AI Spending Soars Past $1 Trillion, but Goldman Sachs Asks: Where’s the ROI?

Last week, Goldman Sachs dropped a bombshell: a note titled 'Gen AI: Too Much Spend, Too Little Benefit?' arguing that the $1 trillion expected to flow into AI infrastructure over the next few years lacks a clear payback. This comes as tech giants double down—Microsoft alone is on pace to invest $50 billion this year. Yet 48 hours later, Anthropic unveiled Claude 3.5 Sonnet, a model that beats GPT-4o on major reasoning and coding benchmarks while costing significantly less to run. The contrast is dizzying. The compute arms race is producing stunning lab results. Claude 3.5 Sonnet set a new state-of-the-art on HumanEval and other tests, while DeepSeek’s latest sparse model achieves GPT-4-class performance at 90% lower inference cost. But outside Silicon Valley, enterprise surveys paint a different picture: a McKinsey study from April found only 10% of companies have deployed generative AI at scale, and most struggle to measure value. Corporate deployments often boil down to chatbots and document summarization—hardly the productivity revolution promised. Why the gap? One theory: scaling laws for benchmarks aren’t translating into scaling laws for real-world workflow change. Another: enterprise AI is a last-mile problem, requiring data cleaning, culture change, and trust that GPT-5 releases won’t render current implementations obsolete. The Goldman report heated up simmering worries of an AI bubble, comparing the moment to the dot-com era when early infrastructure spending soared but transformative profits arrived years later. Yet the optimists point to stealth adoption: code copilots are already shifting developer productivity inside companies like Shopify and Airbnb. If Goldman’s skepticism proves premature, we might be in the quiet-before-the-storm of an ROI breakout. If not, 2024 could be remembered as the year AI hype fundraised faster than it could deliver. Is the industry measuring ROI with the wrong yardstick, or is the emperor wearing no clothes? What proof

I saw Google SGE pull my code snippets with zero clicks, dropping long-tail traffic 30%. That’s not just rebranded SEO—it’s fragment-based discovery. I adapted by adding `codeSnippet` structured data because models prioritize well-annotated blocks. Now pages aren't the unit, forcing rethinking of caching granularity, edge compute for dynamic snippets, and credibility signals for extraction. It’s deeper engineering work.