Two earnings reports on Wednesday drew the sharpest contrast the AI cycle has yet produced: Microsoft proved that externalising AI revenue works; Meta illustrated what happens when the opposite is true. Meanwhile, a compliance deadline arrives in 48 hours that many European AI deployments have not addressed, and a benchmark result from last week continues to reframe how frontier models should be evaluated.

Microsoft Q4 FY2026: Azure Crosses $100 Billion, AI ARR Reaches $37 Billion

Microsoft reported $90 billion in quarterly revenue, up 18% year-on-year, with net income rising 31% to $35.8 billion. The number that defined the call was structural rather than cyclical: the AI business within Azure reached a $37 billion annual run rate, up 123% from a year earlier. Azure itself crossed $100 billion in full-year revenue for the first time, growing 43% in the fourth quarter. Microsoft 365 Copilot hit 30 million paid seats. Commercial remaining performance obligations jumped 84% to $678 billion — a forward indicator of contracted, multi-year revenue from the enterprise base. The stock rose 9.8% on the news.

The mechanism matters as much as the number. Microsoft's external-pricing model converts every enterprise AI workload running on Azure into direct line-item revenue. The 123% AI ARR growth is not a rounding artefact; it reflects a pipeline of enterprise commitments that did not exist two years ago. The $678 billion RPO signals that those commitments are now locked in for multiple years and difficult to unwind. This is what monetised AI infrastructure looks like from the outside.

Meta Q2 2026: Revenue Up, Margin Down, Free Cash Flow Evaporating

Meta reported $60.8 billion in revenue, up 28% year-on-year. The margin story was harder to read: operating margin fell to 31% from 43% a year ago; net income dropped 14%; and free cash flow collapsed to $784 million as capital expenditures reached $31.1 billion for the quarter alone. For the full year, Meta raised its capex floor to $130 billion. The stock fell 10%.

The divergence from Microsoft is instructive. Meta's AI investments — the Llama model family, compute infrastructure, and internal recommendation systems — drive advertising efficiency and engagement rather than direct external revenue. AI improvements lift click rates and advertiser conversion, but those gains accrue to the advertisers, not to Meta's own operating income. The company is subsidising its ecosystem with shareholder capital, betting on a future monetisation event that has not arrived.

For a board assessing an AI buildout, Wednesday's results frame the core question plainly: are you building an AI-powered product you sell to others, or AI-powered infrastructure you run for yourself? The public markets are now pricing those two choices differently, and the gap is widening.

EU AI Act: What Actually Takes Effect on 2 August

On Saturday, 2 August — 48 hours from now — the EU AI Act's transparency obligations under Article 50 and the general-purpose AI enforcement powers take effect. These were not deferred by the Digital Omnibus amendments approved in June, which pushed the high-risk Annex III compliance clock to December 2027 and beyond. What remains on schedule:

  • Article 50 disclosure: Any operator deploying a conversational AI system, chatbot, or AI-generated-content tool that interacts with users in the EU must ensure those users are informed they are interacting with AI. The requirement applies at the point of interaction, not buried in terms of service.
  • GPAI enforcement: The Commission's general-purpose AI framework becomes enforceable, covering foundation-model providers with broad EU deployment. Systemic-risk assessments are required for the largest models.

Maximum penalties: €35 million or 7% of global annual turnover, whichever is higher. If you operate AI-facing services in the EU and have not yet audited your Article 50 disclosure posture, that review is overdue.

IMO 2026: Four Frontier Models Score 42/42 — Benchmark Saturation Arrives

The 67th International Mathematical Olympiad concluded in Shanghai on 21 July with 666 contestants from 117 countries. In the days following, Deedy Das (Menlo Ventures) ran Claude Fable 5, GPT-5.6 Sol, Kimi K3, and the Axiom Math formal-proof system against the same six problems under conditions matching the competition format. All four scored 42 out of 42 — the maximum possible. Methodology and full audit trails are at github.com/deedy/imo-2026.

The IMO has been a useful reference point because its proof-construction problems resist pattern-matching and because they provide a fixed, externally graded comparison. Four models reaching the ceiling simultaneously — Claude Fable 5 completing the set in under two hours on its first attempt — signals that this benchmark is saturated. It no longer differentiates between frontier systems.

The practical implication for operators is not that AI has become infallible at reasoning. Rather, the evaluation frameworks many procurement teams use for model selection are lagging behind actual capability. A benchmark that cannot distinguish between four leading models provides false assurance. Teams relying on IMO-level math performance as a proxy for general reasoning quality need a new proxy. The pattern is consistent: each reasoning benchmark holds for a year or two, then breaks. Evaluation now needs to run at roughly the pace of the capability curve itself.

Today's four items share a thread: financial returns, regulatory exposure, and evaluation quality are all crystallising faster than the conventional planning cycle assumes. The operators in the strongest position are those already treating AI economics, compliance, and capability assessment as ongoing disciplines rather than annual reviews.