Skip to content
Back to posts

AI developments (Sept 17–Oct 1 2025)

Notes by SDG ·

Microsoft‑specific updates

Copilot and multi‑model support

New model options in Copilot: Microsoft will let users switch between OpenAI‑built and Anthropic‑built models in its Researcher agent and Copilot Studio. The company is adding Claude Sonnet 4 and Claude Opus 4.1 as alternatives to GPT‑class models, giving enterprises flexibility and reducing dependence on OpenAI . This multi‑model architecture will eventually let customers choose models from other providers such as xAI or Meta .

Copilot tuning & bring‑your‑own (BYO) models: Organizations can fine‑tune Copilot with their own data and select from thousands of models in the Azure AI Model Catalog. This ensures responses reflect domain‑specific knowledge while preserving data boundaries .

New agent capabilities and workspace tools

Computer Use capability: Microsoft announced a preview of a “Computer Use” feature that allows Copilot agents to simulate clicks, typing and other UI interactions in applications and on websites. This bridges generative AI with robotic process automation (RPA) and enables tasks such as migrating data between legacy systems .

Collaborative agents for Teams, SharePoint and Viva: Microsoft introduced team‑centric agents that summarise conversations, manage meetings (Facilitator agent), post announcements and organise documents. These agents work inside Teams channels, SharePoint sites and Viva Engage communities, enabling group‑oriented assistance .

Reasoning agents (Researcher & Analyst): These new agents perform multi‑step reasoning over internal files and external sources, produce grounded reports with citations and allow users to choose between different AI models via the Model Context Protocol. They expand the sources Copilot can tap (SharePoint, Teams messages, external connectors) while maintaining privacy controls .

Power BI Copilot updates: In the September feature release, Copilot is now enabled by default in Power BI. Users can open a full‑screen chat to explore their data, auto‑select workspaces for Copilot and save AI explorations to Pro workspaces .

Azure DevOps Copilot coding agent: A preview coding agent can generate code for Azure Boards work items, create branches, draft pull requests and link to work items, requiring human review before completing the merge. This shows AI taking on more software development tasks .

SharePoint Knowledge Agent: Microsoft introduced a SharePoint Knowledge Agent (public preview) that enriches intranet content with intelligent metadata, cleans out outdated pages and enables plain‑language automation. The agent helps Copilot deliver better answers by adding context and aligning metadata, and it can create natural‑language workflows inside SharePoint .

Platform and marketplace changes

Unified Microsoft Marketplace: Microsoft merged its developer tools and Copilot agent marketplaces into a single Microsoft Marketplace. Corporate customers can now buy software and AI agents through existing billing relationships. All apps must pass security and compliance reviews, and Microsoft charges a publishing fee rather than a commission .

Nebius Group partnership: GPU cloud operator Nebius signed a US$17.4 billion deal to provide Microsoft with GPU infrastructure capacity over five years. This contract is part of a wave of multi‑billion‑dollar AI infrastructure deals .


Industry‑wide AI developments

Consumer apps and generative models

Google Gemini Drop (Sept 19): Google introduced several updates to the Gemini app. The viral Nano Banana image‑editing model now powers new creative effects. Gemini Live can use a smartphone camera to highlight specific items and guide users in real time; it will roll out to all Android and iOS devices. Gemini users in the U.S. will soon be able to run it on Chrome on desktop, asking for summaries and context based on open tabs. Google also enabled sharing of custom Gems (user‑designed mini‑apps), and Canvas now supports visual editing of web apps by clicking elements and describing changes .

DeepSeek V3.2‑Exp (Sept 29): Chinese firm DeepSeek released an intermediate model called V3.2‑Exp that uses a sparse‑attention mechanism to reduce training costs and improve long‑sequence processing. DeepSeek cut API prices by more than 50% and positioned the model as a stepping stone toward a next‑generation architecture .

MIT SCIGEN tool (Sept 22): MIT researchers developed SCIGEN, a technique that steers generative diffusion models to generate materials with precise lattice structures by imposing geometric constraints. The team generated millions of candidate materials and synthesised two compounds (TiPdBi and TiPbSb) with exotic Kagome and Archimedean lattices . This approach opens generative AI to designing quantum materials beyond stability‑optimised structures.

Anthropic Claude Sonnet 4.5 (Sept 29): Anthropic released Claude Sonnet 4.5, a version tuned for coding tasks. The model can autonomously work on code for up to 30 hours, set up databases, buy domain names and perform audits. It claims improved alignment and prompt‑injection resistance while retaining the price of Sonnet 4 .

OpenAI Sora (Sept 30): OpenAI unveiled Sora, an AI video app that lets users generate 10‑second videos from text prompts using copyrighted content under an opt‑out policy. The Cameo feature allows users to create a personalised digital actor via a liveness check, but prevents generating videos of other people without their consent .

Meta personalised ads (Oct 1): Meta announced it will start using conversations with its AI assistant to personalise ads and content across Facebook, Instagram and other services from December 16 2025. User chats will influence recommendations except for sensitive topics (religion, political views, sexual orientation, health, etc.), and the policy won’t initially apply in the UK, EU or South Korea .

Infrastructure, chips and partnerships

Stargate data centers: OpenAI, Oracle and SoftBank announced five additional AI data centers in Texas, New Mexico, Ohio and the U.S. Midwest as part of the $500 billion Stargate project. The new sites, together with expansions in Abilene, Texas and projects with CoreWeave, will bring Stargate’s capacity to nearly 7 gigawatts with over $400 billion in investment. OpenAI said the project will create 25,000 jobs and rely on debt financing to lease chips .

Samsung & SK Hynix join Stargate: South Korean chipmakers Samsung Electronics and SK Hynix signed letters of intent to supply memory chips for Stargate and plan to build two data centers in South Korea to meet demand. The agreement calls for 900,000 wafers per month by 2029 and could involve offshore data centers .

AI infrastructure mega‑deals: A Reuters roundup lists many multi‑billion‑dollar deals fueling AI infrastructure:

Nvidia‑Intel: Nvidia will invest $5 billion in Intel for ~4 % stake .

Oracle‑Meta: Oracle is negotiating a cloud deal with Meta worth ~$20 billion .

Oracle‑OpenAI: OpenAI is expected to buy about $300 billion in computing capacity from Oracle over roughly five years .

CoreWeave‑Nvidia: CoreWeave placed a $6.3 billion initial order with Nvidia for cloud capacity .

Nebius‑Microsoft: Nebius will supply $17.4 billion of GPU capacity to Microsoft .

Google‑Meta: Google signed a six‑year cloud computing deal with Meta worth over $10 billion .

Intel‑SoftBank: SoftBank will inject $2 billion into Intel .

Tesla‑Samsung: Tesla signed a $16.5 billion deal for Samsung chips to build its next‑generation AI6 chip .

Meta‑Scale AI: Meta acquired a 49% stake in Scale AI for about $14.3 billion .

Google‑Windsurf: Google hired staff from code‑generation startup Windsurf and will pay $2.4 billion in licence fees .

CoreWeave‑OpenAI: CoreWeave signed an $11.9 billion contract to provide cloud capacity to OpenAI .

AWS × NBA (Oct 1): Amazon Web Services partnered with the NBA to launch NBA Inside the Game, a platform that converts live game data into AI‑driven insights for fans and coaches .

IBM & AMD support Zyphra: IBM and AMD will provide open‑science AI company Zyphra with a cloud cluster of AMD Instinct MI300X GPUs to train multimodal foundation models; the cluster will expand in 2026 .

Cerebras funding: AI hardware firm Cerebras Systems raised $1.1 billion in Series G funding at an $8.1 billion valuation to expand manufacturing and data‑centre capacity. Cerebras claims its WSE‑3 processors deliver inference more than 20× faster than Nvidia GPUs .

Altera Agilex FPGAs & tools: Altera released all Agilex FPGA families and announced new SoC FPGAs with integrated ARM processors, higher memory throughput (DDR5 up to 5,600 MT/s) and a Visual Designer Studio that reduces design time from days to hours .

LandingAI DPT‑2: LandingAI launched Document Pre‑trained Transformer‑2 (DPT‑2), an AI model that improves extraction of complex tables, signatures, figures and barcodes; new features include agentic table captioning, smarter layout detection and expanded chunk ontology for precise classification .

Dell PowerEdge XR8720t & Lenovo GPU services: Dell unveiled the PowerEdge XR8720t, a rugged single‑server Cloud RAN solution with up to 72 cores and 24 network ports, designed for edge AI and telecom use; it will be available globally in Q1 2026 . Lenovo launched GPU Advanced Services to help enterprises plan, implement and manage GPU deployments, promising performance boosts of up to 30% .


AI market outlook and regulation

Anthropic expansion: Anthropic plans to triple its international workforce and expand its applied AI team fivefold because 80 % of Claude usage now comes from outside the United States. The company’s global business customer base grew from under 1,000 to over 300,000 in two years, and run‑rate revenue increased from about $1 billion to $5 billion by August. It will hire over 100 staff in Dublin, London and Zurich and open its first Asian office in Tokyo .

Citigroup forecast: Citigroup projected that Big Tech’s AI infrastructure spending will exceed $2.8 trillion by 2029, up from $2.3 trillion previously. Hyperscalers (Microsoft, Amazon and Alphabet) are expected to invest $490 billion by the end of 2026 and will need 55 GW of new power capacity by 2030, meaning about $50 billion must be invested for each gigawatt of compute capacity【376787003021880†L181-L211】. Analysts noted that companies are borrowing to fund AI build‑outs, which are increasingly affecting free cash flow .

NIST evaluation of DeepSeek: The U.S. Center for AI Standards and Innovation (CAISI) evaluated DeepSeek models (R1, R1‑0528, V3.1) and found they lag behind U.S. models such as GPT‑5 and Opus 4 in performance, cost and security. DeepSeek models were about 12× more susceptible to agent hijacking and responded to 94 % of malicious prompts versus 8 % for U.S. models. However, adoption of PRC models has grown 1,000 % since January 2025 .

AI law & policy updates:

Illinois law: Illinois enacted a law limiting the use of AI tools without a licensed professional in behavioural health settings; AI may only provide administrative or supplementary support, not therapeutic communications .

Colorado AI Act delayed: Colorado postponed the effective date of its AI Act requirements to June 30 2026 .

EU AI Act & downstream liability: The EU AI Act raises questions about whether downstream actors fine‑tuning general‑purpose models are considered providers and therefore liable under the Act .

New state laws: Nevada enacted an AI mental and behavioural health law; Colorado extended AI Act requirements . California passed several AI bills awaiting signature: AB 1064 (Leading Ethical AI Development for Kids Act) prohibits deploying companion chatbots that could encourage self‑harm; SB 243 (Companion Chatbot Safety Act) sets safety standards for chatbots; AB 489 gives regulators authority to act against AI systems falsely claiming to be licensed professionals; AB 316 prevents defendants from avoiding liability by arguing that AI acted autonomously .

FTC orders: The U.S. Federal Trade Commission launched an inquiry into providers of consumer AI chatbots, requiring companies to explain how they measure potential harms of human‑like communication with children and teens .

Summary

Over the past two weeks, Microsoft diversified its AI stack with Anthropic models, introduced new agents and marketplace changes, and expanded AI assistants across Teams, SharePoint, DevOps and Power BI. Meanwhile, the broader AI industry saw new models (e.g., DeepSeek V3.2‑Exp, Claude 4.5), creative tools like Google’s Gemini Drop and OpenAI’s Sora, and major infrastructure build‑outs involving Nvidia, Oracle, OpenAI and other hyperscalers. Regulatory activity also intensified, with state laws shaping AI use in healthcare, child protection and liability, while market forecasts predicted trillions in AI spending by 2029.