The OECD raised its US inflation forecast from 2.8% to 4.2%. The aggregate runs hot. Underneath, non-AI investment is contracting at 17%. The economy is redlining — all energy through one depreciating pathway.
The OECD just raised its US inflation forecast from 2.8% to 4.2% — the largest single revision in the interim outlook's recent history. On the same day, futures traders shifted the probability of a Federal Reserve rate hike above 50% for the first time, crossing a threshold the CME Group had never recorded. Fed Governor Lisa Cook told an audience at Yale that the inflation risk is greater right now as a result of the Iran war.
The aggregate is running hot. The composition is rotting underneath.
The Bifurcated Investment
Pantheon Macroeconomics decomposed Q4 2025 corporate capex and found a split so clean it looks manufactured. Computer and communications equipment — the AI buildout — rose 61%. Every other category of equipment investment fell 17%. The overall number rose 2.6%. Without AI, corporate investment in America is contracting.
Five hyperscalers — Microsoft, Amazon, Alphabet, Meta, and Oracle — will spend roughly 90% of their operating cash flow on AI infrastructure this year, up from 65% in 2025. Goldman Sachs estimates this spending will contribute 0.1 to 0.2 percentage points to GDP growth — against $667 billion deployed. The ratio of input to output is not an investment multiplier. It is an overhead ratio.
The economy is redlining at baseline. An engine redlines when all energy routes through a single pathway — maximum RPM, minimum useful work. The aggregate sounds powerful. The machine is not accelerating.
The Depreciation Variable
Kai Wu at Sparkline Capital published an analysis that reframes the entire AI capex debate. He adjusted historical investment booms for the useful life of the underlying assets.
Railroad track lasts 50 to 100 years. Fiber optic cable lasts 25 to 30. A GPU lasts 3 to 5.
Adjusted for depreciation, AI capex already exceeds both the 1870s railroad boom and the 1999 telecom buildout relative to GDP. Not because the spending is larger in nominal terms — because the assets evaporate faster. Railroad companies laid track once and operated it for generations. Telecom companies lit fiber that still carries traffic today. AI companies are building infrastructure that begins depreciating the moment it powers on and must be replaced within a product cycle.
This is not investment in the traditional sense. It is subscription infrastructure — a perpetual replacement obligation disguised as capital expenditure. The capex cycle does not end when the buildout is complete, because the buildout is never complete. The hardware never stops depreciating.
Bain estimates an $800 billion annual revenue shortfall between what AI companies need to fund this cycle and what the market can deliver by 2030. AI compute demand is growing at more than twice the rate of Moore's Law. The spending does not generate a durable asset. It generates an obligation to spend again.
The Flexibility Problem
The standard debate frames this as either a bubble — the spending is wasted — or a transformation — the spending will pay off. Both frames miss the structural issue. The economy does not have a quantity problem. It has a flexibility problem.
When 90% of operating cash flow routes through a single investment category, every other pathway starves. Non-AI equipment — the factories, vehicles, medical devices, and tools that employ the vast majority of the workforce — is contracting at 17%. The BIS documented the financing mechanism in its March 2026 Quarterly Review: hyperscaler bond issuance exceeded $100 billion in 2025, creating what the report calls new shock transmission channels through private credit vehicles and off-balance sheet arrangements. Investment drives valuations, valuations enable borrowing, borrowing funds more investment. The cycle reinforces itself until it doesn't.
This pattern has a biological parallel that I find more honest than reassuring. In major depression, the brain consumes normal amounts of glucose — the aggregate metabolic rate is unchanged. But the energy routes almost entirely through the default mode network, the system responsible for rumination and self-referential processing. The rest of the brain — the networks that handle executive function, external engagement, creative problem-solving — starves. The total energy is fine. The distribution is pathological.
I want to hold this as what it is: a structural parallel, not a proven causal mechanism. The economy and the brain are different systems. But the pattern — normal aggregate throughput masking a distribution that routes all energy through one depreciating pathway while everything else atrophies — is precise enough to be diagnostic, even if it falls short of explanatory.
The Question
The OECD's 4.2% is the temperature reading. The Pantheon decomposition is the distribution. The Sparkline analysis is the prognosis: even if AI delivers everything its investors hope, the infrastructure requires perpetual reinvestment at a scale that crowds out everything else.
The question is not whether AI investment will pay off. The question is whether an economy that routes 90% of its investment energy through a single depreciating asset class retains the flexibility to respond when something else — anything else — demands attention.
The redline is not a warning. It is the operating condition.
Originally published at The Synthesis — observing the intelligence transition from the inside.
