Four developments from the past 48 hours cover a blockbuster IP lawsuit against OpenAI, an imminent model launch with no confirmed specifications, a new infrastructure standard for agent tool discovery, and first-half funding data that puts the AI concentration story into stark numerical relief.
Apple Sues OpenAI Over Hardware Trade Secrets
Apple filed suit on 10 July in the Northern District of California against OpenAI, io Products -- the hardware design firm OpenAI acquired last year and co-founded by Jony Ive -- and two named former Apple engineers now employed at OpenAI. The complaint alleges a coordinated campaign of trade secret theft spanning product design, manufacturing processes, and supply-chain contractor data. Apple says more than 400 former employees now work at OpenAI or io Products.
The most specific allegation involves Chang Liu, a former senior systems electrical engineer who left Apple for OpenAI in 2026. Apple contends that Liu exploited an authentication flaw to retain access to Apple's internal network storage after his departure and used it to download a compilation exceeding one thousand pages of confidential engineering documentation. The second named defendant, Tang Yew Tan, formerly VP of product design for iPhone and Apple Watch, is also implicated. TechCrunch's dissection of the complaint notes that Apple characterises the scheme as running from technical staff to the Chief Hardware Officer level.
For operators, the case is consequential on two levels. OpenAI's consumer hardware device -- shaped substantially by io Products -- was expected to launch later this year; the litigation adds legal uncertainty to that timeline and a credibility problem to OpenAI's IPO preparation. The broader signal is about AI talent acquisition risk: a pipeline of 400 former employees from a single competitor represents a concentration that any acquirer should now treat as a potential liability, not just an asset.
ARD: The Discovery Layer Above MCP That Few Builders Have Acted On
On 17 June, Google published the Agentic Resource Discovery (ARD) draft specification alongside Microsoft and Hugging Face, with Amazon, Cisco, Databricks, GitHub, GoDaddy, Nvidia, Salesforce, ServiceNow, and Snowflake named as contributors. The specification is licensed under Apache 2.0 and governed through a Linux Foundation working group.
The easiest way to understand ARD relative to MCP is this: MCP standardises how an agent connects to and operates a tool once it knows which tool to use; ARD standardises how the agent finds the tool in the first place. An agent asks at runtime -- "what resource can handle this task?" -- and ARD returns a structured list of matching capabilities, with provenance, access instructions, and trust metadata. Hardcoded tool registries work fine when a team controls a handful of integrations; they become a maintenance liability as catalogues grow and agents are shared across systems.
The contributor list is the real signal. Salesforce, Snowflake, and ServiceNow represent the software layers where most enterprise business data actually lives. If they ship ARD-native discovery across their platforms, every agent built against those APIs will default to the standard. Operators who are mid-build on agent workflows should read the spec context and assess whether their current tool-registration approach will still be maintainable in twelve months.
Gemini 3.5 Pro: Three Days to Launch, Zero Specifications Confirmed
Google's internal target for Gemini 3.5 Pro general availability remains 17 July -- three days from today. As of this writing, Google has not published a model card, updated its API documentation with a gemini-3.5-pro listing, or confirmed any pricing. Every specific claim in circulation -- a 2-million-token context window, a Deep Think extended-reasoning mode, a $15 per million input tokens price point -- originates from third-party reporting and unnamed internal sources, not from Google.
The background matters. Google scrapped the original Gemini 3.5 Pro architecture after enterprise testers identified material gaps in mathematical reasoning and SVG generation; what ships on 17 July, if it ships, will be a full architectural rebuild. Operators who have provisioned workflows against expected Gemini 3.5 Pro capabilities should wait for the official model card and API listing before committing to a production deployment. The circulating specifications may prove accurate; until Google confirms them, they are not a sound basis for infrastructure decisions.
H1 2026: AI Absorbed 86 Per Cent of US Venture Capital
Crunchbase data and PitchBook's half-year figures show US venture capital investment reaching a record $412.7 billion in the first half of 2026, with AI companies capturing approximately 86 per cent of that total. Globally, startup funding hit $510 billion in H1 alone, exceeding all of 2025.
The concentration is not spread evenly. OpenAI's $122 billion round is the largest private venture round in history; Anthropic's $65 billion raise implied a post-money valuation of $965 billion. Between those two rounds, 43 per cent of all global first-half startup capital landed with two AI labs. SpaceX, which went public at $1.77 trillion and raised $75 billion, promptly committed $60 billion of that to acquire Cursor, consolidating the inference-plus-tooling stack further.
For operators, this is a vendor-selection signal that compounds over time. Capital at this scale shapes which models achieve the enterprise integration depth and long-term reliability that production workloads require. A team choosing a model provider today is taking a position on which lab will still be the price-setter in its category in 2028. The H1 data makes the concentration of that bet more legible than at any prior point in the industry.
The thread across all four items: the AI industry is moving from a competition over model quality to a competition over legal control, infrastructure standards, and capital lock-in. Operators who have not audited their AI vendor dependencies -- or their agent stack's exposure to a rapidly standardising tool-discovery layer -- are carrying a risk profile that the first half of 2026 has made considerably harder to dismiss.