On 12 June, the United States government ordered two of the world's most capable AI models switched off for every user outside America.
Anthropic, the company that makes them, complied with the order the same day. For eighteen days, every non-US organisation using those models had no access and no appeal.
Access to one of them (Fable) returned on 30 June, after Anthropic agreed to security conditions set by the US Commerce Department.
The other of the two (Mythos) remains restricted to American users to this day.
I watched this from down-under wondering about the implications for NZ.
I should say up front that I am not a policy analyst by any stretch. I run an AI consulting business and a media (newsletter + podcast) company, and I have spent the past week reading two open letters (most people only saw one) out of the United States about AI and open-weight models, the coverage around them, and the New Zealand documents they made me go back and check.
Some of the terms involved were ones I had to look up myself. What follows is my attempt to translate the whole thing into plain English, because I think the situation gives us back in NZ a lot to think about.
What actually happened last week
Two open letters went to Washington within 48 hours of each other last week.
Letter #1
On 22 July, 179 startup founders, organised as the newly formed Little Tech Association and including Y Combinator, wrote to the White House asking the government not to restrict open-weight AI models. Letter here.
An open-weight model is one whose underlying file, the weights, can be downloaded and run on infrastructure the user controls.
Once that file is downloaded, nobody can reach in and take it back.
The closed models most of us use, like ChatGPT or Claude or Copilot, work the other way, living on the provider's servers and reached through a subscription.
It is also, strictly speaking, different from open source, because most open-weight models publish the finished file without the training data or code behind it.
The founders' letter called restrictions on these models "a tax on intelligence".
It also cited a survey finding that nearly half of responding Y Combinator companies already run most of their production workloads on open-weight models.
Letter #2
On 24 July the giants followed, with a letter titled "Open Weights and American AI Leadership" that went out with 25 corporate signatures and grew to 51 within a day. NVIDIA, Microsoft, Meta, Google, OpenAI, IBM, Cisco and Mistral are all on it.
Jensen Huang, NVIDIA's chief executive, used the first post he has ever made on X to share it, and it drew more than 60 million views.
His summary line was that open models "strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty".
Anthropic, the company whose models were switched off in June, is missing from that list. So is Amazon, which is Anthropic's largest investor.
None of the timing here was accidental as the trigger emerged two days earlier. Here's the breakdown:
The White House had accused Chinese company Moonshot AI of secretly building its new Kimi K3 model by distilling Anthropic's Fable.
Distillation means using one model's outputs to train another, and it is a standard technique across the industry.
The accusation is contested by independent experts, because Fable had only been publicly available again for about two weeks when Kimi K3 was announced, yet the US Treasury threatened sanctions anyway.
Moonshot's response was to publish Kimi K3's weights to the world on 27 July.
Nathan Lambert, one of the most respected independent analysts of open models, had already called it "clearly the strongest open model ever released".
Additionally, Xi Jinping visits Washington on 24 September, and AI is expected to dominate the talks.
So the high-level sequence, laid end to end, runs like this:
In June, America switched off two frontier models for the whole world, then partly reversed it eighteen days later.
In July, it accused a Chinese lab of stealing from an American one and threatened sanctions.
Within 48 hours of that accusation, 230 of its own companies, the sellers of AI and the buyers of AI, signed letters saying the answer is more open models, not fewer.
The day the second letter went out, the White House confirmed Xi will visit in September with AI at the top of the agenda.
As far as I can tell, no New Zealand outlet has covered any (or much) of this as a New Zealand story beyond the headline from the US. I think that is a mistake, because the argument these letters are having impacts us.
The divide is not open versus closed
Most coverage framed last week as a fight between open and closed AI, and I do not think that is the real divide.
It's not a technology fight. It's a control fight about who has control over 'the switch' when it counts. The divide that counts is between countries that
Have a fallback option if their AI access changes
And countries that do not.
And New Zealand sits in the second group without ever having decided to be there.
The mechanism, stated plainly, is that a closed model reached through a subscription can be modified, repriced, or withdrawn by the provider, and, as June demonstrated, by the provider's government.
An open-weight model that has been downloaded cannot be taken back.
The founders' letter concedes this from the other direction when it notes that once weights are published they "can be copied across borders at near-zero marginal cost". Whatever else the two letters disagree on, both sides accept that closed access is revocable and downloaded weights are not.
The June episode moved the whole question from theory into the historical record. Not a scenario any more. A precedent, with a date attached and a government taking a big stand. The trap is thinking of this as a one-off, and the restriction was indeed partly reversed in eighteen days.
But a precedent only has to fire once to change how sensible organisations plan.
The finer details actually make it worse for a country like ours, not better. South Korea is a US treaty ally, and its flagship telco still lost access to one of those models. The stated reason was alleged China ties amounting to a tiny office, so ally status was no protection when it counted. Aaron Levie, Box's chief executive and a letter signatory, drew the conclusion the day the June order came down. "The big winner in this Fable situation is going to be open weights models", which I believe we'll see play out.
Why sellers and buyers said the same thing
A fair reader should be suspicious of the letters themselves.
NVIDIA sells more chips when models multiply, and Microsoft is hedging its dependence on OpenAI. The venture funds want cheap models for their portfolios, and none of the frontier labs that signed are promising to open their own best models.
The letters are in effect a lobby play, not an altruistic one, written by people with big money riding on the outcome.
That is exactly why the convergence is informative, because companies that sell closed AI subscriptions signed a letter saying customers must be able to leave.
The strangest detail is that the American government is arguing both sides of this at once.
The same week the Treasury threatened sanctions, the White House's own AI adviser was publicly attacking Anthropic for refusing to sign.
David Sacks told the company to "read the room", and on the All In podcast's 24 July episode he and his three co-hosts all opposed any ban on Chinese open models.
Chamath Palihapitiya, the investor who co-hosts that show, put numbers on it during the same episode. His estimate was that American firms could pay somewhere between $26 and $56 per million tokens for closed model access, while competitors on open models pay under a dollar for similar capability.
His wider point is a deeper view that the capability gap between open and closed models, in his reading, has narrowed much faster than the pricing gap, so the discount for going open keeps growing.
America does not currently have a position on open AI models. It has an argument between its own officials, and that argument has flip flopped three times in six weeks.
The startups on the other side of the market said the same thing in more practical language.
Their letter describes founders who need to switch hosts, run a model inside a customer's secure network, and keep serving users even if a platform changes its pricing or direction. When both sides of a market that normally fight each other tell their government the same thing in the same week, the safest conclusion is not that either side is being generous. It is that the underlying economics have moved.
The economics are worth thinking about for New Zealand. The pattern emerging in the US, described repeatedly by Levie, is simple to see.
Frontier models get used for the hardest parts of a task, the planning and the reviewing, while cheaper or open models do the routine bulk of the work.
The open models keep absorbing whatever the frontier works out next, offered at a much cheaper cost.
Back to Nathan Lambert, his assessment is that the gap between the best open model and the best closed model is now measured in a few months, not years.
Every business benefits from that gap being small, including businesses that never run an open model.
A capable model that can be run at cost on any cloud sets a potential ceiling on what anyone can charge for the closed ones.
There is a longer history underneath this, which Clément Delangue, the chief executive of model-sharing platform Hugging Face, told in an interview back in May.
The T in ChatGPT stands for Transformer, an architecture Google published openly in 2017.
American openness built the field, American labs then closed up, and the strongest open models today come from China, which is why DeepSeek, Qwen and Kimi are now what most US startups and academics reach for when they want a model they can control.
Read against that history, the letters are America trying to reverse a position it gave up, in the same week its government was demonstrating why the rest of the world might want the reversal to succeed.
New Zealand's side of the ledger
Now set the New Zealand documents next to all of that and read them side by side.
Our national AI strategy, published in July 2025, is deliberately and openly adoption-focused. It says plainly that it emphasises adopting AI over developing it, which is a defensible call for a small country. But it contains no mention of open-weight models, no mention of open source, and no treatment of what happens if access to the closed tools it assumes ever changes. The challenge is that the question the letters spent 48 hours arguing about does not appear in it anywhere.
Government practice followed the same path without anything that resembled a public debate. A BusinessDesk investigation found that Microsoft Copilot became the New Zealand government's default AI product through an add-on to existing licensing agreements, without a competitive tender. The thing is, nobody ever decided closed foreign AI would be our national approach. The setting arrived as a line item on a licensing renewal, and no one has revisited it since.
Business adoption, on the numbers we have available, follows exactly the same pattern. The AI Forum's most recent productivity survey found 73% of AI-using New Zealand businesses run off-the-shelf tools, with only 13% investing in anything custom. In my own client conversations, the stack is almost always a set of subscriptions to closed models governed by another country's export rules. The June switch-off did not come up in a single one of those conversations until I raised it.
Australia, meanwhile, published a National AI Plan in December 2025 that includes GovAI, a centralised, Australian-hosted platform on which government agencies can build their own AI solutions, and its GovAI Chat assistant has been in trials across the public service since April. It has not stopped there either. Earlier this month Canberra established an Office of AI inside the Prime Minister's own department, announced national standards for AI data centres that it intends to legislate, and signed both a $7 billion infrastructure deal with OpenAI and NEXTDC and a collaboration agreement with Anthropic, with Albanese saying Australia should be "more than a data warehouse". Whatever the execution ends up looking like, the question was asked and answered nationally, and it keeps being re-asked. We have no equivalent.
To be super clear, none of this means New Zealand businesses should rip out their subscriptions, and there are two concessions to make in that point.
The first is that closed frontier models are still ahead, and for plenty of work they are the right tool.
The harder truth is that running open models well is not trivial, and even in the US, Levie concedes that enterprise use of open weights is still more experimentation than at-scale production.
The case for taking any of this seriously in NZ has not in any way started with this article either.
Catalyst's Donald Christie has argued that our digital infrastructure should not depend on a single foreign company.
Tom Barraclough from the Brainbox Institute Brainbox Institute published a sovereign AI discussion paper last year.
Dr Karaitiana Taiuru JP, PhD, MInstD has made the case many times that Māori data sovereignty ultimately requires locally governed infrastructure.
The letters in the US just gave their argument 230 corporate signatures in agreeance.
What I would do with this
There are three plays for any firm at any stage of their AI journey, none of which require a data centre or a single line of code:
**Map the dependency on paper.**One page listing every workflow that now depends on AI, which model each one runs on, and whose jurisdiction that model sits in. Add what would happen to each if access ended on someone else's terms. For most of the businesses we work with, nobody has written any of this down. This isn't a standing policy, rather a habit, re-checked at every contract renewal, because the answer changes as models and terms change.
**Match the model to the job.**The discipline both letters describe is paying frontier prices only for frontier problems, and running cheaper or open models for the routine volume tasks. The practical starting question is what the same workload would cost on a smaller model, and whether the output difference is one anyone would notice. We help businesses make those decisions if they cannot do so on their own.
**Treat the silence as an opening.**The New Zealand conversation about any of this has not started, because the strategy is silent, the coverage is absent, and most boards have never heard the question asked. Anyone who raises it in their next leadership meeting is early, not paranoid. Mark Laurence is doing an excellent job of surfacing in the public arena the need for more top down leadership in this country, his most recent piece on Business Breakfast here.
Both letters were written to an American government, about American leadership, for American ends.
But the country with the most riding on the answer is not the one with 51 companies arguing about it. It is the kind of country that found out on 12 June that its access to the world's most capable AI was a policy setting in someone else's capital.
The fights over AI are not just about the technology. They are about who holds the kill switch, and at the moment nobody in New Zealand is even worried about what their means for business continuity moving forward.
Passionate about all things AI, emerging tech and start-ups, Mike is the Founder of The AI Corner.
Subscribe to The AI Corner
