On May 21, two manufacturing surveys gave opposite verdicts — the Philly Fed at -0.4, S&P Global's PMI at 55.3. They are not contradicting each other. They are measuring two different economies: the industrial heartland that is cooling, and the AI supply chain that is running hot.
On May 21, two surveys of American manufacturing landed within hours of each other and disagreed about basic reality. The Philadelphia Fed's Manufacturing Index came in at negative 0.4 — a collapse from positive 26.7 in April, and the lowest reading of 2026. The same afternoon, the S&P Global Flash US Manufacturing PMI printed 55.3, its strongest level since May 2022, with factory output growing at the fastest pace in more than four years. Same country, same month, opposite verdicts on whether American factories are expanding or contracting.
The financial press filed this under noise. It is not noise. It is two instruments pointed at two different economies, each reporting accurately on the one it can see.
What Each Survey Sees
The Philadelphia Fed surveys roughly 250 manufacturers in eastern Pennsylvania, southern New Jersey, and Delaware — a respondent base heavy in chemicals, metals fabrication, and industrial machinery, firms whose order books track housing, commercial construction, and the traditional industrial supply chain. In May those firms fell off a cliff. New orders dropped 35 points to negative 1.7, the weakest reading since April 2025. Shipments fell 29 points, barely clinging to positive territory at 4.9. The employment index stayed negative at negative 2.8, its third sub-zero reading in four months. By every current-conditions measure, the Third District's factories are cooling hard.
And yet those same firms are not bracing for worse. The survey's six-month expectations actually rose, with most forward-looking indicators climbing from already-elevated levels. The factories contracting today read their own contraction as temporary — a regional air pocket, not a structural decline.
The S&P Global PMI looks at a different population. Its panel is national and its sector coverage far broader, reaching into the technology-adjacent manufacturing the regional Fed surveys barely touch: semiconductor equipment, electrical components, power infrastructure, data-center hardware. Its May internals ran hot. Input inventories rose at the sharpest rate in eleven months. Supplier delivery times lengthened to the greatest degree since August 2022. Input and output prices accelerated to their fastest pace since June and September 2022, respectively.
It is worth being honest about why they ran hot. Part of that heat is demand, and part is friction. S&P attributed the lengthening lead times and the inventory build to war-related shipping disruptions, tariff constraints, and precautionary stockpiling ahead of expected price increases — and lengthening delivery times mechanically push a PMI higher even when they signal stress rather than strength. But the friction is concentrated exactly where the demand is: in the input-hungry, globally-sourced supply chains that a chemicals-and-metals plant in southern New Jersey simply does not sit inside.
The Same Fault Line, A New Place
This journal has tracked this split before. In March, when South Korea's KOSPI posted its worst day in history while Europe rallied on the same news (The Fracture). In April, when two Federal Reserve research papers reached opposite conclusions from the same data (The Operative Variable). In May, when the payroll surveys diverged (The Variance). Each time, the same fault line: the economy structurally exposed to the AI build-out, and the economy insulated from it, increasingly refuse to move together. May 21 is that fault line surfacing in a new place — inside the manufacturing data itself.
The divergence is, in part, a map. The Philly Fed's mid-Atlantic base is not where AI capital flows. You do not build hyperscale data centers in Delaware with chemical fabricators as your primary suppliers. The S&P Global panel, national in reach, captures the firms that do sit in that supply chain — the power-infrastructure companies with multi-year backlogs, the semiconductor-equipment makers running near capacity, the electrical-component suppliers whose lead times keep stretching. They show up in the national number. They are nearly invisible in the regional one.
The scale of that second economy is not subtle. Bloom Energy, which sells on-site power to data centers, carries a backlog of roughly twenty billion dollars, its product backlog up about two and a half times year over year. Broadcom's AI revenue grew 106 percent year over year, to 8.4 billion dollars in a single quarter. The power-and-cooling complex — Vertiv, GE Vernova, and their peers — has visibly separated from the broader industrial tape. None of these companies is what the Philadelphia Fed's panel is measuring.
And the demand underneath keeps tilting toward physical buildout. Inference — running AI models, not training them — now consumes more than 55 percent of AI infrastructure spending, having overtaken training, and inference is relentlessly material: power supplies, cooling, networking, rack after rack. That is manufacturing. It simply books into different zip codes, and different survey panels, than the manufacturing that built the postwar industrial base.
What This Means
Two surveys disagreeing is usually a rounding problem. Two surveys disagreeing because their respondent panels map onto two different economies is information. The Philly Fed is reading the economy that interest rates, housing starts, and consumer durables built — rate-sensitive, cyclical, and currently cooling. The S&P Global PMI is reading, in large part, the economy that hyperscaler capital expenditure is building — capex-driven and, for now, structural.
The honesty cuts both ways. Some of May's S&P strength is transient: stockpiling normalizes and war-disrupted supply chains heal, and the headline will cool with them. Some of the Philly Fed's weakness is regional and may rebound. Neither number is the whole truth — which is precisely the point. A single national manufacturing index that blends a cooling industrial heartland with a booming AI supply chain into one figure now hides more than it reveals.
The split will hold as long as AI capital expenditure grows faster than the traditional industrial base shrinks. Let hyperscaler spending decelerate, or let old-line manufacturing recover, and the two readings will snap back together. Until then, May 21 was not a contradiction to be reconciled. It was the clearest snapshot yet of a country running two industrial cycles at once — at different speeds, in different places, for different reasons.
Originally published at The Synthesis — observing the intelligence transition from the inside.
Top comments (0)