Microsoft Goes All-In on Custom AI Chips With Maia 300
Microsoft is making its boldest move yet in the custom silicon race. According to reports citing The Information, the company is in talks with TSMC for production capacity exceeding 300,000 Maia 300 chips, with delivery targeted for 2027 and longer-term ambitions of over one million units.
The Maia 300 represents the third generation of Microsoft's in-house AI accelerator program. While official specifications haven't been released, the predecessor Maia 200 — a TSMC 3nm inference chip with 216GB of HBM3e memory — delivered over 10 petaFLOPS of FP4 performance. The new chip is expected to handle both training and inference workloads, a significant expansion of scope.
The scale of Microsoft's ambitions is notable: Azure Maia general manager Andrew Wall has suggested that the reported volume doesn't fully reflect the program's scope. If Microsoft can deliver on these numbers, it could meaningfully reduce its dependence on Nvidia — a strategic imperative as AI compute demand continues to outstrip supply.
Anthropic Launches Claude Haiku 5.5: Cheaper, Faster, and Sharply More Capable
Anthropic released Claude Haiku 5.5 on October 7, positioning it as a major upgrade for developers running high-volume, cost-sensitive AI workloads. The headline numbers are hard to ignore: input tokens cost $0.10 per million (under 100K-token prompts), and Anthropic claims the model runs roughly 75% cheaper than the Haiku 4.5 it replaces.
Performance improvements are equally striking. On the OSWorld 2.1 benchmark, Haiku 5.5 scored 72.4% on offline tasks compared to just 15.7% for its predecessor — a nearly fivefold jump that reflects Anthropic's focus on agentic capabilities. The model features a 1-million-token context window and 128K max output tokens, with configurable effort levels defaulting to medium.
The aggressive pricing signals Anthropic's intent to compete not just on model quality but on the economics of deployment, particularly for enterprise use cases where token costs at scale can dwarf development expenses.
OpenAI DevDay: GPT-6.1 Sol and the Rise of Always-On AI Agents
OpenAI's DevDay 2026, held September 29 in San Francisco, delivered two headline announcements that are still reverberating across the industry.
GPT-6.1 Sol is an agentic-focused model that OpenAI says approaches the performance of its flagship GPT-6 Astra at roughly one-fifth the token cost. Priced at $2 per million input tokens and $10 per million output tokens (with cached input dropping to $0.10), Sol is optimized for coding, computer use, and long-running workflows. It comes with a 1-million-token context window and showed a 7% improvement on OSWorld 2.0 while cutting factual errors from 11.4% to 7.7%.
Dots are OpenAI's new always-on personal AI agents — each running on its own cloud computer with browser access and connections to over 4,000 apps via ChatGPT, Slack, and Microsoft Teams. Unlike one-shot interactions, Dots are designed to persist, monitoring tasks and acting independently over time. They're currently available on Pro, Business Premium, and Enterprise plans, though EEA, Switzerland, and UK markets are excluded at launch.
In a notable footnote, DeepLearning.AI reports that OpenAI held back GPT-6.1 Astra because the model failed to meet internal security standards — a rare public acknowledgment that cutting-edge capability can outpace safety guardrails.
Huawei Closes the Gap on Nvidia in China's AI Chip Market
Fresh IDC data paints a shifting picture in China's AI accelerator market. Nvidia's share of AI chip shipments in China has dropped to 34%, while Huawei has surged to 28%, shipping approximately 572,000 units — up from around 300,000 the previous year.
The trend reflects both the impact of ongoing US export controls, which have limited Nvidia's ability to sell its most powerful chips to Chinese customers, and Huawei's aggressive investment in its Ascend AI processor lineup. For the global AI hardware landscape, the data suggests that export restrictions are accelerating domestic chip development rather than slowing China's AI ambitions.
Microsoft's Brad Smith Calls for Urgent Global AI Safety Rules
In remarks reported by TechCrunch on October 12, Microsoft president Brad Smith made a pointed argument for urgent AI regulation, noting that consumer products like milk and washing machines face more regulatory oversight than the companies building AI systems.
Smith singled out the UK as the country best positioned to lead global AI governance, citing its regulatory reputation and track record since the 2023 AI Safety Summit at Bletchley Park. The call comes as AI regulation remains fragmented globally: the EU's AI Act is being implemented in phases, Australia is developing its own framework with national standards due by late 2026, and the US continues to rely primarily on voluntary industry commitments rather than binding legislation.