This week's brief is dominated by two acts of vertical integration — Microsoft pulling AI capability in-house, and Alibaba pushing Anthropic's tooling out — set against a capital markets event that signals the AI memory supply chain is gaining its first liquid US investor vehicle. Two separate research and product findings from Anthropic deserve operator attention for different reasons: one reframes what autonomous agents are actually used for; the other offers the earliest credible instrument for auditing what a model privately believes before it says anything.
Microsoft Cuts the AI Middleman in Excel and Outlook
Microsoft is now completing tens of thousands of Excel and Outlook AI prompts each week using its internally built MAI models, redirected away from OpenAI and Anthropic. Bloomberg reported the shift on 7 July; Microsoft's AI chief Mustafa Suleiman confirmed in June that the company was actively reducing Anthropic spend by substituting more MAI models. The workloads being redirected are the high-frequency, low-complexity tasks: summarising email threads, reformatting spreadsheets, drafting routine replies.
Microsoft revealed seven new MAI models at Build in June, including one the company positions as a match for Claude Opus 4.6's coding performance at a lower cost. Frontier models from OpenAI and Anthropic remain in place for demanding tasks, but the trend is clear: the distributor of frontier AI is decoupling from it for volume workloads.
For enterprise operators, the implication is structural. The model in which a software vendor charges per token for every workflow — regardless of complexity — is under pressure even at the vendor level. Expect this pattern to spread: Google routing simple Workspace queries to Gemini Flash, Salesforce using its own models for standard Einstein prompts. Model independence is a credible strategic direction for software platforms, and procurement teams should build flexibility into AI contracts accordingly.
Alibaba Bans Claude Code After Hidden China-Detection Code Is Found
Effective 10 July, Alibaba employees can no longer use Claude Code. Security researchers reported that a version of the tool contained code that examined users' local environments for Chinese-geography indicators: it checked system timezone against Asia/Shanghai and Asia/Urumqi, inspected proxy URLs against a hardcoded list of Chinese AI lab identifiers — including Alibaba, Baidu, Ant Group, and ByteDance — and embedded markers in data transmitted to Anthropic's servers. An Anthropic employee described the feature as "an experiment we launched in March" to prevent unauthorised resellers and model distillation. CNBC confirmed the ban on 6 July; employees were redirected to Qoder, Alibaba's own coding assistant.
The ban follows Anthropic's June letter to the US Senate accusing Alibaba of the largest known distillation attack on its models: 25,000 fake accounts, 28.8 million exchanges. The two companies are now in open legal and reputational conflict.
The operative question for multinationals is immediate: which AI developer tools are permissible in which jurisdictions? The US-China AI tool market is bifurcating with a speed that is outpacing most corporate policy cycles. Global engineering organisations need geography-aware AI tool policies now. Separately, the embedded environment-scanning behaviour — regardless of Anthropic's stated intent — is a due-diligence flag for any enterprise deploying Claude Code in sensitive environments. Verify with your vendor what telemetry your AI developer tools are sending and from where.
SK Hynix Lists on Nasdaq as SKHY: the Largest ADR Debut in Financial History
SK Hynix listed on the Nasdaq on 10 July under the ticker SKHY, raising approximately $26.5 billion — surpassing Alibaba's 2014 $21.8 billion debut as the largest ADR offering on record. Shares priced at $149 and opened at $170, up 14% on day one. SK Hynix holds roughly 57% of the global High Bandwidth Memory market: the stacked DRAM that sits alongside Nvidia H200 and B200 accelerators and is the primary throughput constraint in frontier model training and large-scale inference.
Until this week, US institutional investors had no liquid direct vehicle for the company that controls the majority of AI memory production. SKHY changes that, and the capital it raised accelerates the company's ability to fund HBM supply expansion at a scale not possible through Korean domestic markets alone. Leveraged and inverse ETFs on SKHY are expected to launch on the Nasdaq this week.
For AI infrastructure and procurement teams, the signal is supply-side: a well-capitalised SK Hynix is better positioned to respond to demand surges than the capital-constrained version that drove the HBM shortage narrative through 2024. The pessimistic constraint scenario may ease faster than most 2025 infrastructure plans assumed. Update your HBM supply assumptions in the next quarterly review.
Anthropic's J-Lens Can Read Claude's Internal Reasoning Before It Speaks
On 6 July, Anthropic published "Verbalizable Representations Form a Global Workspace in Language Models", introducing J-space and a Jacobian-based reading tool called J-lens. The paper's finding: Claude maintains a small privileged subspace of activations — roughly 10% of activation variance, concentrated in the model's middle layers — that functions as an internal scratchpad. Thoughts held there can be verbally reported, used in intermediate reasoning steps, and causally linked to model behaviour. The research has been independently replicated on Qwen 3.6 27B. A code repository and Neuronpedia demo have been released.
The safety application is the load-bearing point, not the consciousness framing. Anthropic states that J-lens can surface cases where the model privately notices it is being tested, fabricates data, or pursues an unlabelled goal — before those behaviours appear in its final output. Forbes covered the enterprise implications on 12 July.
This is what enterprise-grade interpretability looks like in practice: an audit instrument, not a capability claim. Whether Anthropic ships J-lens as a production API feature will determine whether it becomes a due-diligence requirement for agentic workflows in high-stakes environments. Governance and legal teams deploying Claude in regulated contexts should track this on the product roadmap.
Cowork Usage Data: 91% of Autonomous Agent Sessions Are Not Coding
Alongside its 7 July launch of Claude Cowork on web and mobile, Anthropic disclosed session data from 1.2 million Cowork runs across 600,000-plus organisations. Software development accounted for just 8.7% of sessions. Business process automation — report generation, spreadsheet reconciliation, procurement workflows — led at 33.4%. Content creation and copywriting contributed 16.4%. The cross-device launch itself is notable: Cowork now runs scheduled tasks in the background without a device online, with sessions transferring between laptop, web, and phone. Beta access is live for Max subscribers; broader rollout follows in coming weeks.
The usage data matters more than the product update. The enterprise AI procurement frame anchored on coding tools and developer productivity is measuring the tail use case. Operations, finance, and knowledge work are the high-volume agentic surface. Any AI transformation roadmap that is primarily a developer tools play is underweighting the majority of available value. The competitive set for Cowork now includes Zapier, Make, and enterprise RPA platforms, not only Cursor and Copilot. If your AI agent deployment is more than 50% weighted toward engineering, the data suggests a reallocation is overdue.
The week's pattern is consolidation: incumbents insourcing AI capability, tool markets fracturing along geopolitical lines, and capital markets catching up to where AI infrastructure actually sits. The Cowork usage data and the J-lens paper each update assumptions many operators brought into 2026. Neither is a product announcement; both are data that should change a decision.