Return to Part 0: Table of ContentsPrevious Article: Part 41, Ninety-One Per Cent Adopted, Four Per Cent Changed
Gartner expects global spending on AI models and platforms to reach US$64 billion in 2026, up 63 per cent on 2025 [1] [2]. That is the number making the rounds this week. It is also the wrong number to anchor a renewal decision on, because Gartner's own forecast, published 20 July 2026, does not move as one figure. It moves as three.
Domain-specific and specialised models are growing at 210 per cent. Generalist generative AI models are growing at 117 per cent. AI platforms, the layer that holds the models, the data connections and the governance controls together, are growing at 36.9 per cent [1] [2]. Platform spending is increasing at roughly a third of the generalist rate and a sixth of the specialist one. Whatever story the headline US$64 billion tells, the three rates underneath it tell a different one: the money is moving away from the layer that holds everything, toward the models built to do one thing well.
Two corrections belong in the same breath as that number, not added later as an afterthought. The forecast is three months old at the time of writing, not a fresh release; Gartner published it on 20 July 2026. And US$64 billion is not what Gartner means when it talks about AI spending generally. The firm's own separate total AI spending forecast, updated again in a quarterly release reported in September 2026, puts the whole category at roughly US$2.7 trillion for 2026 [3]. US$64 billion is models and platforms only, something like a fortieth of the wider figure. Treat the two as describing different things, because they do.
Three rates, not one
A platform, in Gartner's own category, is the orchestration layer: the connections, the permissions, the logging and the governance controls an enterprise wraps around whichever models it actually runs. A generalist model is the kind most people picture when they hear "AI," a foundation model built to answer almost anything reasonably well. A domain-specific or specialised model, a DSLM in the industry's own shorthand, is built and tuned for one task: reading contracts, classifying support tickets, summarising claims. All three sit inside the same US$64 billion market Gartner measured, and all three are growing, but not at anything close to the same speed.
Arunasree Cheparthi, senior principal research analyst at Gartner, put the shape of it directly: "Enterprise AI budgets are coming under greater scrutiny, with increased focus on usage efficiency, cost control and measurable outcomes" [1] [2]. That is not a sentence about AI spending shrinking. It is a sentence about AI spending being asked to show its working, category by category, rather than being waved through as one line on a budget.
Why the fastest rate is also the smallest
The 210 per cent figure is the one that will get quoted on its own, and it is also the easiest of the three to overstate. Domain-specific and specialised models grew from roughly US$1.6 billion to roughly US$4.9 billion in Gartner's own figures [1]. That is a real and large increase, and it is also growth measured off the smallest base of the three categories, inside a market Gartner sizes at US$64 billion overall. A rate without its base is a story half told.
I made close to the same point about a different Gartner forecast once before. Writing about Gartner's prediction that a share of enterprises would demote or decommission their AI agents after production incidents, I warned readers not to treat a forecasting firm's own number as settled proof of anything, because the firm has its own stake in the answer. As I wrote in The Hallucination Economy, "Gartner sells advisory work on the problem it forecasts, and the direction of that interest runs towards the problem being large." The caution travels to this week's figures as well as it did to that one. It does not make the three growth rates wrong. It means they are worth reading with their dollar bases attached, not as a verdict handed down from outside the market they describe.
Cheparthi's second comment, reported separately, names where the scrutiny is actually sending the money: "Spending is shifting toward providers who can demonstrate clear value across cost, latency, performance, and reliability... the biggest winners will be vendors that help enterprises manage where and how AI is used across the business" [1] [2]. Two outlets, two quotations, one release. Read together, they describe a market where growth is moving to vendors who can prove a narrow result, not to the vendor holding the platform contract.
What a platform renewal is actually competing against
A platform renewal usually gets signed on momentum: last year's spend, this year's vendor roadmap, a general sense that AI investment should keep climbing. Gartner's own decomposition says that is the wrong comparison to run the decision against. A platform licence is not competing against "AI spending" broadly. It is competing against a generalist-model line growing more than three times as fast, and a specialist-model line growing almost six times as fast, inside the same budget.
That matters because the three categories answer different questions, and an enterprise buyer rarely asks itself which question it is actually paying to answer. A platform buys flexibility: the ability to point many different models at many different problems through one governed connection. A generalist model buys breadth: one capable system that can be redirected at a new task without retraining. A specialist model buys a narrow, measurable result: a model tuned to one workflow, cheaper to run per task, and easier to hold to an evaluation score because the task itself is narrow enough to score.
The procurement mistake sitting underneath this week's number is the same one this series has tracked in different forms all year: buying on momentum rather than on what is actually being measured. A vendor pitching a platform renewal this quarter will, reasonably, point to the AI market's overall growth. The honest question back is which of the three rates the renewal is actually priced against, because a platform contract growing at 36.9 per cent a year is not riding the wave the headline number describes. It is the slowest part of it, and the part most likely to be signed without anyone checking.
None of this says platforms are a bad buy. An organisation running a dozen different models across a dozen workflows needs the governance layer a platform provides, and that layer does not disappear just because the fastest growth sits elsewhere. What changes is the basis for the renewal conversation. A platform earns its contract on what it governs, not on the fact that AI spending in general is going up. Ask a vendor to show that basis before the next renewal, not after it.
The New Zealand read, and what it cannot prove
Prospa's SME Sentiment Tracker, published 26 October 2026 and genuinely inside this week's window, gives New Zealand its own version of the same split. Fifty-nine per cent of New Zealand small and medium businesses now use AI in some form, up from 50 per cent six months earlier. Inside that figure: 33 per cent use it in a specific business function, 19 per cent are experimenting, and 7 per cent describe it as integrated across multiple functions and strategically important [4].
Set the two findings beside each other and a reading suggests itself immediately. The 33 per cent using AI in one specific function looks like Gartner's specialist buyer, the one driving the 210 per cent rate. The 7 per cent integrated across functions looks like Gartner's platform buyer, sitting on the layer growing slowest. It is a tidy parallel, and it needs a flag attached the moment it is stated rather than left to travel on its own.
Gartner's figures measure category-level vendor and software spending growth across a global, enterprise-weighted population. Prospa's figures measure self-reported categories of AI use inside a survey of 501 New Zealand business decision-makers, each running a business under 50 full-time staff. These are two different units, measured by two different instruments, in two different populations. No source measured the connection between them directly. It is a reading, constructed for this piece, between two independently true facts, and it is useful precisely because it is stated as a reading rather than dressed up as something either company found.
The 59 per cent figure itself is not the only one circulating. Other recent figures for New Zealand AI adoption have run as high as 67 per cent and 87 per cent, from different instruments measuring different things in different ways [4]. That spread is worth naming rather than quietly resolving. A sentence that starts "New Zealand AI adoption is..." is underdetermined by more than the gap it is usually used to settle, and the honest answer is that three different instruments currently disagree with each other by a wide margin, not that one of them is the real number.
What buying on one number looks like
The same mistake shows up one layer down, at the point where an organisation picks a specific model rather than a spending category. MiniMax's new M3.1 Flash model scored 42 on the Artificial Analysis Intelligence Index, 13 points above its own predecessor, at roughly half the cost per task [5]. It also carries a widely circulated figure describing a 93 per cent hallucination rate, which sounds damning until the definition underneath it is read: that figure is the share of the model's wrong answers that were outright incorrect rather than partial or a refusal, not the share of all its answers that were wrong. On the separate AA-Omniscience accuracy benchmark, the same model answers correctly 56 per cent of the time, ahead of both of the rival models it is usually compared against [5].
A buyer who reads only the Intelligence Index score, or only the hallucination headline, is making the procurement version of reading Gartner's 210 per cent figure without its dollar base: taking one number out of a decomposed measurement and treating it as the whole answer. The model is simultaneously more accurate than its rivals on one measure and less calibrated about its own failures on another. A one-line summary that picks whichever number fits the pitch being made is not a smaller version of the mistake. It is the same mistake.
A useful discipline, whether the decision is a platform renewal or a single model choice, is to write down four things before signing anything: which of the three growth categories the spend actually falls into, what the dollar base behind any quoted growth rate actually is, who measured the figure being relied on, and whether that measurement was checked by anyone outside the party selling the result. None of the four requires new tooling. All four require someone to ask the question before the contract, not after.
One more figure is worth a single sentence, with its own counterweight attached rather than left to stand alone. Recent US polling found 57 per cent of adults saying AI is a negative influence on society, and 66 per cent opposing new data centres in their area [6]; a separate national survey found AI ranking third among Americans' biggest fears, at 51.9 per cent, but found that weekly AI users feared it noticeably less than people who had never used it [7]. Reported together, sentiment toward AI looks to be worsening in general and improving sharply with direct use, and a sentence that reports only the first half recommends roughly the opposite of what the fuller picture actually supports.
- Gartner's 2026 AI models and platforms forecast, published 20 July 2026, splits US$64 billion into three growth rates: specialised models at 210 per cent, generalist models at 117 per cent, platforms at 36.9 per cent.
- The 210 per cent figure grew off the smallest dollar base of the three, roughly US$1.6 billion to US$4.9 billion; a rate without its base overstates the finding.
- Prospa's New Zealand SME Sentiment Tracker shows a similar functional split, 33 per cent in a specific function against 7 per cent integrated across functions, read here as an analogy to Gartner's categories, not as a measurement of the same thing.
- MiniMax's M3.1 Flash model shows the same error at the model-selection layer: a single headline number, read alone, tells a different story than the full, decomposed result.
- Before any platform renewal or model decision, name the growth category, the dollar base, and who did the measuring.
The open-source dimension of this sits closer to the specialist category than the headline number suggests. A meaningful share of the fastest-growing segment is not proprietary at all: Hugging Face's model hub carries thousands of openly licensed, domain-specific models that an organisation can download, tune and run on its own infrastructure, with no platform vendor standing between the model and the task it performs. That availability is not a side note to the specialisation story. It is one of the mechanisms making the specialist category buyable without a matching platform contract, because the tuning and serving method is published rather than withheld by a single vendor. An organisation choosing a specialist model is increasingly also choosing whether to buy its independence from a platform or rent it, and the open-weight path is the one that keeps that choice open for later.
The implication on the sovereignty side is worth stating directly. If the fastest-growing category of enterprise AI spending is also the category most available in open form, open weights versus closed weights stops being a licensing preference and becomes a sovereign-capacity question: which organisations, and which countries, can run their own specialist models rather than renting continued access to someone else's platform. Export control regimes already treat frontier computing capacity, and increasingly the model weights themselves, as strategic assets rather than ordinary software products. A procurement decision made this quarter on cost grounds alone, specialist model against platform renewal, is also a decision about whether the resulting capability can be inspected, moved, or kept running without a foreign vendor's continued cooperation. That question never appears on Gartner's growth chart. It appears later, when continuity actually matters.
If your next platform renewal is being justified by how fast AI spending is growing overall, which of the three rates underneath that number is the one your vendor is actually asking you to pay for?
This is one of seven weekly series in The Hamberger Report. Subscribe on LinkedIn and the next one arrives in your feed.
The views expressed in this article are entirely my own, informed by morethan 30 years of professional experience in architecture, security, andtechnology leadership in New Zealand. I write as director of Te PonoLimited; the views are personal and do not represent the position of anyclient, any government agency, or the New Zealand government. My commentaryon legislation and policy is analytical, drawing on publicly availablesources and my professional expertise in architecture, security, and AIgovernance, and it is politically neutral.
Andreas Hamberger is a New Zealand leader in Architecture & Security and Associate Member of the Institute of Directors. The Hallucination Economy examines what happens when organisations buy, deploy, report on and are audited against AI claims that nobody outside the producer can check. Through Te Pono he provides independent reviews of agentic AI control planes for organisations moving from pilot to production; contact andreas@thehambergerreport.com for the scope and fixed fee.
This article was produced with AI assistance under my direction. Research, drafting and images pass through a pipeline I built and govern: automated gates for source verification, forbidden language and political neutrality, and my own review before anything is published. The tools include Claude, Gemini and Openart. The frameworks, arguments and editorial judgements are mine and are the same discipline I apply to the AI systems I audit for clients. AI accelerated the work; the thinking, and the responsibility for it, are mine.
[1] CFOtech Australia. "Gartner forecasts AI spending to hit US$64bn in 2026." 20 July 2026. https://cfotech.com.au/story/gartner-forecasts-ai-spending-to-hit-usd-64bn-in-2026
[2] ARN. "Gartner predicts AI models and platforms spend to hit US$64B." 20 July 2026. https://www.arnnet.com.au/article/4199047/gartner-predicts-ai-models-and-platforms-spend-to-hit-us64b.html
[3] CIO. "AI spend will jump 49.5% in 2026, says Gartner." September 2026. https://www.cio.com/article/4225785/ai-spend-will-jump-49-5-in-2026-says-gartner.html
[4] eCommerceNews New Zealand. "New Zealand small businesses adopt AI as hold-outs shrink." 26 October 2026. https://ecommercenews.co.nz/story/new-zealand-small-businesses-adopt-ai-as-hold-outs-shrink
[5] VentureBeat. "MiniMax's new M3.1 Flash beats DeepSeek and Opus 5 on a key business benchmark at lower cost, but with a high hallucination rate." 19 to 20 October 2026. https://venturebeat.com/technology/minimaxs-new-m3-1-flash-beats-deepseek-and-opus-5-on-a-key-business-benchmark-at-lower-cost-but-with-a-high-hallucination-rate
[6] ABC News, The Washington Post and Ipsos. "Most Americans say AI is bad for society, oppose new data centers: POLL." 21 October 2026. https://abcnews.com/Politics/americans-ai-bad-society-oppose-data-centers-abcpostipsos/story?id=137414085
[7] Chapman University. Survey of American Fears, 12th annual wave. 20 October 2026. https://www.eurekalert.org/news-releases/1148421