Nvidia's $91 Billion Bar
Nvidia reports second-quarter results for its 2027 fiscal year after US markets close today, with the numbers due around 4:20pm ET and the call at 5pm. The company has already told investors to expect about $91 billion in revenue at a 74.9% GAAP gross margin; analyst consensus has crept a little higher, to around $92–94 billion. Last quarter Nvidia sold $81.6 billion, and $75.2 billion of that — more than 92% — was data-centre chips. So this is less an earnings report than a reading on whether the AI build-out is still accelerating.
For anyone buying inference in Europe, the print matters even without owning a GPU. Blackwell supply, pricing, and Nvidia's own commentary on demand set the cost floor that every hosted-model provider passes through. Two things to watch: how fast Blackwell is ramping, and what management says about China after Washington's on-again, off-again H200 export rules. A weak quarter wouldn't change which model answers your prompt tomorrow — it would change what that prompt costs six months out.
Agent Fleets Are Tripling. Almost No One Has Scaled One.
Salesforce's second Agentic Enterprise Index, out this month, says the average business on its platform now runs 13 AI agents, up from five in early 2025 — nearly triple — and stands up a new one in about half the time it used to. The figures come from Agentforce usage plus a 4,689-person survey. The counter-number is the real story: McKinsey and S&P Global put the share of enterprises running at least one agent in production at about 31% — but far fewer, well under a fifth by most counts, have scaled orchestrated, multi-agent workflows. Activating agents is easy; getting measurable value out of a fleet is not.
As the count climbs, the governance question sharpens — every agent is one more thing holding credentials, reaching into data, and calling a model somewhere. For regulated European buyers that is where sovereignty stops being abstract: which agent can see which data, and in which jurisdiction that data is processed.
France Puts a Number on the AI-Summary Problem
France's press alliance APIG has filed a complaint with the country's competition authority over Google's AI Overviews and AI Mode — the features that answer a search directly instead of sending the reader to the source. France's media regulator, Arcom, estimates the summaries have cut publisher traffic by 33–38%. Google switched the features on in France on 22 July; the complaint landed on 11 August and points back to commitments Google made after a €500 million French fine in 2021 over news licensing. The broader signal for Europe: the fight over who captures the value AI pulls out of other people's content is becoming a competition-law question, not just a copyright one.
Quick Hits
- OpenAI dropped GPT-5.6 Sol to a promotional $4/$20 per million input/output tokens on 21 August, down from $5/$30 and running through about 21 November — another turn of a frontier price war that keeps routed-inference costs falling. (pricepertoken.com)
- The EU AI Act's high-risk obligations have slipped. Under the Digital Omnibus regulation (EU 2026/1744, in force since 27 July), most standalone Annex III high-risk requirements now fall due 2 December 2027 rather than this month — a reminder that “in force” and “enforceable” are different dates. (EUR-Lex)
- Anthropic reported preliminary revenue of more than $11.5 billion for Q2 and its first quarter of positive operating income — up from about $4.7 billion in Q1. (CNBC)
- Open models keep closing the gap: on the BenchLM leaderboard, open-weight Kimi K3 holds roughly 97% of the top score at an output price about 70% lower — the kind of trade-off that makes self-hosting a real option for European teams. (benchlm.ai)