Kia ora! Welcome to New Zealand’s weekly roundup of AI news and education.
Two things happened off last week's newsletter.
The 6-week coaching slots are now fully booked.
And separately, a few of you said the course felt like a lot to take on.
So I've landed on a leaner, shorter (four-weeks), and more flexible format: same coaching spine, tailored to your workflows and what you want to focus on.
In terms of options, I’ve been coaching people on general Claude fluency, Claude for marketing, Claude for sales, Claude for decision-makers, building a second brain with Claude Code, Claude Code fluency, or becoming your team's AI enablement champion.
If one of those sounds like you, get in touch.
Reminder: We’re hiring a Principal AI Engineer at Allexive. DM me or apply here.
Happy reading ✌️
Did someone forward you this? Sign up!
🇳🇿 New Zealand News
NZ ranks #5 of 121 countries for Claude usage, 4.84x the expected rate. Our usage nearly matches Australia's, except Australia over-indexes on business operations while NZ over-indexes on starting a business.
3 min read
Our take: NZ’s Claude usage is solid, really promising showcase of the average Claude user operating at the frontier in terms of volume. But, are we pointing the tools at the right activities? The fix is not less ambition, it is aiming the same appetite at the business already running. A workflow redesigned this quarter beats a business plan drafted in ChatGPT that never leaves the laptop.
Datacom's profit nearly halved to $20m as it poured money into AI data centres. Revenue hit a record $1.58bn, but a deliberate bet on AI-ready infrastructure, including the new T4 site at Highbrook, cut net profit 46%.
4 min read
Our take: Owning five NZ data centres and calling them "sovereign infrastructure" is a bet that customers will pay a premium to keep data onshore, rather than the cheapest overseas cloud. Borrowings up to $131m in "the most volatile supply chain since Covid" is a lot of risk to carry on a bet about future demand.
A Crown agency asked NZ's leaders to build a $25-35bn AI data centre industry over five years. Invest New Zealand gathered ~100 leaders from power, construction and KiwiSaver funds to back data centres, renewables and transmission.
4 min read
Our take: This is the government picking a lane in the AI economy, and the lane is landlord, not builder. NZ would host the compute for other people's models rather than make its own, which is a realistic read of where a small country can actually compete. It puts NZ in a global bidding war for the same hyperscaler contracts that Ireland, the Nordics and Canada are all chasing with the same story.
NZ joined Project Arcadia, a US-led effort to build a shared Five Eyes AI battlefield command tool. Cabinet approval for funding will be sought, and recent conflicts have seen AI speed up targeting by up to 50 times.
4 min read
Our take: Judges are now looking for AI giveaways like em dashes and stock phrases. That will not work for long, because the tools learn from the criticism and stop making those mistakes.
An EMA survey found 83% of NZ businesses use AI but only 13% have a policy for it. Just 4% of AI initiatives are led by HR, with most driven from the top and little planning for how the work itself changes.
4 min read
Our take: The answer is not to ban AI but to pick which thinking stays human. Keeping some work in-house on purpose is what separates firms that use AI well from ones that just use it.
Heidi, the AI scribe now in NZ emergency departments, stores its data in Australia. The company has committed to moving to NZ-based hosting within 12 months, but says the local infrastructure to run it at full capacity is not ready yet.
2 min read
Our take: A concrete example of data sovereignty. Patient notes from NZ hospitals sit on Australian servers today, and the fix depends on local cloud infrastructure that does not yet exist at the scale a national health tool needs. The safeguards are decent, with 14-day deletion and no reuse for product training, but they are the vendor's own commitments. When a public health system depends on a private tool hosted offshore, the power to enforce those promises is thinner than it looks.
Most young NZ workers now use AI at work, but half say their employer's tools fall short. Deloitte's 2026 survey of 501 New Zealanders found AI adoption is outpacing what workplaces are set up to support.
4 min read
Our take: Only 26% of NZ Gen Zs and millennials have done any AI training, and just 27% of Gen Zs keep seeking it as the tech moves, against 35% globally. So the most-exposed generation is upskilling itself on the job while formal training barely exists.
📚️ Mike’s Takes From The Week
Helping leaders and teams adapt, learn, and scale with AI.
1️⃣ The wrong model "Effort" dial in Claude can cost 7x more tokens (and it has nothing to do with model choice): Anthropic's own numbers show "high" effort burns roughly 7x the tokens of "low" on the same task. The fix is two dials, model (what Claude knows) and effort (how hard it tries), matched deliberately: Sonnet on low for routine work, Opus on high for a proposal review.
Read.
2️⃣ 150 hands went up for "paid for AI training this year". Almost none stayed up for "does your business run differently now?": Speaking at NZEE, the pattern echoed McKinsey's 2026 finding that 90% of companies are experimenting and fewer than 10% capture value. Training is 10-20% of the work; work redesign is the other 80%.
Read.
3️⃣ Thirty AI tools listed eight months ago, most now replaced by Claude: Mockups, code, research, spec docs, slide decks, inbox triage, all now inside Claude. What still needs a second tool: Suno for jingles, Nano Banana for images, ElevenLabs for audio, Veo for video. The monthly test: name the job, not the tool, then make Claude prove it.
Read.
4️⃣ New US data on 21,000 firms says heavy AI adopters grew headcount 10%, not shrank it: Entry-level roles grew even faster, at 12%. Firms that barely touched AI saw no change either way. Meanwhile NZ redundancies get blamed on AI and 37% of Kiwi workers feel guilty using it. The gap is leadership avoiding the harder call, not the tool.
Read.
🎙️ The AI Corner Podcast
This week's guest is Pouria Asghari, Head of Go-to-Market Ecosystems and AI at Halter. Hear:
How Halter uses AI agents to handle customer support queries, retrieve trusted information and escalate higher-risk issues quickly.
How its field sales teams use voice, meeting recordings and AI to reduce CRM admin and improve customer follow-up.
Why strong data, workflows and customer experience matter more than chasing the cheapest AI model.
🌍 Tech Updates From Global
The selected top headlines from each major AI tech company.
Anthropic
Claude Cowork expanded from desktop to web, iOS and Android for Max users, with background work continuing while the device is offline. (Jul 7)
Claude Code desktop added a sandboxed in-app browser Claude can drive itself to open docs, read pages, click and fill forms. (Jul 10)
Claude added Microsoft 365 write actions for drafting email, managing calendar and updating OneDrive and SharePoint, plus Trusted Devices for remote Claude Code steering. (Jul 8)
Reflect launched in beta, a usage dashboard visualising Claude activity across an AI Fluency framework with break reminders and quiet hours. (Jul 9)
Anthropic signed a $19B, 20-year data-centre lease with TeraWulf at a 401MW Kentucky campus, with initial capacity live in H2 2027. (Jul 6)
Anthropic launched Inviting Hard Questions, a public campaign asking people to submit their hardest questions on AI's effect on jobs, society and families. (Jul 9)
OpenAI
ChatGPT Work launched as a GPT-5.6 agent that pulls context from a team's apps and works for hours to produce finished docs, sheets, decks and sites. (Jul 9)
GPT-Live-1 and GPT-Live-1 mini launched as full-duplex voice models that listen and speak at once, becoming the default for ChatGPT Voice. (Jul 8)
Codex merged into the ChatGPT desktop app on macOS and Windows with inline code editing and GitHub PR review in the sidebar. (Jul 9)
ChatGPT Sites entered public beta, letting users build and publish interactive sites, dashboards and internal portals from inside ChatGPT. (Jul 9)
GPT-5.6 reached general availability across ChatGPT, Codex and the API after the government-limited preview, at Sol $5/$30, Terra $2.50/$15 and Luna $1/$6 per 1M tokens. (Jul 9)
OpenAI is retiring its standalone Atlas browser on August 9, folding agentic browsing into ChatGPT and Codex. (Jul 9)
Apple sued OpenAI, its io hardware unit and two former Apple employees over alleged theft of confidential hardware designs and supply-chain strategy. (Jul 10)
Seventeen news organisations including the New York Times asked a court to sanction OpenAI for alleged discovery misconduct in the copyright case. (Jul 9)
Google Photos added Video Remix, a Gemini Omni tool that restyles and relights clips via templates, for AI Plus, Pro and Ultra subscribers in select countries. (Jul 8)
Fill with Gemini in Sheets added 11 languages including Mandarin, Turkish, Polish and Indonesian. (Jul 7)
Cognizant expanded its Google Cloud partnership to deploy Gemini Enterprise to 100,000 associates and certify over 10,000 professionals. (Jul 7)
An EU court ruling gave Google until July 27 to open the Android AI layer Gemini uses, wake words, screen context and on-device data, to rival developers. (Jul 9)
Chinese Frontier Labs
Tencent open-sourced Hy3, a 295B MoE (21B active, 256K context) under Apache 2.0, free on OpenRouter through July 21 and claiming a halved hallucination rate. (Jul 6)
MiniMax announced M3 Pro, a planned 2.7T-parameter open model targeted for Q3, which would be the largest open-source model shipped to date. (Jul 8)
DeepSeek confirmed legacy deepseek-chat and deepseek-reasoner API names retire July 24, pushing developers onto deepseek-v4-pro and deepseek-v4-flash. (Jul 8)
Beijing's Ministry of Commerce opened talks with Alibaba, ByteDance and Z.ai over restricting foreign access to advanced and unreleased Chinese AI models. (Jul 7)
Open-Weight Model Updates
Tencent Hy3 landed day-one on Hugging Face and ModelScope, claiming parity with GLM-5.2 and DeepSeek V4 at roughly half the size except on coding. (Jul 6)
NVIDIA and Hugging Face released Isaac GR00T 1.7, an open commercially-usable vision-language-action model for humanoid robots, into the LeRobot workflow. (Jul 7)
Ollama raised a $65M Series B at nearly 8.9M monthly developers, reporting use inside 85% of the Fortune 500. (Jul 9)
Poolside opened Laguna XS 2.1, a locally-runnable 33B MoE coding model, free during preview on Hugging Face and OpenRouter. (Jul 2)
Mistral confirmed a new open-weight "fat but sparse" MoE family entering early access in July, with no benchmarks or licence disclosed yet. (Jul 4)
Microsoft
GPT-5.6 became the preferred model across Word, Excel, PowerPoint, Chat and Cowork in Microsoft 365 Copilot, amid ongoing OpenAI breakup chatter. (Jul 9)
GitHub Copilot added GPT-5.6 Sol, Terra and Luna across IDE, CLI, mobile, cloud-agent and web surfaces. (Jul 9)
Moonshot's Kimi K2.7 Code model extended to GitHub Copilot Business and Enterprise tiers. (Jul 7)
VS Code 1.128 shipped multi-chat for Claude agent sessions, letting users compare approaches and branch from an earlier turn in parallel. (Jul 8)
xAI
SpaceXAI released Grok 4.5, its first model built with newly-acquired Cursor, at $2 input and $6 output per 1M tokens and pitched as faster and cheaper than Opus. (Jul 8)
xAI rebranded to SpaceXAI five months after the SpaceX merger, with a new logo revealed on X. (Jul 7)
Grok Build shipped a refreshed sessions flow and faster large-repo startup, while Grok Voice gained 21 multilingual voices and voice cloning. (Jul 10)
Amazon
AWS made Anthropic's Claude Sonnet 5 available across its platform, adding the near-Opus default model to Bedrock. (Jul 6)
Amazon WorkSpaces for AI agents reached general availability, letting agents securely operate desktop apps without custom integration. (Jul 6)
Amazon will close Mechanical Turk, SageMaker Ground Truth and Augmented AI to new customers on July 30, with existing customers retaining access. (Jul 11)
Apple
Apple committed over $30B to Broadcom to design and produce 15B-plus US-made custom silicon and wireless chips, including $1.5B into a Colorado facility. (Jul 8)
Meta
AI chief Alexandr Wang reportedly told staff Meta's next flagship model, codenamed Watermelon, has reached GPT-5.5 parity on internal benchmarks, though it is still training and unverified. (Jul 5)
Zuckerberg told staff AI agent development had not accelerated as expected over the prior four months, after Meta cut 8,000 roles while shielding AI teams. (Jul 7)
Meta broke ground on a $13B, 1GW AI data centre in Alberta, its first in Canada and largest outside the US. (Jul 8)
Mistral
Mistral released Robostral Navigate, an 8B model letting robots navigate complex environments from a single RGB camera and language prompts, trained entirely in simulation. (Jul 8)
🤦 ️ AI Fail Of The Week
We all love AI, but it’s certainly far from perfect 🤔 …
✨A few people have asked…
It’s Mike here, I run The AI Corner.
I’m not just into writing about AI. I run Allexive, and we help businesses grow without adding headcount by implementing AI platforms, and building AI systems.
👋 Mike & Erin