1. The Numbers
Broadcom is in talks to raise more than $60 billion, and potentially as much as $100 billion, in debt to finance AI chip capacity for Anthropic and other customers. That is according to Bloomberg’s reporting on 20 August, which cites people familiar with the matter and no named executive. This is not, strictly, a story about a company selling more chips. It is a story about who is willing to lend against the bet that those chips will be needed, and on what terms they will get paid back if the bet is wrong.
We treat the headline figure as Assessed, not Established. Bloomberg’s own account described a senior-secured tranche in the range of $60 billion to $70 billion, with Broadcom guaranteeing part of it, plus a junior tranche of roughly $30 billion — numbers its sourcing characterised as still fluid. CNBC, reporting separately and also citing unnamed sources, put the deal at “upwards of $70 billion,” a materially smaller figure than Bloomberg’s upper bound. That gap between two well-resourced newsrooms working the same story is itself informative: even at the Tier 2 institutional level, the number has not settled. We will treat any single figure quoted this week as provisional until Broadcom, Apollo or Blackstone confirm a structure on the record.
What is Established, because it rests on the companies’ own disclosures, is the deal this new package would extend. On 9 June 2026, Broadcom, Apollo Global Management and Blackstone announced — in press releases from all three firms — a partnership called the AI XPV Platform, an initial $35 billion capital solution led by Apollo, built to finance more than 20 gigawatts of Broadcom custom-chip and networking capacity for frontier AI labs, including Anthropic and OpenAI, through 2028. The first tranche funded part of Anthropic’s previously announced plan to add more than a gigawatt of training and inference capacity, largely at Fluidstack-operated sites. The new financing under discussion is being reported as a scale-up of that same structure, not a new one.
2. The Mechanics
The mechanics matter more than the headline number. In a senior/junior note structure, the senior tranche gets paid first — from the revenue the underlying data centers and chip capacity generate — and carries the lowest interest rate because its risk is lowest. The junior tranche is paid only after the senior lenders are satisfied, carries higher yield to compensate, and absorbs the first real losses if the capacity built with this money does not earn what was projected. Private credit firms such as Apollo and Blackstone specialise in precisely this: originating and holding the junior, higher-yielding risk that banks are reluctant to keep on their own balance sheets, while banks and insurers take the safer senior piece. Broadcom’s own guarantee of part of the senior tranche is the detail worth sitting with — it means Broadcom is not merely selling chips into a financing vehicle and stepping back; it is putting its own credit behind at least part of the debt that pays for those chips, which lowers borrowing costs for the whole structure but also ties Broadcom’s balance sheet more tightly to whether the capacity gets used.
Set against the wave it sits inside, this is not an outlier — it is close to the norm. S&P Global has tracked roughly $220 billion in bonds issued by AI hyperscalers through mid-August 2026, against roughly $12.5 billion in the same period of 2025 — an increase reported at nearly 17-fold. Amazon, Alphabet, Meta, Microsoft and Oracle have all issued investment-grade bonds this year specifically tied to data center buildout, and reporting from Fortune and others has tracked tens of billions more in off-balance-sheet structures, including Meta’s private-credit joint venture with Blue Owl and Oracle’s project-finance arrangements. The Broadcom-Anthropic package is a variant of the same instinct: keep the capacity spending moving faster than operating cash flow alone would allow, and use debt — increasingly private credit rather than public bonds — to bridge the gap.
3. The Case, and the Test
The case for this leverage deserves to be stated plainly before it is questioned. AI compute demand from frontier labs has, so far, run ahead of what any single company’s cash flow could fund organically, and the firms taking this risk are not naive: Apollo and Blackstone are sophisticated credit investors who priced this exposure once already in June and are reportedly prepared to price it again at greater scale. Debt, properly structured with a real senior/junior waterfall, is also the standard tool for financing long-lived infrastructure with a shorter buildout period than payback horizon — this is how power plants, pipelines and toll roads get built, not a departure from established finance practice. If Anthropic and its peers do sustain their contracted revenue growth, this leverage financed capacity years earlier than equity alone would have, and the lenders earned a fair return for taking the risk.
What would make it look reckless in hindsight is narrower than the framing “AI debt bubble” suggests: it depends on whether the specific demand this capacity is built for — Anthropic’s contracted and anticipated compute needs — materialises on the timeline the financing assumes. If it does, the senior lenders are made whole from cash flow, the junior lenders earn their premium, and Broadcom’s guarantee is never called. If demand growth stalls or Anthropic’s own revenue trajectory disappoints relative to its compute commitments, the junior tranche absorbs losses first, private credit funds mark down positions that retail investors increasingly hold indirectly through retirement and insurance products, and Broadcom’s guaranteed portion of the senior debt becomes a real balance-sheet liability rather than a contingent one. Nobody involved in this reporting — Broadcom, Apollo, Blackstone, or the outlets covering it — has offered a public projection of what utilisation rate the underlying capacity needs to hit for the economics to work. That is the number this desk considers the actual test, and it is not yet public. We do not know it.
What to watch: Broadcom reports fiscal third-quarter 2026 earnings on 2 September, its first results call since this financing was reported, and the one most likely to bring an on-the-record figure — either confirming a structure or declining to. A subsequent SEC filing or joint press release from Broadcom, Apollo and Blackstone, in the pattern of their 9 June announcement, would move this from Assessed to Established. Until one of those lands, treat every number in circulation, including the ones in this piece, as a range under negotiation rather than a signed deal.