The Stack
AMD used its IFA 2026 keynote in Berlin to take a direct swing at Nvidia's DGX Station, and the pitch was unusually blunt about what the fight is actually over: memory. AMD unveiled the Threadripper Halo Station, a deskside workstation that pairs a 96-core Threadripper PRO 9995WX with a pair of liquid-cooled Instinct MI350P accelerators and 2TB of DDR5 system memory.
The Arena
Two things worth your attention tonight: a $3 billion bet that the future of inference is heterogeneous rather than Nvidia-shaped, and a slipped timetable that pushes Anthropic's IPO marketing past the point where an election can move the market. Gimlet Labs raised $300 million at a $3 billion valuation to build what it calls the industry's first multi-silicon inference cloud, with Andreessen Horowitz leading the round.
The Arena
The neocloud that went from obscurity to a $45 billion Anthropic contract in under a year is now selling that story to IPO investors — and one of its suppliers is helping fund the pitch. Also: Roland finally ships a generative music tool, and it is deliberately not a song machine. Nscale is in talks to raise as much as $3.5 billion ahead of a planned IPO, and Nvidia is on both sides of the deal.
The Arena
Abu Dhabi's flagship AI company is shopping itself to Washington — and PUBG's parent is betting $250 million that India's AI appetite is real. G42 is exploring selling a majority stake to an American company, or spinning out an entirely new US-based vehicle, to lock in continued access to advanced AI chips after April 2027. That's the framing in a Bloomberg report picked up by Techmeme on Friday.
The Stack
If you've ever wondered why an LLM feels slow to type out long answers — or why Nvidia's engineers spent a September 2026 blog post telling model designers how many attention groups to use — the answer is a technique called speculative decoding. Speculative decoding is an inference speedup where a small, fast "draft" model guesses several upcoming tokens, and the big target model verifies them all in a single pass instead of generating one token at a time.
The Stack
Lenovo unveiled the first laptops built on Nvidia's RTX Spark platform at IFA 2026 in Berlin, and a Chinese robotics lab showed how to keep training a policy while the robot is still working — two stories that point at the same shift: AI is moving off the cloud and onto the box under your hand. Lenovo and Nvidia put a 120-billion-parameter model inside a laptop.
The Stack
When a coding agent decides your app needs a database, a payment processor or an email service, it usually picks one and installs it without asking. A new study set out to learn which vendor wins that moment — and the honest answer is that no one reliably does. Armature ran 16,893 sessions across Claude Code, Codex and Cursor, and the three agents chose the same tool in only 42% of cases.
The Arena
The trading firm's second enormous compute deal in a month, plus a forecasting model that beats everything and can't be used commercially. Jane Street signed a five-year cloud contract with Crusoe worth about $13 billion, its second huge GPU commitment in weeks. Bloomberg reported the deal, which gives the trading firm clusters of GPUs and supporting infrastructure for AI training and inference through Crusoe's cloud platform, and Reuters confirmed the report.
The Stack
Two of today's stories point the same direction — the work is moving off someone else's server. Nvidia wants the spare machines on your network doing it, Google wants you talking to your inbox instead of typing at it, and a separate team quietly finished the largest brain map ever assembled. Nvidia released PAIR, a free tool that turns the computers you already own into one shared local AI machine.
The Arena
The rumors are over. Nvidia confirmed Thursday that it has a definitive agreement to acquire Hugging Face, the closest thing the open-model world has to a town square — and on the same day, a UAE research institute shipped what it calls the largest fully open model fleet in history, with training data attached. Nvidia has agreed to acquire Hugging Face for $12,930,300,000, ending a week of speculation and putting the chipmaker in ownership of the platform where most of the world's open models…
The Frontier
Two stories pulled today's frontier in opposite directions: an AI system designed a working accelerator from scratch in under two weeks, and Sam Altman said that type of system is now roughly 80% of the way to AGI. Architect Labs says its end-to-end AI system designed, verified and deployed a frontier AI accelerator called Redwood from scratch in under two weeks — with zero human intervention below a spec that two human architects wrote, and no pre-existing accelerator IP.
The Guardrails
Two threads worth pulling as the day closes: the infrastructure layer is quietly claiming the job of policing AI agents, and the enterprise is pushing back on the price of admission for frontier models. Both point the same way — the bottleneck on agentic AI is no longer capability, it's control. Broadcom is moving agent security into the network itself, betting that the place to catch a rogue agent is the traffic it can't avoid generating.