The one thing
Microsoft has built a new in-house AI chip to power its cloud
Why it mattersMicrosoft now builds its own chips for Azure and Microsoft 365, so ask what that means for your AI pricing and capacity.
Seventy: Tuesday, 13 October 2026
What happened. Microsoft introduced Maia 300, its latest custom AI chip, built into Azure servers and racks. Executives from Azure hardware, Microsoft AI and Microsoft 365 are presenting it as part of a full-stack approach to running AI efficiently.
67 governments are funding 184 national AI projects, CNAS count shows
Why it mattersGovernments in 67 countries are funding 184 national AI projects, so expect public-sector buyers to demand local compute, models and data.
What happened. A CNAS tracker counts 184 government-backed sovereign AI projects across 67 countries as of mid-2026, covering compute, models and data. Arm says the projects share priorities in chip skills, computing capacity, home-grown tech industries and AI use in public services.
03Models
AWS lets Southeast Asian and Taiwanese customers reach OpenAI's GPT-6 models across Regions
Why it mattersIf you build in Southeast Asia or Taiwan, demand spikes no longer have to throttle your OpenAI-powered features on .
What happened. Amazon Bedrock now serves OpenAI's GPT-6 Sol and GPT-6 Luna to customers in Thailand, Malaysia, Singapore, Indonesia and Taiwan, routing requests across AWS Regions when local capacity runs out. GPT-6 Astra is offered only from the Singapore Region.
Amazon switches AgentCore short-term memory from per-event fees to usage-based billing
Why it mattersAgents that write many small events now pay by data volume, so your memory bill follows usage instead of event count.
What happened. Amazon Bedrock AgentCore now charges for short-term memory by data ingested, retrieved and stored, replacing the old flat rate per event. Ingestion costs $1.00 per GB, and pricing covers three dimensions in total.
05Chips
Microsoft puts its own Maia 200 chip into production, with Maia 300 due 2027
Why it mattersMicrosoft is making its own chip the main engine, so expect Azure pricing to follow Maia economics.
What happened. Microsoft's Maia 200 chip is now running production workloads including Copilot, handling OpenAI models, Microsoft's own MAI models and open-source ones. Maia 300 arrives in 2027, which Microsoft expects to give 40% more tokens per megawatt than leading GPU systems, with multi-gigawatt rollout also starting that year.
06Safety
UK government expands AI-driven cyber defence after pilots fixed over 1,000 vulnerabilities
Why it mattersUK rules on reporting serious AI incidents and near misses are being drafted now, not later.
What happened. The UK government announced on 12 October that it will extend two pilots which found and fixed over 1,000 potential cyber vulnerabilities in government systems. It is also weighing incident and near-miss reporting rules for AI under the Cyber Security and Resilience Bill, and the AI Security Institute is evaluating the newest models.
AWS shows how to host small llama.cpp scoring models on SageMaker
Why it mattersAWS now offers a managed container for serving small routing models on CPU, so you can skip GPU capacity for simple classification calls.
What happened. AWS published a guide to serving decision models, which score a fixed set of choices such as routing a request or picking an agent action, with llama.cpp on Amazon SageMaker AI. It uses an AWS container to serve compressed models on CPU or GPU machines.
08Funding
Salesforce agrees to buy analytics startup Loops to strengthen Tableau
Why it mattersSalesforce is buying metric-explanation analytics for Tableau, so a standalone vendor selling the same capability now competes with your provider.
What happened. Salesforce agreed to buy Loops, a startup whose AI tracks company performance, flags why key numbers moved, and points out where to improve. Salesforce will fold it into its Tableau analytics push.
Apple researchers pool idle Mac GPUs into a shared inference cloud
Why it mattersCompanies with fleets of Macs get a research route to running AI workloads on idle hardware instead of renting GPUs.
What happened. Apple researchers describe Minions, a system that pools idle GPUs on Apple Silicon Macs into a distributed virtual GPU cloud for running machine learning inference. It targets the cost, limited elasticity and queue times of centralized GPU infrastructure.
That's the whole issue.
9 stories, about 3 min 18 sec. Nothing was left out to fill the time.
Words of the day
- ASIC
- A chip designed for one job only, such as running AI models, rather than a general-purpose processor. Google's TPU is one.
- llama.cpp
- An open-source C/C++ runtime for running quantized language models efficiently on ordinary CPUs and GPUs.
- Apple Silicon
- Apple's own ARM-based chips, combining CPU, GPU and shared memory, used in current Macs.
“Someone's sitting in the shade today because someone planted a tree a long time ago.”