AI Is Propping Up GDP Growth. It Isn't Propping Up Profits Yet.

AI infrastructure spending accounted for roughly 75% of all US GDP growth in the first quarter of 2026. Strip it out, and the economy was essentially flat. That figure gets cited constantly as evidence the AI boom is real and economically significant. A second figure gets cited far less: according to MIT's NANDA research, 95% of enterprise generative AI pilots show no measurable impact on profit and loss, despite $30 to $40 billion in enterprise spending on them.

Both numbers are accurate. They are also measuring two different things that keep getting conflated in coverage of whether AI spending is "worth it."

Two phases, one number problem

The top five Western hyperscalers, Amazon, Alphabet, Meta, Microsoft, and Oracle, are committing $725 to $805 billion to 2026 capital expenditure, up roughly 77% from 2025's actual spend of about $410 billion. Sequoia Capital's David Cahn calculates that this level of spending requires roughly $600 billion in annual AI revenue to justify itself financially, a gap he describes as widening rather than closing, with the 2026 shortfall approaching $1 trillion. Microsoft's own numbers illustrate the scale: roughly $13 billion in attributable AI revenue against $89 billion in AI capex, a 6.8x gap.

Goldman Sachs's Jim Covello, the bank's most prominent AI skeptic, has said that after two more years of data he's moved further from evidence of AI spending producing commensurate returns, not closer to it. That's notable coming from inside a bank that has otherwise been broadly constructive on AI investment.

The GDP contribution and the profit gap are not contradictory. They're sequential. Capex spending itself, the construction of data centers, chip fabrication, and power infrastructure, shows up in GDP the moment it happens, regardless of whether the finished product ever generates a return. That's the "canal and railroad" phase: infrastructure spending is GDP-positive by definition, whether or not the infrastructure ever gets used productively. MIT's 95% failure rate is measuring a different, later phase: whether companies actually deploying finished AI products are getting measurable value out of them. Most of that spending is going toward sales and marketing tools, even though MIT found back-office automation is where the actual measurable returns show up.

Put plainly, the current AI economy is GDP-positive as a construction boom, not yet as a technology-adoption boom. That distinction matters for anyone reading the GDP number as evidence the return question is already settled. It isn't. The infrastructure is being built well ahead of proof that the finished product pays for it, and the industry does not yet have a clear model for when, or whether, that payoff phase actually arrives.

Sources: Goldman Sachs · Sequoia Capital · Forbes · Fortune (MIT NANDA) · TECHi · Data Center Dynamics