A Monday defined by the money and the rules, not new models.
TSMC's record quarter (net profit +77% to a single-quarter high, HPC now 66% of revenue, AI outlook raised past 40%) is the loudest rebuttal yet to last week's "is AI capex overbuilt?" wobble — CoWoS packaging is sold out into 2027. On the regulatory front, Brussels ordered Google to open Android to rival AI assistants and share Search data, and the Future of Life Institute's summer safety scorecard gave no frontier lab above a C+. For your engineers specifically, the MCPModel Context ProtocolAn open standard that lets AI apps connect to data sources, tools and workflows through one common interface.MCP ↗ stateless spec finalizes Jul 28 (eight days out) and Claude Code shipped a security/stability update worth pulling.
TLDR
- TSMC posts a record Q2 — net profit +77.4% to NT$706.6B, HPC 66% of revenue, full-year AI outlook raised past 40%, CoWoS sold out (Jul 16) Infra Trending
- EU orders Google to open Android to rival AI assistants and share Search data under the DMA — data-sharing from Jan 2027, Android interop by Jul 2027 (Jul 16) Policy
- FLI Summer 2026 AI Safety Index: no lab above C+, with four top labs walking back earlier safety-pause pledges (Jul 19) Policy
- Hyundai buys SoftBank's remaining
10% of Boston Dynamics ($325M), taking full ownership to accelerate its physical-AI push (Jul 16) Robotics/IoT - MCP's stateless spec (2026-07-28) publishes in 8 days — the handshake and session header go away; remote-server deployments should prep now Agents
Action Items
Things your engineering team should look at or do this week.
- Audit remote MCP servers for the stateless spec, which finalizes Jul 28 (8 days out). The 2026-07-28 release candidate makes the protocol stateless at the core — the handshake and session header are gone, and server-initiated calls (sampling/elicitation) are replaced by Multi Round-Trip Requests (an
InputRequiredResult+ opaquerequestStatethe client re-issues). That removes the sticky-routing / shared-session-store dance horizontal deployments needed, but it's a breaking change. Beta SDKs (Python 2.0.0b1, TS v2, Go, C#) are out; validate against real workloads before the window closes. (MCP blog | SDK betas) - Pull the latest Claude Code. The July update adds the
EndConversationtool, a progress heartbeat for long-running tool calls that previously went silent, safer Bash/PowerShell handling, and new permission prompts fordockercommands carrying daemon-redirect flags that previously ran unprompted — plus message-level OpenTelemetry attributes (message.uuid,client_request_id,tool_source) for tool-provenance correlation. (Claude Code changelog) - If you build search or assistant products for the EU, log the Google DMA timelines. The Commission's specification measures open anonymized Search data to eligible competitors (including AI chatbots with search) from Jan 2027, and let third-party assistants use more Android functions — voice activation, in-app actions — from Jul 2027. New distribution surfaces worth planning around, not urgent compliance. (European Commission)
Models & Tools
- MCP 2026-07-28 stateless spec release candidate (final: Jul 28) Agents — The next Model Context Protocol version drops stateful sessions entirely and adds an Extensions framework, Tasks, MCP Apps, authorization hardening, and a formal deprecation policy (Active → Deprecated → Removed, 12-month minimum). The big win is scaling: remote servers no longer need sticky sessions or a shared session store. If you run or consume remote MCP servers, this is the migration to schedule. (MCP blog | Microsoft on scaling)
- Claude Code ships a security + stability update (~Jul 19) Trending — Beyond the
EndConversationtool and long-running-task heartbeats, the release tightens permission checks (docker daemon-redirect flags now prompt), improves background-session cleanup, strengthens telemetry, and adds reasoning-effort to the subagent status-line payload. A maintenance drop, but the docker-permission and Bash/PowerShell hardening are worth having if agents run shell commands in your environment. (Claude Code changelog | Releasebot) - OpenAI expands ChatGPT custom instructions and adds unified search (~Jul 18–19) Customer AI — Custom instructions jump to 5,000 characters (from 1,500) for Plus, Enterprise, Business, and Education users, and a new unified search spans chats, projects, images, and documents across web, iOS, and Android. Minor for API users, but relevant if your team standardizes on ChatGPT workspaces. (OpenAI release notes | Releasebot)
Industry & Business
- TSMC posts a record Q2 as AI demand accelerates (Jul 16) Infra Trending — Revenue hit a record $40.2B (+36% YoY) and net profit surged 77.4% to NT$706.6B, with gross margin at 67.7%. High-performance computing — the AI-accelerator segment — rose 20% sequentially and now accounts for 66% of wafer revenue; TSMC raised its full-year AI-chip outlook past 40%, said CoWoS advanced packaging is sold out with 1-year-plus lead times, and pledged $100B more for Arizona fabs. The clearest signal that the infrastructure cycle many expected to plateau is instead speeding up. (DigiTimes | Investing.com)
- Hyundai takes full ownership of Boston Dynamics (Jul 16) Robotics/IoT — Hyundai Motor Group will buy SoftBank's remaining ~9.9% stake for roughly $325M, making Boston Dynamics a wholly owned subsidiary after SoftBank triggered a put option from the 2021 deal. Hyundai frames it as speeding execution of its "physical AI" strategy; it plans to deploy Atlas at its Metaplant America facility from 2028 for parts-sequencing tasks. (Bloomberg | Hyundai newsroom)
- Helsing raises $1.8B at an $18B valuation (Jul 13) Trending — The Munich defense-AI firm (software for unmanned systems and drones) closed Europe's biggest-ever defense-tech round in a Series E led by demand that "significantly" exceeded allocation, with JPMorgan, Lightspeed, Iconiq, Goldman Sachs Alternatives, and CPP Investments participating. It also announced a first US manufacturing base in West Virginia — a marker of how fast capital is rushing into defense AI (see also Anduril). (CNBC | Defense News)
- Anthropic launches Claude for Teachers (Jul 19) Customer AI — Verified US K-12 educators get free access to premium Claude tools, teaching-oriented skills, curriculum connections aligned to standards in all 50 states, new education connectors, and AI-fluency training. A distribution/goodwill play into education as the labs compete for early-career mindshare. (Anthropic newsroom | Releasebot)
Policy & Safety
- EU orders Google to open Android and share Search data with AI rivals (Jul 16) Policy — Under the Digital Markets Act, the Commission issued binding specification measures requiring Google to give competing AI assistants equal access to Android functions (voice activation like "Hey Google," in-app actions such as booking a taxi or drafting chat replies) and to share anonymized Search data with eligible competitors, including AI chatbots with search features. Data-sharing starts Jan 2027; Android interop reaches users from Jul 2027. Google's Kent Walker objected on privacy and security grounds. (European Commission | SiliconANGLE)
- FLI's Summer 2026 AI Safety Index: no lab tops C+ (Jul 19) Policy — The Future of Life Institute graded nine frontier labs; Anthropic led at C+ (2.66/4.0), with OpenAI (2.28) and Google DeepMind (2.01) tied around a C. The headline finding is backsliding: Anthropic, OpenAI, Google DeepMind, and Meta have all walked back earlier commitments to pause development at defined risk thresholds and to release only with capability-appropriate safeguards, and all four reversed prior restrictions on military work to pursue defense contracts. Evidence was collected through Jun 3, so recent moves aren't reflected. (Future of Life Institute | TechTimes)