For the last few years, the case for personal, on-device intelligence came with an asterisk. Yes, owning your intelligence is better for privacy, for autonomy and for judgment. But the best models live in someone else’s data center, and the open ones aren’t close.
This week, three pieces of news, from three very different corners of the industry, removed that asterisk. Put them side by side and the conclusion is hard to avoid: the gap has closed.
Signal one: the market has already voted
On October 24, a16z shared a chart from its Charts of the Week on what is happening at OpenRouter, the platform developers use to route requests across many different models. Two numbers stand out.
Token demand on OpenRouter is up about 27x since last November, from roughly 6 trillion tokens a week to roughly 165 trillion. And open-weight models now handle about 72% of those tokens, up from about 28% last November.
In less than a year, on a platform where developers can pick any model with a line of config, the share going to open weights went from barely a quarter to nearly three quarters. Nobody forced that shift. Developers, choosing freely, decided that models they could inspect, host and run anywhere were good enough for most of their work.
Signal two: IBM says the quiet part out loud
A few days earlier, IBM CEO Arvind Krishna told investors on the company’s Q3 earnings call that open-weight models now make brilliant sense for on-premise enterprise use. The Register’s account is worth reading in full.
In 2024, Krishna said, a broad benchmark might score the top closed model at 70 and the open model at 20 or 30. By late 2025 it was 70 to 50. Today, he says, open-weight models deliver “anywhere from 25 percent to 10 percent of the token cost of the frontier models. That is very attractive given that they’re now 90 percent as effective.”
Then he framed the choice every organization now faces: “Would you like something that’s 10 percent of the cost if it’s 90 percent as good as the best, or do you want to pay that 10 times higher cost?”
And he answered the follow-up question himself. How many tasks really need the frontier? “Some do, but the vast majority, summarizing an email, examining a contract, answering a customer call, composing an email, all of those can be done as well with the much cheaper answer.”
That’s not an open-source activist talking. That’s the CEO of IBM, on an earnings call, telling Wall Street that open weights on hardware you control are the rational default.
Signal three: the frontier labs blink
The third signal comes from the frontier labs themselves. Max Woolf (minimaxir) laid out a dated timeline of every OpenAI and Anthropic price change since June, under the title “OpenAI and Anthropic cut their LLM prices by up to 80% within a few months. Why?” It became one of the busiest threads on Hacker News this weekend.
The cuts are real. OpenAI cut GPT-5.6 Luna by 80% in July, and Woolf notes that without the cut it would have lost to DeepSeek V4 Pro on quality per dollar. Anthropic cut cache-read prices by 75% on Fable 5.1, which for agentic coding (where cache reads are 96 to 99% of token spend) works out to roughly a 70 to 75% cut in real cost. Opus 5.5 came in 40 to 60% cheaper in total than Opus 5. And GPT-6.1 Sol, by Woolf’s math, makes that model family cumulatively 75% cheaper.
Woolf’s own conclusion is blunt: “The idea that token prices will suddenly rise is more a fantasy.” One commenter on the thread put the dynamic plainly: the big labs are “competing against free/local/open weight models.” Another summed it up in four words: “There is no moat.”
TINOMO
I’ve been saying it for a while now, so often it became an acronym: TINOMO. There is no moat.
Here’s why. Intelligence degrades the friction of information processing. That saved friction doesn’t stay with whoever made the model. It distributes through the value chain, to the developers, businesses and people who put the intelligence to work. A model is a lead, not a moat, and leads shrink.
In my view there are only two ways to keep a moat. You can iterate at the edge faster and better than anyone else. Or you can excel at hard-to-replicate physical processes: rockets, factories, batteries, chips. I call that the Elon Edge. Bits copy; atoms don’t.
The culture of the big labs leans on the Jevons Paradox: when the cost per unit of a resource falls, total use rises, often by more than the savings. That’s true for everyday intelligence. Cheaper tokens mean more summaries, more code, more agents. But it isn’t true at the frontier edge. Few people can actually use frontier intelligence. Most people don’t have a use case for an Einstein oracle at any price. Their needs are emails, contracts and calls, exactly what open weights now handle.
Good enough and cheap enough
Put the three signals together.
Open weights are now good enough: about 90% as effective, by IBM’s own measure, and enough for “the vast majority” of everyday work. They are cheap enough: 10 to 25% of frontier token cost. And the market has already moved: roughly 72% of OpenRouter tokens. The frontier labs have responded the only way they can, by cutting prices hard.
When the closed models were far ahead, the cloud was a moat. If you wanted the best intelligence, you rented it, on the provider’s terms, with your data flowing through its servers and its business model shaping the results. That trade made a kind of sense when the alternative was a model scoring 20 against 70.
It makes much less sense at 90%. When an open model is nearly as good, costs a fraction as much, and can run anywhere, the question changes. It’s no longer “which cloud do I rent?” It’s “why am I renting at all?”
Cheaper closed APIs don’t change that answer. A price cut makes renting cheaper. It doesn’t make it ownership. Your memory, your context and your judgment still sit inside someone else’s system, under terms that can change with the next policy update or the next IPO. A local model costs nothing per token, and no vendor can raise that price.
From the enterprise rack to the personal node
Krishna was talking about enterprises running AI in “two or three racks.” But the logic doesn’t stop at the enterprise. Every generation of open models gets smaller, faster and more capable at a given size. What runs in a rack today runs on a workstation tomorrow, and on a laptop or a phone soon after. You can already run capable smaller open models on a Mac or a single consumer GPU, though not yet the very largest ones.
Krishna was also honest about the hard part. Enterprises, he said, ask: “Where are my kernel engineers? Where are the people who worry about the underlying network?” That’s the real gap now. It isn’t the quality of the models. It’s the work of deploying them. As I noted in “Seven Billion Intelligences”, Mozilla CTO Raffi Krikorian calls it the missing double-click installer.
That gap is a solvable engineering problem, the kind open ecosystems have solved before. Red Hat made Linux routine for the enterprise. The question for the next few years is who will make the personal node routine for individuals, families and small teams. That’s the work POND exists to do.
Open source is not open use
There’s a catch I want to be honest about: open source is not open use. Many of the best open-weight models are too big to run locally unless you have a significant budget. So in practice most people still reach them through cloud deployments. The weights are free; the hardware isn’t.
That’s why I don’t think POND systems should depend on one giant model doing everything. They should distribute the intelligence workload across a system of interactions: your node, its tools, its memory, smaller specialized models, and peers you choose to work with. Raw model inference is one part of that system, not the whole of it.
That’s how we at Intelligence Systems are building Inti, a POND personal node technology: Intelligence for the Rest of Us. Not an Einstein oracle in a data center, but capable intelligence that lives with you, on hardware you can actually afford.
The POND Principle
The POND Principle describes where this leads: Human-Machine Intelligence (HMI) that is
Personal & Private: intelligence that starts with the person and keeps their data, identity, memory and context under their control. When the model is good enough to run locally, there is no reason to ship your life to someone else’s server.
On-Device & On-Premise: intelligence that runs on hardware the person or community owns. IBM is now making this case for the enterprise rack. The same economics apply to your desk and your pocket.
Nodal & Networked: each person’s environment is a node, and nodes connect voluntarily, peer to peer, sharing only what’s authorized. Open weights mean every node can run the same capable intelligence, so the network is a network of peers, not of renters.
Decentralized & Distributed: a counterbalance to centralized AI, with no single authority owning or controlling the network. A world where 72% of tokens already flow through open weights is a world where that counterbalance is no longer hypothetical.
Why this matters beyond price
This isn’t only about saving money. As I wrote in “We Are Already Superintelligence”, humanity is already a superintelligence: an emergent network of people, language, knowledge and exchange. And in “Superintelligence Strategy: The Perfect Machine”, I argued that the goal isn’t one machine that stands above that network, but intelligence that strengthens it.
A world of rented intelligence concentrates that power in a few companies. A world of owned intelligence, a personal node for every person and connected by choice, distributes it. For a while, the cost and quality gap made the first world look inevitable. This week’s numbers say otherwise. The gap closed, and with it the strongest argument for centralization.
Cloud AI will still have a role. Some tasks really do need the frontier, as Krishna says. But the default is shifting: everyday intelligence runs on hardware you own, and you reach for the cloud by choice, not by necessity.
You can start today
The POND Starter Kit on Hugging Face collects 13 open models for chat, reasoning, coding, vision, speech and memory that run on a laptop, a Mac or a single consumer GPU. It’s a starting point for your own personal node, and we want to hear which models earn their place on your machine.
Join the POND
If you’re a builder, a thinker, a policy-maker or simply someone who wants the intelligence in your life to work for you, come help build what comes next.
Start at pond.is
Join the conversation on Discord, including the weekly PONDcast
Suggest models, projects and standards in GitHub Discussions
Write to us at foundation@pond.is
The gap closed. The personal node is next.
Welcome to the POND. 🐸