Growing Applications
The top of the 'atoms to applications' stack bears all of its weight.
Driven by applications and use cases, AI will thrive or fail. Broad classes of applications have been visible, distantly, since the release of ChatGPT.
In the following four years, much has been attempted, with mixed successes. Agents arrived with the first API to GPT-4, and fell over almost immediately, undone by hallucinations, insufficient context, or poorly implemented goal logic. A lot of things went wrong for a long time, long enough that many began to question whether AI would ever be "good enough" to get the job done.
That all changed at 'the Watershed', when agent 'task horizons' extended beyond an hour, driven by fourth- and fifth-generation language models, in harnesses that gave them the memory, tools and logic needed to complete tasks successfully, whether in software, law, marketing or customer service.
In the year since the Watershed, firms have raced to keep pace with the expansive new capabilities on offer, translating capacity into applications.
Applications are where all the money lands, it's an enduring focus of Noops Research, and the capstone presentation for our AI Investor summit.
This is the coalface, where experimentation and deployment meet speculation and overreach. To navigate these waters, we present Applications in a matrix of knowns and unknowns.
Known Knowns: Justifying Capex
An $11,000,000,000,000 investment. That's the estimate from SemiAnalysis representing the total capex in data centres from 2024 - 2029.
Data centres are stuffed with rapidly depreciating equipment; the performance difference between two generations of Nvidia GPUs - H200 from 2024 and Vera Rubin in 2026 - is thirty-nine times. Any state-of-the-art data centre built just two years ago needs to earn back its investment very quickly. Before it becomes obsolete.
The frontier labs will pay to train their models only on the latest-generation GPUs. They can inference on old GPUs - but will earn far lower margins than if they inference on new ones. The frontier labs need inferencing demand in order to pay for training, and inferencing demand can only be generated by applications.
Software engineering is the current 'killer app' for inferencing. Most of the tokens currently consumed by businesses are used in software development and software maintenance.
Noops Research believes that adoption of AI in software engineering is following a standard 'S-curve' trajectory, and that we're still in the early phase of growth, somewhere between 1 and 10% of the addressable market.
As deployment across software engineering reaches the centre of the S-curve in the next several years, we predict a minimum of 10x growth in token consumption.
Finance is emerging as the second 'killer app' for inferencing, with funds like Jane Street leading the way in AI tooling for staff. The largest consumer of tokens, by time zone, is the US East Coast, headquarters of the global financial sector.
That rollout is only in its very early days, and we expect Finance applications will lead to at least a 100x increase in token consumption as AI tooling reaches the entire sector.
Between software engineering and finance, AI applications are on growth trajectories sufficient to justify existing capex.
Known Unknowns: SaaSpocalypse
Earlier this year, as 'vibe coding' transformed the practice of software engineering, a new word emerged: "SaaSpocalypse".
For SaaS businesses, the question posed by "good enough" AI is existential: If a customer operates a fleet of a thousand agents, do they need a Salesforce? Or Atlassian, Canva, SAP, ServiceNow, Oracle, Xero or Workday?
The market doesn't seem to have an answer to this question. SaaS stocks were all marked down early in 2026, though most recovered much of their drop in valuation.
That should not imply a return to the status quo. "Good enough" agents changed everything.
In April, Salesforce introduced "Headless 360", an API and CLI interface to Salesforce, making agents first-class users of the tool, with CEO Marc Benioff saying,"No Browser Required, Our API is the UI." That's getting out in front of the agentification-of-all-the-things, but does it do more than buy time?
No one knows. Firms like SalesForce and SAP and Xero could become 'systems of record' - where organisations keep their structured data. That's a very different business - but at least it's a business.
In 2026, agents can use the existing tools. Will that matter as a more fundamental transformation of software takes hold over the next decade, and all software becomes 'software factories'?
In 2036, will anyone use a word processor? A spreadsheet? Presentation builder? Email? Group messaging?
Pick a pure software play and ask - when you have a thousand agents at your fingertips - whether anyone needs any of those tools anymore, or whether agents make all of these coordination and communication tasks broadly obsolete.
The applications we use reflect the capabilities of the systems they're built from. Those capabilities have completely transformed - why would applications remain the same?
Most SaaS incumbents are adopting AI as quickly as they can, hoping that being "AI first" buys them breathing room to reinvent their business models for an agentic era.
Unknown Knowns: Software Factories
A new organisational form has emerged in the post-Watershed era: 'greenfields' businesses built with the base assumption that they will be using fleets of agents to service the needs of customers with their own agents.
At the heart of these businesses sits a new thing, only possible with 'good enough' AI: the 'self-healing loop'. The organisation defines a process to provide a product or service, automates the production cycle for that product or service, then monitors the outputs for errors and improvements.
These 'software factories' emerged as organisations explored the capabilities of the new AI tools, the name applied only after the practice itself had cohered.
Melbourne-based Stile Education runs their software engineering process within a self-healing loop. Software gets released to customers with detailed telemetry feeding usage and error data back to Stile Education. As customer errors accrue, they are detected within that telemetry, isolated, and passed along with all relevant information to agentic analysis tools. Those agents first locate specific cause of the problem, effect a repair, test that repair, then merge the fixed software, which again feeds its telemetry back into the system.
The loop never stops monitoring. With the right settings, the software never stops improving. The outcome: an 'anti-fragile' process. The harder and faster it runs, the better it gets.
Investment fund Bracket 22 touts itself as "human insight x artificial intelligence". The comparison with a traditional fund is striking: running costs down from ~$5m in salaries, benefits and space to "$30,000 to $40,000 a year, total." A 125x cost-compression in a greenfields investment fund will inspire imitators.
We're now watching a broader range of established businesses adopt similar practices, including Uber, Cognition and Stripe.
The cost and quality outcomes of software factories are driving their adoption, with greenfields businesses being able to move faster, as they will have no requirement to refit process infrastructure to these emerging practices.
Unknown Unknowns
It's very hard to know the shape of this future. Too much is in flux at present, and the space - with the exceptions of software engineering and finance - is largely unexplored.
Some of the speculative possibilities include:
Agents as counterparties: As 'know-your-agent' (KYA) systems become trusted and reliable, agentic commerce takes off. Stripe are working hard to own as much of this space as possible.
Indemnity: If an agent creates an auditable record to serve as the basis for underwriting, insurers become the auditors of self-healing loops in software factories, better agents and self-healing loops get lower premiums. No one offers this yet, though Armilla AI and Munich Re are pricing on audits.
Verification: Jane Street uses 'backpressure' to ensure its agents and loops stay reliable. At scale, this implies an enormous market for machine-checkable guarantees around agentic outputs, sold to whoever automates the approval gate. The buyer is the CISO who won't let general-purpose agents near the network, and the insurer pricing the policy.
Agent-native media and formats: Humans need text, but we're already seeing models like GPT-6 Astra adopting a mixture of symbols and words that represent a 'compression' of their internal chains-of-thought. As models get smarter, their outputs become less comprehensible.
The Real World Gets a Look In
Applications are not just software. What happens when a manufacturing line gets enough AI tooling to run a 'self-healing loop'? Or a pharmaceutical manufacturer? A farmer?
The real world is much broader than what gets dreamt up on a screen.
It's here that Noops Research admits our biggest blind spot. We're so trained by applications running in their own tidy little software universes we find it hard to know what it will mean when agentic applications have found their way into nearly every manufactured object, either through its production process or in its ongoing operation.
We know the agentic world will be different. It already is. That's why this moment is so pregnant with opportunity.