If you needed proof that the AI industry is sprinting in two directions at once, this past week delivered it. On one end, Meta shipped a keychain-sized AI companion that looks like a toy. On the other, AI labs quietly demonstrated that frontier models can now do original cryptanalytic research good enough to crack codes that stumped human experts for two decades. And underneath both, the infrastructure money keeps getting bigger, with a single cloud contract now running into the tens of billions of dollars. Here’s what actually happened, and why it matters.
Meta wants AI to live on your keychain
At Meta Connect, Mark Zuckerberg unveiled the Muse Charm, a palm-sized wearable built around Meta’s Muse AI agent — the same assistant represented internally by a cartoonish avatar named “Jolly.” According to Zuckerberg, the team’s goal was to take “the whole Muse experience, including the whole real-time voice and avatar stack,” and compress it into something that clips onto a keychain. The device has a small touchscreen for the avatar, a fingerprint sensor to wake it, and voice interaction, with no phone unlock required to talk to it (TechCrunch).
Meta hasn’t announced pricing, but says the Charm is on track to ship in time for the 2026 holidays. The comparisons to Tamagotchi and Labubu-style pocket toys aren’t accidental — outlets covering the launch have leaned hard into the “pocket pet” framing (DailyCoin; Axios). It’s a deliberate positioning choice: AI wearables have become something of a punchline after a string of underwhelming amulet-style devices, so Meta appears to be leaning into “low-stakes and fun” rather than pitching the Charm as a serious productivity tool. It’s also notable timing — Meta is getting a dedicated AI hardware product to market before OpenAI does, even if analysts remain skeptical that novelty devices translate into a durable platform (CNBC).
Two frontier models just cracked Enigma messages that beat human cryptographers for 20 years
Meanwhile, in a much quieter corner of the internet, something genuinely remarkable happened. Turing’s “other test” isn’t the famous imitation game — it’s a reference to his WWII work breaking Nazi Enigma codes. A small handful of wartime Enigma messages have remained unsolved for decades because of transcription errors or unusual encoding quirks, and cryptography hobbyists have kept trying to crack them ever since.
This week, two different researchers used two different frontier models to do exactly that. Developer Carter Leffen prompted OpenAI’s newly released GPT-6 Astra to search an Enigma message archive, and the model conducted its own archival research, picked up on contextual clues, built a working Enigma simulator, and recovered the plaintext of a message that had “baffled researchers since 2005.” Separately, Jack Willis used Anthropic’s Claude Opus 5 to break a different unsolved message by reasoning about known officer signature patterns (TechCrunch).
Veteran cryptologist Frode Weierud, who has spent years working on these archives, said GPT-6 Astra was “behaving like a very professional cryptanalyst,” compressing what would normally take a human researcher weeks or months into about two days. Only seven of these historical Enigma messages remain unbroken. It’s a small, almost whimsical demonstration — but it’s also a genuinely useful signal about where model capability actually stands: this isn’t a benchmark score, it’s an open historical problem that real experts hadn’t solved, done autonomously, end to end.
The infrastructure race keeps getting more expensive
Behind both stories, the money keeps flowing at a scale that’s hard to fully process. Anthropic and Akamai announced a seven-year, $11.6 billion cloud infrastructure agreement, with Akamai supplying compute capacity to support Anthropic’s growing model training and inference demand — and reportedly giving Anthropic a path to a stock stake in Akamai as part of the arrangement (TechCrunch; Akamai). It’s the latest in a string of massive multi-year compute commitments from AI labs, and it lands the same week reports surfaced that Anthropic’s founders are working to secure voting control ahead of a planned IPO.
Not to be outdone, British “neocloud” operator Nscale — a vertically integrated AI cloud platform that bundles software, compute, and power — announced it had raised $3.36 billion in pre-IPO convertible financing, led by Third Point and joined by investors including Nvidia, Apollo, and Citadel. Nvidia alone is committing an additional $1 billion expected to close in mid-November. Nscale says it’s now sitting on more than $103 billion in total contracted value ahead of its planned U.S. listing (TechCrunch; Nscale).
The takeaway
Put together, this week is a decent snapshot of where the AI industry actually is in late 2026: consumer-facing AI is being repackaged as something small, cute, and impulse-buyable, model capability is quietly crossing into “does original expert-level research” territory, and the infrastructure bill underneath it all keeps climbing into the tens of billions per deal. None of these three threads is really separate — the wearables need the models, the models need the compute, and the compute is what all that infrastructure money is chasing. Whether the “pocket pet” approach to AI hardware sticks around past the holidays is genuinely an open question. Whether models keep getting better at open-ended research problems like Enigma-cracking is not — that trend line looks very clear.
No responses yet