AI-Proof - Weekly AI Pulse
A concise summary of the week’s most important AI developments
Executive Summary
The economics of AI are coming under closer scrutiny. Microsoft is rationing scarce Azure capacity between its own products and external customers, Nvidia is reportedly considering an extraordinary guarantee for OpenAI’s next data-centre campus, and investors are increasingly distinguishing between ambitious AI spending and investment supported by visible revenue growth.
At the product level, the market is moving beyond standalone chatbots towards systems that can coordinate models, agents, tools and company data. Microsoft is consolidating its AI products, ChatGPT Voice can now direct longer-running Work and Codex tasks, and Salesforce and HubSpot are giving businesses new ways to build, deploy and govern agents. Notion’s acquisition of ZeroEntropy also shows that speed, search quality and inference cost are becoming important competitive differentiators.
The strategic divide around open AI is widening. Kimi K3 demonstrates how quickly Chinese open-weight models are advancing, while US policymakers debate restrictions and most major technology companies argue against premature controls. Meanwhile, increasingly capable models such as Claude Opus 5 are becoming cheaper to use. The result is more choice for businesses, but also greater complexity around model selection, security, governance and commercial value.
What to Try This Week
1. Teach Claude a workflow by recording it
Skills and automated workflows can sound far more technical than they really are. Claude for Chrome can now learn a repetitive browser workflow simply by watching you complete it. Pick a straightforward task you regularly perform, record the steps and see whether Claude can repeat them. Start with a low-risk process and supervise its first few attempts.
2. Direct an agent using ChatGPT Voice
ChatGPT Voice becomes much more interesting when it is connected to Work or Codex rather than used only for conversation. In the desktop app, try starting a task by voice, asking for a progress update and then interrupting or redirecting it while it works. It feels less like dictation and more like managing an AI colleague.
3. Put Kimi K3 or Claude Opus 5 through a proper test
Kimi K3 and Claude Opus 5 are two of the strongest new models released this week, and both are worth trying on a genuinely demanding task. Give one a long document, a difficult research question or a complex piece of analysis rather than a simple prompt. When testing Kimi through a third-party service, avoid uploading sensitive business or personal information.
Geopolitics, Governance and Big Moves
Microsoft’s internal AI push is straining Azure capacity
Business Insider reported that infrastructure shortages have led Microsoft to prioritise first-party AI services, including Copilot, over some external Azure demand. Microsoft has separately acknowledged that customer demand continues to exceed available capacity. Despite the constraint, Azure revenue grew 43% and passed $100 billion annually, showing how compute allocation has become a strategic choice between supporting Microsoft’s own products and serving cloud customers. (Business Insider)
Notion acquires ZeroEntropy to make workspace AI faster and cheaper
Notion has acquired ZeroEntropy, a specialist in efficient task-specific models, reranking and AI search. ZeroEntropy’s technology had already made Notion’s unified search up to 30% faster, cut reranking latency by 85% and reduced inference costs. The team will form Notion’s new Model Research group, while ZeroEntropy’s models are being released under an Apache 2.0 licence and its standalone products will close on 4 September. (Notion)
Kimi K3 reignites Washington’s push to restrict Chinese AI
Moonshot’s release of Kimi K3 has revived debate in Washington over restrictions on advanced Chinese AI. Options reportedly considered include procurement limits, Commerce Department Entity List additions, cybersecurity advisories and sanctions linked to alleged unauthorised distillation from US models. China rejects the accusations, while Beijing is separately considering tighter export controls on AI models and semiconductor technologies. No comprehensive US ban has yet been announced. (axios.com)
AI industry backs open weights, but Anthropic holds out
A letter now signed by 77 organisations, including OpenAI, Google, Nvidia, Microsoft and Meta, urged policymakers not to restrict open-weight AI prematurely. Anthropic is the notable frontier-lab holdout and has supported tougher action against alleged unauthorised distillation by Chinese developers. The split is not simply open versus closed: OpenAI signed the letter while also lobbying Washington over Chinese model copying. (NVIDIA Images)
Investors reward AI spending only when revenue is visible
Meta reported second-quarter revenue of $60.8 billion, but earnings of $6.18 per share and rising capital spending helped send its shares down about 9%. Microsoft rose 16% after Azure grew 43% and the company kept its spending outlook steady. Amazon then lifted 2026 capital expenditure to $220 billion, yet gained after AWS growth accelerated to 37%. The contrast suggests investors are rewarding spending backed by visible demand. (Reuters)
Nvidia weighs extraordinary $250bn guarantee for OpenAI’s Ohio campus
The Wall Street Journal reported that Nvidia is discussing a guarantee for up to $250 billion of financing supporting OpenAI’s lease at a 10-gigawatt data-centre campus in southern Ohio, led by SoftBank subsidiary SB Energy. Separate talks could finance up to $350 billion of chips. Neither arrangement is signed. The proposed structure could lower borrowing costs, while deepening financial concentration across Nvidia, OpenAI, SoftBank and lenders. (wsj.com)
Tools and Releases
Microsoft to unite Copilot, Code and autonomous agents in one ‘super app’
Microsoft will bring Copilot Chat, Cowork, Autopilots and Code into a single “super app” this quarter, spanning consumer and business users. The move is less about adding another feature and more about reducing Microsoft’s fragmented AI experience. A unified interface could make it easier to move from asking questions to creating software and delegating longer, autonomous tasks without switching products. (Microsoft)
ChatGPT Voice can now direct Work and Codex agents on desktop
OpenAI has expanded GPT-Live on the ChatGPT desktop app for macOS and Windows. In Work and Codex, eligible users can use voice to start, prioritise, interrupt and redirect tasks, while coordinating several agents through one conversation. This moves voice beyond hands-free chat towards an operating layer for agentic work, although availability and usage limits vary by plan and workspace. (OpenAI Help Center)
ChatGPT launches reusable Skills, while Claude learns workflows from recordings
OpenAI has introduced reusable Skills that package instructions, examples and code so ChatGPT can perform recurring tasks more consistently. Skills can be created through chat, an editor or file upload, then shared across a workspace. Anthropic already offers a related capability in Claude for Chrome, which can learn a repeatable browser workflow by watching a recording of the user’s steps. (OpenAI Help Center)
Anthropic launches Claude Opus 5 at half the price of Fable 5
Anthropic launched Claude Opus 5 on 24 July at $5 per million input tokens and $25 per million output tokens, half the price of Claude Fable 5. It brings a one-million-token context window, up to 128,000 output tokens and thinking enabled by default. Unlike Fable 5, Anthropic does not list Opus 5 among the models requiring mandatory 30-day data retention. (Anthropic)
Gemini adds voice control and on-screen reasoning to macOS
Google has added system-wide voice assistance to the Gemini app for macOS. Holding the Fn key enables intelligent dictation in any active window, cleaning up hesitations and corrections before placing formatted text at the cursor. Users can also opt into Gemini reasoning, allowing it to use selected files, images, documents and on-screen context to summarise information, edit content or carry out more complex instructions. (blog.google)
Moonshot releases Kimi K3, the first open 3-trillion-class model
Moonshot AI has released the full weights for Kimi K3, a 2.8-trillion-parameter mixture-of-experts model with 104 billion parameters active per token and a one-million-token context window. Its native multimodal design handles text, images and video, while the open weights allow organisations to run or adapt it themselves. Given its scale, most organisations will still need substantial computing resources or a specialist inference provider. (Hugging Face)
Salesforce opens its platform to AI agents with Headless 360
Salesforce’s Headless 360 makes major platform capabilities available through APIs, Model Context Protocol tools and command-line commands, so agents can work with CRM data and business logic without navigating the Salesforce interface. The programme includes more than 60 MCP tools and 30 preconfigured coding skills. Agents inherit existing permissions and governance, reducing the need to build a separate security model for automation. (salesforce.com)
HubSpot launches a shared control centre for AI agents
HubSpot has launched Agent Hub and Agent Builder in public beta for Professional and Enterprise customers. Agent Hub provides one place to deploy, monitor and manage agents across marketing, sales and service, while Agent Builder lets teams create custom agents and automations using natural language and existing CRM context. The products are included at those tiers, although custom agent activity consumes HubSpot Credits. (HubSpot)
FLUX 3 expands from images into video and native audio
Black Forest Labs has unveiled FLUX 3, a multimodal model that can create and edit images and generate videos of up to 20 seconds with synchronised audio. It supports text-to-video, image-to-video, video-to-video, keyframes and multilingual dialogue. Video is entering gated Early Access through APIs and private weights, while image access is planned to follow. Standard pricing and full technical details have not yet been announced. (Black Forest Labs)
OpenAI cuts Luna API pricing by 80% and launches GPT-Transcribe
OpenAI cut GPT-5.6 Luna API pricing to $0.20 per million input tokens and $1.20 per million output tokens, an 80% reduction. Terra fell 20% to $2 and $12, while Sol stayed at $5 and $30 and gained a faster premium mode. OpenAI also launched GPT-Transcribe at $0.0045 per audio minute, giving Whisper users a new migration path for recorded speech. (OpenAI)
Quick Hits
Europe’s largest round of the week was Spanish: Multiverse Computing announced a Series C of up to €500m at a €1.5bn valuation on 27 July, for technology that compresses models to cut inference cost.
A UK small-cap bought an AI agent business: Tavistock Investments agreed on 28 July to acquire 87.9% of Liverpool-based Plus Group for up to £16m, of which £11.5m is deferred over four years against performance criteria.
The hidden cost of adoption got a number: Freshworks research published on 24 July found 88% of UK IT decision-makers say managing AI complexity has increased their team’s workload, with roughly a quarter of AI budgets absorbed by integration, governance and rework.