Divya van Mahajan

Monthly Synthesis

AI Briefing Synthesis — May 2026

aibriefingsynthesis

Overview

May 2026 marked a broadly acknowledged inflection point: AI transitioned from experimental technology to critical business infrastructure, with the first AI lab posting a profit, multi-trillion-dollar compute deals being struck, and agentic AI crossing from experimentation into early production. Simultaneously, the economics of AI shifted — token subsidies are ending, usage-based pricing is arriving, and the brief window of cheap, unconstrained AI experimentation is closing. Across all articles, the dominant theme is that the time to prepare for agentic AI is now, not later.

Major Topics

AI Economics: Subsidies Ending, Real Pricing Arriving

Anthropic restructured pricing to separate interactive from programmatic (agentic) usage, ending the implicit token subsidy that allowed developers and enterprises to run high-volume workloads at flat-rate consumer prices. The gap between subscription price and actual token value consumed was in some cases 10x–25x. OpenAI is expected to make comparable changes within 12 months. Companies running AI-heavy workflows need to build real cost assumptions into planning now. Salesforce’s introduction of “Agentic Work Units” as a metric signals that the industry is moving toward output-based pricing models.

Agents: The Shift from Assistants to Parallel Workers

AI agents do not reduce work — they expose the infinite backlog of everything an organisation always wanted to do but lacked capacity to pursue. Where AI assistants compressed time, agents effectively replicate the worker in parallel, dissolving natural stopping points that made the workday feel finite. The psychological result: every knowledge worker begins to experience something structurally identical to being a startup founder — exhilarating capability combined with anxiety around judgment, prioritisation, and coordination. New organisational support architectures are needed: technical infrastructure for agents, human support systems for pacing, and new coordination mechanisms.

The “Human Sandwich” Model of Agentic Work

Practitioners at the frontier (e.g. Every/Dan Shipper) report a counterintuitive finding: the more they automate, the more expert human work there is to do. LLMs commoditise whatever competence has been recorded and made explicit, collapsing the value of default AI output (“slop”) while simultaneously creating demand for differentiated expert human judgment. The effective collaboration pattern is neither autonomous agents nor turn-based prompting, but a semi-synchronous model: humans frame work, agents execute, humans evaluate and redirect. Shared team agents outperform personal agents for most enterprise use cases.

Jobs: Doom Narrative Softening, but Displacement is Real in Pockets

The mass-unemployment framing is being challenged from two directions: intellectually (Jevons’ Paradox, historical labour data) and empirically (software engineering job postings rising, startup formation accelerating, AI revenues growing at unprecedented rates). AI will create six broad families of new roles: navigators, continuous support workers, AI-augmented service operators, data and ops specialists, QA/safety/compliance roles, and escalation specialists. However, concentrated displacement in specific communities and job categories is real and underappreciated — the most important risk that the optimistic narrative underweights.

Competitive Landscape: Anthropic Ascendant, Google Fragmented

Anthropic posted its first profitable quarter ($44B annualised revenue), overtook OpenAI in enterprise AI adoption, hired Andrej Karpathy to lead recursive self-improvement research, and secured a $45B compute deal with SpaceX. OpenAI pivoted hard to “work AI” via Codex — now a mobile-first, always-on agentic platform. Google demonstrated real capability progress (Anti-Gravity 2.0, Omni video editing, 900M Gemini MAUs) but is strategically fragmented: Demis Hassabis pursuing a 5–10 year world-model AGI path while internal factions (reportedly including Sergey Brin) urgently demand a coding-agent-driven recursive self-improvement approach. Elon Musk is executing a third-phase pivot from model builder to compute infrastructure provider (XAI dissolving into SpaceX AI).

Infrastructure: A Decades-Long Manufacturing Renaissance

The SpaceX–Anthropic and NVIDIA–Corning deals reframe the AI buildout not as a speculative short-term surge but as a decades-long American manufacturing and infrastructure investment. Compute demand is structurally difficult to overbuild; enterprise AI adoption is still early. Wall Street signalled a transition from bubble anxiety to confident long-horizon investment.

Consumer vs. Enterprise Split — and the Coming Consumer Renaissance

Enterprise AI is where economics and attention are concentrated: API/work users can be 100x more valuable than consumer subscribers by token volume. The consumer AI renaissance is predicted 12–24 months away. Meta ($125B–$145B infrastructure bet) is the primary contrarian, betting on consumer-first. Paths to consumer monetisation (advertising, agentic commerce, AI devices) exist but carry unresolved challenges.

Executive AI Literacy as Strategic Input

A CEO’s personal AI usage quality directly shapes organisational adoption culture. Most executives fall into three failing patterns: informed-but-not-building, building-privately-but-not-deploying, or delegating-without-engaging. Practical recommendations: voice-driven research, AI-assisted scenario planning, four “digital team members” (Research Analyst, Strategic Thought Partner, Communication Expert, Operational Powerhouse). Personal AI fluency is now a strategic imperative.

Token Maximalism and Organisational Experimentation

The transition from assisted to agentic AI represents a fundamental shift with no established best practices — the only way to develop organisational competence is through large-scale experimentation. Companies that embrace this messy, expensive experimentation phase will be substantially better positioned than those that wait for near-term ROI justification. More sophisticated output-based metrics (e.g. Agentic Work Units) represent evolution of this approach, not repudiation.

  • Token subsidies ending — usage-based pricing replacing flat-fee models across all major labs
  • Agents moving from experimentation to early production in enterprise settings
  • AI economics shifting toward scarcity pricing; compute demand outpacing supply
  • Human premium in AI-era work: relationship, accountability, trust, embodied presence not easily AI-substitutable
  • Recursive self-improvement approaching a potential threshold (Karpathy hire, maths problem solved)
  • Consumer AI neglected in favour of enterprise; renaissance expected in 12–24 months
  • Harness layers (Codex, Claude Code, Cursor) becoming as strategically important as models
  • “Doom” job narrative softening; nuanced displacement conversation beginning
  • Wall Street transitioning from bubble anxiety to long-horizon AI infrastructure investment

Emerging Ideas

  • Infinite backlog problem: Agents surface everything an organisation always wanted to do but couldn’t — creating new forms of overwhelm and requiring new organisational design
  • Agentic Work Units: Output-based AI usage metric (Salesforce) replacing raw token counts
  • Interaction models: Thinking Machines Lab’s turn-based architecture critique — 200ms microturns, parallel I/O streams, visually proactive interjection — a new class of AI interaction distinct from current chat paradigms
  • Recursive self-improvement threshold: The possibility that AI training on AI-generated outputs could produce capability jumps faster than annual release cycles — now being actively pursued at Anthropic
  • Foothills of the singularity: Demis Hassabis framing, May 2026 — mainstream acknowledgment from a leading researcher
  • Human sandwich: Semi-synchronous human-agent collaboration model emerging as best practice
  • Harness wars: Labs absorbing agent primitives pioneered by open-source; harness companies building models (Cursor’s Composer 2.5)

Sources