A Chip Company Is Now Acting Like a Bank

Here's the deal: on August 20, Bloomberg reported that Broadcom is in talks with lenders to raise more than $60 billion in debt. The money isn't for R&D, or fabs, or an acquisition. It's to buy custom AI accelerators that Broadcom itself designs — and then lease them to frontier AI labs, Anthropic chief among them.

Read that again, because it sounds backwards. Chip companies make chips and sell them to customers. That's the whole business. But in AI infrastructure right now, that simple transaction has stopped clearing. The volume customers want costs more than customers can raise. Anthropic's stated plan involves over a gigawatt of training and inference capacity, and paying for that out of equity would dissolve the ownership of everyone already on the cap table.

So an intermediary appeared. The company that builds the silicon also arranges the money, borrows against the hardware as collateral, and converts the customer's enormous one-time capital problem into a monthly lease payment. That's the "AI XPV Platform" Broadcom set up with Apollo and Blackstone on June 9, and this $60 billion raise is its second deal.

The numbers give you the scale. The senior secured tranche under discussion runs $60–70 billion, with a junior tranche of roughly $30 billion sitting alongside it. Bankers involved have floated a total as high as $100 billion. The first deal, two months ago, was $35 billion. It's roughly tripled since June.

The Three Parties — Who Builds, Who Funds, Who Uses

Broadcom is a custom silicon company. Unlike Nvidia, which sells its own branded GPUs, Broadcom designs and manufactures accelerators (it calls them XPUs) and networking chips to customer specification. A large share of Google's TPU program runs through Broadcom, and Anthropic and OpenAI are now on the marquee customer list. What makes its role here unusual is that Broadcom isn't just supplying parts — it is backstopping portions of the senior tranche with its own credit. Broadcom carries investment-grade ratings, and lending that rating to the structure drops the borrowing cost materially.

Apollo and Blackstone are the capital. Apollo's June 9 announcement laid it out: Apollo-managed funds and affiliates led an initial $35 billion capital solution, with Blackstone and leading global banks alongside. The appeal is obvious to anyone who runs a private credit book. AI data center equipment throws off contracted cash flow for the life of the lease, the collateral physically exists, and with a Broadcom guarantee attached, the risk-adjusted return looks unusually clean for an asset of this size.

Anthropic is the end user. The June platform announcement specifically references Anthropic's previously disclosed expansion of more than 1GW of compute for training and inference starting mid-2026, deployed at Fluidstack-based sites. Leasing rather than buying keeps a mountain of capex off Anthropic's balance sheet. What it gets instead is a long-dated fixed cost that shows up every month regardless of how the year goes.

There's a fourth name in the paperwork worth noticing. The June platform documents list OpenAI as a target customer too, not just Anthropic. The platform as a whole is aimed at enabling more than 20 gigawatts of compute capacity through 2028. Twenty gigawatts is roughly the output of twenty nuclear reactors. This was never a single-customer arrangement — it's a pipeline built to serve several frontier labs at once.

How the Structure Actually Works

Because the June deal's mechanics were disclosed, we can infer the shape of this one. Bloomberg reported the $35 billion was split into three tranches, two of them senior and backed by Broadcom: a $6 billion A1 note and a $24 billion A2 note. The rest sat in a junior position carrying more risk for more yield.

The deal now under discussion scales that template up:

First deal (announced 2026-06-09) Second deal (reported 2026-08-20, in talks)
Total size $35B $60–70B (up to ~$100B floated)
Senior $6B A1 note + $24B A2 note (Broadcom-backed) $60–70B
Junior Balance ~$30B
Led by Apollo, with Blackstone and global banks Blackstone, Apollo and others
Use of proceeds Acquire XPUs and networking, lease to frontier labs Same
End customers Anthropic, OpenAI (20GW+ target through 2028) Anthropic-centered

Why debt instead of equity? Three reasons stack up.

The asset behaves like debt collateral. AI accelerators and the racks holding them are physical objects. The counterparty is a credit-checked, well-capitalized lab. The term and the payment are written into a contract. This is a completely different risk profile from writing an equity check into a software startup — it's closer to aircraft or real estate leasing, which is exactly the shape credit markets are built to price.

Equity is absurdly expensive right now. Selling equity means permanently handing over a slice of the company's future at today's valuations, which for frontier labs are extraordinary. Debt costs interest and then the relationship ends. Layer a Broadcom guarantee on top to compress the spread, and the gap in cost of capital gets wide enough to determine strategy by itself.

Speed. Building 20GW by 2028 means orders have to be placed now, and orders require money up front. Equity rounds take months of diligence and negotiation. Structured debt with clear collateral and contracted cash flow closes far faster.

There's a quiet assumption underneath all of it, though: the lease payments have to keep arriving for the full term. If AI lab revenue doesn't compound the way the model says, or if today's silicon loses its economics three years from now, that risk lands on the debt holders and on Broadcom, which guaranteed the senior paper.

One more thing worth flagging. Broadcom isn't booking a straightforward sale here — it's arranging for a financing platform to do the buying. Revenue recognition and cash collection separate, and debt fills the gap between them. No semiconductor company has previously designed demand and financing as a single product. Broadcom has effectively bolted a leasing company onto a fabless chip business.

What Each Side Gets

Broadcom gets certainty of demand. When customers pay cash, orders move with the customer's fundraising calendar — they slip, they shrink, they get cancelled. Solving the financing removes that dependency entirely. It's also a way to beat Nvidia without fighting Nvidia. Instead of competing on benchmark performance, Broadcom shifts the axis of competition to who can actually get customers the volume they need.

Apollo and Blackstone get scale. Institutional capital is hunting for stable yield, and AI infrastructure leases offer contracted cash flow with hard collateral behind it. Senior paper carrying a Broadcom guarantee prices close to investment grade. And the sheer size matters — there are very few asset classes where a mega-fund can deploy tens of billions in a single structure.

Anthropic gets compute without dilution. This matters more than it sounds. Frontier lab competition ultimately reduces to how much compute you can secure and how fast. Funding all of it with stock means the company's ownership melts into infrastructure costs. Leasing pushes the expense into opex instead. The tradeoff is a fixed annual obligation that squeezes hard if revenue growth disappoints.

Power and land developers benefit quietly. Twenty gigawatts is a number on paper unless generation, transmission, cooling and sites arrive with it. Once financing of this scale is locked, the contracts downstream start moving in sequence.

The Precedents — One Cautionary, One Reassuring

This model isn't new. Telecom equipment ran the same play around 2000. Lucent and Nortel sold gear to newly formed carriers and lent them the purchase price — vendor financing. Revenue looked wonderful, stocks climbed. Then the dot-com collapse took the carriers down, the receivables never came back, and the losses landed on the equipment makers' books. Lucent never recovered.

The reassuring precedent is aircraft leasing. Airlines lease rather than buy from lessors like GECAS and AerCap, and that structure has worked for decades. The difference is that aircraft have predictable residual value and liquidity — if one airline fails, another takes the plane. The collateral genuinely functions as collateral.

Which one do AI accelerators resemble? Right now, closer to aircraft. Demand exceeds supply, so if Anthropic couldn't use the capacity, there's a queue of buyers who would. The problem is the clock. Semiconductor residual value curves look nothing like aircraft. A plane flies for twenty years; an AI accelerator gets superseded in three to five. If the lease term outruns the generational cycle, that gap becomes a loss.

And data center sites and power contracts are far less portable than airplanes. You can relocate chips; you cannot relocate a twenty-year power purchase agreement in a specific county. The least liquid thing in this structure isn't the silicon — it's the real estate and the electricity.

Speaking of which, the power bill deserves its own line. Running 20GW for five years costs tens of billions in electricity alone. Lease payments are fixed in the contract; power prices are not. When regional power markets move, the lessee absorbs it. Headline deal size tells you little about where the operating margin actually lands.

How Competitors Respond

Nvidia isn't standing still. It already invests directly in neoclouds and startups inside its ecosystem and pre-commits supply, achieving a similar effect. But the flavor differs. Nvidia locks customers in with dominant general-purpose performance and CUDA; Broadcom sells bespoke design bundled with the money to buy it. For a customer who knows their workload precisely and needs enormous volume, the second option can win on total cost of ownership.

Marvell played a different card from the same position. On August 18 it issued Google a warrant worth about $12.2 billion, expanding a custom silicon agreement across the TPU ecosystem. Where Broadcom solved the customer's financing problem with debt, Marvell bound the customer with its own equity. Same problem, opposite currency.

AMD and Intel face real pressure here. A competition that used to run on performance now has a financing variable in it. Guaranteeing tens of billions in senior debt requires a balance sheet and a credit rating that very few companies possess. That's a short list, and it isn't getting longer.

The big three clouds occupy an awkward spot. Renting compute from hyperscalers used to be the default path for AI labs. Now those labs deal directly with chipmakers and private credit funds to build their own footprint. Microsoft, Google and Amazon remain the largest suppliers by far, but if their biggest customers keep moving toward owned capacity, the long-run negotiating balance shifts.

Server integrators like Supermicro and Dell get locked-in volume, which is good news. The catch is that their counterparty becomes a financing platform rather than an AI lab, and single enormous orders are exactly the setup where buyers grind margins down.

So What Actually Changes

If you build on AI APIs, nothing changes at your price page tomorrow. But as this structure spreads, downward pressure on inference pricing increases. Leased capacity sitting idle is a pure loss, so operators push utilization hard, and that shows up as price cuts or batch discounts. OpenAI dropping GPT-5.6 Sol API pricing by 20–33% on August 21 sits on exactly this current.

If you're an investor, track where the risk moved. Broadcom's revenue line will look excellent, but standing behind it is a guarantee written on its own credit. Reading the growth rate alone misses the off-balance-sheet exposure. Two things to watch: how much of the guarantee gets disclosed, and whether the lessees' revenue actually compounds on schedule.

If you work in semiconductors or data centers outside the US, the signal here is financial, not technical. The decisive advantage in large AI infrastructure is shifting from design capability to financing structure. A fabless company can design an excellent chip and still lose the deal because it cannot help the customer pay for volume.

If you follow AI policy, the real story is the 20 gigawatts. That's a grid problem and a land problem before it's a chip problem. US states are already issuing executive orders on data center power allocation. When capital gets committed this fast, the bottleneck migrates from money to electricity and permits.

If you're just a user, this won't touch your day. But it explains a lot about why AI pricing will move the way it moves over the next few years. Lease payments on the hardware being installed right now lock in a cost floor, and consumer pricing can't stray far from that floor for long.

🥄 Three Things You're Probably Wondering

— Isn't this just financial engineering? Not fairly characterized that way. Leasing is a decades-proven technique in aircraft and real estate, the collateral genuinely exists, and the counterparties are real operating businesses. What's true is that risk didn't vanish — it relocated. Broadcom guaranteeing the senior tranche means that if the end customer can't pay, the exposure comes home to Broadcom. You won't see that by reading the revenue line.

— Can Anthropic actually carry this? At its current revenue trajectory, the market appears to think so. The nuance is that lease obligations don't shrink when revenue does. If price competition among frontier labs continues at this intensity, revenue can grow while margin thins — and that combination is the worst case for a leveraged lease structure. Watch the term length and the early-termination terms, not the headline number.

— Is this proof of an AI bubble? Too early to call either way. Debt entering a market isn't itself evidence of a bubble; telecom, power and rail all built out their initial infrastructure on borrowed money and plenty of it survived. What is true is that debt punishes the downside much harder. Equity just loses value; debt that can't be serviced transfers the assets. The metrics that matter here are utilization and lease collection, not deal size.

Sources

Numbers and criteria are as of announcement and may change. Investment calls are yours to make!