AI Spending Is Pushing Rates Up — And Those Rates Are Eating AI’s Returns
Japan's 10-year government bond yield touched 2.945% — the highest since September 1996, thirty years ago (Nikkei, 19 August 2026).
News like this usually gets filed under "bond markets."
It isn't a bond market story. It is a story about the discount rate written into your capex approval.
And this particular story has an awkward loop inside it.
One of the forces pushing yields up is the sheer scale of AI-related investment. And what erodes the economics of that AI investment is the higher yield it helped create.
In other words, the AI capex boom is raising its own hurdle rate. Let's take that loop apart in the language of corporate finance.
What is actually happening
The facts first. All figures below are as reported by Nikkei.
Japan: 10-year at 2.945% (18 Aug, a 30-year high); 30-year at 4.155%, approaching May's peak of 4.200%
United States: 30-year above 5.3% (17 Aug, a 19-year high); 10-year at 4.7%
Europe: France above 4% (a 17-year high); Germany at 3.2% (a 15-year high)
This is not a Japanese phenomenon. It is simultaneous and global. Nikkei groups the causes into three.
1. Geopolitics and inflation. The US–Iran negotiating deadline lapsed without extension, and crude has stayed on an upward track. A bond pays a fixed coupon, which makes it structurally vulnerable to inflation. So bonds get sold.
2. Fiscal concern. The US deficit reached $1,798.8bn in the ten months to July — already above the previous full fiscal year. With midterms in November, there is little appetite for tightening fiscal discipline. France faces the same pressure ahead of its 2027 presidential election.
3. Large-scale AI-related corporate issuance. Hyperscalers are funding data-centre build-out in the bond market. Alphabet decided this month to issue $25bn of dollar-denominated debt.
The third point is the interesting one.
Crowding out — but running through the tech sector
Textbook crowding out is a government issuing so much debt that it squeezes out private borrowers. Here the direction is reversed.
Tech credit no longer sits below sovereign credit. Highly rated, very large, and higher-yielding corporate paper arrives in size. Money that used to sit in government bonds moves across. Supply and demand for sovereigns loosens, prices fall, yields rise.
A portfolio manager at Columbia Threadneedle told Nikkei that the competition between US Treasuries and tech corporate bonds is especially intense at maturities beyond thirty years.
That detail matters. The competition is concentrated at the ultra-long end. Which means the assets most exposed are the ones funded longest and recovered slowest. Data centres are exactly that.
Separately, the Bank of Japan's "Summary of Opinions" from its July meeting reportedly featured repeated observations that global AI-related demand and expansionary fiscal policy could add to demand and push prices higher. AI has been discussed as deflationary on the supply side, through productivity. On the demand side — power, equipment, construction, semiconductors — it is inflationary. In the current phase, it is the second effect that is showing up in rates.
So the loop looks like this:
AI investment → enormous funding demand (bond issuance plus power and equipment inflation) → higher rates → higher WACC → a higher hurdle rate for AI investment
The boom is raising the bar it has to clear.
Now the capex-approval part
How does this reach a corporate decision? In issue 3 (NPV vs IRR) I wrote that the discount rate is, in practice, WACC — the hurdle rate. Let's open it up.
WACC = E/V × Re + D/V × Rd × (1 − tax rate)
Re = risk-free rate (Rf) + β × equity risk premium
Notice that Rf appears in both Re and Rd. When government yields rise, the cost of debt rises and the cost of equity rises. A rate move hits WACC twice.
Some numbers (all figures illustrative; they do not represent any real company).
Assumptions: D/V = 30%, E/V = 70%, β = 1.1, equity risk premium 5.5%, credit spread +1.0%, effective tax rate 30%
Low-rate period (Rf = 1.0%) Re 7.05% / Rd pre-tax 2.00% / Rd after tax 1.40% / WACC 5.4%
Now (Rf ≈ 2.9%) Re 8.95% / Rd pre-tax 3.90% / Rd after tax 2.73% / WACC 7.1%
A 1.9-point move in the risk-free rate lifts WACC by roughly 1.7 points. In relative terms, the hurdle rate is about 30% higher.
Plenty of companies still run a fixed internal hurdle rate — 8%, 10% — set years ago and never revisited. Through a decade of low rates, that was conservative. It may not be any more.
Which projects fail first — a duration problem
A higher WACC lowers the NPV of every project. But not by the same amount.
Compare two (figures illustrative):
Project A: replacement capex outlay −1,000 / annual FCF +260 / 5 years / undiscounted 1,300 / IRR 9.43%
Project B: AI / data-centre type outlay −1,000 / annual FCF +110 / 15 years / undiscounted 1,650 / IRR 7.03%
Project B returns more cash in total (1,650 vs 1,300). It just returns it later.
At WACC 5.4% Project A +113.3 / Project B +111.5
At WACC 7.1% Project A +63.2 / Project B −4.4
In the low-rate world the two were effectively tied (+113 vs +112). Either one added about the same value.
After the same 1.7-point move, A survives with a 44% haircut and B goes underwater.
The reason is plain: B's IRR is 7.03%, just below the new 7.1% WACC.
And B's IRR is low not because the business is worse, but because the cash comes back later.
Rate sensitivity here follows the same logic as bond duration. The more back-loaded the cash flows, the longer compounding works on them, and the more violently they respond to a change in the discount rate. Data centres, power infrastructure, long-cycle R&D, decarbonisation capex, and any acquisition where most of the price is terminal value — all long-duration assets.
Higher rates do not kill the worst projects first. They kill the longest ones first.
This is, I think, the most commonly misread part in practice. "Rates are up, so let's tighten capex" sounds disciplined. But an even tightening does not land evenly: the projects that fail are disproportionately the long ones. Left unmanaged, the portfolio drifts short all by itself.
So why hasn't AI capex stopped?
Are the hyperscalers being irrational? Probably not. Their arithmetic contains two variables the example above leaves out.
First, the cost of not investing. Declining Project B does not leave you at zero. In a world where a competitor secures the capacity, the customers, the models and the standard first, your own future cash flows do not stay flat — they fall. Once that decline is in the base case, the comparison changes.
Second, real options. Once you have secured the grid connection and the land, a data centre can be expanded in stages. The first building is less a standalone NPV than the right to build the next ten. And option value rises with uncertainty, not despite it.
Both arguments are also excellent excuses. When too many projects get exempted from the hurdle rate because they are "strategic," capital discipline quietly stops existing.
So the question is not "should we invest in AI or not." It is "which AI investments clear the hurdle now that the hurdle has moved."
Five things to change when rates rise
1. Refresh the internal hurdle rate. When was it set? What was Rf then? What is it now? If it hasn't moved, is that deliberate conservatism, or neglect?
2. Split WACC by currency. Levels and moves differ across Japan, the US and Europe. One group-wide WACC applied to overseas projects produces systematic errors.
3. Run sensitivity on the discount rate, not just on revenue. What does ±1 point do to NPV? For long-dated projects, that sensitivity is itself a decision input.
4. Look at the portfolio by duration. Project by project, you cannot see that the failures are clustering in the long-dated bucket. Sort by payback period and check the mix is the one you intended.
5. Discuss funding and investment in the same room. Treasury moves the denominator (WACC); the business units set the numerator (FCF). When rates are moving, debating those in separate meetings is itself the risk.
The chain
Profit → working capital → operating cash flow → investment → FCF → WACC → NPV / IRR → capital allocation → enterprise value.
Issue 2 covered FCF, issue 3 covered NPV and IRR, and a later issue covered ROIC > WACC.
What this week's news moved is the denominator in all of them. When the denominator moves, projects whose numerators haven't changed at all are quietly repriced.
So when you read that long-term yields are at a 30-year high, try reading it as:
"Is our investment threshold still assuming the rates of thirty years ago — or of five years ago?"
On yield alone, bonds look attractive. Yet Nikkei closes with a life insurer's view that with no clear trigger for rates to fall, there is no need to rush in. If that is how buyers see it, then for issuers — the operating companies raising the money — it means planning on the assumption that this level persists for a while.
A discount rate is not weather. It is an assumption. When the assumption changes, you rewrite it.
―――
Note: figures on yields, fiscal balances and bond issuance, and market participants' comments, are as reported by Nikkei, 19 August 2026, "Interest rates rising in a global chain."
Note: the WACC, NPV and IRR examples (Rf, β, capital structure, Projects A and B) are illustrative and do not represent any real company.
This article is for information purposes only and is not investment advice or a recommendation to buy or sell any security.
Japanese version: https://note.com/ample_dog4434/n/nd87949807dbb
Know the news. Read the business behind it.
