CPI Didn’t Break the Dollar-Watt Loop. Now Comes the Grid.

July inflation delivered no shock: US CPI rose 0.1% in July and 3.4% from a year earlier, while core CPI rose 0.2% and 2.5%. That weakens the case for an immediate inflation-driven Fed hike, but it does not erase the financing pressure already visible in AI infrastructure. The next test arrives on August 17, when six US grid operators must answer FERC’s large-load show-cause orders.

The 30-second decision

The July CPI print did not confirm SIAIntel’s hotter-inflation branch. It confirmed the center case.

The Bureau of Labor Statistics reported exactly the four headline numbers around which the pre-data consensus had clustered: headline CPI at +0.1% month over month and +3.4% year over year, and core CPI at +0.2% month over month and +2.5% year over year. The release therefore did not produce the broad inflation reacceleration that would have made a September rate increase substantially easier to argue.

That matters because SIAIntel’s August 11 pre-CPI labor-market analysis made a deliberately conditional call: a shrinking labor force kept rate-hike risk alive, but a hotter and broader CPI print was needed to turn that risk into a stronger policy trigger.

Signal verdict: PARTIALLY CONFIRMED. The “weak payrolls do not automatically rule out tightening” mechanism remains intact. The hot-CPI trigger did not fire.

The more important new signal is elsewhere. AI infrastructure is beginning to face a double repricing:

1. Dollar repricing: lenders are charging more for contract-renewal, customer and capital-structure risk. 2. Watt repricing: regulators are moving toward making large loads bear more of the grid costs created to serve them.

The CPI did not break that loop. It simply failed to accelerate the first leg.

Signal Verification | What SIAIntel got right — and what it did not

A credible intelligence record has to preserve misses as carefully as hits.

On August 11, SIAIntel wrote that the July employment report did not itself justify a hike. Payrolls had fallen by 23,000, the labor force had contracted by 264,000 and unemployment had still declined to 4.1%. Our thesis was that labor-supply contraction made the weak payroll number less straightforward than the headline implied.

We then defined four CPI paths. The second path — CPI near expectations, leaving the September meeting open rather than settled — is the one that materialized.

What did not happen matters just as much. Core CPI did not print at 0.3% or above. There was no broad upside surprise. We therefore do not classify today as “Signal Confirmed” for a September hike, and we do not rewrite yesterday’s article to make it look more accurate after the fact.

That distinction is the point of SIAIntel’s prediction-feedback discipline: the value is not guessing one number. It is identifying a mechanism before the confirmation window and then recording what the evidence actually did.

What July CPI actually says

The BLS July CPI release shows a cooling headline with a still-positive service core.

Energy fell 1.5% month over month, with gasoline down 2.9%. That was the most important downward force on the headline. Food rose 0.1%. Shelter rose 0.1% and accounted for roughly two-thirds of the monthly all-items increase, according to BLS. Services excluding energy services rose 0.2% month over month and 3.0% year over year.

The detailed CPI table also shows electricity up 0.1% on the month and 4.2% from a year earlier.

This is not a clean “inflation is solved” print. But neither is it a new inflation shock. Reuters characterized the report as matching expectations and softening the case for a September rate increase.

For this article, that creates an important analytical boundary:

Fact: July CPI was benign relative to the hot-risk scenario. Inference: the macro discount-rate shock did not intensify today. Not established: that financing costs for AI infrastructure will now fall materially.

Those are different claims.

The Dollar leg is not the Fed alone

It is tempting to define the “Dollar” side of the Dollar-Watt Loop as the policy rate. That is too narrow.

For capital-intensive AI infrastructure, the dollar leg includes:

• the risk-free rate;
• corporate and project-credit spreads;
• lender protections and covenants;
• customer credit quality;
• the length of contracted revenue;
• the life of the debt;
• the life of the data-center lease;
• the residual value of GPUs and other infrastructure;
• the cost of refinancing when the original customer contract expires.

A neutral CPI print can reduce one macro risk without solving the other eight.

This is why CoreWeave matters. Its latest financing shows that the credit market can tighten inside an AI demand boom.

CoreWeave’s growth is real — and so is the interest burden

CoreWeave’s second-quarter results show why simplistic “AI boom versus AI bust” framing is inadequate.

Reuters reported that Q2 revenue reached about $2.58 billion, up 112% from a year earlier, while revenue backlog stood at $104.2 billion and the company raised its 2026 capital-spending plan to $35–39 billion. Demand is not the missing variable.

At the same time, the company remains intensely capital dependent. In the first quarter, CoreWeave’s official results showed $536 million of net interest expense on $2.078 billion of revenue. Q2 reporting indicates debt costs rose further.

The conclusion is not that CoreWeave is failing. The conclusion is that a company can have record AI demand, expanding backlog and rising equity value while the financing layer becomes more demanding.

That is precisely what a mature infrastructure market looks like: investors stop treating “AI” as a substitute for underwriting.

The credit-market clue is in DDTL 4.0, 5.0 and 5.5

CoreWeave’s financing sequence gives us a rare window into how lenders differentiate AI infrastructure risk.

On March 31, the company closed an investment-grade $8.5 billion DDTL 4.0 facility. Its floating-rate tranche was priced at SOFR + 2.25%, with a fixed-rate tranche around 5.9%, and the facility matures in March 2032.

On May 18, CoreWeave closed a publicly syndicated $3.1 billion DDTL 5.0 facility. It was meaningfully oversubscribed and tightened during syndication to SOFR + 4.50%. It supports infrastructure dedicated to two non-investment-grade customer contracts and matures in roughly 5.5 years.

Then came the $2.6 billion DDTL 5.5 financing. The Financial Times reported that initial demand was weak enough for terms to be sweetened: pricing moved to SOFR + 5.5 percentage points, the debt was offered around 96–97 cents on the dollar, and lender protections were tightened. Orders subsequently rose to nearly $9 billion.

This is not a clean time-series comparison. The facilities differ in collateral, customers, ratings and structure. DDTL 4.0 is not directly comparable with DDTL 5.5.

But the signal survives that caveat: the market is differentiating AI credit risk more aggressively.

The hidden risk is a clock mismatch

The most important detail in the DDTL 5.5 story is not the headline yield.

It is time.

The FT reported that the underlying customer contracts are generally three to five years, while CoreWeave’s data-center leases can extend up to 15 years. Fitch cited that duration mismatch as a credit risk. Lenders therefore have to ask a question that GPU scarcity once allowed the market to postpone:

What happens when the customer contract ends but the infrastructure obligation does not?

A lender is not merely underwriting today’s Anthropic, Jane Street or other contract. It is underwriting the probability that capacity can be renewed or re-contracted at attractive economics several years from now.

That is a fundamentally different asset than a short-lived “AI trade.”

The DDTL structure does not prove distress. It proves that renewal risk now has a price.

This is the Dollar-Watt Loop becoming more mature

SIAIntel introduced the Dollar-Watt Loop as a framework for understanding how power access becomes a financial asset.

The first version of the loop was:

Watt → data center → lease → debt → capital

Today’s evidence adds a second layer:

Contract duration → lender protection → refinancing risk → required return

In other words, the market is no longer only asking whether a megawatt can be energized. It is asking whether the cash flow attached to that megawatt survives long enough, at a high enough price, to service the capital stack.

That is the Dollar repricing.

Now comes the Watt repricing.

August 17 is the grid test

On June 18, the Federal Energy Regulatory Commission issued Section 206 show-cause orders to all six regional grid operators under its jurisdiction: PJM, MISO, SPP, CAISO, ISO New England and NYISO.

The operators were told to justify existing tariffs or propose reforms governing how data centers, manufacturing facilities and other large loads connect to the transmission system.

FERC’s commissioners were unusually explicit about the economic problem. Commissioner David Rosner said the orders require Cost Recovery Agreements designed to ensure that a large load pays its fair share of infrastructure built to serve it, even if the project does not come online as planned. The stated purpose is to prevent those costs from being shifted to residential customers.

The procedural deadline is now close. The six RTOs/ISOs must file their show-cause responses by August 17, 2026, absent an approved delay.

This is not yet a final national tariff. It is the point at which the six regional answers become visible.

That distinction matters. We are not forecasting that FERC will impose one uniform data-center charge on August 17. We are forecasting that the filings will reveal how much of the incremental grid-risk burden is moving toward the large load itself.

Virginia is the live laboratory

Virginia already shows what that transfer can look like.

The Virginia State Corporation Commission’s data-center initiatives created a separate GS-5 rate class for large loads. Among the safeguards:

• new large-load customers must take and pay for service for at least 14 years;
• large loads must pay at least 85% of transmission and distribution costs incurred to serve them each month, regardless of actual usage;
• customers without sufficient credit can face collateral requirements covering up to 60% of minimum charges;
• Dominion must submit alternative cost-allocation proposals intended to reduce cost shifting.

A separate Virginia debate is going further. On July 23, Governor Abigail Spanberger’s office highlighted a filing urging regulators to apply a “but-for” test: if a transmission project would not be needed but for a data center, the data center should bear the cost rather than the broader customer base.

Virginia is not a preview of the exact FERC outcome. State retail ratemaking and federal transmission jurisdiction are not the same thing.

But economically, the direction is unmistakable: the party creating the marginal infrastructure requirement is being asked to carry more of its cost and commitment risk.

The Double Repricing

Put the two legs together and a new AI infrastructure equation appears.

Dollar repricing

The project must clear a higher-quality underwriting test:

customer durability + contract length + credit spread + refinancing risk + residual asset value

Watt repricing

The project must clear a harder physical-cost test:

generation + transmission upgrade + interconnection + minimum payment + collateral + curtailment/reliability obligations

The AI campus that looked attractive when the model assumed cheap refinancing and broadly socialized grid upgrades can look very different when both costs are assigned more precisely.

This does not mean AI capex stops.

It means the hurdle rate becomes location- and contract-specific.

Why AI stocks can rise while project economics tighten

This is one of the most important apparent contradictions in the current market.

CoreWeave shares jumped after its strong Q2 growth numbers. That is rational if investors are pricing faster revenue growth, improving utilization or greater long-term demand.

At the same time, lenders can demand a higher spread on a specific financing. That is also rational if that facility has shorter customer contracts, weaker credit quality, more renewal risk or a less protective collateral package.

Equity and credit do not have to tell the same story on the same day.

For SIAIntel, that divergence is useful. It tells us the AI boom has entered a stage where demand is no longer enough to explain valuation. Capital structure matters.

The same is becoming true of power. “We have 500 MW planned” is not equivalent to “we have 500 MW financed, interconnected, allocated and economically protected.”

What would confirm the Watt leg on August 17?

SIAIntel will score the six filings against a pre-declared test.

CONFIRMED

The Watt leg is confirmed if multiple RTO/ISO responses materially increase large-load responsibility through mechanisms such as:

• direct cost-recovery agreements;
• stronger deposits, collateral or exit fees;
• minimum-payment or capacity-reservation obligations;
• direct assignment of incremental transmission upgrades;
• rules that protect existing ratepayers when a planned load is delayed, downsized or cancelled;
• explicit financial responsibility for co-located or electrically proximate arrangements.

PARTIALLY CONFIRMED

The signal is partial if the filings mainly improve study speed, transparency, forecasting or reliability rules but leave material cost-allocation risk broadly shared.

REJECTED

The Watt leg is rejected if the six responses preserve the economic status quo with no meaningful transfer of incremental infrastructure or cancellation risk toward large loads.

This is the advantage of a dated thesis: August 17 can test it.

Causal Chain | From CPI to compute economics

The mechanism now looks like this:

CPI near consensus → no new inflation shock → Fed uncertainty remains, but the hot-data trigger does not fire → AI infrastructure credit continues to be priced on its own risks → customer-contract duration and refinancing terms matter more → FERC/Virginia push grid cost causation toward large loads → project WACC and all-in delivered-megawatt cost become more differentiated → marginal AI campuses reprice by location, customer quality and contract structure.

The important insight is that the final six arrows can continue even if the first arrow is benign.

That is why today’s CPI did not break the Dollar-Watt Loop.

Household, company and investor impact

Households: The grid-policy fight is ultimately about who pays for infrastructure. If regulators successfully isolate data-center-driven costs, residential customers are better protected from stranded or underused network investments. If they do not, rapid load growth can continue to migrate into bills.

Companies: AI developers and data-center operators need to model more than power price per megawatt-hour. Long-term minimum charges, collateral, interconnection obligations and transmission cost recovery can alter project economics before the first GPU is installed.

Banks and private-credit investors: Customer concentration and contract duration become central. A five-year revenue contract supporting a much longer infrastructure commitment is not automatically unsafe, but it demands a credible re-contracting and residual-value case.

Equity investors: Revenue growth and backlog remain important, but they should be separated from financing quality. A company can win demand and still experience a higher marginal cost of debt.

Utilities: Stronger financial commitments from large loads can reduce stranded-cost risk, but they do not eliminate the physical challenge of building enough generation and transmission quickly.

Country Lens | Why this matters beyond the United States

The US is becoming a live experiment in converting data-center power demand into explicit credit and tariff obligations.

Europe, the Gulf and Asia face different regulatory systems, but the same physical equation applies: large AI campuses need land, firm power, transmission capacity, financing and customers whose contracts last long enough to support the infrastructure.

For Türkiye, the relevant lesson is not to copy Virginia’s GS-5 tariff. It is to understand that bankable megawatts require a complete allocation of risk: who funds the substation, who guarantees minimum demand, who bears cancellation risk, who owns the grid upgrade and who pays if the campus arrives late.

Countries that answer those questions clearly may attract AI infrastructure even when their nominal electricity price is not the lowest.

Counter-Thesis | Why the squeeze may be overstated

There are three strong counterarguments.

First, AI demand may remain so strong that customer contracts renew at equal or better economics. If compute scarcity persists, today’s maturity mismatch can prove manageable.

Second, deeper institutional participation can ultimately lower funding costs. CoreWeave itself argues that HPC-backed financing is maturing into a scalable asset class, and successful syndications expand the investor base.

Third, better large-load tariffs can accelerate projects rather than merely make them more expensive. If a developer accepts explicit cost responsibility in exchange for faster, more predictable grid access, the reduction in delay risk can offset part of the higher direct cost.

The Double Repricing thesis is therefore not “AI infrastructure is unfinanceable.” It is:

AI infrastructure is becoming financeable only after more of its real risk is explicitly priced.

Break-This-Thesis | What would make us wrong?

The thesis weakens materially if several of the following occur together:

• new AI/HPC financings return to materially tighter spreads without stronger lender protection;
• customer-contract terms lengthen enough to match debt and infrastructure obligations;
• GPU residual values and secondary-market demand prove unusually durable;
• FERC’s August filings leave large-load cost responsibility largely unchanged;
• Virginia-style minimum charges and collateral structures fail to spread to other jurisdictions;
• grid buildout accelerates enough that transmission scarcity ceases to be a binding project constraint.

Conversely, the thesis strengthens if lenders demand more contract-specific protection while regulators simultaneously assign more upgrade and cancellation risk directly to large loads.

5-Day / 30-Day / 90-Day Watchlist

Next five days

• August 13: US PPI, a second inflation-cost check.
• August 17: PJM, MISO, SPP, CAISO, ISO-NE and NYISO show-cause responses.
• Compare each filing against the pre-declared CONFIRMED / PARTIAL / REJECTED test.

Next 30 days

• September 4: August US employment report.
• September 11: August CPI.
• September 15–16: FOMC meeting.
• Watch whether new AI infrastructure loans price tighter or require additional covenants.

Next 90 days

• Track large-load tariff filings and FERC orders following the responses.
• Track CoreWeave and peers for new debt, lease-backed structures and customer-contract duration.
• Track whether transmission and interconnection costs become explicit line items in AI-campus announcements.

Analyst Intelligence Box

Signal class: AI infrastructure double repricing Macro verdict: July CPI center case confirmed; hot-CPI trigger not confirmed Prior SIAIntel signal: Dollar-Watt Loop — power access becoming a financial asset Dollar evidence: contract-specific spread, covenant and renewal-risk pricing in AI infrastructure debt Watt evidence: FERC cost-recovery framework + Virginia large-load safeguards Next hard test: August 17, 2026 show-cause responses Current verdict: DOLLAR = ACTIVE / WATT = OPEN TEST Confidence: High that risk allocation is becoming more explicit; medium on the magnitude of the final tariff cost shift

Source Boundary | What is fact and what is SIAIntel inference?

Verified facts in this file come from BLS, FERC, Virginia SCC, CoreWeave disclosures and clearly attributed financial reporting.

SIAIntel inference: the simultaneous repricing of financing risk and grid-cost responsibility constitutes a “Double Repricing” phase of the Dollar-Watt Loop.

Not yet verified: the exact August 17 tariff proposals, the ultimate FERC-approved replacement rates, future CoreWeave contract renewals, or whether financing spreads will remain elevated.

The article does not predict default, a collapse in AI demand or a guaranteed September Fed move.

SIAIntel Bottom Line

The most important result of today’s CPI release is what it did not do.

It did not produce the broad inflation shock required to turn yesterday’s labor-force paradox into a clean September-hike call. That part of the thesis remains unconfirmed.

But the AI infrastructure story did not disappear with a benign CPI print.

Credit investors are already distinguishing between customer quality, contract duration and refinancing risk. Regulators are simultaneously asking who should pay for the transmission and interconnection required by massive new loads.

The first phase of the AI infrastructure boom priced megawatt scarcity. The next phase is pricing the real cost of financing and connecting those megawatts.

On August 17, the Watt side gets its first six-market federal test.

---

Editorial Note

This SIAIntel Premium file is a living intelligence analysis. It records prior claims before new evidence arrives, separates verified facts from inference, and will be updated only when a declared trigger is observed.

This article is for intelligence and educational purposes only. It is not investment, legal or individualized financial advice.