The Week You Could Finally Own the Model

NVIDIA ships an open 74-billion-parameter model and the laptops to run it, Sam Altman and Bernie Sanders fight over open weights, Meta goes closed, and memory prices keep the bill climbing.

For two years the advice was simple: rent the smartest model you can afford and don't think about the hardware. This week the other option showed up at retail. NVIDIA released an open model built to run on a laptop, the laptops ship Friday, and a free piece of software turns the PCs you already own into a shared inference pool. The same week, the CEO of the biggest closed-model company said society will have to accept "some fairly severe cyber incidents" from open models, and a U.S. senator answered with a call to ban superintelligence outright. Owning your AI just got easier and more political at the same time. Join us at AI Tangle as we untangle this week's happenings in AI.

On Oct. 13 NVIDIA laid out its RTX Spark launch week, and the centerpiece isn't the laptop. It's the model. Nemotron 3.5 Bolt 74B-A9B, released Oct. 15, is a mixture-of-experts model with 74 billion parameters, about 9 billion of them active per token, packed into 24GB with 2-bit quantization that NVIDIA says was trained in from the start so quality holds. The company calls it fully open: weights, training data and recipes. It runs at launch in llama.cpp, vLLM, Ollama, Unsloth and LM Studio, and the agent frameworks Hermes and OpenClaw can call it directly. NVIDIA claims it's faster than popular dense models and smarter than mixture-of-experts models with a similar footprint. Those are the vendor's numbers. No independent benchmarks exist yet.

The hardware lands the next day. RTX Spark laptops ship Friday, Oct. 16, led by Microsoft's Surface Laptop Ultra, which starts at $2,599 with 24GB of unified memory and tops out at 128GB in the full configurations from ASUS, Dell, HP, Lenovo and MSI. The third piece is software you don't pay for: NVIDIA's Personal AI Router, which left beta this week, discovers the PCs on your local network, installs the inference engines and sends each request to whichever machine can serve it. Pair it with a smaller model NVIDIA is shipping Oct. 23 for voice and device assistants, and you have, for the first time, an open model, a machine sized for it and a way to share the load across an office, all from one vendor in one week.

Here's the business read. We spent Thursday's Deep Dive arguing you should rent the frontier and own the rest. This is what owning the rest looks like when it reaches a product catalog: a laptop under $3,000 running a model you can inspect, on data that never leaves the building, with the overflow routed to the desktop in the next office. The catch is the one the rest of this issue is about. The best model in the world is still closed, the memory in that laptop costs more every quarter, and the politics of open weights got loud this week. Buy the stack for the jobs that need privacy and repetition, keep a cloud account for the hard ones, and don't let anyone tell you it has to be either.

Read the full story →

The share of enterprise AI tokens now served by open-weight models, up from 23% a year earlier, in Enterprise Technology Research's September survey of 200 enterprises. Respondents expect 41% within 12 months, and 69% of the companies running open models in production name cost savings as the reason.

Source: Techstrong.ai

Sam Altman told Politico's Decoded podcast that open source models will bring a "coming tidal wave of cybersecurity problems," and that society will "have to just accept some fairly severe cyber incidents from open models, in exchange for the liberty that comes with that." On Oct. 12 Sen. Bernie Sanders answered on X: "No, Mr. Altman. We won't accept a severe AI cyberattack in exchange for 'liberty,'" followed by a call to pause advanced AI and permanently ban artificial superintelligence, the substance of the bill he and Rep. Greg Casar introduced last month. Sanders has the public behind him on trust, if not on the ban: a Quinnipiac poll of 1,202 adults taken Sept. 24 to 27 found almost three-quarters have little or no trust in AI company leaders, and 86% want independent safety standards even if they slow development. For business, the point isn't who wins the argument. It's that the open model you download this week is now a political object, and the compliance questions about where it runs and what it's allowed to do will arrive before the regulations do.

The lab that gave the world Llama shipped Muse Wonder and Muse Wonder Full on Oct. 12, its first models it calls frontier-class, and neither comes with weights. Wonder costs $4 per million input tokens and $20 per million output, in line with OpenAI and Anthropic, and the pair was trained at Meta's Prometheus campus in Ohio. The timing is the story: the same week NVIDIA went all-in on open weights, the company that used to lead on them walked away, and Enterprise Technology Research's survey shows why it mattered, with Llama's share of production open-model deployments falling from 55% to 35% in a year while Google's Gemma took the lead. If your open-model plan was built on Llama, this is the week to re-read it.

Microsoft's Scott Guthrie announced Maia 300 on Oct. 12, a third-generation accelerator built for long-context, memory-heavy inference, and said the company will scale the Maia family to multi-gigawatt deployments starting in 2027. Microsoft expects 40% more tokens per megawatt than leading GPU systems, a claim it hasn't benchmarked publicly, and says Maia 200 already runs Copilot in production. Owning the intelligence isn't only a laptop decision. The biggest renters of NVIDIA silicon are all building their own, measured in tokens per megawatt, because power is the constraint they plan around. When your cloud provider owns the chip, the price of a token is a knob they control, which is one more reason to keep a local option for the work that runs all day.

Barclays reports that DDR4 contract prices rose 24% in three months through October, with DDR5 up 20% over the same stretch and DDR5 spot prices up another 8% in October alone, Wccftech reported Oct. 12. The bank's note says Micron's DRAM supply stays constrained through 2028 and that new customer agreements now track market pricing instead of the old floor-and-ceiling structures, which is a polite way of saying nobody is locking in a discount. Every machine in this issue, from the 128GB RTX Spark to the 512GB Mac Studio Apple promised for late October, is priced off that curve. If the own-it math works for you today, it works better than it will in January. We dig into what the memory run does to every purchase you make in Thursday's Deep Dive.

Anthropic's alignment science lead said there's better than a 10% chance AI kills everyone within a decade, and the number went everywhere. Kreps, who runs Cornell's Tech Policy Institute, argues the decimal point is the problem: unlike nuclear risk, AI has no established causal chain to put a probability on, and a single p(doom) figure hides the policy trade-offs it's supposed to inform. If you're going to own a model, it's worth knowing what the doom debate can and can't measure.

Read it →