Built Like a Railway, Priced Like Software

Original Lens and publication context

The second half of the AI trade belongs to the cost of capital. The build must be funded before the benefits arrive — and the market has begun pricing the difference.

AI is still framed as a software cycle, but the physical build underway is an infrastructure boom closer to the railway era — and infrastructure booms raise the cost of capital while software cycles lower it.
Hyperscale capex funding is leaking off private balance sheets into expanding credit issuance and, now, the equity channel — a private-sector supply leg competing with fiscal borrowing for the same waning duration bid.
Soft June payrolls (three-month average down to 111k from 188k) and a mildly dovish Sintra performance from the Fed chair compress July hike risk; any tightening is optional, shallow, and capped at two to three moves.
Equity leadership is rotating out of the AI capex suppliers — semiconductors and memory — and into the hyperscale spenders and rate-sensitive broadening trades, the fourth such rotation of this AI cycle.
The macro desk caps long-end rallies while the equity desk expects front-end hike pricing to unwind — both are right at their own tenor, and the reconciliation is a steeper curve.

I. Trigger

Three developments converged in the first week of July, and together they flipped the framing of the AI trade from where the year began. Micro is once again driving macro — but the question has shifted from how large the productivity dividend will be to how the build that precedes it gets funded.

The first trigger was the labour market. June payrolls took the air out of the July hike balloon. The three-month average of nonfarm payrolls has slipped to 111k from 188k, the acceleration story has faded, and the tick lower in the unemployment rate was nullified by a simultaneous drop in participation. Downward revisions and a negative June print for hospitality leave health services as the only genuine job creator. The softness is corroborated across the panel: vacancies are soft, the labour-market differential is ticking lower, hiring intentions are muted, and — critically — there is no pickup in wages.

EXHIBIT 1 — WHO IS ACTUALLY HIRING. THE RED LINE IS HEALTHCARE — THE ONLY SECTOR CONSISTENTLY ADDING JOBS. THE BLUE AND ORANGE LINES ARE THE REST OF THE PRIVATE ECONOMY, HOVERING AROUND ZERO. IN PLAIN WORDS: STRIP OUT HOSPITALS AND CARE HOMES, AND AMERICA IS BARELY CREATING JOBS. THAT IS WHY A JULY RATE HIKE SUDDENLY LOOKS UNNECESSARY.
Supplied archive exhibit · original labels. Read alongside the editorial corrections.

EXHIBIT 2 — THE UNDERLYING TREND IN JOB GROWTH. THIS SMOOTHS OUT THE MONTHLY NOISE. TREND JOB CREATION COLLAPSED TO NEAR ZERO IN MID-2025, RECOVERED THROUGH 2026, BUT SITS AT A MODEST 75–120K A MONTH — AN ECONOMY THAT IS GROWING, NOT OVERHEATING. NOTHING HERE FORCES THE FED'S HAND.
Supplied archive exhibit · original labels. Read alongside the editorial corrections.

The second trigger was Sintra. Chair Warsh gave away little — a Greenspan-era communication style is clearly the objective — but the balance of his remarks tilted dovish. Inflation risks were characterised as having come down, and his emphasis on the employment side of the dual mandate was more pronounced than at the June FOMC. His AI commentary, notably, focused on the destination: the supply-side impact that could eventually lower neutral rates. It is precisely that framing this report challenges.

The third trigger came from the market's own internals. The high-beta momentum basket — read: memory and chip names — suffered its largest two-day drawdown since the pandemic, ending a historic run in semiconductors that had persisted since the end of March. The catalyst structure matters more than the drawdown itself: one of the largest hyperscale platforms announced it would begin selling excess compute capacity to outside customers, the first public admission from inside the spending complex that capex trajectories can bend.

EXHIBIT 3 — THE MOMENTUM BASKET BREAKS. THE 'MOMENTUM BASKET' BUNDLES THE MARKET'S HOTTEST STOCKS — THIS CYCLE, CHIPS AND MEMORY. IT DROPPED 17.3% IN TWO DAYS, THE SHARPEST BREAK SINCE THE 2020 PANDEMIC CRASH (DASHED ARROW). WHEN THE MOST CROWDED TRADE BREAKS THIS HARD, LEADERSHIP USUALLY CHANGES.
Supplied archive exhibit · original labels. Read alongside the editorial corrections.

The first hyperscaler to sell its spare compute told the market what discounted cash flow models could not: the capex assumptions have a ceiling.

II. Flow

The dominant framing of AI as a productivity story has supported the rates outlook at points this year: disinflation eventually, cheaper compute, better margins, possibly fewer jobs, a lower neutral rate down the road. That framing is likely right about the destination. It places far too little weight on the journey.

What is underway is not a software cycle. It is a capex cycle — one closer in physical footprint to a railway build than to the dotcom expansion that keeps getting drawn as the comparison. The distinction is not rhetorical; it is the entire rates story. The software cycle lowered equilibrium rates because it needed very little capital to scale. Infrastructure booms raise them, because they compete for scarce capital — steel, land, power, grid connections, skilled construction labour, cooling, and above all, funding.

EXHIBIT 4 — THE BUILD KEEPS GETTING BIGGER. LEFT PANEL: TECHNOLOGY INVESTMENT AS A SHARE OF US GDP IS AT AN ALL-TIME HIGH — ABOVE EVEN THE DOTCOM PEAK. RIGHT PANEL: CONSENSUS HYPERSCALER CAPEX ESTIMATES FOR 2027 HAVE JUMPED FROM $614BN TO $942BN IN JUST EIGHT MONTHS. EVERY UPGRADE TO THE BUILD IS AN UPGRADE TO THE FUNDING IT NEEDS.
Supplied archive exhibit · original labels. Read alongside the editorial corrections.

Hyperscale capex assumptions have moved in only one direction, and the funding is now visibly leaking out of the private balance sheet. Credit issuance tied to the build is large and still expanding, and the equity channel is now being tapped alongside it. The implication for duration is direct: the fiscal supply story that has weighed on DM long ends all year now has a second, private-sector leg competing for the same waning duration bid.

EXHIBIT 5 — HOW THE BUILD IS BEING PAID FOR. BOND ISSUANCE BY THE FIVE BIGGEST CLOUD COMPANIES IS NEAR $170BN ALREADY IN 2026 — ROUGHLY TEN TIMES A NORMAL YEAR — AND NOW SPANS DOLLARS, EUROS, STERLING, AND YEN. IN PLAIN WORDS: THE AI BUILD HAS OUTGROWN INTERNAL CASH FLOW AND MOVED ONTO THE BOND MARKET, WHERE IT COMPETES WITH GOVERNMENTS FOR THE SAME INVESTORS.
Supplied archive exhibit · original labels. Read alongside the editorial corrections.

The equity market is running its own referendum on the same question. The rotation out of the AI capex spenders' suppliers and into the spenders themselves is the cleanest signal available that the equity side has begun pricing return on invested capital rather than narrative. The divergence between hyperscale platforms and the semiconductor complex — the latter wholly dependent on the former — had become unsustainable; these groups ebb and flow in opposite directions near the end of such divergences, and reconciliation arrives when the spenders tone down capex guidance or signal a change in direction. The excess-compute announcement was exactly that signal.

EXHIBIT 6 — SUPPLIERS CANNOT OUTRUN THEIR CUSTOMERS FOREVER. THROUGH Q2, THE CHIP SUPPLIERS (YELLOW) KEPT RALLYING WHILE THEIR BIGGEST CUSTOMERS, THE HYPERSCALERS (BLUE), FLATLINED. THE GAP IS UNNATURAL: THE SUPPLIERS' REVENUE IS THE CUSTOMERS' SPENDING. ONE OF THE TWO LINES HAD TO GIVE — AND IT IS NOW RESOLVING.
Supplied archive exhibit · original labels. Read alongside the editorial corrections.

EXHIBIT 7 — CHECK WRITERS VERSUS CHECK RECEIVERS. THE 'CHECK WRITERS' (HYPERSCALERS, RED — THE ONES PAYING FOR CHIPS) FELL ALL SPRING WHILE THE 'CHECK RECEIVERS' (SEMIS, GREEN — THE ONES COLLECTING) SOARED. A MARKET WHERE THE PAYER FALLS AND THE COLLECTOR RALLIES IS ASSUMING SPENDING NEVER SLOWS. THE EXCESS-COMPUTE ANNOUNCEMENT BROKE THAT ASSUMPTION.
Supplied archive exhibit · original labels. Read alongside the editorial corrections.

Beneath the AI complex, a broader rotation is compounding the flow. The economy entered a new expansion after the rolling recession completed in April 2025, and new expansions deliver classic operating leverage: revenue growth returning to cost-efficient companies, with earnings surprising to the upside in the laggard cohorts. That broadening began in November 2025, was interrupted by the Iran war in late February — oil surged, and pricing for cuts flipped to hikes — and has resumed since crude peaked in May. The steep fall in oil has stabilised rates, which is itself a driver of the rotation into lagging areas: consumer discretionary goods, transports, regional banks, and biotech. Structurally, this favours the equal-weighted index over the cap-weighted index.

EXHIBIT 8 — THE AVERAGE STOCK VERSUS THE GIANTS. BOTH PANELS COMPARE THE EQUAL-WEIGHTED S&P 500 (EVERY COMPANY COUNTS THE SAME) WITH THE CAP-WEIGHTED INDEX (THE GIANTS DOMINATE). A RISING LINE MEANS THE AVERAGE STOCK IS WINNING. THE BROADENING BEGAN IN LATE 2025, PAUSED WHEN THE IRAN WAR SPIKED OIL, AND HAS RESUMED SINCE OIL PEAKED IN MAY.
Supplied archive exhibit · original labels. Read alongside the editorial corrections.

EXHIBIT 9 — LEADERSHIP CHANGE, VISIBLE IN PRICES. CONSUMER DISCRETIONARY, TRANSPORTS, AND REGIONAL BANKS HAVE ALL OUTPERFORMED THE INDEX SINCE MID-MAY — PRECISELY WHEN OIL PEAKED AND RATE FEARS EASED. THIS IS WHAT A LEADERSHIP CHANGE LOOKS LIKE WHILE IT IS HAPPENING, NOT AFTER.
Supplied archive exhibit · original labels. Read alongside the editorial corrections.

EXHIBIT 10 — ANALYSTS UPGRADING, NOT JUST PRICES RISING. EARNINGS REVISION BREADTH FOR CONSUMER DISCRETIONARY GOODS (LEFT) AND TRANSPORTS (RIGHT) HAS SWUNG POSITIVE — MORE ANALYSTS RAISING PROFIT FORECASTS THAN CUTTING THEM. PRICE MOVES BACKED BY RISING EARNINGS ESTIMATES TEND TO LAST; MOVES WITHOUT THEM TEND TO FADE.
Supplied archive exhibit · original labels. Read alongside the editorial corrections.

Figure 01 · Conceptual framework

Why an AI buildout reaches the yield curve

Reading the figureThe source’s cost-of-capital mechanism. It is a hypothesis about transmission, not a measured forecast of interest rates.

III. Risk

The physical bottleneck. If the stated capex numbers are realised, the build collides with power availability, grid capacity, skilled construction labour, and cooling. Each is a potential headwind to inflation normalisation. A capex boom that tightens real-resource markets while the central bank is trying to re-anchor inflation credibility is an uncomfortable combination for duration.

The unresolved verdict. The burden of proof now sits with whether the productivity gain emerges, and how long it takes. Until it does, more weight belongs on the known — more competition for capital, private and public, credit and equity — than on the promised destination.

The policy risk. A hike remains possible. It would burnish the Fed's inflation-fighting credibility at seemingly minimal cost. But it is optional rather than necessary — the energy move is working in the right direction, sequential inflation is expected lower from here, helped by PCE methodology changes — and there is a deeper question of whether hikes are even the right tool for a capex-driven cycle. The rate-sensitive sectors hikes actually reach, housing above all, are already visibly struggling. If the Fed moves, a sustained cycle beyond two to three hikes is hard to construct.

The commodity analogy. The memory subsector is the most commodity-like corner of the semiconductor complex, and its parabolic advance this year tracks the precious-metals-equity analog with uncomfortable fidelity: parabolic move, commodity underlier, wide price swings. That analog suggests the correction in semiconductors has further to run, led by memory. Peak rate of change in revisions breadth — not peak cycle — is the operative diagnosis; this is the fourth such correction since the AI cycle began in November 2022, and the prior three resolved as resets, not endings.

EXHIBIT 11 — THE SILVER TEMPLATE. THE STRANGEST CHART OF THE YEAR, AND ONE OF THE MOST USEFUL. THE SEMICONDUCTOR INDEX (BLUE) HAS TRACKED SILVER-MINING STOCKS (YELLOW) WITH ROUGHLY A FOUR-MONTH LAG — BOTH ARE COMMODITY BUSINESSES PRONE TO PARABOLIC RUNS AND HARD RESETS. SILVER STOCKS HAVE ALREADY ROLLED OVER. IF THE ANALOG HOLDS, CHIPS FOLLOW — LED BY MEMORY.
Supplied archive exhibit · original labels. Read alongside the editorial corrections.

EXHIBIT 12 — INSIDE THE MACHINE: THE MEMORY SQUEEZE. THE BILL OF MATERIALS FOR THE NEXT-GENERATION AI RACK: TOTAL COST NEARLY DOUBLES TO ABOUT $7.8M PER RACK — AND THE MEMORY LINE ALONE RISES 435% (BOXED). MEMORY IS WHERE COST INFLATION INSIDE AI HARDWARE IS CONCENTRATED, WHICH IS WHY MEMORY STOCKS WENT PARABOLIC — AND WHY THEY FALL HARDEST WHEN CAPEX EXPECTATIONS BEND.
Supplied archive exhibit · original labels. Read alongside the editorial corrections.

EXHIBIT 13 — MEMORY BECOMES A QUARTER OF THE MACHINE. THE SAME STORY AS SHARES OF THE WHOLE: MEMORY GROWS FROM 7% OF RACK COST TO 26% (CIRCLED) ACROSS THREE HARDWARE GENERATIONS. A COMPONENT THAT BECOMES A QUARTER OF THE MACHINE BEHAVES LIKE A COMMODITY SQUEEZE — SPECTACULAR ON THE WAY UP, BRUTAL ON THE WAY DOWN.
Supplied archive exhibit · original labels. Read alongside the editorial corrections.

IV. Price

The bond market still carries roughly one and a half hikes through the first quarter of 2027. Against a backdrop of falling energy prices, peaking tariff-related inflation, and contained services and housing inflation — with core CPI likely to hold below three percent — that pricing looks adequately hawkish at the front end, and arguably too hawkish. A Fed that stays on hold is a positive surprise for equities via lower policy-rate expectations, and the recent outperformance of rate-sensitive cohorts may be the early signal, just as energy-equity underperformance front-ran the fall in crude.

EXHIBIT 14 — WHAT THE BOND MARKET EXPECTS, MEETING BY MEETING. 'BPS PRICED' IS THE TIGHTENING BUILT INTO MARKET PRICES BY EACH FED MEETING DATE: ABOUT 38BP — ROUGHLY ONE AND A HALF 25BP HIKES — BY MARCH 2027, AND NOTHING AFTER. IN PLAIN WORDS: MARKETS EXPECT AT MOST ONE OR TWO HIKES, THEN DONE. THAT IS THE BAR THE DATA HAS TO BEAT.
Supplied archive exhibit · original labels. Read alongside the editorial corrections.

EXHIBIT 15 — THE JULY HIKE BALLOON DEFLATES. PRICED 2026 FED TIGHTENING IN MOTION: EXPECTATIONS DROPPED SHARPLY ON THE WEAK JUNE PAYROLLS AND HAVE SETTLED NEAR 30BP (CIRCLED) — BARELY MORE THAN ONE HIKE. TWO WEEKS AGO THE DEBATE WAS JULY; NOW IT IS WHETHER THE FED MOVES AT ALL.
Supplied archive exhibit · original labels. Read alongside the editorial corrections.

Further out the curve, the cost-of-capital argument caps the upside. Rallies in the belly and long end of DM curves now run into two supply legs — fiscal and private — bidding for the same shrinking pool of duration demand. Rates markets need to price the build before they price the benefits, because there is materially more certainty today on the former than the latter.

Within equities, the pricing asymmetry favours the spenders over their suppliers. Hyperscale platforms have already absorbed their period of underperformance as capex expectations ramped against them; the high capex-to-sales factor is topping after a torrid year, and any tempering of capex guidance from here transfers relief to the platforms and pressure to the chip complex. Biotech carries its own rate kicker: in elevated-and-falling rate regimes the group has historically delivered strong returns, and a ramping M&A cycle compounds the case.

EXHIBIT 16 — HEAVY SPENDERS VERSUS LIGHT SPENDERS. THIS FACTOR BUYS THE COMPANIES WITH THE HIGHEST CAPITAL SPENDING RELATIVE TO SALES AND SELLS THE LOWEST. ITS ENORMOUS RUN THROUGH 2026 IS NOW PAUSING. IN PLAIN WORDS: THE MARKET HAS STOPPED REWARDING SPENDING FOR ITS OWN SAKE.
Supplied archive exhibit · original labels. Read alongside the editorial corrections.

EXHIBIT 17 — THE ROIC QUESTION, LIVE IN PRICES. THE SAME HIGH-CAPEX GROUP MEASURED AGAINST THE ENTIRE TOP 1,000 STOCKS: IT HAS STARTED TO UNDERPERFORM. INVESTORS HAVE BEGUN ASKING WHAT THE RETURN ON ALL THIS SPENDING ACTUALLY IS — AND ASKING IT WITH THEIR MONEY.
Supplied archive exhibit · original labels. Read alongside the editorial corrections.

EXHIBIT 18 — WHERE RATE RELIEF LANDS HARDEST. BIOTECH'S AVERAGE QUARTERLY PERFORMANCE IN FOUR RATE REGIMES. ITS BEST ENVIRONMENT BY FAR IS HIGH-AND-FALLING RATES: +4.4% PER QUARTER ON AVERAGE (BOXED). THAT IS PRECISELY TODAY'S SETUP IF THE FED STAYS ON HOLD WHILE INFLATION COOLS — WITH A RAMPING M&A CYCLE AS A SECOND ENGINE.
Supplied archive exhibit · original labels. Read alongside the editorial corrections.

Figure 02 · Qualitative comparison

Two parts of the curve, two questions

Two parts of the curve, two questions
Segment	Main question in the report	Relevant evidence
Front end	What will the central bank do next?	Inflation, employment and policy decisions
Belly and long end	What return attracts duration capital?	Issuance, real resources and investor demand
Equities	Who turns spending into earnings?	Cash flows, margins and capital discipline

Reading the figureThis separates the source’s policy and financing mechanisms without assuming either dominates every market regime.

V. Position

All positioning is expressed at sector and factor level, consistent with Buy Side Lens policy.

SLEEVE	SECTOR-LEVEL POSITION	HORIZON	DRIVER
Rates — DM curves	Fade rallies in the belly and long end; the front end is adequately priced — hold front-end exposure, resist duration extension.	H2 2026	Cost of capital / dual supply legs
Equities — AI complex	Rotate semiconductor and memory exposure toward hyperscale platform exposure; treat the semis correction as ongoing, led by memory.	3–6 months	Revisions-breadth peak / capex guidance
Equities — broadening	Overweight consumer discretionary goods, transports, biotech, and regional banks; prefer equal-weight over cap-weight index exposure.	H2 2026	New-expansion operating leverage
Energy	Crude has peaked; avoid chasing energy equity strength — its earlier underperformance correctly led the commodity lower.	H2 2026	Oil supply / rate stabilisation
Watchlist	Inflation prints into September; hyperscale capex guidance revisions; power, grid, and construction-labour bottleneck data.	Rolling	Decides the hike question

VI. Two Desks, One Question

This report consolidates two sell-side views published in the same week — the Goldman Sachs macro desk's H2 roadmap and Morgan Stanley's equity strategy note. They approach the same underlying question from opposite ends of the capital structure. Companion desk notes covering each in full detail accompany this report. The comparison below is the map.

	THE MACRO DESK (GOLDMAN SACHS)	THE EQUITY DESK (MORGAN STANLEY)
The question	Who funds the build — and what does that do to interest rates?	Who leads the equity market while the build is being questioned?
Core call	The AI capex cycle raises the cost of capital during the build; DM duration rallies are capped.	The broadening resumes; rotate out of semis into hyperscalers and the lagging average stock.
The Fed	Could hike — optional, not necessary; any cycle stops at 2–3 moves; prefers fading duration rallies.	Stays on hold all year; rate pricing is too hawkish; a hold is a positive surprise for equities.
AI framing	A railway-scale infrastructure boom; price the build before the benefits.	A mid-cycle reset — the fourth correction since 2022, not the end of the capex cycle.
Expression	Rates: hold the front end, fade rallies in the belly and long end.	Equities: hyperscalers over semis; discretionary goods, transports, biotech, regional banks; equal-weight over cap-weight.
Key evidence	Record hyperscaler credit issuance and equity taps; physical bottlenecks in power, grid, labour, cooling.	Peak revisions breadth in semis; the silver analog; Meta's excess-compute sale; broadening in prices and earnings revisions.

Where they agree. Both desks treat the hike as shallow and optional, both treat the semiconductor correction as real rather than noise, and both place the capex-funding question at the centre of the second half. Neither calls the end of the AI cycle.

Where they collide — and why both can be right. The apparent contradiction is on rates: the macro desk caps duration rallies while the equity desk expects rate expectations to fall. Resolve it by tenor. The equity desk's argument lives in the front end — the priced hikes that never arrive get unwound, which is bullish two-year paper and rate-sensitive equities. The macro desk's argument lives in the long end — the dual supply of fiscal and private duration caps ten- and thirty-year rallies regardless of what the Fed skips. Both legs together produce one clean macro read: a steeper curve.

EXHIBIT 19 — THE ROTATION BEGINNING, CLOSE UP. LATE JUNE IN DETAIL: THE HYPERSCALERS (RED) BOUNCED OFF THEIR LOWS AT EXACTLY THE MOMENT THE SEMIS (GREEN) STUMBLED FROM THEIR HIGHS (BOXED). THE GAP THAT DEFINED THE SECOND QUARTER HAS STARTED TO CLOSE — FROM BOTH DIRECTIONS AT ONCE.
Supplied archive exhibit · original labels. Read alongside the editorial corrections.

VII. The Lens View

The destination may still be the software-cycle one — disinflation, better margins, a lower neutral rate. But the journey is a capex cycle with a railway-scale physical footprint, and it has to be funded. That competition for capital — private and public, credit and equity — is the story of the second half, and it is what caps how far belly and long-end rallies can extend. In the front end, payrolls have made the hike optional rather than necessary; the inflation prints will decide September. And where the two desks appear to disagree, the reconciliation is the trade: front end rallies as phantom hikes unwind, long end stays pinned by supply — the curve steepens.

We keep describing AI as a software cycle while building it like a railway. Price the build before the benefits — there is more certainty in the former than the latter.