Two Bottlenecks: Sovereign Refinancing and AI Infrastructure — Sep 27, 2026

U.S. Treasury yields moved as follows from September 22 to September 25.

Source: U.S. Treasury daily par yield curve

“The long end moved more” refers to the larger moves in both the 10- and 30-year yields relative to the two-year, not to a comparison between the 10- and 30-year maturities. The 2s10s curve steepened by about 11 basis points over the four sessions. Investors are repricing long-duration funding, not only the next policy meeting.

The Federal Reserve raised its target range to 3.75%–4.00% on September 16. Its FOMC statement described activity and domestic spending as resilient, productivity and capital investment as strong, and inflation as elevated. The Fed’s policy-action history shows that this was the first increase since July 2023.

The September projections put the median appropriate policy rate at 4.1% at year-end. Compared with the current range midpoint of 3.875%, that would be broadly consistent with one more 25-basis-point increase if the median path were realised exactly. It is a conditional projection, not a commitment.

AI investment supports earnings and activity, but it also raises demand for equipment, construction, power, and skilled labour. Strong growth, including AI capital spending, can coexist with higher discount rates.

Central-bank policy alone does not explain long yields. Governments and companies both need large amounts of funding, and most government borrowing rolls over existing debt rather than financing new spending.

The OECD Global Debt Report 2026 expects governments and companies to borrow $29T from bond markets in 2026, $4T or 17% more than in 2024. Refinancing existing debt will account for 78% of OECD government borrowing. With central banks holding less government debt and more price-sensitive investors absorbing supply, the market may require higher yields.

The report was published on March 4, 2026 and primarily uses data through the end of 2025. It therefore does not incorporate the later Middle East supply shock or the Fed’s September increase. The $29T and 78% figures are best treated as the starting structure of the refinancing problem, not as a live forecast marked to current yields.

The U.S. fiscal position reinforces the pressure. The CBO’s 2026–2036 outlook projects net federal interest outlays of $1.0T in 2026, equal to 3.3% of GDP, rising to $2.1T, or 4.6% of GDP, by 2036. The CBO baseline is not re-marked to market every day. If long yields remain above its assumptions for longer, the reasonable directional risk to interest expense is upward.

For investors, long yields are not an abstract macro variable. They reduce the present value of distant cash flows and raise the hurdle rate for projects with long payback periods, including data centres and power generation. Revenue growth must be judged together with the upfront capital required to produce it.

Energy prices connect the sovereign and AI stories. The IEA’s September 2026 Oil Market Report estimates that global oil production fell 1.6 mb/d in August to 100.1 mb/d, with more than 10 mb/d of Gulf output still shut in. Observed inventories declined by 507 mb from February through August. North Sea Dated crude averaged $91 per barrel in August and reached $113.48 on September 9.

In a separate September 18 analysis, the IEA warned that sustained Gulf constraints and further inventory depletion could require higher prices and additional demand reductions to close the gap. This article does not use a single softer oil session as evidence that the shock has ended. Supply flows and inventories need to normalise.

A renewed energy shock can raise inflation expectations, long yields, freight costs, and electricity costs together. AI infrastructure can face strong revenue demand and rising input costs at the same time.

Microsoft said it added another gigawatt of capacity in FY2026 Q4 and remains on track to roughly double overall capacity in two years. Quarterly capital expenditure was $41B including finance leases.

Calendar-2026 capital-expenditure expectations declined from about $190B to approximately $175B. According to Microsoft’s FY2026 Q4 earnings discussion, the change reflects lease classification after an update to data-centre useful lives, not lower investment. Finance leases are included in capital expenditure; operating leases are not. Microsoft said its underlying investment expectations were unchanged outside that classification effect.

Reported capital expenditure is therefore incomplete on its own. Microsoft’s FY2026 Form 10-K disclosed $329.1B of additional leases, primarily for data centres, that had not yet commenced at June 30, 2026. Some arrangements remain subject to contractual conditions and are scheduled to commence from FY2027 through FY2033. This is not one year’s cash spending or a current liability. It is a long-term commitment that had not yet been recognised as a lease liability because the leases had not commenced.

Meta’s Q2 2026 release guides to $130–145B of 2026 capital expenditure, including principal payments on finance leases.

Adding Microsoft’s approximately $175B and Meta’s $130–145B produces $305–320B. The two companies define capital expenditure differently, so this is not a comparable industry total. Economic commitments may also be understated when capacity is procured through operating leases that are excluded from reported capital expenditure. The sum is useful only as a directional indicator of scale.

GPUs and servers cannot expand usable compute on their own. Substations, transmission, generation, cooling equipment, fibre, and network hardware must arrive on time.

The IEA’s Energy and AI outlook expects global data-centre electricity consumption to almost double from 485 TWh in 2025 to 950 TWh in 2030. Electricity use by AI-focused data centres triples over that period. The IEA also notes that bottlenecks across the value chain make more aggressive near-term scenarios less likely despite booming investment.

The constraint is already visible in U.S. power forecasts. The EIA’s September 2026 Short-Term Energy Outlook projects U.S. electricity sales at a record 4,135 TWh in 2026 and 4,211 TWh in 2027, driven by data centres and manufacturing.

Texas illustrates why scope matters. The Governor’s directive requires an audit of data centres advancing through ERCOT’s interconnection process. ERCOT is considering about 474 GW of connection requests, more than five times its record peak, and roughly 90% of new power requests are from data centres. This is not a blanket halt on every data-centre project in Texas. ERCOT’s Batch Zero large-load framework primarily groups projects of 75 MW and above, while separate information requests cover data centres in the 25–75 MW range.

The two bottlenecks compete for capacity in the same bond market. The OECD report says nine major hyperscalers raised $122B from bond markets in 2025, nearly half of global technology-company issuance. Those nine companies are forecast to spend a cumulative $4.1T from 2026 through 2030. AI investment is increasing technology companies’ reliance on external funding while governments are also bringing large refinancing needs to market.

The cash-conversion pressure is already visible.

Microsoft’s $19.6B is a simple proxy calculated from its official cash-flow statement: $55.4B of operating cash flow less $35.8B of cash additions to property and equipment. It differs from Microsoft’s separately disclosed $41B of capital expenditure including finance leases. Meta’s $0.78B is the company’s non-GAAP FCF measure. The figures are not comparable enough for a cross-company ranking. They are useful for assessing the year-over-year effect of the buildout within each company. Sources: Microsoft FY2026 Q4 earnings tables and Meta Q2 2026 results.

These figures do not establish that either company is financially fragile. Revenue still grew 18% and 28%, respectively. They show that revenue growth alone no longer measures the capital productivity of the AI buildout.

Not every company associated with AI has the same economics. The company spending the capital and the supplier paid to relieve a constraint can have very different cash-flow profiles.

The useful sequence is:

1. Demand quality — Does backlog convert into revenue and cash?
2. Pricing power — Can higher equipment and power costs be passed through?
3. Capital intensity — How quickly are capital expenditure and new lease commitments growing for each additional dollar of revenue?
4. Duration of scarcity — Is the shortage temporary, or does it involve multi-year permitting and grid work?
5. Balance-sheet capacity — Can the investment plan continue with long yields around 5%?

Power, cooling, networking, and data-centre operations are not peripheral to AI. But scarcity alone does not make the related equities cheap. Orders can grow while working capital and capital expenditure rise even faster, delaying cash available to shareholders.

The levels below are illustrative monitoring markers, not forecasts.

No probabilities are assigned. A one-day breach does not change the thesis by itself. The markers are a consistent way to monitor several signals over time, not a target-price model.

• The 10- and 30-year Treasury yields and the 2s10s slope
• The Fed’s next decision and any change in its year-end policy path
• Actual Treasury issuance costs and OECD sovereign refinancing volumes
• Crude and refined-product prices, global inventories, and Gulf export flows
• Hyperscaler capital-expenditure guidance, uncommenced lease commitments, and bond issuance
• FCF conversion at Microsoft, Meta, and other hyperscalers
• Data-centre grid-connection outcomes and delivery times for power and cooling equipment

Two funding competitions are taking place in the same bond market. Governments must refinance large debt loads at higher rates, while technology companies are committing unprecedented capital to AI infrastructure and relying more on external funding. One raises the discount rate; the other aims to raise future growth.

“AI will grow” is not a complete investment case. Investors still need to account for capital expenditure, leases, and working capital required to create each additional dollar of revenue. If long yields stay elevated, the market is likely to reward companies that relieve real constraints and produce cash, not simply those with the strongest AI narrative.

The defining question is whether returns on incremental AI capital can stay above a rising cost of capital.