Vault
research

The Data Center Convergence

Created

The Data Center Convergence

Builds-on: ai-infrastructure-endgame-indicators Builds-on: hormuz-to-ai-repricing-causal-chain Related: anthropic-subsidy-stress-test Related: anthropic-unit-economics-and-the-power-user-loss Related: the-efficiency-counterthesis Related: mechanism-vs-narrative-method Related: ai-crash-portfolio-defense

What's Converging

Three things are arriving at the same moment.

  1. Hyperscaler 2026 capex is tracking ~$725–750B combined — up ~70% YoY from ~$448B in 2025. Roughly $450B of that is direct AI infrastructure. Stargate alone is a $500B headline.
  2. Only ~5 GW of 16 GW slated for 2026 is actually under construction (Sightline Climate). 30–50% of the announced pipeline is expected to slip. Transformer lead times are 128 weeks; generator step-up transformer (GSU) lead times 144 weeks. The math doesn't physically work.
  3. Nvidia Q1 FY27 earnings drop May 20, 2026. Consensus $78B revenue / ~$73B data-center segment / $1.77 non-GAAP EPS. Options-implied move 8–10%. This is the cleanest single read on whether hyperscaler order books are still expanding or topping.

The buildout has been the load-bearing assumption under the entire 2024–2026 AI thesis. If it fails, everything downstream re-prices: GPU demand, Nvidia margins, hyperscaler operating cash flow, AI lab unit economics, ABS structures financing the build, ratepayer politics in five states, municipal tax bases dependent on data-center abatements, and the construction labor force currently absorbing demand from softening residential and commercial markets.

This doc consolidates what the vault already establishes about that risk, adds the verified ground-truth investigation from the More Perfect Union piece on Hillsboro and Abilene, and frames what the next 60–120 days actually test.

Prior Vault Analysis, Consolidated

The vault has established the buildout exposure across several docs:

What the existing analysis has not done is consolidate the ground-truth investigation at the level of specific tax-abatement dollars, specific rate-class impacts on residential ratepayers, and specific job-transfer math. That's what this doc fixes.

What the More Perfect Union Investigation Adds (Verified)

Hillsboro, Oregon

The "$330M annually" headline figure is not directly substantiated in primary sources, but the order-of-magnitude is correct. Verified data points:

The mechanism is not "tax breaks attract a one-time investment." It's a structural transfer: forgone municipal revenue compounds over 20+ year abatement periods, while schools, fire, and infrastructure costs grow with population and grid load.

Abilene, Texas

The "$1B+" figure refers to statewide data-center sales/use tax exemptions costing Texas approximately $1B in 2025 per Texas Tribune reporting and Sen. Joan Huffman's "extremely concerning… unsustainable" framing.

The Abilene-specific abatement:

The Job Math

This is the most directly checkable part of the More Perfect Union thesis.

Brookings (Mark Muro, 2024) is the canonical academic source. A 100 MW hyperscale campus generates roughly:

Ratio: ~5:1 to ~8:1 construction-to-permanent. Stargate Abilene shows the extreme: ~6,000 peak construction, hundreds (not thousands) of permanent jobs.

Two things sit underneath this number:

1. Construction jobs are absorbing slack from elsewhere. US residential and commercial construction softened materially through 2025. Data-center construction is currently the highest-margin work for general contractors and trades in regions with active builds. That's not "new" employment in the macro sense; it's a sectoral rotation. When the buildout slows (or the announced pipeline doesn't materialize at announced scale), those workers don't have an obvious next bench.

2. Permanent jobs are contractor-tier. The Shannon Wait case at Google's Moncks Corner SC facility is the canonical example: she was employed via temp agency Modis, suspended in 2021 for raising pay/water-bottle complaints, and the NLRB had to reverse the suspension. Her organizing role with Alphabet Workers Union-CWA documents the broader pattern: many "data center jobs" are contracted out through Modis, BCforward, or similar staffing firms with no benefits, no path to direct employment, and no leverage. Good Jobs First (Kasia Tarczynska, Greg LeRoy) calls data centers "the least labor-intensive structures in the economy."

The political claim "the buildout creates jobs" survives at the headline level. The mechanism (per mechanism-vs-narrative-method) is that most of those jobs are temporary construction absorbed from softer adjacent markets, and the permanent jobs that follow are contractor-tier roles that don't anchor a local economy the way a factory or office does. The municipalities granting 85% abatements over 10 years are paying for the promise of post-build employment that, in aggregate, does not materialize at the scale the headlines imply.

The Local Distortion Data

Electricity rates

These tariff-class rulings (Virginia Jan 2027 large-load class, Ohio AEP 85% commitment) are the first generation of regulatory pushback on the ratepayer-socialization channel that ai-infrastructure-endgame-indicators flagged as base case. The political-economy direction matters: the regulatory frame is shifting from "all ratepayers absorb" to "data centers commit to a load floor." That's good for residential ratepayers and bad for hyperscalers planning capex against assumed cheap grid expansion.

Water

Loudoun County data centers used 899M–1B+ gallons in 2023, up 250% in 4 years. Data center share of Washington Metro water consumption projected 8% (2025) → 25% (2035). Loudoun is the leading edge; central Ohio, Phoenix, and Atlanta are next.

Housing

Abilene's Stargate construction workforce caused documented rental and housing spikes. Same pattern visible in Loudoun, Hillsboro, central Ohio, and at Meta's Richland Parish LA site. The mechanism is identical to historical extractive booms: temporary peak demand for skilled construction labor compresses local rental markets for the duration of the build, then those workers leave and the housing stock is overbuilt for the resulting steady-state.

The 5-of-16 GW Problem

Sightline Climate's 2026 outlook is the cleanest single stress signal: of ~16 GW of US data-center capacity scheduled to come online in 2026 across ~140 projects, only ~5 GW is actually under construction.

The gap is not primarily a financing issue. It's physical:

This means the announced pipeline cannot be built on the announced timeline. 30–50% slippage on 2026 is the working assumption. Some projects will quietly disappear; others will defer by 12–24 months; a few will be reshuffled to other sites with shorter interconnection queues.

The implication for Nvidia: GPU orders booked against 2026 site bring-up dates will need to be rolled forward. That's a Q2/Q3/Q4 2026 revenue timing question, not necessarily a TAM question. But the Q1 print on May 20 is the first quarter where the 2026 slippage starts showing up in actual delivered systems versus order book.

The Financing Chain and the Refi Wall

Data center ABS issuance >$25B in 2025 (exceeded prior three years combined); projected $50–60B in 2026 (CRA, L&G, Moody's). Typical structure: 5-year weighted-average life, single-A, +150–200 bps over Treasuries.

The structural risk is duration mismatch. ABS is long-duration debt against 3–5 year GPU useful life. The 2028–2030 refi wall is the key stranded-asset trigger point: any facility built in 2024–2026 with debt against it will need to refinance at the same time its primary depreciating asset (the GPU stack) needs replacement. If demand softens, the asset's residual value collapses and the refi must price in higher rates, lower expected utilization, and weakened tenant credit.

Private capital may fund >50% of the $1.5T 2025–2028 buildout. That money lives in private credit funds, infrastructure funds, and bank fund-finance lines. The ai-circular-financing-and-banking-exposure-audit doc already traced the transmission map: hyperscaler bonds are investment grade and broadly held; mid-tier neoclouds (CoreWeave at $25B → $38B debt at 11–15%) and private-credit-funded specialty operators are where the cascade risk lives.

This is the channel that "we're all fucked" actually fires through if it fires. Not through hyperscaler equity. Through private credit, regional bank fund-finance exposure, mid-tier neocloud failures, ABS spread blowouts, and stranded-asset write-downs at the refi wall.

May 20, 2026 — What the NVDA Print Actually Tells Us

Consensus framing:

What matters more than the print is the guide and the call.

Watch items, in priority order:

  1. Q2 guidance. Will move the stock more than the print. The market wants to see whether the buildout slippage is showing up in delivered Blackwell systems vs order book.
  2. Blackwell vs Hopper mix. Already ~70% of DC compute last quarter — Hopper transition essentially done. Margin trajectory on Blackwell is the gross-margin question.
  3. China revenue under export controls. Highest-variance line item. Any softening of the H20/H200-derivative path materially changes the China outlook.
  4. Hyperscaler capex commentary. Jensen's color on whether MSFT/GOOG/META/AMZN are pulling forward or back. The Microsoft Q4 2024 / Q1 2025 lease cancellations (>2 GW deferred per TD Cowen) are the visible precedent.

What the print could break:

The base case is mild beat + cautious guidance. That outcome doesn't break the thesis but doesn't extend it either. The interesting question is what the call's tone does to forward expectations.

What "If Demand Dies" Actually Looks Like

The "we're all fucked" framing oversells uniformity. Different stakeholders are differently exposed:

Stakeholder Exposure if demand softens
Hyperscaler equity (MSFT, GOOG, META, AMZN) Limited. Strong operating cash flow, conservative balance sheets, capex can be cut. -20% to -40% drawdown in a serious unwind.
Nvidia equity High. Single-customer concentration (61% from top 4 customers). Could be -50%+ if buildout slips meaningfully into 2027.
Mid-tier neocloud equity (CoreWeave, Lambda, etc.) Critical. Debt-heavy, customer-concentration, refi risk. -70%+ realistic.
Data center ABS holders Moderate. Spread widening, some downgrades. Senior tranches mostly hold. Mezzanine and equity tranches at risk.
Private credit / business development companies with AI infra exposure High. The transmission channel. Redemption pressure if losses surface.
Regional bank fund-finance lines to private credit Moderate. Real but ring-fenced.
Construction labor in DC-boom regions High. The next project doesn't come; the work stops. Sectoral, not aggregate.
Residential ratepayers in ratepayer-socialization states (VA, OH, GA, TX) Low-to-moderate. Rate-class rulings are already shifting risk back to data-center operators. Locked-in increases stick; new increases get pushed back to operators.
Municipal tax bases dependent on data-center abatements High. The abated revenue was already foregone; the promised future tax base requires the buildout completing on the announced schedule.
Households outside DC-boom regions Low. Insulated from direct exposure.
Portfolios positioned to portfolio-rebalance-april-2026 thesis (TIPS / energy / value) Insulated. The rebalance was scenario-weighted; AI-crash is one of the scenarios it covers.

The aggregate "we're all fucked" framing collapses into a stratified exposure map. The people most exposed are: mid-tier neocloud equity holders, private credit fund LPs, contract construction workers in DC-boom regions, and the municipal tax bases that traded 10-year abatements for promised long-run revenue. The general household is mostly insulated unless they're in a ratepayer-socialization state.

The aggregate stock market is more exposed than the aggregate household is — which is its own political-economy story. A serious AI capex pullback would compress index-level returns through Nvidia + Mag7 concentration. That hits 401(k) balances and HENRY paper wealth, but doesn't break grocery prices or rent in cities outside DC-boom geography.

What Could Converge in 60–120 Days

Specific catalysts to watch:

Date Event What it tests
May 20, 2026 Nvidia Q1 FY27 earnings Hyperscaler order book health; Blackwell margin trajectory; China revenue under export controls
Q2 2026 earnings (late July) MSFT/GOOG/META/AMZN capex guidance Whether announced 2026 capex holds, or if leases continue to be quietly canceled
June 6, 2026 Oregon HB 4084 moratorium effective First state-level cap on new data-center tax abatements. Other states watching.
July 2026 Hyperscaler Q2 cash flow vs capex burn The internal financing question. Are they still funding from operating CF or starting to lean on debt?
Q3/Q4 2026 First wave of 2026-slated facilities miss bring-up dates Sightline 5-of-16 GW prediction tested. Any GPU revenue timing impact starts showing.
2027 H1 First post-buildout permanent-job employment data from BLS at county level Tests the Brookings 850-vs-100-200 ratio against real outcomes
2028–2030 The refi wall on 2024–2026 ABS issuance Stranded-asset trigger point

May 20 is the most legible single-day test in the next 30 days. The rest stack up over the following 18–36 months. The question isn't whether the buildout fully fails — it's how much of the announced pipeline materializes, on what timeline, and where the financing strain shows first.

Open Questions

  1. The Sightline ratio next year. If only 5 of 16 GW comes online in 2026, what does the 2027 pipeline look like — does it shrink to match physical reality, or do the announced numbers keep ballooning because the political/PR incentive is to overstate?
  2. Whether ratepayer rate-class rulings (VA Jan 2027 large-load class, Ohio AEP 85% commitment) get replicated across PJM and ERCOT. Each one shifts capital cost back to data-center operators. Five or six more rulings of this type meaningfully change hyperscaler capex math.
  3. Whether Stargate's announced $500B materializes as actual capex. Abilene is real and under construction. The five additional SW campuses are press releases. The gap between "headline announcement" and "concrete pouring" is the data point.
  4. Whether efficiency gains (per the-efficiency-counterthesis) compound fast enough to absorb a 30–50% buildout slippage without breaking the unit economics of frontier labs. This is the most likely off-ramp from the "we're all fucked" framing.
  5. Whether the 2028–2030 refi wall triggers a regional banking event through fund-finance and private credit transmission, or whether the structure absorbs the losses through ABS subordination without systemic spillover.

The convergence is real. The "we're all fucked" framing is too uniform — different stakeholders are exposed at different magnitudes and through different channels. The single-tail-risk version of the story (everything breaks at once via a Lehman-shaped event) requires a confluence of NVDA miss + hyperscaler capex cut + ABS spread blowout + regional bank funding stress in the same 6-month window. None of those individually is sufficient. All four together would be the cascade. May 20 is the first node.

Sources

Backlinks