TL;DR
5 takeawaysMicrosoft announces Maia 300, its third-generation in-house AI chip, and says Maia deployments will scale to multiple gigawatts from 2027 as the primary inference system for its products
Sources: MicrosoftThe Verge
HCLTech signs a Global Premier Partnership with Anthropic and will set up a dedicated Anthropic Business Unit and Center of Excellence for Claude deployments
Sources: HCLTech via PR Newswire
Ambi, founded by former Plaud exec Peter Mo, raises $26 million and ships apps for iPhone, Mac and Apple Watch that can listen in the background
Sources: TechCrunch
Short: Meta prices Muse Wonder at $4.00 per million input tokens and $20.00 per million output tokens on its Model API
Sources: TechCrunchMeta
Short: Odyssey opens a free public research preview of its Odyssey-3 world model
Sources: The Decoder
The Details
Story 01 / 03
Microsoft’s Maia 300 aims to make in-house silicon its main AI inference engine
- What happened
- Microsoft announced Azure Maia 300, the third generation of its custom AI accelerator, in a blog post by Scott Guthrie. Microsoft says it will scale the Maia family to multi-gigawatt deployments starting in 2027, “making Maia the primary inference system for Microsoft products and services.” Maia 300 compute chiplets are built on TSMC’s N2P process, with eight HBM4 stacks giving 18.0 TB/s of peak memory bandwidth within a 950-watt TDP. A rack holds 96 accelerators in a 167-kilowatt envelope. Microsoft says it expects Maia 300 to deliver 40% more tokens per megawatt than leading GPU-based systems, and up to 7x the rack-scale token throughput of Maia 200. Production deployments begin in 2027. Maia 200 is already serving Microsoft Copilot, with GitHub Copilot and Microsoft Foundry to follow by end of year, according to Microsoft.
- Why it matters
- The Verge says the efficiency gains will let Microsoft rely mainly on its own chips rather than ones from Nvidia and AMD. Microsoft AI CEO Mustafa Suleyman told The Verge that running Microsoft’s models on Maia is “a critical focus” for his team.
- Enterprise impact
- Azure customers may see more Microsoft services, and eventually Foundry model serving, running on Maia. Watch how pricing and availability compare with GPU-backed options.
- Startup impact
- Inference cost per token is the number to track. If Microsoft’s efficiency claims hold, cheaper serving inside Azure could shift where startups host models.
- What to focus on
- More details promised at Microsoft Ignite next month, and independent numbers once Maia 300 reaches production in 2027.
Discuss this story
Comment with your GitHub account. One thread per story.
Story 02 / 03
HCLTech and Anthropic sign a Global Premier Partnership
- What happened
- HCLTech announced a Global Premier Partnership with Anthropic focused on safety, cybersecurity and governance. It will establish a dedicated Anthropic Business Unit with a Center of Excellence that builds agentic AI solutions and provides secure sandbox environments for enterprise Claude deployments through forward-deployed engineers. HCLSoftware becomes an Anthropic ISV partner and will embed Claude and Claude Code into select products, including HCL BigFix and HCL AppScan. HCLTech will also integrate Claude and Claude Code into its AI Force engineering platform, with policy guardrails and human oversight for high-impact changes, the company says.
- Why it matters
- Large systems integrators are becoming a main channel for frontier models into banks, hospitals and manufacturers. Anthropic President Daniela Amodei said HCLTech manages “those kinds of systems” for such customers.
- Enterprise impact
- HCLTech clients get a packaged route to Claude with sandboxes, governance and staff training included.
- Startup impact
- Integrator partnerships raise the bar for agent startups selling into the same accounts: expect buyers to ask for comparable guardrails and oversight.
- What to focus on
- Which industries the new business unit targets first and when the HCLSoftware integrations ship.
Discuss this story
Comment with your GitHub account. One thread per story.
Story 03 / 03
Ambi raises $26M for an AI assistant that listens all day
- What happened
- Ambi, built by former Plaud exec Peter Mo, said it has raised $26 million across multiple rounds led by Luminous Ventures and ODM Huaqin, TechCrunch reports. It is releasing apps for iPhone, Mac and Apple Watch that work like meeting notetakers but can optionally keep listening to capture context through the day, with an agent that offers suggestions during meetings. A desktop app that watches users’ work and a hardware device with a screen, speaker, camera and microphones are planned.
- Why it matters
- Ambient capture raises consent questions for people nearby. Mo said ambient capture is off by default and can be turned off at any time, and the company says it doesn’t store recordings and encrypts data, though it didn’t say what encryption it uses. Mo said disclosure is a problem the industry hasn’t solved.
- Enterprise impact
- Always-on assistants on employee devices need clear recording and consent policies before anyone switches them on.
- Startup impact
- Context capture is a crowded race; trust features like visible indicators may matter as much as accuracy.
- What to focus on
- Whether Ambi’s apps and device add indicators that alert other people they are being recorded.
Discuss this story
Comment with your GitHub account. One thread per story.
Worth a Glance
Muse Wonder pricing lands
TechCrunch reports Muse Wonder costs $4.00 per million input tokens and $20.00 per million output tokens on the Meta Model API, while Muse Wonder Full stays limited to Muse Code subscribers. Meta says the models were trained at its Prometheus data center.
TechCrunch(opens in a new tab)Meta’s research notes on Muse Wonder
Meta’s research blog post on the new models.
Meta(opens in a new tab)Odyssey-3 world model preview
Odyssey opened a free public research preview of Odyssey-3, a real-time world model; developers can apply for API access.
The Decoder(opens in a new tab)
One Thing Not to Miss Today
Inference is where the money goes
Microsoft is betting its own silicon on serving models, not training them. Track tokens per megawatt alongside benchmark scores.