The View From Data Centre Alley: Why New Zealand's AI Edge Will Be Won on Data, Not Models

I'm writing this from Alexandria, Virginia an old riverside city just across the Potomac from Washington, D.C. This stretch of Northern Virginia is also one of the nerve centres of the modern state: thick with federal agencies and the contractors who orbit them. And about forty minutes up the road sits Loudoun County's "Data Centre Alley" the densest cluster of data centres on earth, through which an estimated 70 percent of the world's internet traffic passes every day. The United States spent something like US$425 billion building data centres in 2025 alone.

I’ve used the time here watching how AI is actually being adopted across US agencies and businesses, and deepening my own grounding in data science and AI.

You'd think that living this close to the engine room of the internet, and across the river from the people writing the rules for it, the big lesson would be about scale: more compute, more power, bigger models. It isn't. The thing I keep noticing, and the thing I think New Zealand organisations should pay close attention to, is far less glamorous.

The AI race is quietly bottlenecked on data. Not the “having” of it, the“readiness” of it.

The lesson hiding behind the hype

The clearest version of this is coming out of the US federal government. After several years of AI enthusiasm, the consistent message from agencies here is that you have to move slowly with your data before you can move fast with AI. You can't point a model at a filing cabinet. The White House's own national AI framework, released this March, now talks explicitly about making government datasets "AI-ready" which is a quiet admission that an enormous amount of public information simply isn't.

That's the part that never makes it onto a conference slide. Behind every impressive AI demo is a far less impressive reality: decades of records trapped in paper, microfilm, and scanned PDFs that are really just photographs of text. Information sitting in incompatible systems with no consistent labelling. Archives so dark that no algorithm however clever can read them. AI cannot reason over what it cannot access. Poor-quality data doesn’t just slow you down; it caps what’s possible at all.

You can see the whole argument in one institution just up the road. The US National Archives holds more than 13 billion pages of records and until recently had digitised barely two percent of them. It’s now pouring money into a dedicated digitisation centre, not to make tidy scans but to turn paper intomachine-readable, well-described data its new AI tools can actually search. As its own archivists put it, simply scanning more pages isn’t enough; people have to be able to find what’s in them. That distinction, between digitised and genuinely usable, is the whole game.

The shift I notice here is subtle but important. Records aren’t simply scanned for storage; they’re indexed, structured and enriched so they can be queried and fed into decisions later. The organising question has moved from “how do we store this?” to “how will this be used?” and that one change reshapes every choice about formats, metadata and classification.

Why this matters back home

New Zealand actually moved well in 2025. It launched its first national AI strategy in July, with a deliberately light-touch, get-on-with-it tone. The ambition is real, and it's showing: public-sector AI use more than doubled in a single year.

But ambition is the easy part; the foundations take longer. New Zealand’s advisers have been candid that the real constraint isn’t appetite for AI it’s whether the underlying data is ready for it. The description that stuck with me was of organisations carrying the AI ambition of 2026 on data foundations laid for an earlier era: legacy systems, information spread across silos, inconsistent metadata. And this is nobody’s failing in particular. It’s the natural state of almost any organisation that has spent decades accumulating records government agencies, councils, health providers, law firms, museums and banks alike. The records piled up long before machine learning was on anyone’s radar, which is exactly why getting them ready is the work that now matters most. This isn’t a New Zealand problem, or a government one: Gartner expects organisations to abandon 60% of AI projects through 2026 for want of AI-ready data, and MIT’s 2025 research found that 95% of generative-AI deployments delivered no measurable return with the failure almost never the model itself.

It also stops being a job owned solely by the records team or IT. Done well, digitisation sits across compliance, operations and innovation at once which is exactly why it tends to stall when it’s treated as a back-office task.

The opportunity most people are underestimating

Here's where I think New Zealand has a genuine, slightly counter-intuitive opening.

The temptation is to believe the AI advantage goes to whoever buys the flashiest model or builds the biggest data centre. From where I'm sitting close enough to both the data centres and the policymakers to watch how this actually plays out I'd argue the opposite. That flashy layer is fast becoming a commodity. Anyone can rent it. The durable advantage goes to whoever has done the unglamorous foundational work: getting their information out of boxes and into structured, searchable, machine-readable form, with metadata good enough that a system canactually find and trust what it needs.

And that work is “tractable”. It doesn't require a hyperscaler or a million-dollarbuild. It's the kind of thing a focused organisation can actually finish and it's the difference between an AI project that quietly stalls and one that works.

Where I'd start

If I were advising a New Zealand organisation eyeing AI right now, I'd suggest starting “before” the AI:

-Find out where your data actually lives. Most organisations are surprised how much sits in paper, image-only PDFs, or systems nobody has opened in years.

-Treat digitisation as infrastructure, not admin. Converting records into clean, structured, machine-readable data is the on-ramp to everything else not a back-office chore.

-Prioritise your highest-value records first. You have to make your most important information usable.

-Get your metadata and standards right early. Consistency is what lets a modelor a person find and trust the answer.

-Keep data sovereignty in view. Where your data lives, and who can read it, matters more in an AI world, not less.

All this extraordinary infrastructure exists to do one thing: move and process information. And it is only ever as good as the information you feed it. New Zealand doesn't need to win the race to build the biggest model. It needs to win the quieter race to get its data ready because that's the part nearly everyone underestimates, even here, a stone's throw from Washington and a short drive from the busiest data centres on earth.

There’s a local wrinkle, too. New Zealanders’ trust in government use of AI is still cautious, even as agencies move quickly. Getting the data foundation right accurate, well-described, properly governed isn’t only a technical step; it’s part of how that trust gets earned. You can’t be transparent about information you can’t reliably find.

The encouraging part is that this is a solvable problem. The digitisation expertise to do it well already exists in New Zealand. The real question is whether organisations treat that work as the foundation it is and start now, while the gap is still closeable.

‍

Guest Blogger: Deb Petre

Sources

KPMGNew Zealand, “Addressing the AI gaps to foster trust in government”:https://kpmg.com/nz/en/insights/government-public-sector/addressing-the-ai-gaps-to-foster-trust-in-government.html

NewZealand’s Strategy for Artificial Intelligence: Investing with Confidence(MBIE, 2025):https://www.mbie.govt.nz/business-and-employment/economic-growth/digital-policy/new-zealands-ai-strategy-investing-with-confidence

Gartner,“Lack of AI-Ready Data Puts AI Projects at Risk” (2025):https://www.gartner.com/en/newsroom/press-releases/2025-02-26-lack-of-ai-ready-data-puts-ai-projects-at-risk

MITProject NANDA, “The GenAI Divide: State of AI in Business” (2025)

HarvardBusiness Review Analytic Services & Cloudera, AI data-readiness study(2026):https://www.cloudera.com/about/news-and-blogs/press-releases/2026-03-05-only-7-percent-of-enterprises-say-their-data-is-completely-ready-for-ai-according-to-new-report-from-cloudera-and-harvard-business-review-analytic-services-reveals.html

IBM,“The Biggest AI Adoption Challenges for 2026”:https://www.ibm.com/think/insights/ai-adoption-challenges

‍