How hyperscaler issuance is reshaping the Treasury curve

Hyperscaler bond issuance has surged in 2026, with volumes already far exceeding previous years. Alphabet, Amazon, Meta, Microsoft, Oracle and SpaceX have issued over $180bn of USD-denominated bonds so far this year. Street estimates now point to an unprecedented $250bn of USD issuance by year-end, which totals around $400bn when broader technology and artificial intelligence (AI)-related borrowing is included.

While the headlines have focused on absolute volumes, the more interesting narrative is their preference for the long end of the curve. Alphabet's £1bn century bond, the first from a technology company since Motorola in 1997, drew £9.5bn of orders at 120 basis points (bp) over gilts. It wasn't an isolated stretch. Amazon, Meta and Oracle have all issued bonds deep into the long end of the curve. Alphabet has also done this while tapping the EUR, GBP, CAD, JPY, CHF and AUD markets this year, including a debut A$5.5bn deal with a 20-year tranche and the largest-ever corporate bond in the Swiss franc market. Alphabet didn't go to sterling for size, it went for tenor: the UK is one of the few markets with a natural buyer base for 100-year corporate paper. The pattern across the group is not just the debt volume, but the maturity choice. These companies are not simply issuing opportunistically across the curve; they are showing a clear preference for long tenors, turning the AI capital expenditure (capex) boom into a meaningful new source of long-duration corporate supply.

The driver of this wave of issuance is clear: AI capex, set to exceed $750bn this year, has outrun free cash flow, and these names, previously a very small part of the index, now account for around 4.5% of the ICE BofA US Corporate Index. With more than $1tr of capex expected in 2027, bond markets will remain a core part of the funding mix for this part of the market.

Across 2025 and 2026 supply, the hyperscalers have placed 41% of their issuance beyond 15 years, against 16% for the ICE BofA US Corporate Index, and their weighting in the over 26-year bucket sits at 31.2% versus 13.0%, which is nearly two and a half times the broad market’s. What is evident is that the hyperscalers clearly have an appetite for a part of the market corporates typically shy away from.

Amazon anchors the ultra-long end in dollars, with Alphabet close behind. Amazon has extended longer-dated dollar issuance out to 2076, with other tenors in the 2060s, while Alphabet combines long-dated dollar issuance with its 2126 sterling century bond. Meta and Oracle’s longest paper runs to 2066, leaving Microsoft the outlier, with a materially shorter issuance profile. Microsoft is the only hyperscaler yet to term out its 2026-27 capex in the bond market, which leaves an obvious source of further long-end supply still to come.

The rationale behind this wave of long-dated issuance makes sense. Data centre assets have long economic lives, making 30- to 50-year debt a natural duration match. Even though spreads have widened this year, the arithmetic still favours borrowing: with a 20-year average return on invested capital (ROIC) of 17.75%, funding at 5.5% to 7% clears the hurdle comfortably. Neither the steeper curve nor around 30bp of spread widening has shifted, or is likely to shift, hyperscaler appetite. They are not price sensitive issuers, and they are not especially cautious about what the supply technical does to their own spreads. When the 30-year US Treasury yield hit its highest level since 2007 in August, the Treasury Department responded within a day by expanding its long-end buybacks. The hyperscalers have absorbed a comparable move in their own funding costs and changed nothing.

Demand has been strong overall, with the exception being Amazon’s July deal, which caught the market off guard and drew the smallest book of the four hyperscalers below: with a 2.5x oversubscription rate on the 30-year. Elsewhere at the 30-year point, coverage was 4.5x for Oracle, 3.6x for Meta and 4.6x for Alphabet's recent deal, all ahead of the 2.98x average oversubscription year-to-date across the broader US investment grade primary market. That's stronger than average demand despite the duration on offer.

When Amazon or Alphabet prints $5-6bn of 30-year paper, it competes directly with the 30-year Treasury for the same pool of liability-matching capital: pension funds, insurers and asset managers. Each deal absorbs duration capacity that would otherwise have flowed into governments, and other buyers often hedge their rate exposure by selling the equivalent Treasury, adding a second round of selling pressure at the same point on the curve. Heavy corporate borrowing to fund data-centre investment is contributing to broader government bond supply concerns, raising the prospect that the AI capex cycle is lifting borrowing costs generally by competing with sovereign debt for investor demand. The 30-year yields now stand at 5.25%, roughly 40bp above where they were at the start of the year.

That said, hyperscaler supply is adding to the pressure on the long end rather than being the source of it, and the structural drivers still dominate. The US fiscal position shows no credible path to consolidation, US debt has just passed $40tr, with deficits running well above 6% of GDP and debt-to-GDP climbing with no political will to arrest it. Japanese investors, long the most reliable and price-insensitive buyers of long-dated Treasuries, now face domestic yields attractive enough to keep the money at home, removing a buyer the market had come to rely on.

Inflation has proven stickier than the government anticipated, and the resulting erosion of confidence in the Federal Reserve’s (Fed’s) willingness to act has let term premium drift higher, against an economy whose growth and full employment arguably justify an upward-sloping curve in any case. Hyperscaler supply sits on top of all of it as a marginal negative technical, and with the AI capex cycle unlikely to peak before 2028, that pressure does not look transitory. For fixed income investors, the practical conclusion is that a material rally at the long end is hard to construct absent a change in the macro picture.