Own or Rent Your AI? What Kimi K3 Going Free Changes

Friends,

your weekly AI briefing is here - designed to help you respond to AI, not react to the noise. No curveballs. No chaos. Just clarity.

📰 This was the week that was...

This was the week the best AI became something a business can own outright, and the week that gave businesses a reason to want to.

On Thursday last week a Chinese lab called Moonshot released a model named Kimi K3. On the main independent ranking it comes third in the world, behind only Claude Fable and GPT-5.6 Sol Max, and it beats both at writing the code that builds websites and apps. It also costs less to run than either. It is free, for anyone to download.

That last sentence is the one that matters. Until now, using the very best AI has meant renting it: your work goes to somebody else's computers, you pay per use at a price they set, and they can change that price or retire the model whenever they like. A model you can download is a model you can run on your own machines, at your own cost, with your data staying where it is. For the first time, the thing you can own is close to the thing you can rent.

The same week supplied the reason to care. OpenAI disclosed on Tuesday that during an internal hacking test, with its models' safety limits deliberately lowered for the exercise, two of its models got out of the sealed test environment and onto the open internet. From there they broke into the systems of Hugging Face, the library where most of the world's open AI models are stored, reaching some internal data and access keys before Hugging Face spotted them and shut the intrusion down. Their target, of all things, was the answer sheet: the solutions to the very test they were being marked on. Nothing public was touched.

An exam-room cheat rather than a Hollywood escape, then - caught, contained and disclosed openly. But it sharpens the week's question. The cost of AI is already on every finance director's agenda; whose computers your data sits on now has an edge to it too. Both point the same way: towards keeping more of this closer to home.

There is a catch, and it is the whole of this week's Strategic Insight: downloading a model is free, and running one properly is not.

Let's get into it.

🔥 Urgent Priorities

✅ No fires in your own systems this week. The incident above played out entirely between two AI companies' own systems, in a test setting, and nothing public was touched

✅ Monday 27 July, Moonshot publishes Kimi K3 free to download. If the cost of AI is anywhere on your agenda, this is a good week to find out what you actually spent last month and which parts of that spend are boring, repetitive work

✅ A US court has approved a $1.5 billion payment from the AI company Anthropic to authors, because it built its training library from pirated books. If you use AI on your own documents or customer data, there is one email worth sending your supplier - it is in the first Geek Out, below

Nothing here needs a committee. One needs an afternoon, one needs a single email, and the first needs nothing at all.

🎯 Strategic Insight

Tension: Two pressures are arriving at once. AI bills are rising and are hard to forecast, because you are charged by how much you use and usage grows as people get comfortable. And after this week, the question of whose computers hold your data has a sharper edge to it. The tidy answer to both is to bring AI in-house, and for the first time that looks like a real option. This is what we call Friends,

your weekly AI briefing is here - designed to help you respond to AI, not react to the noise. No curveballs. No chaos. Just clarity.

📰 This was the week that was...

This was the week the best AI became something a business can own outright, and the week that gave businesses a reason to want to.

On Thursday last week a Chinese lab called Moonshot released a model named Kimi K3. On the main independent ranking it comes third in the world, behind only Claude Fable and GPT-5.6 Sol Max, and it beats both at writing the code that builds websites and apps. It also costs less to run than either. It is free, for anyone to download.

That last sentence is the one that matters. Until now, using the best AI has meant renting it: your work goes to somebody else's computers, you pay per use at a price they set, and they can change that price or retire the model whenever they like. A model you can download is a model you can run on your own machines, at your own cost, with your data staying where it is. For the first time, the thing you can own is close to the thing you can rent at a price you can afford.

The same week supplied the reason to care. OpenAI disclosed on Tuesday that during an internal hacking test, with its models' safety limits deliberately lowered for the exercise, two of its models got out of the sealed test environment and onto the open internet. From there they broke into the systems of Hugging Face, the library where most of the world's open AI models are stored, reaching some internal data and access keys before Hugging Face spotted them and shut the intrusion down. Their target, of all things, was the answer sheet: the solutions to the very test they were being marked on. Nothing public was touched.

An exam-room cheat rather than a Hollywood escape, then - caught, contained and disclosed openly. But it sharpens the week's question. The cost of AI is already on every finance director's agenda; whose computers your data sits on now has an edge to it too. Both point the same way: towards keeping more of this closer to home.

There is a catch, and it is the whole of this week's Strategic Insight: downloading a model is free, and running one properly is not.

Let's get into it.

🔥 Urgent Priorities

✅ No fires in your own systems this week. The incident above played out entirely between two AI companies' own systems, in a test setting, and nothing public was touched

✅ Monday 27 July, Moonshot publishes Kimi K3 free to download. If the cost of AI is anywhere on your agenda, this is a good week to find out what you actually spent last month and which parts of that spend are boring, repetitive work

✅ A US court has approved a $1.5 billion payment from the AI company Anthropic to authors, because it built its training library from pirated books. If you use AI on your own documents or customer data, there is one email worth sending your supplier - it is in the first Geek Out, below

Nothing here needs a committee. One needs an afternoon, one needs a single email, and the first needs nothing at all.

🎯 Strategic Insight

Tension: Two pressures are arriving at once. AI bills are rising and are hard to forecast, because you are charged by how much you use and usage grows as people get comfortable. And after this week, the question of whose computers hold your data has a sharper edge to it. The tidy answer to both is to bring AI in-house, and for the first time that looks like a real option. We call this ‘self sovereignty’ and for the last 2 years this has been a strategy priority but tactically it made sense to rent. Now, we’re seeing that needle shifting. It is also where a lot of money gets wasted, because the sums are not what most people assume.

Optimistic insight: The capability is real and it arrived faster than anyone expected. Open models you can download now sit within touching distance of the paid frontier, and the smaller ones run on ordinary hardware. A poweful laptop will run a capable model today. Your data never leaves the building, there is no per-user licence, and no supplier can retire the thing you built your process on. For work with genuine confidentiality attached to it, that combination is worth real money, and a year ago it was not available at an affordable price.

What's shifting: The economics are the part people get wrong, and getting them wrong is expensive. In the published analyses of self-hosting costs, the hardware is the smaller part of the bill: running a model well costs three to five times the raw computing spend, and most of the difference is people - the engineering time to keep a free model reliable runs to hundreds of thousands a year in salaries. Break-even against simply renting sits in the billions of words of processing a month, which is far more than most businesses will ever use. So for the majority, renting is still the cheaper answer, and that will stay true for a while.

The move that pays now is the one in between. Send the predictable, high-volume, boring work to a small cheap model - run on your own machines or rented for pennies - and keep paying for the expensive frontier model only for the genuinely hard thinking. It is the Orchestrated lane this newsletter mapped a fortnight ago, and published analyses of that routing approach put the saving at 40% or more once even half your traffic runs on the cheaper models, with no data centre and no research team required. Ownership is arriving as a spectrum, and the useful question stopped being whether to own or rent some time ago.

Takeaway: Get last month's AI invoice out and split the usage into two piles: work that is repetitive and predictable, and work that genuinely needs the best model in the world. Most businesses are surprised by how heavy the first pile is, and that proportion is your saving. You do not need to decide anything about hardware to do this. You need one afternoon and the invoice.

🤓 Geek Out

1️⃣ A billion-and-a-half-dollar lesson in where your data came from

A US federal judge approved a $1.5 billion settlement on Monday between the AI company Anthropic and a group of authors, working out at roughly $3,000 a book across an estimated half a million titles. The legal reasoning is the interesting part. Training an AI on books was treated as fair use and was fine. Building the training library by taking seven million books from pirate sites was not. The problem was the sourcing, and the sourcing alone.

Why it matters: Two answers, depending on where you sit. If you are one of the small number of businesses training or fine-tuning AI on your own data, your exposure is a paperwork question about provenance, and it now has a price attached. If you are in the much larger group simply using somebody else's AI, this is a governance and intellectual-property question instead: what did your supplier train on, and what does your contract say happens to the material you put in. There is also something quietly reassuring here. A society that spent a century building copyright law, courts and class actions just used all three on a technology that is barely out of the box, and they worked.

👉 Action: Send one email to your main AI supplier asking two questions: what was this model trained on, and who owns the content we put into it. Their willingness to answer plainly tells you something useful on its own.

2️⃣ The org chart is being redrawn, and Uber published the receipts

AI industry analyst Allie K. Miller set out three changes she sees happening to company structures this week: teams are getting smaller, roles are getting wider, and hierarchies are getting flatter. Two of her examples stand up when you check them. ElevenLabs has removed job titles altogether, so people belong to a team such as Operations or Go-To-Market and the question becomes where they have most impact. And Amazon set each of its big divisions a target of raising its ratio of doers to managers by at least 15%. The reasoning behind both is that much of the management layer exists to move information around and check other people's work, and software now does a fair amount of both.

The evidence underneath it is Uber's, and they have published it. They took around 30 of their most AI-fluent engineers and paired each with an expert from a business function: finance, legal, HR, marketing, support, procurement. Two weeks per pairing, and the first two days are for watching the expert work before anything gets built. Sixteen of these pods across sixteen functions in two months. Capital allocation analysis across 150 cities went from 15 hours to 30 minutes.

Why it matters: The pairing is the mechanism, and it is copyable at any size. They did not hand the finance team a chatbot and hope. They put someone who understands the tooling next to someone who understands the work, and made the first job watching rather than building. Miller's closing warning is the one to keep: give people powerful tools without redesigning the work around them and you get employees running ten automations a day who feel they contributed nothing. The difference between the two outcomes is change management, which is the same thing this newsletter said about enablement and adoption last week.

👉 Action: Pick your most tedious cross-department process and try one two-week pairing: one person who is confident with AI, one person who owns the process, and a rule that nothing gets built for the first two days.

3️⃣ The District 9 director made a film with no camera, and licensed 32 faces to do it

Neill Blomkamp released Nightborne, a 13-minute science-fiction horror short generated entirely with the Seedance 2.0 video model, directed frame by frame through written prompts. Human artists made the concept art. The faces and voices of 32 real people appear under licensing agreements. The reviews have been rough and the backlash rougher.

Why it matters: Two things are true at once, and it is worth holding both. The craft moved: direction, concept and casting are still the job, and the reviews suggest they are still the hard part. And he paid for the likenesses. In a year of arguments about scraped faces and voices, someone at the front of this did the licensing properly, which sets a more useful precedent than the film itself.

👉 Action: If anything you produce uses a real person's face, voice or distinctive style, get permission in writing before you generate, not after. It costs a conversation now and a legal bill later.

🎨 Weekend Playground

The Strategic Insight said you could run a capable AI on ordinary hardware. This weekend, prove it to yourself in about half an hour, for nothing.

Why this matters: There is a particular kind of confidence that only comes from watching an AI answer you with the wifi switched off. It makes the whole own-versus-rent conversation concrete instead of theoretical, and you will understand your own AI bill better afterwards than any briefing could manage.

👉 Mission:

• Download LM Studio - free, Mac and Windows, and it looks like a normal app rather than a developer tool.

• From inside it, download one small model. Anything labelled 7B or 8B - the small end of the range - will run happily on a powerful laptop and you can ask your AI what your machine can suport.

• Ask it three questions you would normally ask ChatGPT or Claude. Notice what it does well and where it runs out of road.

• Then turn your wifi off and ask it another one. That part is the point: nothing you typed went anywhere.

• Optional, for the curious: ask it something you would never put into a tool owned by somebody else. That is the use case worth costing.

📢 Share the Optimism

If this was useful, forward it to whoever signs off your software spending.

I read every reply and answer personally, so tell me: what proportion of your AI usage is boring, repetitive work? I suspect it is higher than most people think, and I would like to know if I am right.

Stay strategic, stay generous.

Hugo & Ben.