The AI ROI Reckoning: IBM, Galloway, Marcus Sound the Same Alarm
The week Anthropic filed its S-1, five independent reports landed the same verdict: enterprise AI spending has structurally outrun measurable return. The convergence is the signal — voices ranging from an NYU professor who called the dot-com crash to an IBM CEO who profits from continued spending all describe the same financial break. The AI cost audit has moved from VC whisper to mainstream earnings-call concern.
What the Source Actually Says
Galloway's Prof G Markets live show breakdown (The AI Corner, June 4) anchors the cluster. The thesis: an MIT professor found 95% of enterprise AI projects connect to no return a CFO can name. The data is specific — Uber exhausted its entire 2026 AI budget before Q1 ended; Stripe burns ~$100K per day on tokens; Salesforce is on track for $300M in Anthropic spend this year; one Anthropic employee ran a $150K Claude Code bill in a single month. Three structural failures drove the overrun: 2023-era compute modeling, underestimated inference volume, and unlimited employee usage with no budget ceiling. Token leaderboards compounded it — Meta and Amazon reward employees for consuming the most tokens, substituting activity for output.
Corroboration came from an unexpected direction. Gary Marcus amplified IBM CEO Arvind Krishna's capex math: the AI buildout requires $6–8T; recovering it over seven years needs $1–2T in new annual AI revenue. Krishna — an infrastructure vendor who benefits from continued spending — says that revenue "does not exist." Marcus added the billing layer: hyperscalers switched from flat-rate to usage billing not by strategy but because unlimited pricing was unsustainable ahead of their IPOs. "Tokenmaxxing is dead," he wrote. A Bain survey via Bloomberg supplied the bluntest summary: "The technology worked. The value didn't arrive."
Galloway's prediction: a 50–70% valuation correction within 24 months. His trigger is a non-tech Fortune 500 CEO announcing a quiet AI pullback on an earnings call, followed by NVIDIA's first earnings miss — NVIDIA has beaten estimates 15 consecutive quarters, making the streak itself the pressure. His labor-market calibration: sustaining current valuations through labor replacement requires 5–7 million layoffs in 2–3 years; actual 2026 AI-attributed layoffs are ~50,000.
Strategic Take
The CFO question — "what return can you name?" — is now mandatory in every AI engagement. When IBM's own CEO says the revenue math doesn't close and Galloway gives a 24-month correction window, the burden of proof has shifted to the AI buyer. Prepare the ROI answer before the meeting starts.
