Google announced Gemini Spark at I/O 2026 — a background agent that runs continuously across Gmail, Docs, Calendar and Meet without needing to be invoked. It books meetings, drafts responses, summarises threads and reorganises your schedule based on stated priorities. Early testers described it as the first AI product that felt genuinely proactive rather than reactive. Google is positioning it as the successor to Google Assistant, which is being quietly wound down.

Microsoft reported its Q3 2026 earnings this week, with its AI business now annualising at $37 billion — up 123% year over year. Copilot is embedded across the entire Microsoft 365 suite, and weekly active usage now rivals Outlook, which has historically been one of the most-used apps in enterprise software. CEO Satya Nadella called AI "the most fundamental shift in how people work since the PC." Every major enterprise software category is now being repriced around AI capability.

In February 2026, OpenAI quietly introduced advertising into ChatGPT with a $200,000 minimum spend — effectively limiting it to large brands. This week they dropped that floor entirely, opening a self-serve ad platform to any business. Industry analysts are now projecting $2.5 billion in ad revenue for 2026. The move mirrors what Google did with AdWords in 2000: democratise access to the platform and let volume do the work. ChatGPT now has over 800 million weekly active users — a larger audience than most social platforms.

Bloomberg published internal screenshots of a redesigned Siri this week — a standalone app with a persistent chat interface, full conversation history, file upload support and the ability to take actions across iOS. It runs Apple's own on-device model for most tasks and routes to Google Gemini for complex queries under the deal announced last year. The design is a significant departure from the current voice-first Siri, suggesting Apple has concluded that text-based AI interaction is where users actually want to be.

Figure AI confirmed this week that its Figure 02 robots are operating on BMW production lines in South Carolina — not in a supervised pilot but in unsupervised commercial deployment. Tesla simultaneously announced Optimus units are active on its Fremont factory floor, handling parts sorting and basic assembly. Both companies cited cost trajectories: current per-unit economics are roughly equivalent to a mid-senior human worker when annualised over five years, but are expected to fall 60–70% over the next 24 months as manufacturing scales.

TSMC's investor day this week confirmed what supply chain insiders had been whispering for months: every advanced node is sold out through 2027, and the real constraint is no longer the chip itself but advanced packaging — specifically CoWoS (chip-on-wafer-on-substrate), which is used to stack HBM memory on top of logic chips. There are only a handful of companies in the world that can do this at scale, and all of them are running at full capacity. Lead times for CoWoS capacity are now being measured in years.

Cursor, the AI-native code editor built on VS Code, hit $500 million in annual recurring revenue this week — a milestone it reached faster than Stripe, Notion, Figma or any previous developer tool in history. Enterprise contracts now account for the majority of revenue, with companies like Stripe, Airbnb and Morgan Stanley standardising their engineering organisations on it. The company has not taken outside funding since its Series B and is reportedly profitable. JetBrains and GitHub Copilot are both losing ground.

The European Commission formally opened an investigation into xAI's Grok this week under the EU AI Act, focusing on whether the model produced political content without adequate safeguards during the French legislative election cycle in April. It is the first enforcement action under the Act targeting a non-EU AI provider. xAI has been given 30 days to respond with documentation on its content moderation systems, training data provenance and output monitoring. A fine of up to 3% of global revenue is possible if violations are confirmed.

Salesforce announced another round of layoffs this week, cutting approximately 1,000 roles across sales operations, customer support and legacy product teams. The company was explicit in its filing: headcount is being reallocated to AI product development and Agentforce engineering. SAP, Adobe and ServiceNow have made near-identical announcements this quarter. What is notable is the framing — these companies are not calling these cost cuts. They are calling them structural realignments. The distinction matters for how employees, investors and regulators interpret what is happening.

GitHub's 2026 Developer Survey, released this week, found that open software engineering roles have fallen 53% from their 2022 peak — the steepest sustained decline in the history of the profession. Junior and associate roles have been hit hardest, down over 70% in some markets. When hiring managers were asked to explain the reduction, 67% cited AI coding tools as the primary driver, compared to 18% who cited economic conditions. The survey also found that developers using AI tools are completing tasks 40–55% faster than those who are not, suggesting the productivity gains are real and companies are adjusting headcount accordingly.

Micron's CFO confirmed on this week's investor call that every unit of high-bandwidth memory the company will produce in 2026 is already committed to customers — primarily Nvidia, AMD and a small number of hyperscalers. Buyers who did not secure allocations earlier in the year are now being told to wait until Q1 2027 at the earliest. HBM is the specialised memory stacked directly on top of AI accelerators, and without it, the most powerful chips in the world cannot function. The constraint is not fabrication — it is the extraordinarily complex bonding and stacking process that only a handful of facilities in the world can perform.

SK Hynix announced this week that it is beginning construction on a new fabrication facility in Cheongju, South Korea, dedicated entirely to producing HBM4 — the next generation of high-bandwidth memory. The facility will cost an estimated $7 billion and is not expected to produce its first commercial yield until late 2027. HBM4 offers roughly double the bandwidth of the current HBM3e generation and will be essential for the next wave of AI accelerators from Nvidia and AMD, both of which are already in discussions with SK Hynix about allocation.

Samsung's latest attempt to qualify its HBM3e chips for use in Nvidia's H200 and B100 accelerators has failed for the second consecutive quarter, according to people familiar with the testing. The failure is primarily related to power consumption and heat dissipation at sustained loads. This is a significant problem for Samsung, which has been counting on Nvidia as an anchor customer for its memory business. It is also a significant problem for the overall supply picture: Samsung represents roughly 30% of global HBM production capacity, and that capacity is effectively unavailable to the market's largest buyer.

Anthropic closed a $3.5 billion funding round this week, pushing its valuation to $61 billion. The round was led by a consortium of sovereign wealth funds including the Saudi Public Investment Fund and Abu Dhabi's Mubadala. Amazon, which has already committed $4 billion to Anthropic as part of a cloud partnership, did not participate in this round. The capital will primarily fund compute — Anthropic's training runs for its next frontier model are estimated to cost between $500 million and $1 billion. The company is the only major AI lab that has publicly committed to not pursuing AGI without safety guarantees in place.

Nvidia briefly crossed a $4 trillion market capitalisation this week, a figure that exceeds the GDP of Germany. The milestone came after the company reported quarterly revenue of $44 billion — up 78% year over year — driven almost entirely by data centre GPU sales. The H100 and H200 remain the dominant AI training chips, and the next-generation Blackwell architecture is selling out as fast as it ships. Nvidia now captures an estimated 85% of the global AI accelerator market by revenue.

ServiceNow reported this week that it has crossed $50 billion in annual recurring revenue — the first SaaS company in history to reach that milestone. The growth is being driven almost entirely by its AI layer, Now Assist, which automates IT service management, HR workflows and customer operations. The company raised its full-year guidance for the third consecutive quarter. CEO Bill McDermott attributed the acceleration to what he called "AI monetisation at enterprise scale" — customers are paying meaningfully more per seat for AI-enabled versions of products they already use.

Goldman Sachs published an updated labour market analysis this week, revising its 2023 estimate upward. The new report projects that AI will automate the equivalent of 300 million full-time jobs globally over the next decade — up from the 200 million estimate two years ago. The revision is driven by faster-than-expected adoption of agentic AI in white-collar work, particularly in legal, financial analysis, software development and customer service roles. The report notes that the impact is already measurable in hiring data in several of these categories.

Internal data leaked to The Information shows GitHub Copilot lost approximately 40% of its enterprise accounts in Q1 2026 — the majority of them migrating to Cursor. The losses are concentrated among companies with more than 500 engineers, where Copilot's integration model has struggled to match Cursor's native AI workflow. Microsoft, which owns GitHub, has responded by announcing Copilot will be deeply embedded into VS Code itself in a forthcoming update — effectively conceding the standalone market and betting on the IDE lock-in.

The US Senate passed the AI Transparency Act this week in a 67-33 vote, sending it to the House. The bill would require any AI system used in consequential decisions — hiring, lending, healthcare, criminal justice — to disclose that AI was involved in the decision and provide a mechanism for human review. It does not regulate the models themselves, only their deployment in specific contexts. The tech industry is largely supportive, seeing it as preferable to more restrictive alternatives that were circulating earlier in the year. Civil liberties groups are more divided.

Meta released Llama 4 this week under an open-source licence, and independent benchmarks suggest it outperforms GPT-4o on most standard evaluations and matches Claude 3.5 Sonnet on reasoning tasks. The release includes both a base model and an instruction-tuned variant. The implications for the AI industry are significant: a model of this capability being freely available means any company, researcher or developer in the world can now run frontier-level AI without paying API fees. It also puts pressure on OpenAI and Anthropic to justify their pricing models.