Evidence over rhetoric: scarcity still shapes AI progress

Scarcity hasn’t vanished; it moved houses. Look — saying AI abundance is the default is the intellectual equivalent of declaring the pantry full when someone’s locked the door.

The Business Times column gets one big thing right: it refuses to take “AI abundance” on faith. The original scarcity — raw compute and storage — eased because cloud providers built massive capacity and open models spread. That’s real. But declaring victory there and generalising to “abundance” everywhere skips the hard part: what actually constrains value.

Here’s what nobody tells you: scarcity is now about different things — high-quality, proprietary data; the engineering know-how to turn models into reliable products; safe deployment at scale under regulatory scrutiny. Those aren’t minor frictions; they’re strategic chokepoints. If you’ve ever tried to get a model into a production incident-management process without breaking everything else, you know the constraint isn’t “Do we have a GPU?” It’s “Can this thing behave predictably when the stakes are high?”

Give me a break — data isn’t free just because models can be trained on public corpora. Companies still hoard the cleaned, tagged, proprietary datasets that make a model competitive in a vertical like finance or healthcare. Cleaning data, annotating edge cases, integrating privacy-preserving pipelines — that’s human work and organisational discipline, not raw silicon. In operations, when we shifted a bottleneck from machines to processes, the PowerPoint looked optimistic but the project timelines didn’t magically shrink; the constraint just got messier and harder to brag about.

That’s why the article’s demand for “sharper evidence” matters. Hype about abundance is cheap. Evidence should show up where pain actually moves: in wages for ML engineers, in the prices and contract terms for curated datasets, in the budgets for production engineering and assurance. If we’re really in an abundance era, those pressure points should ease or change character, not just relocate and rebrand.

Spare me the hype about “AI everywhere” when the economic signals still treat high-quality data and integration talent like scarce assets. Investors who swallow the abundance story whole will fixate on marginal model tweaks and ignore the compounding advantage of boring work: data partnerships, domain expertise, compliance operations, and the glue code that keeps systems from falling over. That’s where durable value hides when models themselves feel interchangeable.

Regulators get misled in a different way. Talk up abundance too loudly and you create two bad paths: either panic about an unstoppable flood of cheap AI or complacency that assumes ubiquity equals maturity. Neither stance matches reality. The scarcity has shifted into governance capacity, audit mechanisms, and the people who can actually interrogate model behaviour against legal and ethical standards. Those are not abundant, and pretending otherwise invites policy mistakes.

The column does understate one thing: there are pockets where abundance is already visible. Look — open-source models and cheaper inference really can democratise creativity and prototyping. In those niches, the constraint moves to imagination and user time, and that’s exciting. You can have thousands of people testing ideas that used to require a specialist team.

But give me a break — democratising prototyping doesn’t equal enterprise-grade abundance. Once you cross into mission-critical territory, new scarcities show up fast: verification, latency guarantees, remediation processes, legal clarity. The integration work alone can dwarf the “cheap model” story. You can prototype in an afternoon and still spend months getting audit, risk, and security to sign off on real deployment.

That’s the real question: what would actually count as evidence that scarcity has eased instead of just migrated? The article is right to push for measurable shifts in constraints, but it could press harder by naming the signals that matter: contract lengths and pricing for data licences, growth in roles focused on production ML engineering, and changes in regulatory filings where model risk shows up as a distinct category. Until those indicators move, “abundance” is a thesis, not a condition.

Wake up — talent markets will bifurcate. One cohort will stick to low-barrier creativity with off‑the‑shelf models, riding whatever interface is hottest this quarter. Another will specialise in the plumbing that turns models into accountable systems: data contracts, observability, failure handling, cross‑functional reviews. Training programmes that obsess over model architecture but skip deployment, monitoring, and change control are grooming people for the wrong scarcity.

Here’s what nobody tells you: abundance‑sellers often conflate output volume with accessible value. More output doesn’t equal more utility when the bottleneck is trust, interpretability or compliance. If customers can’t use the outputs safely — or can’t explain them to a regulator, a board, or a court — abundance collapses into rework, manual checks, and politically constrained pilots.

The Business Times piece lands the crucial punch: stop treating “AI abundance” as self‑evident and start asking where constraint has actually moved. Watch how job postings for production‑grade ML roles and procurement contracts for data and integration services evolve; that’s where the new walls of scarcity will quietly show up.