This week: Anthropic’s valuation approaches $1 trillion, OpenAI launches a $4 billion consulting arm to embed engineers inside your organisation, and China blocks a major US AI acquisition on political grounds. Robot hands hit full human dexterity. Claude agents now learn between sessions. The pace is not slowing. Here is what every SME decision-maker needs to know.
THIS WEEK’S TOP 3 MUST-KNOW CONCEPTS
Anthropic Could Be Worth $1 Trillion
CAPITAL
Anthropic is weighing a fundraising round that could push its valuation near $1 trillion and has committed $200 billion over 5 years to Google Cloud. These are not abstract numbers — they reveal the scale of capital now required to stay competitive at the frontier of AI. The providers you build on today will consolidate rapidly. SMEs need to make deliberate platform choices now, not when the market forces them to.
OpenAI DeployCo: The Implementation War Begins
STRATEGY
On May 11, OpenAI launched DeployCo with $4 billion from McKinsey, Bain, Capgemini, Goldman Sachs, and TPG. The model: place specialised AI engineers inside client organisations to redesign workflows and build production systems. OpenAI is no longer just selling model access — it is competing with your management consultants. SMEs that build internal implementation capability now will avoid dependency on expensive external partners later.
China Blocks Meta’s AI Deal: The AI Cold War Is Real
GEOPOLITICS
The Chinese government rejected Meta’s acquisition of Manus — the leading AI automation startup — and ordered the deal unwound. This is not a routine regulatory block. It is a government restructuring the AI market on political grounds. A bifurcated global AI ecosystem with separate US and Chinese stacks is now structurally underway. Any SME with exposure to Asian markets needs to audit which AI tools are permissible in which jurisdictions and what their vendors’ counterparties will accept.
OTHER MAJOR STORIES THIS WEEK
Claude Mythos and Project Glasswing
CYBERSECURITY
Anthropic launched Project Glasswing — a defensive cybersecurity coalition built around Claude Mythos Preview, its most capable model to date. Anthropic has deemed Mythos too powerful to release publicly: it has already identified thousands of zero-day vulnerabilities and achieved full control-flow hijack on ten fully-patched targets in testing. Launch partners include AWS, Apple, Cisco, CrowdStrike, Google, JPMorganChase, Microsoft, and NVIDIA, with $100M in usage credits committed. For SMEs, the signal is clear: AI-driven vulnerability discovery now outpaces traditional patch cycles entirely. If your business runs any customer-facing software or payment infrastructure, your current security posture assumes a threat environment that no longer exists.
Salesforce Rebuilds Its CRM as AI Agent Infrastructure
ENTERPRISE AI
Salesforce launched Headless 360 at its TDX developer conference, turning the entire platform into infrastructure for AI agents. Every capability — customer data, workflows, business logic, compliance rules — is now exposed as an API or MCP tool, so AI agents can read, write, and execute across Salesforce without a human ever opening a browser. The company is simultaneously moving from per-seat licensing to outcome-based pricing. For SMEs running sales, service, or marketing on Salesforce: the CRM interface you log into daily will increasingly be replaced by agents operating underneath it. The practical question is whether your Salesforce data is clean enough for an agent to trust.
Robot Hands Reach Full Human-Level Dexterity
ROBOTICS
Genesis AI unveiled GENE-26.5 on May 6 — a robotic arm trained on over 200,000 hours of human hand movement data that demonstrated genuine human-level manual dexterity in a single public showcase. It cooked a 20-step meal, cracked eggs one-handed, solved a Rubik’s Cube, and played piano. This removes one of the last remaining barriers to full physical automation: fine motor control in unstructured environments. The implications for any SME in manufacturing, food service, logistics, retail fulfilment, or hospitality are direct and immediate. The question is no longer whether robots can do what your workers do. For a growing range of tasks, they can. The strategic question now is at what pace deployment will accelerate in your sector, and what your workforce planning horizon should be in response.
Claude Agents Now Dream to Improve Themselves
AGENTS
Anthropic launched Dreaming for Claude Managed Agents in research preview — a scheduled background process that reviews past agent sessions, extracts behavioural patterns, identifies failure modes, and curates agent memory automatically without human intervention. Early adopter Harvey Law reported a 6x increase in task completion rate after implementation. For SMEs using AI agents, the implication is direct: agents no longer require constant manual prompt tuning. They improve with use.
QUICK DIGEST
Meta Acquires Humanoid Robotics Startup
ROBOTICS
Meta acquired Assured Robot Intelligence, bringing three elite roboticists into its superintelligence labs to build foundation models for whole-body humanoid control.
OpenAI Adds Welfare Safeguard to ChatGPT
SAFETY
OpenAI introduced Trusted Contact — an optional feature that notifies a chosen adult if ChatGPT detects self-harm concerns. The first welfare safeguard built into a general-purpose AI assistant.
US Government to Pre-Test AI Models Before Launch
REGULATION
Google DeepMind, Microsoft, and xAI signed pre-deployment evaluation agreements with CAISI, the US AI safety standards body — extending 2024 arrangements with Anthropic and OpenAI.