AI News Briefing

The Big Picture

Meta shipped a frontier model, which is a sentence nobody could write about Meta for 14 months. Muse Wonder Full lands at 56 on the Artificial Analysis index — two behind Claude Opus 5.5, level with Sonnet 5.5, eight clear of the Muse Spark 1.3 it replaces. The Terminal-Bench 4.0 jump, 33% to 57%, is the number that should make people sit up: that is hours-long agentic work, not a trivia quiz. Then comes the asterisk, and it is a big one. The model posting those scores is not on the API. It ships only inside Muse Code at $50 a month, while the weaker Wonder — 52, four points down — is what developers can actually build on at $4/$20 per million tokens. Meta rebuilt its lab and then put the result behind a subscription, which is a strange way to win back the developers who left when Llama stopped getting a successor. Meanwhile the money turned twitchy on cue: Broadcom’s Hock Tan told a Singapore audience that investors are gambling in a casino, and Asia obliged, with the Kospi down more than 6%. The capability curve and the confidence curve are no longer pointing the same way.

Chip stocks slide worldwide after Broadcom’s CEO compares AI investors to casino gamblers

Speaking at a Singapore event Monday, Broadcom CEO Hock Tan said “a lot of investors have no clue what they are investing in” and likened the AI trade to gambling, per Bloomberg. Asian markets sold off Tuesday: South Korea’s Kospi fell more than 6% and Japan’s Nikkei 225 nearly 3%, led by memory makers SK Hynix and Samsung. In Europe ASML dropped 3.3%, BESI as much as 6%, and Infineon and STMicroelectronics more than 5%. Tan, whose company still forecasts AI chip revenue growing from about $58B this fiscal year to $230B by fiscal 2028, also warned that access to capital — not chips or memory — could become the next constraint on AI growth. Intel’s Lip-Bu Tan shared the panel and made a similar point.

The warning is coming from inside the house: Tan sells the accelerators, so when he says the buyers of his customers’ stock don’t understand the trade, the market has no comfortable way to discount it.

FT: Chinese tech firms bought over a million Nvidia gaming and workstation chips this year

Chinese groups have bought more than 1 million Nvidia gaming and workstation GPUs in 2026 — mostly RTX 5090 and RTX Pro parts that fall outside US export controls — to cover a shortfall in AI compute, the FT reports. Alibaba, ByteDance and Tencent are named as the main buyers, with combined capex possibly more than doubling to roughly Rmb700B ($104B) this year. Industry estimates put China’s 2026 domestic AI chip supply near 3.5 million units against demand above 5 million; the consumer cards are slower than Huawei’s Ascend 910C on most tasks but adequate for simpler inference work. Nvidia has said its share of China’s AI accelerator market is effectively zero and its guidance assumes no China data-center revenue.

The FT’s arithmetic is the point: even with a million smuggled-in-plain-sight consumer GPUs, China is still roughly 1.5 million accelerators short — which tells you the export controls are shaping what gets bought, not how much.

Anthropic meets investors Wednesday with its S-1 still not public, IPO slipping to mid-November

Anthropic is reportedly hosting prospective investors at its San Francisco headquarters on October 14, but its S-1 has still not appeared on EDGAR. Reports now point to a listing as early as mid-November, with formal marketing possibly starting the week of November 9 and trading before Thanksgiving — none of it confirmed by the company. The targeted valuation is above $2 trillion, roughly double the $965B post-money mark set by May’s $65B Series H, on an offering that could raise about $100B. A leaked prospectus reported by Reuters shows 2025 revenue near $4.6B against $386M in 2024, a net loss around $42B of which more than $34B is fair-value accounting on liabilities rather than cash burn.

The prospectus flags its own tail risk: the Defense Department has designated Anthropic a national-security supply-chain risk, a disclosure public investors would be buying along with the revenue curve.

Meta ships Muse Wonder, its first frontier model — but the strong one skips the API

Meta released Muse Wonder and Muse Wonder Full on Monday, its first frontier-class models and the first flagship since Llama 4. Artificial Analysis scores Wonder Full at 56 on its Intelligence Index v4.3 at max reasoning effort and Wonder at 52, against 58 for Claude Opus 5.5 and 53 each for GPT-6 Astra, Gemini 4 Argon and Claude Fable 5.1. Wonder Full places top-three on six of the index’s ten evals, with its largest gain on Terminal-Bench 4.0 (33% to 57%). Meta is pitching “active memory,” which carries user information across sessions, and “active focus” for keeping multi-step agents on track. Wonder is on the Meta Model API at $4/$20 per million tokens; Wonder Full is restricted to Muse Code subscribers at $50/month and up. Meta’s own card reports 85.6 on MemoryBench against 81.5 for Opus 5.5.

Artificial Analysis’s framing is the one that matters: frontier performance, but only the weaker model is available via API — so the number Meta is advertising is not the number developers can buy.

Microsoft unveils Maia 300 with 288 GB of HBM4, aimed at multi-gigawatt 2027 rollouts

Microsoft detailed Maia 300, its third-generation in-house AI accelerator, on Monday with production targeted for 2027. Compute chiplets are built on TSMC’s N2P node in a multi-chiplet package, with eight HBM4 stacks for 288 GB at 18 TB/s and a 950 W SoC rating, up from the Maia 200’s 750 W. Reported compute is up to 13.8 PFLOPS at FP4 and 6.9 PFLOPS at FP8. A rack holds 96 accelerators in a 167 kW envelope — 27.6 TB of HBM, about 1.7 PB/s aggregate bandwidth — scaling to superpods of 128 racks and 12,288 accelerators. Microsoft claims up to 7x the rack-scale token throughput of Maia 200 and roughly 40% more tokens per megawatt than leading GPU systems; both figures are company-reported and unreproduced.

Microsoft says Maia 300 is meant to become its primary inference chip for its own products — which is the part Nvidia reads, since Copilot-scale inference moving in-house is demand that never shows up in a GPU order.