AI Brief: Meta Tops Benchmarks, Amazon's $7B Bet, Microsoft's Maia 300, Anthropic's IP Defense, IMF Jobs Warning

The week's biggest AI moves landed on several fronts at once: Meta staked a new claim at the frontier, Amazon moved to lock up a key AI software startup, Microsoft doubled down on custom silicon, Anthropic went on the offensive against IP theft at scale, and global economic institutions began issuing their starkest warnings yet about AI's impact on the workforce.

Meta Launches Muse Wonder Models, Claims to Beat Claude and GPT-6 at the Frontier

Meta unveiled two new models — Muse Wonder and Muse Wonder-Full — that the company says outperform Anthropic's Claude Opus 5.5 and OpenAI's GPT-6 Astra on key reasoning and multimodal benchmarks. The launch marks a significant strategic pivot: Meta, long the champion of open-source AI, is now tying its frontier capabilities to paid subscribers rather than releasing weights freely.

The move signals that Meta's leadership believes it has caught — or surpassed — the frontier, and that the economics of staying there require a revenue model. For years, Meta's open-source releases kept pressure on closed competitors by commoditizing yesterday's top-tier models. The shift toward paid access suggests the company sees diminishing returns in that strategy now that it has models it considers genuinely competitive at the top of the stack.

The benchmark claims will face independent scrutiny, and the AI industry has learned to treat self-reported leaderboard wins with healthy skepticism. But the directional signal is clear: the era of Meta as a principled open-source counterweight to OpenAI and Anthropic has given way to a three-way competition among closed or semi-closed frontier labs.

Amazon in Talks to Acquire AI Startup Decart for $7 Billion

Amazon is in advanced discussions to acquire Decart, an AI startup best known for its chip-agnostic inference software and 3D world-modeling capabilities, at a reported valuation of $7 billion. Decart was previously courted by Anthropic, making Amazon's move a direct play to keep a strategically valuable asset out of a key competitor's hands.

The acquisition target's chip-agnostic architecture is the core of Amazon's interest. AWS has invested heavily in its Trainium line of custom accelerators, but software that can run efficiently across hardware vendors would allow Amazon to reduce its dependence on Nvidia while also making Trainium a more attractive platform for third-party developers. Decart's 3D world models add a longer-term bet on physical AI and simulation, areas where AWS has lagged behind competitors with deeper robotics and embodied AI portfolios.

If the deal closes, it would be among the largest pure AI startup acquisitions to date and would further concentrate frontier AI capabilities inside the hyperscalers — a trend regulators in the US and EU have been watching with growing concern.

Microsoft Unveils Maia 300 AI Chip, Plans One Million Units by 2027

Microsoft announced the Maia 300, its third-generation in-house AI accelerator, built on TSMC's 2-nanometer process. The company claims the chip delivers 40% more tokens per megawatt than leading GPU alternatives and achieves 7× the rack-scale throughput of its predecessor, Maia 200.

Microsoft plans to deploy more than one million Maia 300 units across its Azure data centers by the end of 2027, a deployment scale that would make it one of the largest custom silicon rollouts in the industry's history. The chip is designed primarily to run inference workloads for Microsoft's own AI products — including Copilot and Azure OpenAI Service — rather than to be sold directly to customers.

The announcement is the clearest signal yet that Microsoft intends to reduce its infrastructure dependency on Nvidia, whose H100 and Blackwell chips have become a bottleneck and a budget line item that affects every major AI company. By controlling its own silicon at scale, Microsoft can optimize cost-per-token for its highest-volume workloads and gain negotiating leverage in its ongoing relationship with Nvidia.

Anthropic Cuts Distillation Attacks by 90–99%, Bans Tens of Thousands of Accounts

Anthropic announced that new safeguards have reduced so-called distillation attacks — attempts by external actors to extract Claude's capabilities by harvesting its reasoning traces at scale — by between 90 and 99 percent. The company also said it has banned tens of thousands of accounts engaged in systematic capability extraction.

Anthropic named Chinese AI startup Z.ai as a significant perpetrator, accusing it of using proxy networks to route large volumes of queries through Claude's API in order to train competing models on its outputs. The new protections restrict access to extended reasoning traces and add behavioral signals that flag and block high-volume extraction patterns before they complete.

The announcement is notable both for its directness — publicly naming a specific company is unusual in the AI industry's typically diplomatic disputes — and for what it reveals about the scale of the problem. Tens of thousands of banned accounts suggests that distillation-as-a-service has become a significant enough operation to warrant coordinated defensive measures rather than one-off enforcement. Expect other frontier labs to announce similar countermeasures as the practice grows.

IMF Warns AI Is a 'Slow-Moving Tsunami' Threatening 40% of Global Jobs

Speaking at the IMF and World Bank Annual Meetings in Bangkok, IMF Managing Director Kristalina Georgieva delivered one of the institution's most direct warnings to date about AI's labor market consequences. She described the technology as a "slow-moving tsunami" and said that 40% of global jobs face significant AI-driven disruption — with the exposure concentrated heavily in higher-income economies.

The IMF's analysis finds that 60% of jobs in advanced economies involve tasks now reachable by AI automation, compared with roughly 23% in low-income countries. The disparity reflects both the higher share of cognitive, service, and professional work in wealthy nations and the slower adoption of AI tools in economies with less infrastructure and lower capital investment in the technology.

Georgieva urged governments to accelerate workforce retraining programs and social safety net reforms before displacement peaks. The IMF's framing — a slow-moving wave rather than a sudden shock — is intended to convey that governments have a window to act, but only if they treat preparation as urgent now. The political challenge is that the window may close faster than institutions typically move.