Beyond the AI Hype and Doom

This image was created using an AI image creation program

Confused, you soon will be, in this week’s episode of ‘Hype cycle’.

The bull case: efficiency may save everything

The primary counter to the “AI hardware shortage forever” thesis is that AI itself gets more efficient faster than supply expands.

On 27 January 2025, DeepSeek released R1. A frontier-grade reasoning model trained for an alleged $5.6 million versus the roughly $100 million OpenAI reportedly spent training GPT-4.

Nvidia lost $588.8 billion in market value in a single day, the largest single-day loss for any stock in history. [1]

DeepSeek’s techniques are now being adopted broadly. Mixture-of-experts architectures that activate only a fraction of the model’s parameters per query.

Aggressive distillation of large models into smaller ones. Reduced floating-point precision. The trend toward smaller, more efficient models means that per-task GPU demand could fall dramatically even as the number of tasks grows.

There are two competing interpretations. The bullish view, often called Jevons paradox, says efficiency gains will spur far more usage than they save in compute.

Nvidia’s Jensen Huang has consistently argued this is the historical pattern with all computing. The bearish view says that if efficiency improves faster than hyperscaler spending commitments come due, much of the data-centre build will sit idle, like 1990s fibre.

The sceptics

Daron Acemoglu, MIT professor and 2024 Nobel laureate in Economics, is a well-qualified bear. His research estimates that generative AI will deliver no more than a 0.66% increase in total factor productivity over a decade, and possibly less than 0.53%. [2] He projects 1.1 to 1.6% cumulative GDP growth from AI over ten years, against industry forecasts of 7% from Goldman Sachs.

Acemoglu’s distinction matters for managers: he believes AI will affect roughly 5% of the economy (data summarisation, pattern-matching, mid-level clerical work) with average cost savings of around 14 to 15%. Multiply those and you get roughly 0.75% productivity uplift. He explicitly warns that the current trajectory is wasteful, with too much spending chasing labour-replacing automation rather than worker-augmenting tools. [3]

Gary Marcus is the most prominent technical sceptic. His 2025 predictions, that hallucinations would not be solved, that agents would be unreliable, that LLM economics would remain dubious, were substantially confirmed by year-end. [4]

His view is that current architectures are fundamentally incapable of robust reasoning, and that the LLM industry may end up a $30 to $40 billion industry that costs double or treble that to run.

Both Acemoglu and Marcus emphasise that scepticism is not pessimism about AI as a category. It is scepticism that current capital flows match likely returns over the relevant time horizon for shareholders.

What it means for smaller companies

The squeeze at the top of the stack punches down. Reporting from early 2026 indicates that Microsoft and other hyperscalers are diverting GPU stockpiles to internal teams and tier-one customers, leaving smaller AI startups to fight over scraps. Microsoft’s tiered access system requires non-tier-one customers to commit to renting at least 1,000 Blackwell GPUs for at least one year, contracts starting at tens of millions of dollars. [5]

OpenAI’s own CFO admitted in early 2026 that the company is making tough trade-offs on things it cannot pursue because it lacks sufficient compute. [6] If OpenAI is making tough trades, the typical pre-Series-B startup is facing existential ones.

The AI startup scene is bifurcating. A small number of companies with privileged compute access. A long tail with effectively none. This is the opposite of the 2010s cloud era, where AWS democratised infrastructure for small companies. It is more like 1880s railroads: whoever owns the right of way owns the future.

Three things to plan around

If you take only one strategic lesson from this series, it should be this: the rate-limiting step for your AI ambitions in 2026 to 2028 is unlikely to be the AI itself. It will be one of three things.

Compute access: If you do not have an existing hyperscaler relationship at scale, you will pay more, wait longer, and have less negotiating power than you assume. The hardware is spoken for. Plan your AI strategy around what you can get, not what you wish you could buy.

Organisational integration: As MIT and McKinsey both make clear, the binding constraint inside your firm is workflow redesign, data quality and operating-model change. Not the model. [7,8] Companies that win are ones that redesign work. Companies that lose are ones that bolt AI onto unreformed processes. The 5% of pilots that succeed are the ones where someone did the hard, unglamorous work of integrating AI into how people operate.

Realistic agentic expectations: Anything your vendor is selling as an “AI agent” today is almost certainly a tightly bounded copilot pretending to be more. [4] Plan accordingly. Pilot in narrow, well-instrumented domains. Assume failure rates that current AI thought-leaders publicly admit.

The position worth holding

The shovel sellers are doing extraordinarily well. Some of the gold miners will too, the ones who pick narrow, valuable seams and do the unglamorous work of integration. Most will not.

The most valuable competitive position in the AI economy right now is not to be the most aggressive buyer, the loudest deployer, or the boldest forecaster on the earnings call. It is to be the person who understands that the bottleneck has moved, and that the gap between what AI is doing in CEO scripts and what AI is doing in production will probably widen, not close, over the next two to three years.

That is not a bearish position, it’s a realistic one. In a market where 95% of pilots fail, 6% of companies capture all the value, and the world’s three largest memory firms are sold out through 2027, realism is the alpha.

Where does your organisation sit on the spectrum between AI rhetoric and AI reality? I’d be interested to hear from managers navigating the gap between what the board expects and what the technology can deliver.

References

[1] Hypotenuse AI, “What Is DeepSeek R1: How It Achieves High Performance with Less Computing Power.” https://www.hypotenuse.ai/blog/what-is-deepseek-r1-and-why-is-it-making-waves-in-ai

[2] NBER, Daron Acemoglu, “The Simple Macroeconomics of AI” (Working Paper 32487). https://www.nber.org/papers/w32487

[3] Project Syndicate, Daron Acemoglu, “Don’t Believe the AI Hype.” https://www.project-syndicate.org/commentary/ai-productivity-boom-forecasts-countered-by-theory-and-data-by-daron-acemoglu-2024-05

[4] Gary Marcus (Substack), “Six (or Seven) Predictions for AI 2026 from a Generative AI Realist.” https://garymarcus.substack.com/p/six-or-seven-predictions-for-ai-2026

[5] BigGo Finance, “Silicon Valley Sees New GPU Crunch: Microsoft and Cloud Giants Prioritize Internal Needs” (April 2026). https://finance.biggo.com/news/AkI9wp0B5edQG9E453EX

[6] Tomasz Tunguz, “The Beginning of Scarcity in AI.” https://tomtunguz.com/ai-compute-crisis-2026/

[7] Fortune, “MIT report: 95% of generative AI pilots at companies are failing” (August 2025). https://fortune.com/2025/08/18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo/

[8] McKinsey & Company, “The State of AI: Global Survey 2025.” https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai