Are We in an AI Bubble? Layoffs, ‘Wizard of Oz’ Startups, and the Cost of Scaling

AI bubble or necessary correction? Layoffs, ‘Wizard of Oz’ startups, and the cost of scaling

A widely shared Reddit post argues that we could be in a new AI bubble. It points to large layoffs at major firms, early-stage AI products quietly powered by humans, and a growing view that raw model scaling is hitting diminishing returns. For UK readers, it raises practical questions: should you slow AI spending, double down, or simply get more disciplined?

Here’s what the post claims, what it might mean for the UK, and how to assess AI bets with a cooler head.

What the Reddit post claims

The post links several trends and asks whether we’re watching an AI-fuelled labour-to-capital shift:

• Layoffs tied to AI spending: It cites cuts at UPS (48,000), TCS (12,000), Amazon (14,000 corporate roles), Verizon (13,000+), HP (up to 6,000 through 2028), and reported reductions at Apple. It frames these as funding AI infrastructure.
• “Wizard of Oz” startups: Founders of Fireflies.ai reportedly admitted their early “AI” was sometimes powered by humans. In design, this is known as Wizard-of-Oz prototyping, where humans simulate the product before the tech is ready.
• The scaling wall: The post attributes to Ilya Sutskever the view that the “age of scaling” is over. For a decade, bigger models plus more compute equalled better performance; that curve may be flattening.

“Money is flowing out of payroll and into data centers.”

Important note: these are claims in the Reddit post, not independently verified here. Correlation is not causation. Some layoffs may reflect broader restructuring, cyclicality, or post-pandemic over-hiring as much as AI investment.

Source: Reddit discussion

The labour-to-capital shift: what’s actually changing

The core point is hard to ignore: spend is moving from payroll to infrastructure. Training and serving frontier models require expensive GPUs, energy, data centre capacity, and specialist engineering.

For UK organisations, that manifests as higher cloud bills, sticker shock on GPU pricing and availability, and pressure to justify recurring inference costs. In parallel, regulators are turning up the heat on responsible deployment. The UK ICO’s guidance on AI and data protection prioritises transparency, fairness, and accountability – standards many “move fast” pilots do not meet.

Reference: ICO – AI and data protection

Is the “scaling wall” real – and does it matter?

Scaling laws describe how performance improves as we increase model parameters, data, and compute. They have been a reliable guide, but at growing cost. If returns from bigger models flatten, we should expect a shift towards smarter training (better data curation, architectures, and retrieval), domain-specialised models, and stronger product integration rather than just more GPUs.

For buyers, the implication is practical: prioritise solutions where outcomes are measurable and defensible today – not promises of magic tomorrow.

Background: Scaling Laws for Neural Language Models (Kaplan et al.)

The “Wizard of Oz” problem: how to spot it

Wizard-of-Oz prototyping is a legitimate method in UX and product discovery. It helps teams validate demand before the tech exists. The trouble starts when marketing blurs the line and customers buy “AI” that’s actually human labour behind a UI, without clear disclosure.

Risks include inconsistent quality, limited scalability, data protection issues if data is exposed to human contractors, and surprise unit costs. If you’re buying AI, demand transparency on the level of human-in-the-loop used today and expected over time.

Primer: Wizard-of-Oz prototyping (NN/g)

Implications for UK leaders and teams

For business leaders

• Focus on unit economics, not hype: Model the total cost of ownership, including inference, context window expansion, vector databases, observability, and red-teaming. Tie spend to hard KPIs.
• Start with narrow, high-ROI workflows: Document-heavy tasks (drafting, summarising, claims triage) and structured ops (ticket routing, QA checks) often return value fastest.
• Treat AI as a compliance topic: Run Data Protection Impact Assessments, document lawful bases, and ensure meaningful human oversight for high-stakes use cases.
• Reskill before you resize: Redeploy staff into data stewardship, prompt engineering, and AI operations. The best gains come from human-AI teams, not blanket cuts.

For developers and data teams

• Architect for cost and control: Combine smaller models with retrieval augmented generation (RAG – fetching relevant documents at query time) before jumping to frontier APIs.
• Measure and iterate: Track hallucinations, latency, and per-task costs. Align model choices to benchmarked needs, not vendor prestige.
• Automate where it fits the grain: Use AI as a copilot for spreadsheets, docs, and tickets. For example, you can connect ChatGPT to Google Sheets to standardise reporting without complex infrastructure.

Due diligence checklist for AI tools

• Human-in-the-loop: What percentage of outputs involve humans today? How will that change at scale?
• Data handling: Where is data stored and processed? Who can see it? Is data used for model training by default?
• Reliability: What are the uptime, latency, and quality guarantees? Do they offer task-level accuracy metrics and audit logs?
• Model provenance: Which models are used, and why? Can you bring your own model or swap vendors?
• Cost transparency: Clear pricing for inference, context window sizes, and add-ons like embeddings and vector search.
• Security and compliance: DPIA support, ISO 27001, SOC 2 where relevant, and mechanisms for subject access and deletion requests.
• Roadmap honesty: Distinguish shipping features from aspirations. Ask for references in your industry.

Could this be a bubble? Signals to watch

• Unit economics improving: Lower inference costs, better caching, smaller but capable models.
• Shift from demos to durable workflows: Case studies that show multi-quarter productivity gains, not just pilots.
• Vendor consolidation: Fewer, stronger platforms with clearer value – or a long tail of niche specialists built on open components.
• Compute realism: More attention on energy, datacentre siting, and model efficiency, not just parameter counts.
• Regulatory clarity: Practical guidance on high-risk AI uses and procurement standards that buyers can actually implement.

The bottom line

Yes, there’s froth. Some companies will overbuild infrastructure, overpromise capability, and underdeliver value. The cure isn’t to freeze – it’s to get disciplined. Anchor AI investment to specific workflows, auditable outcomes, and transparent operating costs. Demand clarity on human-in-the-loop. Treat compliance as a design constraint, not an afterthought.

AI is neither a magic wand nor a doomsday machine. For UK organisations, the winners will be the ones who combine cautious engineering, honest procurement, and patient change management – and still ship useful things this quarter.