Anthropic Raises 36B Debt to Buy Google AI Chips 2026

Anthropic just lined up $36 billion in debt to rent its own chips back from a shell company. That sentence alone should make any investor sit up. The most valuable private AI company on earth is no longer funding its compute with equity or cash flow. It is borrowing against the future resale value of silicon it does not yet own.

What Actually Happened

Apollo Global Management and Blackstone are arranging a roughly $36 billion private credit package to buy Google custom chips, called tensor processing units or TPUs, on behalf of Anthropic. The structure is deliberately indirect. A special-purpose vehicle borrows the money, takes an equity slice, purchases the TPUs, and then leases the hardware to Anthropic for use in data centers spread across New York, Texas, Louisiana, and Indiana. Anthropic never holds the chips on its own balance sheet. It simply pays rent, and the lenders own the asset that produces the rent.

The debt is sliced into tranches that read like a structured-finance textbook: approximately $6 billion of A1 notes, $25 billion of A2 notes, and $4.5 billion of B notes. The senior portions carry an unusual layer of protection. Broadcom, which co-designs the very TPUs being financed with Google, is providing a residual-value support agreement on roughly $31 billion of the senior debt. In plain terms, Broadcom is promising the chips will be worth something at the end, which is what makes lenders comfortable handing tens of billions to a company that did not exist at this scale eighteen months ago.

The timing is not random. Anthropic filed a confidential S-1 with the SEC in early June at a reported valuation near $965 billion against roughly $47 billion in annualized revenue, after closing a $65 billion Series H that vaulted it past OpenAI's private valuation. Investors in the chip deal were asked to commit within the week, with closing expected days later. This is a company assembling every possible source of capital at once: equity from a megaround, public markets through an IPO filing, and now the largest chip-financing debt transaction ever attempted.

The lease structure carries a consequence that is easy to miss. Anthropic's largest single infrastructure commitment will live in footnotes rather than on the face of its financial statements, at the precise moment public investors are first learning to read those statements. The rent payments become a long-dated fixed obligation that competes with every other claim on Anthropic's future cash, due whether or not the underlying models keep winning. A company losing billions today is locking in years of payments for hardware it does not own, which is exactly the kind of operating leverage that magnifies both the upside and the downside of the entire bet.

Why This Matters More Than People Think

For two years the AI capital story has been told in equity rounds. OpenAI raised $122 billion. Anthropic raised $65 billion. Those headlines trained everyone to think of compute as something you buy with investor cash. This deal rewrites that assumption. It treats AI compute as a financeable, leaseable, securitizable asset class, the same way airlines finance jet engines and landlords finance towers. When you can borrow $36 billion against chips instead of diluting shareholders, the ceiling on how fast a lab can scale moves dramatically higher.

The second shift is about Nvidia. These are Google TPUs, not Nvidia GPUs. Nvidia CEO Jensen Huang recently said his company's $10 billion investment in Anthropic was "probably the last" before the lab goes public, and the subtext is now obvious. Anthropic is actively diversifying its compute supply away from a single vendor that also funds it, and Google plus Broadcom are the beneficiaries. A $36 billion commitment to TPUs is the loudest vote of confidence the custom-silicon camp has received against the Nvidia monoculture.

There is a competitive-timing angle too. Anthropic is buying the right to scale aggressively at the exact moment its rivals are capital-constrained by their own commitments. OpenAI is digesting Stargate, Meta is funding Llama recovery, and every lab is fighting for the same scarce power and packaging capacity. By securing tens of billions in compute now, on terms that do not dilute existing shareholders, Anthropic converts financial engineering into a training-run advantage. The labs that master creative financing, not just model architecture, may end up with the larger clusters, and larger clusters have so far been the most reliable predictor of frontier capability.

The third implication is structural risk migration. By routing the purchase through an off-balance-sheet vehicle and a lease, Anthropic keeps $36 billion of debt away from its own financials right as public investors begin to scrutinize them ahead of an IPO. The risk does not vanish. It moves to Apollo, Blackstone, the note buyers, and ultimately Broadcom's guarantee. Who actually eats the loss if TPUs depreciate faster than the model assumes is the question that determines whether this is clever engineering or a hidden fault line.

The Competitive Landscape

Every frontier lab now faces the same compute-financing problem, and each is solving it differently. OpenAI leaned on the Stargate buildout and a record private raise plus Microsoft's cloud. xAI tapped SpaceX-adjacent capital and its own megarounds. Meta and Google fund compute directly off some of the strongest balance sheets in corporate history. Anthropic, lacking a trillion-dollar parent, has had to financialize the hardware itself, and in doing so it may have created the template smaller labs copy next.

Broadcom is the quiet winner threaded through the entire transaction. It co-develops Google's TPUs, and now it is also underwriting their residual value. That dual role turns Broadcom into something between a chip supplier and a credit insurer for the AI buildout, a position that deepens its leverage over both Google and the labs that buy Google silicon. For a company whose data-center revenue is already compounding past $100 billion in annualized terms, becoming the backstop of last resort for AI compute debt is a remarkable expansion of influence.

Google itself gains something subtler than a hardware sale. Every TPU that Anthropic runs is a TPU not bought from Nvidia, and a public, $36 billion endorsement of Google silicon is the strongest marketing the company's chip group has ever received. It pressures Microsoft and Amazon, whose own custom accelerators, Maia and Trainium, have struggled to win frontier training workloads at this scale. The lab that many viewed as Google's rival is now the anchor tenant validating Google's chips for the rest of the market, an irony that reshapes how every hyperscaler pitches its silicon to the next wave of AI startups.

The cleanest historical parallel is not the dot-com fiber glut but aircraft finance. Airlines almost never buy planes outright; they use special-purpose vehicles and enhanced equipment trust certificates, leasing jets whose residual value is guaranteed by manufacturers and insurers. That market works because aircraft hold value for decades. The entire Anthropic structure imports this playbook wholesale, with one enormous difference: a wide-body jet flies for thirty years, while a cutting-edge AI accelerator can be eclipsed by a better chip in eighteen months.

Hidden Insight: AI Compute Is Being Securitized Before Anyone Proved It Holds Value

Here is what almost no one is saying clearly. This deal does not just finance chips. It quietly asserts that a specialized AI accelerator is a durable, collateral-grade asset, and the entire $36 billion edifice rests on that unproven claim. Aircraft, real estate, and shipping containers became financeable because decades of data showed predictable depreciation curves. There is no such history for TPUs. The market is being asked to underwrite residual value for an asset class that is barely three years old and improving at a pace that makes last year's hardware look obsolete.

The residual-value support agreement is the tell. Lenders demanded that Broadcom, not Anthropic, stand behind roughly $31 billion of the senior debt, because Anthropic's own credit as a cash-burning, pre-IPO startup could not carry it. That detail reveals the real architecture: this is less a bet on Anthropic and more a bet on Broadcom's willingness to guarantee that Google's chips retain worth. The creditworthiness of the AI buildout is being borrowed from a semiconductor vendor's balance sheet, which is a stranger and more fragile arrangement than the clean headline suggests.

The bear case, however, is straightforward and worth stating plainly. If a dramatically better accelerator ships in 2027, or if AI demand growth slows even modestly, the resale value of today's TPUs could collapse well below the lease assumptions, and the residual-value guarantees become real liabilities rather than paper comfort. Critics argue that this is circular financing dressed in structured-finance respectability: chip vendors guarantee chip value so lenders fund chip purchases that enrich chip vendors. The risk the market is underpricing is correlation. In a downturn, chip values, AI revenue, and lab solvency all fall together, and the guarantees that look bulletproof today would be tested all at once.

There is also a governance dimension that the structured-finance framing obscures. An SPV that owns billions in chips and leases them to a single tenant creates a tangle of incentives among Apollo, Blackstone, Broadcom, Google, and Anthropic, each of which has a different view of what should happen if the lab stumbles. Aircraft leasing solved these conflicts over fifty years of bankruptcies and repossessions that built a body of law. AI compute has no such precedent, no established repossession market, and no playbook for what a lender even does with 200,000 specialized accelerators if it ever has to seize them. The legal scaffolding is being invented in real time around tens of billions of dollars.

Zoom out and the deeper signal becomes visible. Private credit has quietly become the shadow banking system of the AI boom. When equity markets eventually tighten, debt structured like this is what keeps the buildout going, and it concentrates AI infrastructure risk inside Apollo, Blackstone, and the pension funds and insurers that buy their notes. The next financial-stability conversation about AI may not be about model safety at all. It may be about how many retirement portfolios are now indirectly long the resale price of a Google chip.

What to Watch Next

In the next 30 days, watch whether the deal closes at the full $36 billion or shrinks during syndication, and track the pricing on the A2 and B notes. If lenders demand wider spreads than the initial terms, that is the market signaling doubt about TPU residual value. Also watch for any credit rating on the senior tranches; a public rating would tell you exactly how agencies model the depreciation of AI silicon, which is information the entire industry currently lacks.

Over the next 90 days, the key marker is Anthropic's S-1 going public and the IPO timeline firming up. Pay close attention to how the prospectus discloses this off-balance-sheet vehicle, because securities lawyers and short sellers will dissect the lease obligations. Watch too for copycat structures: if OpenAI, xAI, or Mistral announce their own chip-financing SPVs, it confirms that debt-funded compute has become the new default rather than a one-off Anthropic maneuver.

By the 180-day mark, the real test is whether a secondary market for used TPUs actually develops and at what prices, because that is the only thing that validates or destroys the residual-value math. Track Broadcom's disclosures for any reserves taken against its guarantees, and watch regulators and the rating agencies for early commentary on private-credit concentration in AI. If Washington or the Fed begins asking how exposed insurers and pensions are to this paper, the story shifts from financial innovation to systemic concern faster than most expect.

Finally, watch the reaction from the other side of the table. Insurers and pension funds are the ultimate buyers of investment-grade private credit, and their appetite for AI-linked paper will reveal how the conservative end of finance views this boom. If allocators quietly cap their exposure to chip-backed notes, future deals get more expensive and the debt-funded scaling model slows on its own. If they pile in chasing yield, the structure spreads across the industry and the systemic stakes climb with every new SPV. The behavior of the least glamorous institutions in finance will tell you more about the durability of the AI buildout than any model benchmark.

Anthropic did not just borrow $36 billion. It convinced the market that a chip which may be obsolete in eighteen months is collateral as solid as a skyscraper.

---