This July, Amazon issued a $25 billion corporate bond sale in the public market. Total orders submitted by buyers amounted to only 1.6 times the issue size, and the bonds broke below their issuance price on the very first day of trading. Four months earlier, for another bond issue by Amazon, total orders were still 3.4 times the issue size. Although both batches of bonds were ultimately sold out in full, the cost of borrowing for tech giants continues to climb.
Start with the overall borrowing volume across the market. According to statistics from BNP Paribas, as of August 10, these tech giants have issued a cumulative total of approximately $220 billion in bonds this year, compared with only $12.5 billion in the same period last year. The list compiled by Bloomberg is even more specific: the six companies—Amazon, Alphabet, Meta, Oracle, NVIDIA, and SpaceX—have already issued more than $200 billion in total US dollar bonds this year; by comparison, bond issuance by highly rated tech companies totaled only $13 billion over the same period last year.
The pace of bond issuance across individual companies has been equally intense. After issuing approximately $37 billion of bonds in March, Amazon followed up with another $25 billion in July; after issuing about $32 billion of bonds in February, Alphabet issued up to $25 billion again in August; NVIDIA in June issued its first corporate bond since 2021, with the size also reaching $25 billion; SpaceX also completed a $25 billion bond issuance this year.
All of these bonds were sold out, with no failed auctions in the public market. The key issue is not whether they sold, but that the terms offered by buyers are tightening.
The ratio of buyer order volume to the issuer’s borrowing size intuitively reflects the intensity of competition for capital; this metric is the subscription multiple. When Amazon issued bonds in March, the subscription multiple still reached 3.4x; by the time it issued bonds again in July, the subscription multiple dropped to 1.6x, shrinking by more than half in just four months.
New bond yields must be higher than a company’s existing outstanding bonds for buyers to be willing to take them; the portion of yield by which new bonds exceed existing bonds is called the concession. Although Alphabet’s August bond sale was sought after overall, yields still had to be 0.1 to 0.15 percentage points higher than its existing bonds to clear the market. Across the entire investment-grade bond market, the median concession has risen from 0.0225 percentage points last year to 0.12 percentage points this year, expanding more than fivefold within a single year.
When a newly listed bond drops below its issue price, it is called breaking below par. In a normal year, roughly two-thirds of newly issued bonds see their prices strengthen after listing; but among the 91 comparable tech giant bonds tracked this year, 78 bonds have yields higher than their levels at issuance, with the median yield shifting upward by 0.22 percentage points. The $25 billion new bond sales issued respectively by NVIDIA, SpaceX, and Amazon all broke below their issuance prices on the very first day of trading.
The additional interest compensation tech companies must pay buyers when borrowing is also widening exponentially. The premium a corporation pays above the risk-free US Treasury benchmark rate is called the spread. Amazon’s latest long-dated bond is about 1.2 percentage points above Treasuries, compared to roughly half that a year ago. Across the entire industry, the spread on tech corporate bonds is around 0.9 percentage points, which is about 0.1 percentage points wider than the spread on investment-grade corporate bonds across the overall market. Historically, tech giants, backed by healthy balance sheets and abundant cash flow, always enjoyed narrower issuance spreads than the market average; now, for the first time, they have systematically flipped to the wider side.
The credit fundamentals of these large companies have not run into trouble. Rating agencies to date have not downgraded any of them, and buy-side institutions also say they are not worried that Amazon or Alphabet will fail to repay. Neil Sutherland, head of US fixed income at Schroders, was quoted as saying that this is not a credit issue for highly rated tech companies.
The real constraint comes from risk control rules in buyer portfolios. In the public bond market, the primary capital buying and holding large company bonds long-term comes from pension funds and insurance companies. For diversification purposes, these institutions typically set an asset allocation cap of 2-3% on any single company in their portfolio, known as a single-issuer limit. No matter how outstanding a company is, its name cannot exceed this percentage threshold on an account statement. Choi, a portfolio manager at Capital Group, put it bluntly: clients do not want to open their statements and discover that 10% of their portfolio is concentrated in a single bond.
The same company issues bonds three or four times a year, tens of billions each time. The first wave of buyers exits after filling their 2-3% limits, and orders flow to the next tier of institutions, which also fill their limits and exit. The more bonds that are issued, the fewer institutions remain with available limits. The interest rate at which the final batch of debt clears depends on what price the buyers who still have room are willing to ask. This group of institutions taking the final orders is called marginal buyers, and their asking price is the marginal financing cost of AI infrastructure.
This mechanism explains why the rise in borrowing costs precedes credit defaults. A credit default is an ex-post, discrete event that occurs after a company’s fundamentals collapse, whereas the exhaustion of limits is an ex-ante, continuous process. It does not require any tech company to make an operational mistake; it only requires the frequency and scale of debt financing to run fast enough.
At the same time, overall demand across the bond market has not dried up. The average yield on the current investment-grade corporate bond index remains at around 5.4%, essentially on par with its long-term historical average. Capital has not left the bond market, but facing relentless borrowing by tech giants, buyers’ asking prices will rise accordingly. George Catrambone, head of fixed income at DWS, put it directly: It’s not a blank check. If these companies keep tapping the market over and over again, concessions are going to get larger and spreads are going to get wider.
The pressure of tightening buyer portfolio limits does not come solely from within the tech sector. The same pension funds and insurance institutions managing long-term capital must both absorb massive bond volumes from tech giants and face heavy supply from US Treasuries.
This year, the US fiscal deficit is approximately $1.9 trillion, accounting for 6% of GDP. The auction yield on the US Treasury’s latest 30-year bond sale reached 5.22%, the highest level since 2001. According to an analysis by Reuters, intensive bond issuance by tech giants is one of the main factors pushing up Treasury yields; conversely, ongoing Treasury issuance is also raising the borrowing threshold for tech giants. Vivek Paul of the BlackRock Investment Institute called it a battle for capital rarely seen in recent years.
Ultimately, the available capacity in buyers’ hands is a shared pool of capital rather than mutually independent drawers. The massive expansion of Treasury supply makes institutional quotas, already rapidly consumed by tech giants, even tighter.
If buyer limits are genuinely real, the most direct test is to observe the actual moves of the borrowing corporations. Facing a capacity ceiling in the public market, tech giants have proactively altered their financing paths, and four detour channels have emerged simultaneously.
The first path is proactively controlling supply. When Alphabet issued up to $25 billion in bonds in August, it confirmed to the public that this offering was the company’s final dollar bond issuance in the public market this year, stepping on the brakes in public channels.
The second path is switching currencies to access new buyer capital pools. While issuing US dollar bonds, Amazon added an additional bond offering of approximately €10 billion and is planning the first Swiss franc bond issuance in company history. Different currencies connect to different institutional investors locally in Europe and Switzerland; this operation is equivalent to dispersing the holding limits that previously weighed on dollar buyers into institutional pools of other currencies for absorption.
The third path is bypassing the public market and targeting long-term capital directly through special purpose project vehicles. An affiliate of BlackRock issued a $12.3 billion dedicated bond (ultimately clearing at $12.5 billion) for Meta’s data center in El Paso, Texas (with a planned capacity of up to 1 GW). This bond adopted a single maturity structure due in 2048 and was intentionally allocated during underwriting exclusively to buy-and-hold pension and insurance funds, steering clear of fast money. The spread on this project bond was about 2.9 percentage points over Treasuries (initial price talk was 2.875 percentage points), more than double Amazon’s direct bond issuance spread (about 1.2 percentage points). This price differential represents the shadow price paid when the same AI infrastructure funding leaves a mega-cap balance sheet.
The fourth path is having chip vendors step in directly to provide commercial credit. Broadcom began negotiations with a group of lenders on August 20 to raise over $60 billion in debt financing for clients like Anthropic, with the total scale potentially reaching $100 billion, and Blackstone and Apollo participating through private credit channels. This transaction builds on Broadcom’s $35 billion collaboration in June with Apollo and Blackstone to support Anthropic’s computing power expansion. Prior to this, NVIDIA collaborated with Wall Street financial institutions to set up a $500 billion AI financing framework and provided a $105 billion residual value guarantee for an Ohio data center project. Their roadmap is consistent: using the commercial credit of chip vendors to substitute for exhausted holding limits in the public bond market.
Comparing these four paths, tech giants are creating new buyer capacity outside public bond channels through four mechanisms: self-imposed supply caps, multi-currency diversion, project asset structuring, and vendor commercial guarantees. Every single detour serves as direct evidence of the existence of the limit wall.
The first exception is companies teetering on the edge of a rating cliff. Among the six debt-issuing giants, Oracle’s financial condition differs most drastically from the other five: its free cash flow was negative in the previous fiscal year, it carries roughly $260 billion in data center lease commitments, and long-term debt increased from $96 billion to $149 billion. S&P estimates that its Debt/EBITDA ratio will peak at 4.4x within two years, with any level exceeding 4.5x triggering another downgrade. If Oracle falls into junk (high-yield) status, investment-grade indexes would exclude it, followed by forced selling from passive funds. That is another form of capacity cliff—one stemming from credit, not from limits.
The second boundary is the statistical scope. The aforementioned $220 billion only counts high-grade bonds issued in public markets. Goldman’s comprehensive tally shows AI-related debt (including loans and private credit) has reached $489 billion this year; Nikkei estimates the hidden debt of the five tech giants (including leases and off-balance-sheet items) at $1.65 trillion, swelling eightfold in four years. The limit thesis applies only to the public bond channel, because this channel is publicly priced daily, allowing us to observe buyer absorption signals.
The third variable is the race between supply and revenue. Goldman expects the five giants to issue $250 billion in bonds in 2026, rising to $400 billion in 2027. If AI revenue materializes rapidly (OpenAI’s annualized revenue has now surpassed $40 billion), or if tech giants slow their capital expenditures, supply pressure will ease on its own; at the current borrowing pace, however, the test posed by this wall next year will only be harsher.