The AI Winter That Didn't Come

The AI winter argument had logic behind it. It went something like this:

The initial wave of generative AI excitement (late 2022 to mid-2023) was driven by novelty. Enterprises experimented broadly. Pilots proliferated. But most pilots would fail to scale. The gap between demo and production would become obvious. Executive enthusiasm would cool. Budgets would tighten. And the industry would enter a correction - not necessarily a crash, but a significant pullback in investment and adoption.

It was a reasonable hypothesis. It had historical precedent. The AI winters of the 1970s and 1990s followed exactly this pattern: breakthrough, hype, disillusionment, retreat.

But 2024 didn't follow the script.

Enterprise AI spending in 2024 didn't contract. It expanded. Dramatically. Across every sector we track, AI budgets grew 50-100% year-on-year. And critically, the nature of spending shifted. Less experimentation. More production infrastructure.

The money didn't just increase - it matured. Pilot budgets became platform budgets. Innovation funds became operational line items. The shift from discretionary to committed spending is the strongest signal that enterprise AI has crossed the experimentation threshold.

The percentage of NZ/AU enterprises with production AI deployments roughly doubled from 2023 to 2024. Not huge in absolute terms - from roughly 15% to roughly 23%. But the trajectory is clear and accelerating.

More importantly, the organisations with production deployments started compounding. Their second capability deployed faster than their first. Their third faster than their second. The infrastructure investment started paying off. This is the compound advantage in action, and it's the strongest argument against an AI winter.

This is what made 2024 different from previous hype cycles. In past AI winters, the technology hit a capability ceiling. The promises outran what the technology could deliver, and the gap was structural - it couldn't be closed with incremental improvement.

In 2024, the technology kept getting better. Claude 3 arrived in March, closing the gap with GPT-4 and introducing genuine competition. GPT-4o launched in May, bringing multimodal capabilities to a broader market. Claude 3.5 Sonnet pushed coding and analysis capabilities further. Each release made enterprise AI more capable, more reliable, and more cost-effective.

When the technology is improving faster than expectations are rising, the bubble doesn't burst. It inflates on substance, not air.

But let's be honest about what the sceptics correctly identified:

This hasn't changed. The majority of AI pilots in 2024 did not progress to production. The reasons are the same as always: lack of data readiness, missing governance frameworks, organisational resistance, and projects designed to demo rather than deploy.

The sceptics were right that most enterprises were doing AI badly. They were wrong that this would stop AI adoption. What happened instead: the enterprises doing AI well accelerated, creating competitive pressure that pulled others forward.

AI vendors continued to overpromise in 2024. Implementation timelines remain unrealistic. Accuracy claims are inflated. Total cost of ownership is understated. The vendor problem the sceptics identified is real.

But enterprises are getting better at vendor management. The naive optimism of 2023 has been replaced with informed scepticism. Procurement processes are adapting (slowly). Technical evaluation is improving. The vendor problem isn't solved, but it's being managed.

The sceptics predicted that talent shortages would constrain adoption. They were right. The talent gap is the single biggest brake on enterprise AI deployment in NZ/AU. But it hasn't stopped adoption - it's slowed it. Organisations are working around the gap through partnerships, upskilling, and selective hiring.

AI governance hasn't kept pace with AI deployment. Most enterprises deploying AI in production don't have governance frameworks that match the risk. The sceptics were right to flag this as a systemic issue. The consequences haven't materialised yet, but they will.

Three structural differences between 2024 and previous AI hype cycles:

Previous AI waves required massive infrastructure investment before any value appeared. This wave offers immediate individual productivity gains. A knowledge worker using Claude or GPT-4 is measurably more productive today, with zero infrastructure investment. This creates grassroots demand that sustains investment even when enterprise initiatives struggle.

AI adoption creates competitive pressure in a way that previous technology waves didn't. When your competitor deploys AI and reduces processing time by 40%, you can't afford to wait. The fear of falling behind is a more powerful driver than the hope of getting ahead.

The AI capability curve is still steep. Every quarter brings meaningfully better models, better tools, and lower costs. In previous AI winters, the technology stagnated. In 2024, it accelerated. It's hard to lose faith in a technology that's visibly improving every month.

The AI winter sceptics weren't wrong about the risks. They were wrong about the timing and the trajectory. The risks they identified - scaling failures, vendor problems, talent gaps, governance deficits - are all real. They just haven't been sufficient to overcome the momentum.

What could change this:

A major AI failure with public consequences. A high-profile AI system causing significant harm - financial loss, safety incident, privacy breach - could trigger regulatory reaction and enterprise retreat. This hasn't happened yet, but the governance deficit makes it possible.

Model capability plateauing. If model improvements slow significantly, the "it'll be better next quarter" narrative breaks down. The current improvement rate won't continue indefinitely. When it slows, expectations will need recalibrating.

Cost escalation. Enterprise AI costs at scale are significant and growing. If costs rise faster than value, the economic case weakens. API pricing, compute costs, and talent costs are all factors.

My assessment: none of these will trigger a winter. They may trigger a correction - a period of more sober investment and more rigorous evaluation. But the fundamental value of AI in enterprise is proven. The question isn't whether enterprises will use AI. It's how fast, how well, and how responsibly.

The winter that was predicted isn't coming. A more interesting season is: the autumn of realism. Not cold enough to stop growth. Cold enough to kill the weakest projects. And that's probably healthy.