the shipper's facility, no drones were used, and the sightings began before the source was lost. The timeline rules out this explanation.

Commercial-drone regulation is a meaningful competitive constraint for the United States relative to China.

US visual-line-of-sight and airspace rules can slow deployment, while China has approved large pilot zones and logged tens of millions of unmanned flight hours. That gap is real even if it does not imply sabotage.

A foreign government would stage conspicuous drone activity to turn US voters and regulators against drone liberalization.

The theory requires coordinated flights, secrecy, and a predictable policy reaction, yet Friedberg offers no detection, attribution, or operational evidence. Ordinary aircraft and reporting contagion explain the observations with far fewer assumptions.

A hostile actor seeking maximum economic disruption would choose a more central technology than delivery drones.

The incentive comparison is sensible, but covert-action target selection is unknowable without intelligence evidence. It works as a rebuttal to proportionality, not as proof of another plot.

The China-drone-psyop theory lacks enough evidence to take seriously.

Subsequent multi-agency findings identified mundane sources for the reports and no anomalous activity, leaving the theory without factual support.

Enterprises will routinely route tasks across many models instead of standardizing on one provider.

Menlo's reports describe multi-model adoption and changing provider shares, although operational complexity may keep the typical production roster far below 30 or 40 models.

Converging model quality and falling token prices will largely erase market value at the provider layer.

Competition clearly compresses unit economics, but current evidence shows demand, subscriptions, enterprise integration, and capital formation growing fast enough to create enormous provider value anyway.

Open models and multiple frontier competitors will keep raw model pricing close to infrastructure cost plus a limited margin.

The competitive mechanism is sound, though scarcity in frontier compute, proprietary capabilities, and enterprise guarantees can preserve premiums for periods of time.

Demand growth and software differentiation can outweigh falling token prices and let model providers multiply revenue.

OpenAI's reported user and enterprise growth through 2026 strongly supports this mechanism, even though durable provider profitability remains less certain than revenue growth.