Rates Near 5% Are the Real Test for the AI Buildout

The market spent last week worrying that politicians might slow down data center construction. The bigger threat to the AI buildout arrived on Tuesday morning in a government inflation report, and it has nothing to do with politics. It is the price of money.

For three years, the AI infrastructure trade ran on one assumption that almost nobody stated out loud: that capital would stay cheap enough, and patient enough, to fund hundreds of billions of dollars of data centers before they earn a return.

That assumption is now being tested.

On November 10, the October consumer price index came in hotter than expected. The 2-year Treasury yield jumped 17 basis points in a day, the 10-year closed at 4.98%, and futures markets moved to price a December rate hike as close to certain. The Federal Reserve, which raised rates in September for the first time since 2023, now looks likely to do it again.

Meanwhile, the AI buildout has quietly changed how it is paid for. In 2023, the hyperscalers funded data centers out of free cash flow. In 2026, Alphabet, Amazon, Meta, Oracle and CoreWeave are funding them with record bond issuance, new share sales, convertible notes and preferred stock.

What the market is fundamentally mispricing right now is that the AI trade has become a rates trade. The companies best positioned for the next five years may not be the ones with the most GPUs. They may be the ones that never need to ask the bond market for permission.

This note works through what changed, who is exposed, and what to watch before the Fed meets on December 8–9.

---

1. The Macro Shift: Inflation Is Not Done, and the Fed Is Hiking Again

What the October CPI report actually said

The Bureau of Labor Statistics reported that the consumer price index rose 0.4% in October on a seasonally adjusted basis, and 0.3% excluding food and energy (BLS, November 10, 2026). Economists surveyed by Dow Jones had expected 0.3% and 0.2% (CNBC).

The details matter more than the headline:

Source: BLS and CNBC, seasonally adjusted month-over-month changes.

Two observations.

First, core inflation rose 0.3% for the second month in a row, matching September. Compounded over a year, a 0.3% monthly pace is about 3.7%, well above the Fed's 2% target.

Second, look at electricity: up 0.9% in a single month. That is not a coincidence in a year when data centers are competing with households for power. The AI buildout is starting to show up in the inflation data the Fed responds to.

One important caveat. The BLS did not publish 12-month inflation rates for October, because it did not collect survey data in October 2025 during the government shutdown (BLS). Anyone quoting a precise year-over-year figure for October is estimating, not reporting.

The Fed has already turned

This is not a hypothetical tightening cycle. It has started.

• September 16, 2026: the Federal Open Market Committee raised its target range by a quarter point to 3.75%–4.00%, its first increase since July 2023, in a 12–0 vote under Chair Kevin Warsh. Sixteen of eighteen participants projected another hike this year (CNBC).
• October 28, 2026: the committee held rates, but two members, Lisa Cook and Beth Hammack, dissented in favor of another quarter-point increase. The statement said simply: "Inflation remains elevated" (Federal Reserve).

After the October CPI report, traders fully priced a quarter-point hike at the December 8–9 meeting, and the odds of a half-point move rose to about 14% from roughly zero, according to CME FedWatch data reported by CNBC. Governor Christopher Waller had said before the report that a 0.3% core reading could move him toward a half-point increase. Dallas Fed President Lorie Logan said she sees a need for two more quarter-point hikes (CNBC).

The bond market's reaction

Source: U.S. Treasury daily par yield curve. The bond market was closed on November 11 for Veterans Day.

The 30-year Treasury is already above 5%. The 10-year is two basis points away.

Here is where things get interesting. Long-term yields matter more for the AI buildout than the Fed's policy rate. Data centers are long-lived assets financed with long-dated debt. The relevant cost of capital for a 15-year data center is not the overnight rate. It is the 10- and 30-year yield plus the company's credit spread. Both are moving the wrong way.

---

2. Who Is Paying for the Buildout? Three Funding Models

The single most useful way to sort AI infrastructure companies right now is not by technology. It is by how they pay for their capital expenditure.

Model 1: Self-funded

Microsoft is the clearest example. In the September quarter it generated $58.0 billion of operating cash flow against $39.4 billion of additions to property and equipment, leaving roughly $18.6 billion of free cash flow in one quarter. It held about $64.5 billion of cash and short-term investments against roughly $35.5 billion of long-term debt including the current portion (Microsoft 10-Q).

Microsoft can build through a rate shock without asking anyone for money. Higher rates raise its discount rate like everyone else's, but they do not threaten its ability to execute.

Model 2: Strong balance sheet, increasingly using capital markets

Alphabet, Amazon and Meta remain enormously profitable, but their CapEx now exceeds what their operations generate in some periods, and all three are tapping outside capital.

• Alphabet: free cash flow was negative $9.1 billion in Q3 and $19.7 billion over the trailing twelve months. Long-term debt rose from $46.5 billion at year-end 2025 to $125.1 billion at September 30. In the first nine months of 2026 it also raised about $42.9 billion by issuing common stock and $19.1 billion through mandatory convertible preferred stock (Alphabet Q3 2026 results).
• Amazon: trailing-twelve-month free cash flow was negative $42.4 billion, with $173.1 billion of long-term debt against $114.2 billion of cash and marketable securities (Amazon 10-Q).
• Meta: Q3 capital expenditures, including finance-lease principal, were $34.91 billion against $32.79 billion of operating cash flow, for negative free cash flow of $2.12 billion. Long-term debt was $80.92 billion against $82.45 billion of cash and marketable securities. Meta also announced a new non-voting Class C common stock, effective January 2027, which it said "will provide Meta the ability to finance its growth" (Meta Q3 2026 results).

These are not distressed companies. But they are no longer immune to the cost of capital. Alphabet is already issuing equity to fund infrastructure, and Meta has created a new share class so it can. That is a meaningful shift for shareholders.

Model 3: Dependent on capital markets

Oracle and CoreWeave are where rates bite hardest.

• Oracle (fiscal Q1 2027, quarter ended August 31): revenue rose 30% to $19.3 billion, cloud infrastructure revenue rose 121% to $7.4 billion, and remaining performance obligations reached $664 billion. But capital expenditures were $28.5 billion in the quarter, free cash flow was negative $5.4 billion, and trailing-four-quarter free cash flow was roughly negative $28.7 billion by our calculation. Notes payable and borrowings totaled about $125.3 billion, and Oracle sold $20 billion of common stock through an at-the-market program in the quarter (Oracle Q1 FY27 results).
• CoreWeave (Q3 2026): revenue rose to $3.66 billion from $1.37 billion a year earlier, and backlog reached about $132.5 billion. But GAAP operating income was $207 million while net interest expense was $846 million, producing a net loss of $709 million. Debt on the balance sheet rose to roughly $46.9 billion at September 30 from about $35.1 billion at June 30. In the quarter it also raised about $4.2 billion of convertible notes and launched an at-the-market equity program (CoreWeave Q3 2026 results).

Debt is long-term debt (Microsoft including the current portion) or total borrowings as each company reports it; definitions differ and figures are not perfectly comparable. A dash means we did not pull a comparable cash figure for this note. TTM figures for Oracle are our sums of the company's quarterly table. Sources linked above.

The real question investors should be asking is: if the 10-year settles above 5% and credit spreads keep widening, which of these companies has to slow down, and which can keep spending?

---

3. The Financial Reality: Rates Hit the Buildout in Three Places

Channel 1: Interest expense

CoreWeave is the cleanest illustration. Its Q3 net interest expense of $846 million was more than four times its GAAP operating income of $207 million. Annualizing that quarter's interest and dividing by its average debt across the quarter implies a borrowing cost of roughly 8%. That is our rough calculation; reported interest includes items such as amortized financing costs.

Now consider a purely illustrative scenario. If CoreWeave's entire ~$46.9 billion of debt cost one percentage point more, annual interest would rise by roughly $470 million. That compares with full-year 2026 adjusted operating income guidance of $1.275–1.375 billion. In practice, rates reset only as debt matures or floating-rate facilities adjust, so the hit would arrive gradually. But the direction is unambiguous: every rate hike moves the breakeven for a debt-funded GPU fleet further out.

For Oracle, the same one-point illustration on about $125 billion of borrowings would mean roughly $1.25 billion a year of extra interest over time. That is manageable for a company with Oracle's revenue, but it lands at the same moment its free cash flow is deeply negative.

Channel 2: Credit spreads

The hyperscalers have become one of the largest groups of borrowers in the US corporate bond market. According to LSEG's FTSE Russell, hyperscaler investment-grade issuance reached $218.1 billion in 2026 by early October, more than double the $107.3 billion for all of 2025. Hyperscalers now make up 4.7% of the FTSE US Broad Investment-Grade corporate index, up from 2.7% two years earlier (LSEG, October 29, 2026).

Investors are demanding more to absorb that supply:

• The group's option-adjusted spread has widened about 31 basis points since December 2025, while the broader index widened about 1 basis point.
• Hyperscaler bonds, which used to trade tighter than the index, now trade wider than it.
• Oracle's spreads are about 100 basis points wider than its peers'.

So the cost of capital is rising twice: through the Treasury yield underneath, and through a wider spread on top.

Channel 3: Dilution

The least appreciated cost is landing on shareholders. If you look closely at the 2026 financing activity, the buildout is increasingly being paid for with equity.

Alphabet issued about $42.9 billion of common stock in nine months plus $19.1 billion of mandatory convertible preferred, which converts into shares. Oracle sold $20 billion of stock in one quarter. CoreWeave launched an at-the-market program. Meta created a new class of non-voting stock explicitly to finance growth.

When a company funds CapEx by selling shares, existing owners pay for the infrastructure through a smaller claim on future earnings. That matters most when share prices are under pressure, because each dollar raised costs more shares.

---

4. The Contrarian Opportunity: The Market Still Trades "AI" as One Thing

Last week's selloff treated AI infrastructure as a single bet: everything with data center exposure fell together on fears of political limits on construction. Rates are going to force a more discriminating market.

Insight 1: Balance-sheet strength is becoming a competitive advantage, not just a safety feature

In a cheap-money world, a company that funds itself internally has no edge over one that borrows. In a 5% world, it does. Microsoft can keep building at a pace a debt-funded competitor cannot match without paying more for every dollar. I would much rather own the company that never needs the bond market than the one that needs it every quarter.

Insight 2: The equity risk premium is thin

FactSet puts the S&P 500's forward 12-month P/E at 17.7 as of November 6, below its 5-year average of 19.8 and 10-year average of 19.1 (FactSet). That sounds reasonable, until you invert it.

A 17.7 P/E is a forward earnings yield of about 5.65%. The 10-year Treasury closed at 4.98% on November 10 (comparing that with the November 6 P/E is approximate, since both move daily). The extra return investors are being offered for owning stocks instead of risk-free government bonds is about two-thirds of a percentage point. Meanwhile, three-month Treasury bills yield 4.30%.

That cushion is being supported by exceptional earnings: FactSet reports blended Q3 earnings growth of 31.4%, with 84% of reporting companies beating estimates. The market is pricing in continued earnings strength, but the underlying economics suggest that any disappointment now has less valuation support beneath it.

Insight 3: Long-duration growth stocks are more rate-sensitive than they look

A simple, purely illustrative model shows why. Value a stream of earnings with a discount rate minus a growth rate:

Hypothetical numbers for illustration only, using value = 1 ÷ (discount rate − growth rate). Not a forecast for any company.

The same half-point move does nearly twice as much damage to the high-growth company. That is the arithmetic behind why AI's most expensive, most growth-dependent names tend to fall hardest when yields jump.

Insight 4: Power prices cut both ways

Electricity CPI rose 0.9% in October. For power generators that sell into rising markets, that is revenue. For regulated utilities, it is a political problem, and higher Treasury yields also squeeze the gap between their allowed returns and their cost of capital. I would not lump "AI power plays" together: a generator with existing assets and pricing power is in a very different position from a utility that must borrow heavily to build new capacity at today's rates.

Where expectations may be too high

The most vulnerable combination right now is a company with negative free cash flow, rising debt, equity issuance and valuation that depends on years of future growth all at once. CoreWeave currently fits all four. Strong revenue growth and record backlogs do not change the fact that those businesses need open, affordable capital markets for the next several years.

---

5. Valuation: What a Higher Discount Rate Does

For this note, the more important valuation question is not whether any one stock looks cheap on forward earnings. It is how much each business depends on the gap between its return on new capital and its cost of capital.

Our qualitative assessment based on the disclosures cited above.

I would treat forward P/E comparisons across these groups with caution. A 25x multiple on a self-funded company and a 25x multiple on a company that must keep raising capital are not the same price, because the second one's future earnings carry financing risk the multiple does not show.

---

6. The Five-Year Outlook: Three Rate Scenarios

Bull case: inflation proves temporary

Assumptions: the October reading reflects one-off energy and food pressure plus shutdown-related distortions. Core inflation cools in early 2027, the Fed stops after December, and the 10-year settles back into the mid-4% range.

Implications: credit spreads stabilize, the dilution question fades as markets reopen on good terms, and investors refocus on backlog conversion. Capital-market-dependent names would likely rebound hardest from a lower base.

Most likely to benefit: CoreWeave and Oracle, followed by the hyperscalers.

Base case: higher for longer

Assumptions: one or two more quarter-point hikes, a 10-year around 5%, and hyperscaler spreads modestly wider. The buildout continues, but the cost of financing it is permanently higher than in 2023–2025.

Implications: the largest platforms keep spending; weaker borrowers slow down, consolidate or pay up. More equity issuance across the sector. Returns increasingly separate by balance-sheet strength rather than by AI exposure.

Most likely to benefit: Microsoft, and, among the rest, whichever companies return to positive free cash flow first.

Bear case: a funding squeeze

Assumptions: inflation keeps re-accelerating, the Fed hikes faster (Waller's half-point scenario), the 10-year moves well above 5%, and bond investors balk at another $200 billion-plus year of hyperscaler supply.

Implications: capital-market-dependent operators face refinancing stress and deeper dilution; some projects are delayed or cancelled; the hyperscalers trim 2027 CapEx growth. Ironically, lower CapEx would improve free cash flow for the strongest platforms, even as it hurts suppliers.

Most vulnerable: CoreWeave and other neoclouds, then Oracle, then the equipment suppliers most exposed to a pause in orders.

We are not assigning price targets to these scenarios. The point is that the dispersion between funding models widens in every one of them except the bull case.

---

7. Positioning: How I Would Think About It

This is a framework for thinking about rate exposure within an AI infrastructure view, not a recommendation for anyone's situation and not a set of position sizes.

How the framework changes the approach

When the risk-free rate is 4.3% on three-month Treasury bills, the hurdle for owning any risky asset is higher than it has been for most of the past fifteen years. In this environment, I would:

• Prioritize funding resilience over growth rates within the AI infrastructure theme.
• Build exposure gradually rather than all at once, especially ahead of the December 8–9 Fed meeting and the November CPI report on December 10, either of which can move yields sharply.
• Re-evaluate after each earnings cycle, focusing on free cash flow and financing activity rather than backlog headlines.

---

8. What Could Prove This Thesis Wrong?

1. The inflation scare fades. October's report has no 12-month comparison, energy and food drove much of the surprise, and one month is not a trend. If core inflation cools quickly, rates could fall and this whole concern recedes.

2. Earnings outrun rates. Q3 S&P 500 earnings grew 31.4% on a blended basis. If AI-driven earnings growth stays that strong, it can absorb a higher discount rate.

3. The Fed stops sooner than markets expect. Not every official is hawkish. Some participants signaled they want more data before supporting further hikes. A pause after December would ease pressure on long yields.

4. AI-driven productivity is disinflationary. If AI meaningfully lowers costs across the economy over the next several years, it could eventually push inflation and rates down, the opposite of the current dynamic.

5. CapEx discipline arrives early. If the hyperscalers slow spending growth in 2027, their free cash flow would improve quickly, reducing their need for outside capital and dilution.

6. Bond investors keep absorbing supply. Demand for high-quality hyperscaler debt has so far been deep. If it stays that way, spreads may stabilize even with heavy issuance.

The metrics to monitor

1. The 10-year Treasury yield relative to 5%, and the 30-year above it.
2. The December 8–9 FOMC decision and the new projections for 2027.
3. The November CPI report on December 10.
4. Hyperscaler credit spreads versus the broader investment-grade index.
5. Financing activity in each quarterly report: new debt, at-the-market share sales, convertibles and preferreds.
6. CoreWeave's interest coverage: operating income relative to interest expense.
7. Oracle's free cash flow in its fiscal Q2 report.
8. Nebius's Q3 results on November 12, for a second read on neocloud financing.

---

The Investment Verdict

The AI infrastructure story is not broken. Demand, backlogs and revenue growth are all extraordinary. But the story has entered a phase where how a company pays for growth matters as much as how fast it grows.

My highest conviction is in the businesses that can fund their buildout internally. The most asymmetric risk is in the businesses whose models depend on borrowing at rates that are rising, from investors who are already demanding more, while issuing shares to fill the gap. And the thing the market is underestimating is that the bond market, not Washington, may end up deciding the pace of the AI buildout.

The next five years could reveal that the decisive competitive advantage in AI infrastructure was never the chips or even the power. It was the cost of capital.

In a world of 5% Treasury yields, the most valuable AI asset may be a balance sheet that never has to ask for money.

Data: BLS CPI release and U.S. Treasury yield data (November 10, 2026); Federal Reserve and CNBC on FOMC decisions; FactSet Earnings Insight (November 6, 2026); LSEG FTSE Russell (October 29, 2026); company earnings releases and filings as linked. Derived figures are our calculations and are labeled as such.