Kia ora! Welcome to New Zealand’s weekly roundup of AI news and education.
This week feels like the market shifted.
The AI bubble bears have been quiet for some time, not because the risks disappeared, but because demand is now showing up in earnings, contracts, accounting changes, bank programmes, SME revenue data and enterprise adoption (even down under!).
Critically in the last week, we’ve seen Big Tech earnings reach dizzying heights. What I’m most interested in is what this tells us about the AI bubble debate.
Secondly, if you’re not across, the Musk vs Altman court case is one for the ages. We’ll also be covering that trial with regular updates on Mike’s LinkedIn.
Happy reading ✌️
Did someone forward you this? Sign up!
Final call: Join the Waboom Claude Cowork & Code Workshop and learn how to actually use AI in your day-to-day work .
In this hands-on session, you’ll walk away able to:
Save hours each week on emails, reports, and admin
Build AI tools that actually take work off your plate
Instantly create quotes, proposals, and marketing content
Go from idea → working tool without needing a developer
Trusted by 1,000+ professionals across New Zealand for hands-on AI training.
📅 Auckland (in-person) — 6 May
🎟️ $199 + GST with code AICORNER
📅 Virtual (NZ-wide) — 7 May
🎟️ $149 + GST with code AICORNER
🇳🇿 New Zealand News
A Deloitte Access Economics study commissioned by 2degrees and released on 28 April found NZ SMEs that adopted AI earned approximately NZ$400,000 more in FY25 than comparable non-adopters, and the average large adopter earned NZ$59.1m more. Eighty-two per cent of NZ businesses report some AI use, but most of that use is features bundled into existing tools (Office, Workspace, CRM) rather than standalone deployments. The report is the first NZ-specific dollar estimate of the within-country adoption premium at firm level.
Full report
Our take: The NZ$59.1m large-business figure is a marketing-friendly number, and it is also the number where the methodology footnotes do most of the work. Self-reported productivity, no control for selection bias (firms that adopt AI early are usually better-run anyway), and FY25 timing means the comparison year is the year AI tooling became table-stakes. Contrast this data against last week's MYOB survey said NZ mid-sized businesses trail Australia on every AI maturity pillar.
ANZ NZ, ASB, BNZ and Kiwibank all joined Visa's Asia Pacific Agentic Ready programme on 30 April, running simulated agent-initiated transactions in controlled environments using Visa's existing tokens, identity verification and risk-management infrastructure before any consumer rollout. It is the first time NZ's big four retail banks have moved together on agentic commerce. The programme is now live across 10 APAC markets. Visa Country Manager NZ David Peacock said the focus is on ensuring agentic commerce is safe and secure for New Zealanders as more examples of agent-initiated payments become a reality.
3 min read
Our take: The big four moving together is the part that should land with NZ retailers and SaaS pricing teams. Agent-led purchasing displacing direct merchant interactions changes the surface area where pricing, promotion and brand decisions get made. When the customer is an AI agent buying on its principal's behalf, the marketing playbook that worked for the human shopper does not survive contact with the agent's preference function.
TUANZ released its sixth annual Digital Priorities Report on 29 April, rating NZ's tech adoption a stagnant 6/10. The report warns the innovation gap with comparable small nations is widening rather than closing. The report draws on interviews with around 30 NZ CIOs and CTOs, and explicitly names Singapore, Denmark and Finland as the comparator set. CIO conversation has shifted away from experimentation toward high-volume practical AI applications, but the topline grade has not moved.
3 min read
Our take: The shift from experimentation to high-volume practical applications is the part NZ CIOs should be acting on this quarter. Pilot programmes have lost their novelty premium and budgets are tight (the ANZ Business Outlook fell 43 points in a month). The CIOs winning right now are the ones who already have one or two production AI workloads earning their keep, not the ones still deciding which framework to standardise on.
BNZ formally changed its software capitalisation accounting on 29 April, writing down NZ$352m of capitalised software and taking a NZ$253m post-tax profit hit at the half-year, citing AI's effect on shortening the useful life of banking technology. Parent NAB applied the same change in its own books on the same day. ANZ, Westpac and ASB now face sector pressure to explain their own AI-related software impairment positions before the next half-year results.
3 min read
Our take: This is the first time a major NZ bank has put a number on AI as a write-off, not just an investment. Until now the AI capex story has been one-directional: budgets up, vendors selected, pilots launched. BNZ has named the other side of the ledger, which is that the software those investments replaced is depreciating faster than the original five-to-seven-year capitalisation assumed.
Nelson AI Sandbox opened a free hands-on AI learning corner inside Halifax Cafe on 1 May, co-founded by ex-Microsoft Silicon Valley executive Richard Brudvik-Lindner who relocated to NZ in 2010 and became a secondary school teacher. The Sandbox has trained roughly 500 staff and volunteers across 210 NZ non-profits in two years, funded by the Rātā Foundation. The cafe specifically targets Nelson's elderly residents and low-wage workers.
4 min read
Our take: Microsoft's NZ training pledge made the headlines two weeks ago: 200,000 Kiwis by 2028. The Nelson AI Sandbox has trained 500 people across 210 non-profits with no Microsoft press release, no US CEO visit, and a budget that fits inside one philanthropic foundation. Scale is not the same as reach, and reach is what literacy work actually requires.
📚️ Mike’s Takes From The Week
Helping leaders and teams adapt, learn, and scale with AI.
1️⃣ The behaviour-change tax goes to zero when AI meets people where they work: Erin took over the AI Corner podcast workflow with one prompt. No training session, no playbook handover. A pre-built Skill walked her through Descript, drafted the LinkedIn post, queued the Spotify upload. The handover tax most teams never pay.
3-min read.
2️⃣ Most have the SaaS and shadow AI conversation backwards: Steven Baker (Node / VendorSage) reckons 85-90% of AI implementations fail for the same reason 70% of SaaS rollouts do, and it is not the tech. The real source of truth nobody owns is how work actually gets done.
3-min read.
3️⃣ NZ's productivity paradox: 82% use AI, GDP per capita still falling: Deloitte says SME adopters earned $400k more in FY25, large businesses $59.1M more. Yet NZ sits 63rd of 67 OECD countries for productivity. Half the early-adopter room at the M2 AI Summit hasn't touched agentic AI.
3-min read.
4️⃣ Training prepares people for AI. A lighthouse team prepares AI for people: McKinsey says 80% of companies see no bottom-line impact from AI. The fix is not better training or more licences, it is a small team shipping production infrastructure inside real workflows. Ramp's 99% adoption proves the pattern.
8-min read.
🛠 Latest Builds and Finds
Helping advanced builders stay at the frontier of AI.
1️⃣ Refero, a design system library built for AI agents to read. Best in class curated brand references for shipping apps with colours, type, spacing, components, and a DESIGN.md file per style. This is the direction every static reference site has to go if it wants to stay useful (assets agents can ingest and apply, not just thumbnails humans browse). (link)
2️⃣ Karpathy on the shift from vibe coding to agentic engineering. His framing of Software 1.0 (handwritten code), 2.0 (learned weights), and 3.0 (natural-language programs that drive models) is the cleanest way to explain what's actually changing in how software gets built. The talk to send anyone who still thinks "AI coding tools" is a single category. (video)
3️⃣ An eight-layer mental model for agentic AI architecture. Neha mapped out the layers anyone designing an agentic system has to think about: orchestration, agents, tools, memory, monitoring, reliability, governance. The diagram is a useful checklist. Most teams building agents skip three or four of these (usually monitoring, reliability, governance) and discover the gap in production. (link)
🌍 Tech Updates From Global
The selected top headlines from each major AI tech company.
OpenAI
Microsoft partnership amendment on 27 April ended Azure exclusivity, capped OpenAI's revenue-share payments and removed Microsoft's legal hook over the $50B Amazon deal, clearing the path to multi-cloud distribution and an IPO. (Apr 27)
GPT-5.4 and Codex went live on Amazon Bedrock 24 hours after exclusivity ended, with Bedrock Managed Agents in preview as AWS positioned itself as exclusive third-party cloud distributor for OpenAI's enterprise platform Frontier. (Apr 28)
Musk v Altman federal trial opened in Oakland on 28 April with Musk seeking $130B+ damages and asking the court to unwind OpenAI's October 2025 public benefit corporation conversion, Day 4 testimony framing himself as "a fool" for funding what became an $800B for-profit. (Apr 28)
GPT-5.5 became Codex's default model with multi-environment app-server sessions, ChatGPT shopping added side-by-side product comparison and reverse-image search, account security added Yubico passkey support, and one Google Drive app replaced separate Drive/Docs/Sheets/Slides connectors. (Apr 30 to May 2)
Anthropic
AWS confirmed on 27 April that Anthropic is now training its most advanced foundation models on AWS Trainium/Graviton co-engineered at silicon level with Annapurna Labs, with Claude Cowork deployable inside Bedrock and Claude Platform on AWS "coming soon". (Apr 27)
Opened its Sydney office on 27 April and named ex-Snowflake SVP Theo Hourmouzis as ANZ General Manager, deepening ties with Commonwealth Bank, Quantium, Canva and Xero ahead of a likely NZ enterprise push. (Apr 27)
Showcased the three-year Google partnership at Google Cloud Next 2026 on 28 April with Shopify's Sidekick agent on Claude+Google Cloud and Palo Alto Networks reporting 20-30% engineering velocity gains as the named ROI proof point. (Apr 28)
Reportedly received preemptive offers for a $50B round at $850-900B valuation with a 48-hour investor allocation deadline (separate from last week's $65B hyperscaler raise), board to decide in May. (Apr 29)
Google / DeepMind
DeepMind unveiled its multimodal AI co-clinician on 30 April, beating GPT-5.4-thinking-with-search 63 to 30 in blind primary-care evaluations across 98 queries. (Apr 30)
Quietly signed a Pentagon AI deal on 28 April letting the DoD use Google models on classified work for "any lawful governmental purpose", contradicting earlier management promises and triggering a letter from 600+ Google and 100+ DeepMind staff. (Apr 28)
Q1 earnings on 29 April pushed Google Cloud past $20B/quarter for the first time, with Alphabet revenue up 22% YoY and the cloud unit consistently growing faster than overall company revenue. (Apr 29)
Microsoft
Microsoft 365 Copilot crossed 20M paid seats (up from 15M in January) and signed Accenture for 740,000 seats (its single largest deal), with 50,000+ seat customers quadrupled and queries up ~20% QoQ. (Apr 29)
Azure grew 40% YoY with commercial bookings up 230% YoY in Q3 FY26, driven by multiple $100M+ enterprise commitments, with OpenAI now accounting for 45% of Azure's $625B remaining performance obligations. (Apr 29)
Microsoft 365 E7 (Frontier Suite) hit general availability 1 May, bundling M365 E5, Entra Suite, M365 Copilot and Microsoft Agent 365 into a single top-tier SKU transactable in the CSP channel. (May 1)
Amazon / AWS
Amazon Quick desktop AI assistant launched at "What's Next with AWS" 2026 with Free/Plus tiers and native integrations to Google Workspace, Zoom, Airtable, Dropbox and Microsoft Teams, with 3M, GoDaddy, AstraZeneca, BMW, Mondelēz, NFL and Southwest as early adopters. (Apr 29)
Andy Jassy's Q1 FY26 disclosure on 30 April put AWS AI revenue run-rate above $15B (260x AWS's first three years), with 125,000+ Bedrock customers running most inference on Trainium and over $225B in Trainium revenue commitments, Trainium 4 reserved ~18 months out. (Apr 30)
Meta
Q1 2026 revenue beat at $56.3B (+33% YoY) and net income $26.8B was overshadowed by 2026 AI capex guidance lifting to $125-145B (from $115-135B) on component-pricing and data-centre costs, sending the stock down ~7% after-hours. (Apr 29)
NVIDIA
Stock dropped 4.6% on 30 April from $209.25 to $199.57 after hyperscaler earnings highlighted custom silicon (Trainium, TPUs) and Meta attributed higher capex partly to "component pricing" rather than GPU demand, before recovering past $5T market cap by 1 May. (Apr 30)
xAI
Grok 4.3 API rolled out fully on 30 April at $1.25/M input and $2.50/M output tokens with a 1M-token context window and the first native video input across the developer console, an aggressive 40% input price cut on prior pricing. (Apr 30)
Musk admitted under cross-examination on 30 April that xAI "partly" used distillation on OpenAI models to train Grok and called it "standard practice", the first on-record admission by a major lab founder with potential IP and regulatory implications. (Apr 30)
Apple
Q2 FY2026 revenue beat at $111.2B (+17% YoY) on 30 April with Services at $30.97B in Cook's first earnings since naming Ternus as successor; Cook confirmed Siri's Google collaboration "going well" and flagged "significantly higher" memory costs ahead due to the AI-driven supply crunch. (Apr 30)
✨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