Version 5

What

Why it matters

Open questions

• SemiAnalysis reports effective Opus 4.7 costs near $0.99/million tokens for optimized agentic workloads [6] — are the enterprises reportedly pulling back (Microsoft, Uber) using actual blended costs or sticker prices in their ROI calculations?

• SemiAnalysis's own token spend equals roughly 30% of employee compensation and they argue research firms, hedge funds, and law firms are heading toward similar ratios [7] — is there external evidence from those industries that this migration is underway at scale?

• McKinsey's $3-5 trillion retail commerce projection requires brands to expose inventory and pricing via machine-readable APIs to remain agent-discoverable [3] — is there observable evidence that retailers are moving toward this infrastructure at scale?

• Outcome-driven agent products like Scout route high-stakes changes (money, external integrations) to human approval before deployment [14] — does this human-in-the-loop constraint limit achievable ROI in ways that the cost-benefit comparisons between agents and human workers do not capture?

Narrative

Timeline

• 2026-05-30: Goldman Sachs 24x AI agent token forecast publicized; reports note Uber and Microsoft are already reconsidering expensive agent deployments on cost grounds. [1]
• 2026-06-19: Eric Schmidt quote — 'if you really want to make money, found an agentic AI company' — begins circulating widely. [23][27]
• 2026-06-24: DeepMind integrates computer use into Gemini 3.5 Flash with adversarial prompt-injection training and confirmation gates for irreversible actions. [12]
• 2026-06-25: OpenAI publishes 'How agents are transforming work'; Codex reported at 99.8% of internal output tokens, up from below 10% a year earlier. [19][2]
• 2026-06-25: OpenAI internal data: non-developer individual agent use grew 137x and organizational use grew 189x since August 2025; 70.2% of users submitted requests exceeding one hour of human work. [2]
• 2026-06-25: Microsoft upgrades Copilot in Excel with external data connectors to FactSet, Morningstar, and others; SKILL.md-defined finance workflows and Plan mode with pre-edit declarations. [11]
• 2026-06-25: McKinsey projects AI agents could mediate $3-5 trillion of global retail consumer commerce by 2030, requiring brands to expose machine-readable APIs to remain agent-discoverable. [3]
• 2026-06-25: Notion announces shutdown of Notion Mail on September 22, 2026, citing that most users rely on AI agents for email rather than dedicated clients. [13]
• 2026-06-25: Multiple reports confirm Microsoft and Uber find AI agent deployments more expensive than equivalent human workers. [8][9][10]
• 2026-06-27: SemiAnalysis: Anthropic ARR grew from $9B to over $44B; gross margins rose from 38% to over 70%; AI labs now capture most value in the AI stack. [4]
• 2026-06-27: SemiAnalysis: B300 GPU software optimization alone achieves 14x throughput gain on DeepSeek R1; model-lab gross margin expansion is structural, not a temporary pricing anomaly. [5]
• 2026-06-27: SemiAnalysis: effective Opus 4.7 cost ~$0.99/million tokens due to 300:1 input/output ratios and 90%+ cache hit rates; tasks that took analyst hours now complete in minutes for a few dollars. [6]
• 2026-06-27: SemiAnalysis discloses internal token spend equals ~30% of employee compensation; employees average ~5B tokens/month, over 5x Meta's per-employee average. [7]
• 2026-06-29: Scout launches as a KPI-driven agent product: users specify a business goal in plain English; the system builds and tests agents autonomously, with human approval gates for changes involving money or external integrations. [14]

Perspectives

Tensions

• Goldman Sachs and McKinsey project multi-trillion-dollar economic impact from agent adoption by 2030; Microsoft and Uber report that current agent deployments cost more than equivalent human workers. [1][15][3][8][9][10]
• SemiAnalysis reports effective Opus 4.7 costs near $0.99/million tokens for optimized agentic workloads; enterprises citing cost as prohibitive appear to be using a different cost basis — whether sticker pricing or suboptimal deployment configurations is not specified in available reports. [6][8][9][10]
• Microsoft is both pulling back on expensive AI agent deployments on cost grounds and simultaneously expanding Copilot in Excel with new finance workflow capabilities — the two positions may reflect different product lines or cost assumptions within the same company. [1][15][11][8]
• OpenAI's research frames agents as broadly transformative across professional roles including legal and finance; the internal token data showing Codex at 99.8% of output suggests current value is concentrated in software development. [19][2]
• SemiAnalysis argues AI labs now capture most value in the stack and that margin expansion is structural; the Goldman Sachs thesis implies cloud infrastructure providers will also benefit as demand grows — whether lab value capture comes at infrastructure providers' expense is unresolved. [4][5][1]