Meta Goes Frontier, Microsoft's AI Chip, and SpaceX's $40B Bet

01 of 04

Meta Finally Swings for the Fences With Muse Wonder

On Monday Meta's Superintelligence Lab released Muse Wonder and Muse Wonder Full, the company's first real shot at a frontier model since Alexandr Wang took over its AI reorg in August 2025. Until now Meta's Muse lineup was Muse Spark, a lightweight model built for speed. Wonder is the big one.

The specifics, from TechCrunch and Meta's model card:

• Price: $4 per million input tokens and $20 per million output tokens on Meta's API.
• Context window: 1 million tokens.
• Access: Muse Wonder is on the Meta Model API and Muse Code. Wonder Full is only for Muse Code subscribers and is not on the API.
• Competition: Meta says both land in the same range as OpenAI's GPT-6 Astra and Anthropic's Claude Fable 5.1, and that Wonder beats Fable, Opus, and Astra on MemoryBench, a test of how well a model remembers and pulls together what you have told it.
• Where it was trained: Meta's Prometheus data center in central Ohio, which came online this year.

The selling point is what Meta calls "active memory." Research scientist Kavya Kopparapu says in the launch video that the goal is a model that remembers what matters to you "over days, weeks, months, and beyond."

The model card also has some numbers worth chewing on. Wonder Full scored 88.4% on CyberGym, a cyber capability test, refused 98.7% of risky bio prompts on BioTIER in English, and posted a 0% harmful action rate on Agentic Misalignment. Meta has not said whether the weights are open, and the full safety report is "coming soon."

The memory pitch is smart, and it is also exactly the thing that should make you nervous coming from Meta. A model that remembers everything about you for months, built by the company whose whole business is knowing everything about you, is a lot of trust to ask for. Also, Meta used to be the open-weights champion. Shipping a frontier model with no word on the weights and a safety report that is not out yet is a real change in personality, and I do not think it is an accident.

02 of 04

Microsoft's Maia 300 Is a Bet on Needing Nvidia Less

Microsoft took the wraps off Maia 300 on Monday, the third generation of its in-house AI chip. The deep dive comes from Rani Borkar, who runs Azure hardware, and Saurabh Dighe, and the specs are real hardware, not a teaser:

• Memory: eight HBM4 stacks per chip, 288 GB of capacity and 18 TB/s of bandwidth.
• Process: the compute sits on TSMC's N2P node.
• Rack: 96 accelerators in a liquid-cooled 167 kW rack, double the density of Maia 200.
• Scale: up to 128 racks tied together optically into a superpod of 12,288 chips.
• Claims: up to 7 times the rack-scale token throughput of Maia 200, and more than 2.5 times the performance per watt.

It is built for inference, meaning running models rather than training them, and Microsoft says it will power Copilot, GitHub Copilot, and Microsoft Foundry. CNBC reports Nvidia and AMD stay in the picture for training. Reuters and others report Microsoft plans to deploy more than a million Maia 300 chips in 2027, though that number is not in Microsoft's own post, and unlike Amazon's and Google's chips, you cannot rent these directly.

Here is where I land on it. Every big cloud is building its own chip for the same reason: Nvidia's margins are their cost. Running a chatbot for hundreds of millions of people is mostly inference, so that is exactly where a homegrown chip pays off first. I am not going to get excited about "up to 7x" against their own last chip with no benchmark details attached. The number I care about is whether a Copilot seat gets cheaper for a small business, and I will believe it when the invoice shows it.

03 of 04

Microsoft's Brad Smith Wants London, Not Washington, Writing the AI Rules

Microsoft president Brad Smith went on a BBC podcast and said the "moment has arrived" for global AI safety rules, and that the U.K. is better placed than the U.S. or the EU to lead. His reasoning: "In Washington, D.C., people don't look at the U.K. and they don't look at London the way they look at Brussels. That's a strength."

He was not vague about what he wants. He backed three things: a legal duty to notify an organization if rogue AI agents breached its systems, approval from outside evaluators before AI labs ship a product, and kill switches that can pause or slow AI if humans lose control. On the voluntary pledges the industry keeps signing, he said they are a "good step," but "it's the companies that are not as responsible that you need to worry about."

London is already moving. On Monday, the U.K. government announced it is stepping up AI defenses. AI Minister Kanishka Narayan and Cyber Minister Liz Lloyd laid out the plan: extend two government pilots that used frontier AI to find and fix more than 1,000 potential vulnerabilities in government systems, look at toughening the Cyber Security and Resilience Bill to cover AI loss-of-control risk at hospitals and utilities, work out how to respond to serious incidents involving autonomous AI agents, and set up a way to report AI near misses. Narayan's message to the labs was blunt: "if you cannot adequately control the risks, do not release the system."

Context matters here. Politico reported in September that the White House asked OpenAI and Anthropic to hold new models back from British testers until a U.S. review. On Monday the U.K.'s AI Security Institute confirmed it can still access and test the latest models, according to TechCrunch.

Look, it is easy to roll your eyes when a trillion-dollar company asks for regulation. Big players love rules they can afford and startups cannot. But I have spent this whole week writing about AI agents filling out forms and poking at websites nobody told them to touch. A rule that says you have to tell people when a rogue agent got into their systems is not a crazy ask, and neither is near-miss reporting. That is how aviation got safe. Whether Washington lets London set the agenda is another story, and I would not bet the ranch on it.

04 of 04

SpaceX Wants to Borrow $40 Billion for Nvidia Chips

The Financial Times reported that Elon Musk's SpaceX is in talks to raise about $40 billion in debt to pay for Nvidia chips, according to a Bloomberg follow-up carried by China Daily Asia. The breakdown is roughly $10 billion in bank loans and $30 billion in investment-grade debt. Apollo Global Management is leading the financing, and Pimco is one of the investment managers looking at it. The deal would be expected to close in 2027.

Sources described it as "among the biggest-ever debt financings for the AI buildout," and also said talks are early and "could end without a deal being completed." SpaceX, Nvidia, and Apollo did not comment. SpaceX shares slipped 1.2% to $169.79 after hours on the news.

What gets me is the word "debt." For two years the AI buildout ran on cash piles and equity from the richest companies on earth. When a rocket company starts lining up $30 billion in bonds to buy GPUs, the financing has moved into a different kind of risk. Chips lose value fast. Loans do not. If the demand holds, this is a smart way to grow without diluting anybody. If it does not, somebody is holding paper backed by very expensive, very used hardware. I am watching who buys those bonds.