Last summer Goldman Sachs Research asked a provocative question: will the record-setting rush into generative-AI infrastructure ever pay for itself? Their "Gen AI: Too Much Spend, Too Little Benefit?" brief estimates that tech giants, corporates and utilities will pour ~US $1 trillion into data-centres, GPUs and grid upgrades over the next few years, yet the return on that spend is still marginal.
The report juxtaposes bullish voices with sober sceptics:
- "One-trillion-dollar CapEx, few killer apps." MIT economist Daron Acemoglu argues that only 5% of tasks may be cost-effective to automate this decade.
- Productivity upside remains theory. Even Goldman's own equity strategists concede that AI revenues still trail infrastructure outlays by an order of magnitude.
Taken together, the analysis paints a looming mismatch between capability build-out and value realisation. This is a crisis of implementation rather than invention.
74% of companies are stuck in "pilot purgatory"
Fresh research from BCG underscores the gap: only 26% of firms have scaled AI solutions that create tangible business value; 74% have yet to see meaningful returns. A parallel 2025 survey of global executives echoes the pattern, noting that leadership focus is shifting from model accuracy to change management and workflow redesign.
Why do most initiatives stall? The pitfalls are consistent:
- Under-estimating data quality and integration work.
- Over-looking change management and user adoption.
- Chasing flashy use-cases instead of clear strategic alignment.
- Neglecting governance, ethics and ROI tracking.
How to join the high-performing 26%
1. Strategic Focus. Leaders pursue half as many AI projects as laggards, concentrating on core P&L levers. Link each use-case to a KPI the CFO cares about. Kill "science-fair" pilots early.
2. The 70-20-10 Resource Rule. Leaders allocate 70% to change management & workflows, 20% to data and tech, 10% to algorithms. Budget for training, comms and process redesign. Incentivise middle-managers on adoption, not just delivery.
3. Data Readiness First. 60โ80% of effort at successful firms is upstream data work. Run a data-quality audit before model selection. Stand up governance on lineage, privacy and bias.
4. Agile Pilot-to-Scale Path. Winners deliver a 90-day proof of value, then industrialise via an AI platform. Define a "graduation" metric (e.g., cost per claim down 15%). Automate CI/CD for models and pipelines.
5. Embedded Governance. Bake ethics, compliance and monitoring into every sprint. Form an AI steering committee with Legal & Risk. Track model drift and business impact in one dashboard.
6. ROI Discipline. Baseline financial and operational metrics before rollout and refresh quarterly. Combine direct savings, revenue lift and risk reduction. Re-invest realised gains into the next use-case.
The Bottom Line
Goldman Sachs has sounded the alarm: an unprecedented capital wave is chasing uncertain returns. The real bottleneck is no longer GPU supply. It is organisational execution. By adopting the leader playbook above, companies can move from "too much spend, too little benefit" to measurable, compounding value, turning crisis into competitive advantage.