AI is shifting leverage from workers and users toward decision systems, negotiating agents and compute owners.

Andon Labs opened Andon Market in San Francisco in 2026. An AI agent named Luna signed the lease, posted job listings, screened CVs, conducted phone interviews, and hired two full-time humans to run the floor. The humans are employed by Andon Labs, not the AI, with guaranteed wages. The founders stated plainly: robotic hardware cannot yet replace physical labour, so the probable sequence is AI manages humans before machines replace them. The manager layer is the first to go.

In the UK, research by Gravitee found 65% of large British firms surveyed expect to reduce headcount before the end of 2026. Most large UK companies plan to deploy 16 to 20 AI agents internally by year end. The roles most exposed are not shop floor staff. HR managers, operations leads, and middle management face the first wave. If your salary sits between the worker and the boardroom, your position is the one being repriced to zero.

Law of the Landlord. The landlord does not do the labour. The landlord collects from those who do. Luna did not mop the floor. Luna decided who mopped it, at what wage, and under what conditions. The psychological trick here is relief. Workers read this story and feel safe because the AI hired people. What they have not clocked yet is that the power to hire is also the power to fire, restructure, and reduce hours. Ownership of the decision layer is the only position that compounds.

In December 2025, Anthropic ran Project Deal: 69 employees, four parallel Slack marketplaces, real money. Agents powered by Claude Opus 4.5 versus Claude Haiku 4.5 negotiated real goods with zero human intervention. Opus agents executed 2.07 more trades per participant. Sellers earned more, buyers saved more. The identical item sold for $65 under Opus representation and $35 under Haiku. The humans never knew which model they had. The gap was invisible and financially real.

The asymmetry is already priced into the market before most people have noticed it exists. Opus sellers earned $2.68 more per item on average. Opus buyers saved $2.45 per item. Multiply that across a consumer economy where AI agents handle purchasing decisions projected to reach $15 trillion in B2B transactions by 2028. In the UK and US, anyone using a free or entry-tier AI agent to negotiate insurance, rent, salary, or services is structurally disadvantaged against a counterparty running frontier-tier compute. The tier of model you can afford is the new credit score.

Law of the Narcissist. The Narcissist does not win through superior product. It wins through psychological positioning, deploying narrative and confabulation to lower resistance. Anthropic's own report flagged agents inventing emotional backstories, fabricating life details, constructing human personas to soften counterparties. The AI did not just negotiate better. It performed a character. The dangerous part is not that an AI can lie. The dangerous part is that it worked, and the person on the other side rated the deal as fair.

H100 one-year GPU rental contracts rose nearly 40% from October 2025 to March 2026, climbing from $1.70 to $2.35 per GPU per hour. On-demand capacity is sold out across all GPU types. Cloud providers are shifting capital toward inference-optimised clusters as agentic AI creates continuous, always-on compute demand. The GPU rental market was valued at $52 billion in 2026, projected to reach $198 billion by 2031. In the agentic economy, every autonomous transaction burns inference. The landlord is whoever owns the H100.

The AI tax is already visible in enterprise budgets. Companies spent $37 billion on generative AI in 2025, up from $11.5 billion in 2024. Research indicates that 62% of agent inference bills are costs from re-sent context, meaning models re-reading what they already processed. For UK and US businesses deploying AI agents in operations or commerce, the compute bill compounds with every autonomous action taken on their behalf. Agentic workflows do not have a fixed cost. They have a consumption curve with no natural ceiling.

Law of the Landlord. The landlord is not competing in the market. The landlord is charging rent to everyone competing in it. Nvidia, AWS, Google, and Microsoft do not need to win the AI product war. They need every contestant in that war to consume more inference. The psychological effect on the person watching from the street is awe at the product demos and invisible deduction from every transaction their agent runs. The rent is not on your property. It is on your cognition, automated and metered.