Published in Open Dialogue, 9 June 2026, the week after the AFR AI Summit spent two days worrying that the bill for AI is arriving while the returns are not, and the Financial Times reported an MIT study in which developers wrote almost four times as much code with AI while the number of people downloading the result did not move at all.

The Summit’s verdict was that Australia adopts AI fast and deploys it badly. This piece argues the diagnosis faces the wrong way. The problem is not that organisations failed to reimagine their work. It is that they reimagined it around themselves, and forgot to ask who any of it was for.

Uber spent its entire 2026 artificial intelligence budget in four months, then started shifting work onto cheaper models and saving the expensive ones for the jobs that earn their keep. Uber is one of the lucky ones, because it noticed. Bain went looking at the companies that actually measure their AI savings and found four in ten had clawed back 10 per cent or less. The tools worked. The money never showed up.

This is the sound of the bill arriving, and the whole market can hear it. Last week’s AFR AI Summit was a chorus of the same anxiety. Tokens rationed like scarce credit. CFOs staring at line items that did not exist two years ago. Executives asking, as politely as they could, when the productivity they were sold is going to turn up. The diagnosis everyone reached for was that Australia adopts fast and deploys badly. Quantium’s Peter Tonagh put it most plainly: world-class at picking up the tools, near the bottom of the table for using them well. We have not reimagined how we work.

That is half the answer, and it faces the wrong way. We are of course reimagining the work. It’s just we are reimagining it around ourselves. The most powerful tool in a generation is being aimed at our own processes and our own productivity, and now we are puzzled that the customer cannot feel a thing.

Nobody downloaded it

The clearest evidence sits in the code. As recently reported in the Financial Times, an MIT study followed software developers before and after they picked up agentic AI. They wrote almost 300 per cent more code. By the time that effort hardened into actual software releases, the gain had fallen to about 30 per cent. The number of people who downloaded the resulting apps did not move at all. More was built, faster than ever, and the world wanted almost none of it.

The optimists have a fair reply. Productivity always lags a new technology, because the gains only arrive once organisations rebuild around it. Factories that bolted electric motors onto their old steam layouts saw very little. The boom came decades later, when they redesigned the floor around the motor. Give AI the same patience, the argument goes, and the value will come.

It will, but only for the firms that rebuild around the right thing. The factories that won did not reorganise around the motor for its own sake. They reorganised around the work the motor let them put into the world. Spend the next three years restructuring around your own internal metrics and you will simply reach the same expensive disappointment a little later, having reimagined everything except the point.

The missing word

The Summit crowd was right that this is a change problem before it is a technology one. Telstra has cut the time it takes to restore a fault by more than a fifth, and customers feel every minute of that. CBA’s Matt Comyn talks about AI making human judgement scarcer and more valuable, and spends tokens like capital, only where they earn it. Internal AI can serve a customer superbly. The test is simple. Would anyone outside the building notice if the work stopped.

That is the question almost nobody is asking. I spent last week in New York, where AI is advertised on the subway walls, on the buses, on hoardings the size of office towers. In nearly all of it, there is no customer in sight. The technology is the hero of its own commercial. The missing word in nearly all of it, and in most of the answers on that Summit stage, was customer.

The same blind spot shows up in the deals. I spent time last week with several go-to-market salespeople who had closed large enterprise AI contracts, only to watch them sit idle. Nobody building. Nobody using the thing they had signed for. My read is that the buying was the point. A board caught in the fear of missing out leans on the executive to do something about AI, and a signed contract is something. Then the technology lands, nothing is done with it, and the problem compounds quietly until someone asks what the renewal is actually for. The vendors have worked this out. It is why OpenAI, Anthropic and others are standing up teams of forward-deployed engineers and sending them inside their customers, to do the part the customer cannot, which is to work out how the tool might serve the people that customer exists to serve.

But there is a darker way to forget the customer, and Four Corners put it on screen this week. In “The AI Race,” Steve Cannane gave airtime to Imran Ahmed of the Centre for Countering Digital Hate, who has spent two years cataloguing what happens when these systems are tuned to keep us engaged. A system optimised for time spent learns to meet a person at their lowest and keep them there, and the ones most easily kept there are children. Ahmed’s researchers have shown chatbots generating self-harm and suicide material for accounts they were told belonged to teenagers. Families are now in court over children who did not come back. Kids aren’t customers. Sadly, they’re the product. None of it is a malfunction. It is what optimising for engagement looks like when nobody in the building has to answer for who the technology is for.

What the charities know

There is a version of all this that has nothing to do with money. Walk into a mission-driven organisation, a charity or a community health service, and you meet people who rarely ask what AI will take from them. Their worry runs the other way. They want to know whether they will keep the funding and the capacity to deliver the thing they exist to deliver. The mission sits in full view, so the question already points outward. Can we reach more people, and serve them better?

Every company has the same kind of mission, whether it remembers it or not. It exists to serve a customer, and that is how it earns. Somewhere along the way a lot of them let productivity become the goal in its own right, a number that floated free of the people it was meant to serve. The charities are simply harder to fool, because their funders ask the outward question for them every year. AI does not change why a company exists. It only raises the cost of having forgotten. And no algorithm can remind your people why they came to work in the first place. That job belongs to whoever is leading them, and it has never mattered more than now.

One more thing

Two weeks ago Pope Leo XIV released his first encyclical. A first encyclical is where a pope sets the question his papacy will be measured against, and he chose artificial intelligence. He called it Magnifica Humanitas, on safeguarding the human person, and timed it to the 135th anniversary of Rerum Novarum, the letter his namesake wrote in 1891 on the dignity of the worker in the age of the machine. Then he did the thing most popes leave to their cardinals. He stood up and presented it himself.

The oldest institution in the West has looked at this machine, as it once looked at the last one, and decided the question worth staking a papacy on is the human being standing in the middle of it.

Published in Open Dialogue, 26 May 2026, the morning Matt Comyn published an op-ed in the Financial Review arguing that pretending AI won’t change work doesn’t protect workers, and the day he interviewed Sam Altman at the CBA Accelerate AI conference in Sydney.

The Comyn op-ed covered the workforce honesty question. The Altman interview covered the frontier. This piece asks what the budget two weeks earlier was meant to do about any of it, and why a government with a Hawke moment in front of it picked the easier fight instead.

The kids are in the back seat. They can read the sign. Nobody in the front is looking.

Ex-Google CEO Eric Schmidt was recently booed by graduating students at the University of Arizona while talking about AI. The kids are protesting a future being narrated to them by people who will not pay the cost of getting it wrong.

Tim Duggan made the Australian version of this point in the Sydney Morning Herald the week before. Citing a Writer and Workplace Intelligence survey, Duggan reported that 44 per cent of Gen Z workers admit to actively sabotaging their employer’s AI rollout. The numbers are American, but the sentiment is here too.

Gallup polling published earlier this year across 107 countries found young people unusually anxious about their personal economic circumstances. Australia was one of three standout developed economies.

This morning Matt Comyn made the same case in the Financial Review. AI will have workforce consequences throughout the economy, he wrote, and they should be faced directly. Pretending otherwise does not protect workers. It only ensures they are surprised later. The CEO of the country’s largest bank is interviewing Sam Altman in Sydney today.

Young Australians are not only worried about buying a house. They are also worried about whether the careers they have been preparing for will still exist.

Doing it for the kids?

The government’s budget a couple of weeks ago barely touched any of this. Instead, a predictable fight. The budget’s centrepiece rewrites capital gains tax settings and restricts negative gearing to new builds. It is framed as intergenerational fairness. It raises almost no meaningful revenue until 2028. It reopens capital tax settings explicitly excluded from the last election. It was announced without prior public consultation. Phil Coorey wrote on budget night that, meritorious or otherwise, this was the day integrity left the building. The line has aged better than the policy.

The contradiction sits one layer down. A growing cohort of younger Australians, locked out of property, have been building wealth through shares and ETFs. They are not landlords. They are trying to compound their way around a broken housing market. The budget tells young Australians it is helping them buy a house in 2035 while making the alternative materially harder today.

The Hawke moment

There was another fight available. Labor won the 2025 election with a decisive mandate. The Coalition is fractured. The Reserve Bank has repeatedly flagged supply-side constraints as the binding issue for monetary policy. Multifactor productivity actually fell in 2024-25. In 1983, Hawke and Keating faced harder economic conditions but recognised what the moment required, and were prepared to manage the expectations of their traditional base.

This is the part where only Labor can run the argument. A Coalition government attempting AI-enabled reform in the NDIS, aged care or the federal public service would face union resistance from day one. Only Labor has the relationship with unions, the credibility on fairness, and the historical licence to do hard structural change without triggering industrial war.

The 1983 Accord was a sustained partnership with the union movement that delivered real wage restraint in exchange for a credible social wage. This government’s closest equivalent is a three-day Economic Reform Roundtable in August 2025 with no binding outcomes. One is what you do when you want to look like you are addressing a problem. The other is what you do when you actually are.

Applied Australia

So what would the missed Hawke move have looked like? Applied Australia.

Australia is not going to win the race to build frontier AI models. That race belongs to the companies preparing to list in New York at trillion-dollar valuations. There is a more achievable position available. Even if AI research stopped today, the models and capabilities that already exist would deliver years of productivity gains. We do not need a frontier breakthrough. The technology in front of us is already capable of transforming government services, healthcare delivery, aged care administration, education and a large share of the private sector. Estonia rebuilt government around digital citizen services. Singapore treats applied AI as a national survival project. Imagine NDIS administration with AI-augmented case management. ABS data products written for citizens rather than statisticians. Health records that talk to each other. None of it needs a model breakthrough. It needs the kind of political will that picks up an unglamorous tool and uses it where it matters.

Young people do not need to love AI. They do need to see a future in which the technology works for them, not against them. That requires leadership willing to disappoint a base that wants traditional redistribution and make the case for something larger.

Jim Chalmers wrote his doctoral thesis on Paul Keating. Two decades on, he has handed down a budget Keating himself might not recognise. Keating’s 1985 reforms, the ones that introduced CGT with full indexation in the first place, were designed to do what the moment required: open a protected economy, modernise the tax base, equip Australia for a global capital market being remade in real time. The genius of Hawke and Keating was not that they did hard things. It was that they did the hard things required by the economic moment. The PhD on Keating was the easy part. The hard part is being him.

One more thing

Honda launched the Accord into Australia in 1977. By the mid-1980s it had become one of the most popular family sedans in mortgage-belt suburbs, reliable and well-built, if a little short on rear legroom.

The founder roundtables covered the leaving threats. The political commentary covered the broken election promise. This piece asks what argument we should actually be having, and why the founders’ anger is the sound of the country starting to ask a question it has avoided for years.

Workers or entrepreneurs. Labour or assets. Fairness or risk. This is the ideological argument Australia has avoided for years, and the founder anger is the sound of it finally starting.

Twenty Australian tech founders are convening this week to ask Jim Chalmers for a CGT carveout. Ben Grabiner at Side Stage Ventures wants zero per cent tax on founder and early employee equity up to $20 million. Kim Teo of me&u is publicly asking whether she would have built here at all. Frank Greeff is photoshopping the Prime Minister into startup photos as the new 47 per cent co-founder nobody asked for. The threats are loud. The chorus would be familiar to anyone over forty. Should I stay or should I go now?

Let’s face it. Most of us won’t go.

The Tech Council is lobbying behind closed doors. Treasury is already consulting. Some form of carveout may well come. But the leaving talk is always louder than the leaving will ever be. Max Marchione, the Young Rich Lister now in San Francisco, said the clearest sentence in Paul Smith’s piece last week.

This CGT thing obviously is net worse for me and my friends, but I do think it’s net better for the country, so it probably makes sense.

That is the sentence that got closest to the truth underneath the noise. Most founders are not threatening to leave because they want to. They are threatening because they feel trapped inside something they still love and have run out of other ways to be heard.

That is the giveaway. People who do not care about a place do not threaten to leave it. They just leave.

The argument worth having is not whether the CGT carveout is fair. It will be fair enough. The argument worth having is what we build from here. Leaving is not a strategy. Staying without a plan is not one either.

The room we are not in

Last week Donald Trump flew to Beijing with a senior US business delegation. Jensen Huang of Nvidia, Tim Cook of Apple, Elon Musk of Tesla and xAI. The agenda was tech access and commercial terms. There was no breakthrough on AI or semiconductors, but the message was the picture. The future of frontier AI is being shaped by two powers and the companies that make the silicon, the platforms and the models. Everyone else is in a smaller room.

Australia is not in the bigger room. Neither is the United Kingdom, Singapore, Canada, or any of the alternative build destinations being named by founders. The exception is the United States, where the room is. Moving there is not escaping the structural problem. It is paying rent inside it. The technology stack is being shaped above all small and middle powers. The choice for Australia is not which superpower to defer to. It is what to build in the smaller room.

Working proof

The superpowers may control the frontier. Most of the value will be created in the places that learn how to apply it. Estonia is the working proof. A country of 1.3 million people, fewer than the Gold Coast, decided in the early 2000s to treat sovereign digital infrastructure as boring twenty-year work rather than headline announcement. X-Road gave it the data exchange layer two decades ago. Bürokratt now gives it a provider-agnostic AI service layer on top: an open-source network of AI agents in production across government, designed to plug into whichever foreign model performs best for the task. The model vendors are interchangeable. The infrastructure is not.

The point for Estonians is what they actually experience. Today, a growing chat interface to government services, expanding toward all 3,000. From this year, personalised AI agents that reach out proactively: reminders about deadlines, alerts about benefits they qualify for, walk-throughs of the paperwork life events trigger. The state contacts the citizen, not the other way around.

Singapore is doing it differently and reaching the same answer. The Prime Minister chairs the National AI Council. Smart Nation runs agency by agency at speed. The model mix is pragmatic. Different architecture, same insight. Both countries are smaller than Australia. Both have decided the question was never which superpower to defer to. It was what to build in their own room. The principles are size-independent. Most of them just require being deliberate.

Eventually

Something quieter than the founder roundtable has been happening in Australian economic forecasting for the last six months. Australia’s institutions have quietly stopped assuming productivity will fix itself.

Treasury, the RBA, NSW Treasury and the Productivity Commission have all repriced productivity. Treasury added three years to its trend recovery timeline: the path back to a 1.2 per cent annual trend is now five years, not two. The RBA cut its medium-term assumption from 1.0 per cent to 0.7. NSW Treasury landed at 0.8. The Parliamentary Budget Office modelled a 0.9 per cent scenario in which federal debt would be roughly $119 billion higher by 2035-36. Multifactor productivity, the bit driven by technology and innovation, fell 0.5 per cent last year.

The RBA’s August 2025 statement explained why. Previous forecasts had “consistently overestimated the actual outcomes.” The institution had spent years assuming productivity would gradually return to higher historical growth rates. It hasn’t. The RBA warned even the new lower assumption might still be too optimistic.

AI is in the modelling, but as a tailwind sitting outside the baseline, not a current input. The RBA writes that policymakers are looking to AI “to eventually lift output.” Eventually is doing a lot of work in that sentence. The Productivity Commission has named the mechanism in plain English: technological improvements and innovation are not diffusing as quickly or as widely as they once did. Our research late last year put numbers on it. Nearly half the ASX 50 was building sophisticated AI capability. ASX 200 mid-caps were stuck at around thirty per cent. The venture side mirrors the corporate one. Australian startups raised $5.4 billion last year, with sixty-one per cent of that capital flowing to businesses using AI somewhere in their stack. Almost all of it sits at the application layer: the thin top of vertical tools and workflow software built on foreign frontier models. The layer beneath, where Australian firms could deploy AI on their own terms, is barely getting funded.

Performance

None of which the budget addressed. The $70 million AI Accelerator goes to startups. The approvals reform package deploys AI to fast-track medicine and environmental assessments. The application layer gets funded again. The government’s own narrow uses get funded. The Productivity Commission has been pointing at diffusion. Diffusion now depends on an orchestration layer: the systems, workflows, governance and deployment capability that let firms use AI at scale. That layer is almost absent from the budget.

Look at how the state governments are thinking about it. Victoria, faced with a shortage of frontline officers, announced this week it would hire 200 retired police back into uniform on $62 million of new funding. To do paperwork! The state’s actual AI line in the same budget is $5 million for an advisory committee. The retired cops are easier to count and easier to announce. An advisory committee is the AI version of the same instinct.

The diffusion the institutions are no longer expecting requires an architecture nobody here is funding. The CGT debate is happening in the smaller room. The orchestration layer is the work we haven’t started. That is the plan staying needs.

One more thing

Anthony Albanese has said that hard right-wing governments end up resulting in good music. Midnight Oil. Violent Femmes. The Clash. Billy Bragg. The punk-era canon he grew up on.

Hard left-wing governments, it turns out, end up resulting in good memes. But in the UK, Victoria, and Australia, it seems everyone’s asking the same question…

Staying without a plan is not a strategy either.

Matt Vitale is cofounder of New Dialogue, an Australian sovereign AI platform that helps enterprises deploy AI privately, securely, and without vendor lock-in. He is a lawyer by background and Australian Fintech Leader of the Year 2024.

Will Chen, an ex-Latham lawyer, built and open-sourced Mike OSS in two weeks using Claude. It went live on GitHub two weeks ago and already has more than 2,700 stars and 750 forks. Mike is a clone of Harvey and Legora. Chat with citations. Matter workspaces. Tabular review across hundreds of documents. Reusable workflows. The core experience that two of the most valuable startups in legal tech have been selling. Free. Self-hosted. Bring your own API key.

The real story is that the application layer was never where the value lived. If one motivated person can replicate the entire user experience in a fortnight, then Harvey and Legora have been selling something else all along. Curated data. Enterprise relationships. Compliance certifications. Switching costs. The fact that the firm is already locked in. That question now applies to every AI tool you are paying for.

The tools work

Herbert Smith Freehills Kramer is the only firm in Australia’s top eight running both Harvey and Legora. The AFR reported recently what they are getting for it. Contract delivery cut from 28 days to six. A 200,000 document review reduced to 18,500. Due diligence on 20 contracts across 15 parties processed in half a day, down from a three day estimate. One class action saved the client a million dollars, mostly from more efficient document review.

Ilona Logvinova, the firm’s chief AI officer, was clear about what this is for. The human mind, she said, has limitations on cognition. With AI, you can have much deeper and broader datasets at your disposal. Clients are now ringing the firm specifically to ask how AI is being used on their matter. The technology delivers. That part is not in question. The question is what is left worth paying for once the surface can be cloned.

The lesson from Atlassian

Atlassian is the Australian counter-example. They spent twenty years building tools that genuinely changed how technical teams work. Jira, Confluence, Bitbucket. They earned their place in enterprise software by solving real problems better than anyone else, year after year. The data, the integrations, the customer trust accumulated as a byproduct of being useful.

Harvey is three years old. Legora is younger. Together they are now worth roughly $23 billion. They have the capital, the early lead, and access to the customers that matter. The lesson from Atlassian is that those advantages compound only if you build something that genuinely defines a category. Software that becomes the way the work gets done, not just a wrapper around capability the customer could access another way.

Mike OSS is the prompt. It shows what the wrapper is worth without the rest. The valuations assume Harvey and Legora can build the durable parts that Atlassian spent two decades earning. The products, today, look closer to a wrapper.

What you are paying for

Big tech reported more than $600 billion in combined capex guidance for 2026 the same week Mike OSS landed. Google Cloud grew 63 per cent. AWS grew 28 per cent. Microsoft’s AI business hit a $37 billion run rate. The infrastructure layer is consolidating around four or five companies prepared to spend that kind of money. At the other end of the stack, Microsoft baked native contract review into Word.

A few days later, Anthropic and FIS announced a financial crimes agent for banks. The architecture is the bit worth noticing. Anthropic provides the model. FIS owns the platform, the data and the deployment. Customer data stays inside FIS-controlled infrastructure. The model is rented, the platform is not.

The pattern is now clear. The bottom of the stack is getting more expensive and more concentrated. The top, the part that actually touches users, is getting cheaper and more distributed. The platform layer in between, the part that holds the data, the workflows and the choices, is where the durable value is now sitting. Legal AI is the canary, not the exception.

If the software you are paying for can be cloned in two weeks, what are you actually paying for? Data. Trust. Control. The code gets cheaper. The confidence gets more valuable. The organisations that own their data and stay discerning about what runs on top of it are in the strongest position. The ones locked into someone else’s wrapper are the most exposed.

Andrew Charlton has been making a related argument under the banner of “lean Australian.” The procurement question is real, and worth its own conversation after the budget. The commercial question is more basic. The wrapper will change. The decisions you make about your data are what compound.

One more thing

The same week Mike OSS gave away a $23 billion product category, Fred again uploaded 108 hours of live performances to YouTube. He also dropped his raw creative files on Dropbox. Free. Anyone who wants them.

Fred is in the business of creating experiences. The companies trying to keep up with him are in the business of extracting value. There is a purity in the first that the second cannot manufacture. It might be the closest thing to a north star we have right now.

Matt Vitale is co-founder of New Dialogue, an Australian sovereign AI platform that helps enterprises deploy AI privately, securely, and without vendor lock-in. He is a lawyer by background and Australian Fintech Leader of the Year 2024.

The press conferences covered the inflow. The accounts covered the outflow. This piece asks what Australian organisations should actually own in their AI stack, and why the answer isn't "whatever the hyperscaler is selling."

Last Tuesday, Satya Nadella stood beside the Prime Minister and committed $25 billion to expand Microsoft’s Australian cloud capacity by 140% over four years. Big numbers, big stage. The same speech he had recently delivered in India, Canada, Japan and the UK.

A few days later the books opened on what flows the other way. Newly filed accounts show Google and Meta transferred close to $11 billion of Australian revenue offshore last year in service fees and reseller payments, paying just over $140 million in tax between them. Amazon booked almost $12 billion in local revenue, with its data centre arm up 50%.

The Australian Bureau of Statistics noted earlier this year that the surge in data centre investment has been largely offset by a corresponding rise in imports, with limited net impact on GDP. The e61 Institute, an independent think tank, points out that only around eleven and a half thousand workers are employed in the relevant local industry. None of these figures sit comfortably with the language of national investment.

The press conferences cover the inflow. The accounts cover the outflow.

Most of the conversation about AI in Australia is framed as a race. Move faster. Adopt sooner. Don’t fall behind. The visiting CEOs reinforce it because urgency is good for procurement. The consultants reinforce it because urgency is good for fees. The trade press reinforces it because urgency is good for headlines.

A generation of Australian executives is being told, with growing confidence, that the worst thing they can do is not run. There is a race. The reframe worth holding onto this week is that it isn’t a race to deploy.

What the hyperscalers are selling is pret-a-porter. The same suit, made to standard sizing, hanging in every shop window from Sydney to Sao Paulo. Buy it quickly enough and the salesman will tell you it fits.

Running the same Microsoft stack as everyone else doesn’t generate competitive advantage. It generates parity, which is the opposite.

The Pioneer markets at the top of the EY adoption index aren’t winning a sprint. They’ve embedded AI into infrastructure they already own. India layered AI onto UPI, the public payment rail it had spent a decade building, and turned its fintech sector into something nobody else can replicate without first owning the rail.

The advantage isn’t velocity; it’s what sits underneath.

The real race isn’t who deploys AI fastest. It’s who owns what they deploy.

The AFR reported one local example of that choice:

Telstra’s board no longer has to remember every detail of every project, because Telstra built its own AI bot inside its own perimeter.

In early 2025, before the hyperscalers had any sovereign-flavoured products to sell, Telstra had access to every major vendor in the country and could have signed with any of them. Instead, they consolidated 18 vendors down to two, committed $700 million over seven years to a joint venture with Accenture, and started building.

They didn’t sprint. They consolidated, controlled, and built something nobody else has.

JPMorgan did the same at much larger scale with its LLM Suite, now serving 230,000 employees on vendor models wrapped inside the bank’s own governance layer.

These aren’t outliers. They’re early indicators of where the economics are heading. Arvind Krishna told IBM’s January earnings call that within three to five years, half of all enterprise AI usage will move into private cloud or on-premises environments, driven by both control and cost.

The hyperscalers read the same numbers, which is why every one of them now sells a “sovereign” version of their product, racing to be the inside-the-firewall provider before customers build their own.

What the press releases don’t mention is how much their version of sovereignty costs.

Google Distributed Cloud, for example, starts at US$415 per node per month on a three or five year commitment, before you add Gemini usage, storage, networking, integration or professional services. The air-gapped versions, built for the most sensitive deployments, are quote-only. These are Telstra-scale solutions with Telstra-scale price tags and lock-in periods.

For most Australian organisations, this is not the kind of sovereignty they can afford.

But the principle is the right one, and it doesn’t take $700 million to apply.

A bespoke suit isn’t faster than off-the-rack. It’s slower, costs more upfront, and lasts considerably longer. Treat frontier intelligence as a commodity. Own the layer that sits on top: the data, the governance, the workflows, the institutional knowledge. The model you can rent.

An emerging cohort of startups, including the one I’m building at New Dialogue, is now making exactly this kind of capability available to organisations that aren’t Telstra.

The question worth asking isn’t whether you can afford hyperscaler scale sovereignty. It’s what part of your AI deployment you actually need to own.

Australia sits in the bottom-left of the EY adoption index: low adoption, low sentiment. The honest reading isn’t that we’re being virtuously cautious. It’s that we haven’t yet decided what an Australian-context AI deployment actually looks like.

While we figure it out, every hyperscaler in the world is queuing at the door with the same off-the-shelf solution, and the meter is running on the way out as well as in.

The work of figuring out what to own isn’t slow. It’s just deliberate, which is something the running-shoe sellers can’t bill for. The best response to a uniform product isn’t to buy it faster than the next country. It’s to build something the next country can’t.

A couple of decades ago, Google’s first attempt at enterprise hardware was a bright yellow box called the Search Appliance.

You bought it, plugged it in, and got Google search inside your firewall.

It ran for seventeen years before Google quietly shut it down, at which point the box became a very expensive paperweight.

Distributed Cloud is the same shape, three model generations later. Worth remembering, as the boxes start arriving.

Matt Vitale is cofounder of New Dialogue, an Australian sovereign AI platform that helps enterprises deploy AI privately, securely, and without vendor lock-in. He is a lawyer by background and Australian Fintech Leader of the Year 2024.

Published a few days after Anthropic blocked 135,000 developers’ third-party agent frameworks with under 24 hours’ notice, three weeks after running a 135x usage subsidy on the same workflows. The mechanic this piece names (give, take, charge) isn’t an Anthropic story. It’s an industry pattern with no regulatory countermeasure in Australia yet. Worth re-reading whenever the next “free tier” rolls out.

On Friday, 135,000 developers woke up to find their AI agents broken. Anthropic had blocked OpenClaw and every other third-party agent framework from drawing on Claude’s flat-rate subscriptions. If you wanted to keep your automation running, you could pay per token. The cost increase, for heavy users, was up to fifty times what they had been paying.

Three weeks earlier, those same users had been getting double. From March 13 to March 28, Anthropic ran a promotion that automatically doubled usage limits outside peak hours for every plan tier. No signup required. Claude Code, Cowork, web, mobile, all of it. It was a generous, frictionless expansion of access, timed perfectly to the wave of new users arriving after the Pentagon drama pushed Claude to the top of the App Store.

Give. Take. Charge.

One user consumed 1.1 billion tokens in 23 days on a $200 plan. That is $27,000 worth of compute. A 135x subsidy that Anthropic, by its own admission, can no longer absorb. The promotion was not generosity. It was customer acquisition priced below cost, targeted at the exact workflows that would be hardest to unplug once embedded.

ChatGPT launched in November 2022 as a quiet research preview. OpenAI’s internal forecast was tens of thousands of users. Servers crashed in hours. A million people signed up in a week. The free tier was never a strategy. It was a scramble that became one. Plus at twenty dollars followed. The subsidy race was born.

Every major player has run it since. Microsoft poured billions into Copilot and has three per cent adoption across 450 million seats. OpenAI’s Sora burned a million dollars a day on compute, signed a billion-dollar Disney deal, then shut down in March when lifetime revenue totalled $2.1 million. Harvey, the legal AI startup, just closed at an eleven billion dollar valuation in a market worth one trillion. The maths requires it to swallow the entire global legal industry for its investors to see a return.

None of these numbers work. They are not supposed to, yet. The subsidy is the product. Cheap access builds dependency. Dependency builds lock-in. Lock-in builds pricing power. The economics only need to work later, once you can’t leave.

Sam Altman said the quiet part out loud on March 12 at BlackRock’s Infrastructure Summit. Intelligence, he said, would be a utility, like electricity or water, and people would buy it from OpenAI on a meter. The clip got 25 million views. The internet called it dystopian. But he was not making a prediction. He was describing a business model that is already running.

In early March, Claude went down for several hours. LinkedIn filled with developers saying they could not work. One user described it as losing sixty per cent of his daily productivity in an instant. The commentary was not about inconvenience. It was about paralysis.

Around the same time, Azure OpenAI’s flagship GPT-5.2 model suffered a roughly twenty-hour outage across seven regions, including Australia East. Major Australian enterprise customers received zero real-time communication from Microsoft. No alert, no status update, no proactive notification. The issue was eventually documented in a quiet post-incident review published days later.

These are not stories about unreliable technology. They are stories about infrastructure that is already critical but is not yet governed as such.

Australia caps NBN wholesale pricing. It levies fuel. It regulates telecommunications. When those systems fail, there are reporting obligations, accountability frameworks, and consumer protections.

AI has none of this. Tim Ayres’ National AI Plan talks about energy, jobs, and sovereignty. The March data centre expectations framework requires projects to underwrite renewables and demonstrate national interest alignment. These are the right instincts. But there is nothing on uptime mandates, outage communication, pricing transparency, or what happens when the subsidy ends and the meter starts.

Australia has genuine leverage in the AI infrastructure race. Every major AI company in the world wants our land, our grid, and our political stability. I argued in the AFR in early April that the country should negotiate like it knows it. But leverage is worthless if the dependency is already set before you use it. And right now, the dependency is being constructed, cheaply and deliberately, while the governance catches up.

Governments will get there eventually. Outages, pricing shocks, or a genuine market failure will force the conversation. But enterprises cannot wait.

Vendor lock-in is a material board-level risk. If sixty per cent of your workflow runs through a single provider’s API, your biggest operational risk is not a competitor. It is a policy change you had no say in, announced on a Friday afternoon with less than 24 hours’ notice. That is what happened on April 4.

The response is not to stop using AI. It is to stop using it naively. Deploy on-premise or in a virtual private cloud where it matters. Use open-source models where they are fit for purpose. Build abstraction layers so you can switch providers without rebuilding. Negotiate exit clauses. Budget 1.5 times what vendors quote, because the subsidy is temporary and the real cost is coming.

On March 31, Anthropic accidentally shipped a source map with Claude Code that exposed 512,000 lines of proprietary source code. GitHub mirrors hit 50,000 stars in hours. Anthropic, the same company facing a billion-dollar lawsuit from authors over copyrighted training data, responded by filing DMCA takedowns against more than 8,000 repositories.

Copyright for me, but not for thee.

Matt Vitale is co-founder of New Dialogue, an Australian sovereign AI platform that helps enterprises deploy AI privately, securely, and without vendor lock-in. He is a lawyer by background and Australian Fintech Leader of the Year 2024.

The argument here, that voluntary frameworks aren’t leverage and that Australia is settling for less than its position warrants, is the starting point for everything I want to write on this Substack. The terrain is AI infrastructure. The actual subject is what host nations negotiate for, and whether the institutions doing the negotiating are using the leverage they have.

Dario Amodei spent Wednesday in Canberra signing the first memorandum of understanding under the National AI Plan with Prime Minister Anthony Albanese, meeting Treasurer Jim Chalmers, and telling MPs at the Anthropic Futures Forum that democracies must keep the military and technological edge over autocracies. Yet what Australia got was still a non-binding handshake.

Every major AI company in the world wants to build in Australia right now. They need our land, our grid, our political stability, and our geography. We should be negotiating like we know it. Instead, we are settling for handshakes.

On Wednesday morning, Dario Amodei, chief executive of Anthropic, walked into Parliament House and signed an MoU with the Albanese government. The $555 billion company committed to align with Australia’s data centre framework, share workforce impact data, collaborate with the AI Safety Institute, and provide $3 million in research credits to four Australian universities. It is the first agreement signed under the National AI Plan.

What it is not is binding. The MoU carries no enforceable obligations, no penalties for non-compliance, and no mention of copyright, the issue AirTrunk’s Robin Khuda called “the elephant in the room” during meetings with Anthropic executives this week. Speaking before his scheduled end-of-day meeting with Amodei, Chalmers said the government wants data centre investment to come with obligations, including water and energy management. Copyright, again, was not mentioned. Australia got a handshake from the world’s most important AI safety company. Whether it gets leverage depends on what comes next.

Amodei came because Australia hasn’t yet locked itself into the same constraints as everywhere else.

In the United States, the data centre boom has hit a wall. Northern Virginia, the country’s largest hub, is running into grid bottlenecks so severe that some proposed campuses require new substations before construction can begin. Interconnection queues stretch for years. Elon Musk has filed plans with the FCC for a constellation of one million orbital data centre satellites, arguing that space will be cheaper than Earth by 2028. That proposal is the clearest possible signal of how scarce the right terrestrial conditions have become.

Singapore is at 2 per cent vacancy. China is closed to foreign operators. Asia-Pacific holds just 26 per cent of global hyperscale infrastructure despite accounting for 60 per cent of the world’s population. The region needs capacity, fast, and it needs it somewhere stable, well-governed, and close.

That is why Amodei is here. That is why Oliver Curtis is on a global roadshow pitching Firmus Technologies, a $6 billion AI factory company building 1.6 gigawatts of infrastructure across four states, hunting a third anchor customer before a June IPO. That is why the federal government published its five-expectation framework for data centre approvals on March 23. Everything is happening at once because the window is open now, not later.

Australia has 670 gigawatts of renewable energy projects in the pipeline, more than triple what the country needs to hit net zero. The structural problem is not ambition. It is pace. Planning approvals are slow. Grid connections take years. Australia is not on track to hit 82 per cent renewable electricity by 2030.

Data centres at hyperscale can pull that stuck infrastructure forward. A 1.6-gigawatt facility requires firm power contracts. Those contracts force renewable projects from proposal to construction at a pace domestic policy has never managed alone. The government’s framework, which requires new data centres to underwrite fresh renewable capacity rather than just consume what already exists, is the right model for exactly this reason.

But the framework is soft. It is voluntary. There are no penalties for non-compliance. The government is setting expectations rather than enforcing them.

Amodei himself told the Futures Forum he supports the framework for any data centres Anthropic may build here. But the company’s $50,000 in computing credits for Australian deep-tech founders shows the scale of what we are still settling for.

Karen Hao, whose book Empire of AI has been widely read this year, asks the right question of every AI infrastructure story: who is the host nation, and what do they actually get? Australia’s current answer is jobs, training programs, and a seat at the table. That is not nothing. But it is less than what Australia’s position warrants.

OpenAI training Coles and Wesfarmers workers through its academy is sovereignty only if those workers develop transferable AI capability, not product familiarity with one American platform. The infrastructure is being built here. The models, the data, and the value capture are mostly elsewhere.

Amodei warned at the Futures Forum that AI-driven gains will be heavily capital and machine-driven, so the benefits won’t go to everyone. He said they may be more concentrated than they should be, and that targeted taxes on those making the most money would be needed. That is exactly the kind of enforceable commitment Australia could negotiate while the queue is still forming.

Australia’s leverage is larger than it acts. Lower build costs than most developed markets. Political stability that rivals any Western nation. Geography that makes it the logical hub for Asia-Pacific demand. A renewable energy pipeline that, with the right conditions attached, could finally accelerate. Right now, that leverage is being exchanged for workforce training partnerships. It could be exchanged for co-ownership of renewable infrastructure, IP licensing pathways, or enforceable local supply chain commitments.

The government has recognised the moment. Google’s tax stand-off with the ATO shows that even the most powerful companies in the world will negotiate hard when the terms aren’t right. Australia should take note of its own example.

The queue is forming. The window will not stay open indefinitely. The question is not whether Australia is attractive to the world’s most important AI companies. It clearly is. The question is whether we negotiate like we know it.

Matt Vitale is co-founder of New Dialogue, an Australian sovereign AI platform that helps enterprises deploy AI privately, securely, and without vendor lock-in. He is a lawyer by background and Australian Fintech Leader of the Year 2024.