Blackstone 36B Loan Builds Anthropic AI Chip Army 2026

Anthropic is about to take on more debt for computer chips than most countries borrow in a year, and almost none of it will show up on its balance sheet. Blackstone and Apollo are assembling a $36 billion private credit package to buy Google TPUs and lease them back to the AI lab, a structure that lets the borrowing sit anywhere but on Anthropic's own books. For a company that just raised $65 billion in equity at a near-trillion-dollar valuation, the more revealing number is this one, because it shows how the AI buildout is really being paid for.

What Actually Happened

Apollo Global Management and Blackstone are shopping a $36 billion debt deal to a wider pool of investors, with both firms planning to keep large portions of the loan on their own books rather than syndicate all of it away. The money will not flow to Anthropic directly. Instead, the borrowed capital buys Google's custom tensor processing units, the TPUs that Anthropic agreed to deploy under a deal Broadcom helped broker, and a dedicated leasing vehicle then rents that hardware back to the lab across data centers in New York, Texas, Louisiana, and Indiana. Anthropic gets the compute; the lenders get the paper.

The mechanics matter because they keep tens of billions in obligations off Anthropic's financial statements. The chips are owned by the financing vehicle, Anthropic pays lease payments as an operating expense, and the debt belongs to the lenders and their co-investors. Broadcom, which co-designs Google's TPUs, is backstopping payments on the largest tranches, a credit enhancement that makes the senior debt palatable to insurers and pension funds that would never lend to a money-losing startup directly. Investors were asked to submit orders this week, with the transaction expected to close as early as next week, an aggressive timeline that signals heavy demand.

This is, by the accounting of the people arranging it, one of the largest private credit deals ever assembled and the biggest chip-financing debt transaction in history. It dwarfs the leveraged buyouts that defined private equity a decade ago and exceeds the debt raised for most national infrastructure projects. The structure borrows directly from how telecom towers, aircraft fleets, and power plants get financed: an asset with predictable cash flows, wrapped in a special-purpose vehicle, funded with debt that the operating company never has to claim as its own. What was once an exotic engineering technique for airlines is now the financing backbone of frontier AI.

Why This Matters More Than People Think

The headline most outlets ran was the dollar figure. The real story is that AI compute has officially become an asset class. When Blackstone and Apollo are willing to put $36 billion of their own and their clients' money behind GPUs and TPUs, they are betting that AI accelerators now behave like infrastructure: long-lived, cash-generating, and financeable against contracted revenue. That is a profound shift from 2023, when the same chips were treated as fast-depreciating gear that no sane lender would touch without a personal guarantee. In three years, silicon went from collateral nobody wanted to collateral the largest credit funds compete to underwrite.

For Anthropic, the off-balance-sheet treatment is the entire point. The company is preparing an IPO at a $965 billion valuation, and public-market investors scrutinize debt loads with a severity that private rounds never imposed. By routing the chip purchases through a leasing vehicle, Anthropic keeps its reported leverage low while still controlling the compute it needs to train and serve models like Claude Opus 4.8. It gets the hardware without the balance-sheet penalty, and it preserves the clean financials that command a trillion-dollar story. A debt-laden balance sheet at IPO would invite questions; an operating lease tucked into the footnotes does not.

The broader implication is that the AI buildout has outgrown equity. No venture round, however large, can fund the hundreds of gigawatts of data center capacity the frontier labs say they need. Equity is expensive and dilutive, and there is only so much of it before founders and early backers are washed out. Debt, secured against chips with a clear resale market and lease payments from a fast-growing tenant, is how the next phase gets paid for. The capital markets are now doing for AI what they once did for railroads, fiber, and cell towers, turning a speculative buildout into a financeable one.

There is also a signal for the rest of the industry buried in the timing. Anthropic is doing this on the eve of going public, which means it has chosen leverage over additional dilution at the exact moment its equity is most valuable. That is a confident bet that its revenue will compound faster than its lease obligations. If it is right, the structure looks like financial genius. If demand for Claude plateaus, those fixed lease payments become an anchor that equity never would have been.

The Competitive Landscape

Anthropic is not alone in reaching for leverage, and that is the point. OpenAI has stitched together its own web of compute commitments, including arrangements with Oracle and a sprawling set of data center partners, and Meta has signaled it will raise its 2026 AI capital expenditure toward $145 billion. Microsoft, Amazon, and Google fund their buildouts from operating cash flow, an advantage the pure-play labs simply do not have. For Anthropic and OpenAI, structured debt is the only way to match hyperscaler spending without diluting founders and early backers into oblivion, which makes financing innovation as strategic as model innovation.

The named players here, Apollo and Blackstone, are not AI companies at all. They are the largest alternative asset managers in the world, and they have spent years building private credit arms precisely for moments like this. Apollo manages credit at a scale that rivals major banks, much of it tied to its insurance affiliate Athene, and Blackstone's credit and insurance platform has become its fastest-growing business. By underwriting AI infrastructure, they are extending the same playbook they ran on energy pipelines and commercial real estate into the hottest asset class of the decade, and collecting fees that dwarf anything in traditional lending.

The historical parallel is the telecom boom of the late 1990s, when companies like WorldCom and Global Crossing borrowed enormous sums to lay fiber against projected demand. Some of that debt funded infrastructure the world genuinely needed, and some funded a glut that took a decade to absorb and bankrupted the borrowers. The difference this time is that AI demand is already manifest in revenue: Anthropic's run-rate has climbed past $47 billion by some accounts, a level of actual cash the fiber barons never had at the moment they borrowed. Whether that revenue grows fast enough to service $36 billion in chip debt is the question every credit committee had to answer before signing.

There is also a competitive twist few have noticed. By financing through Google's TPUs rather than Nvidia GPUs, Anthropic is diversifying away from the chip that powers most of its rivals, and Google is using cheap capital to lock a marquee customer onto its silicon for years. Nvidia, which sells into nearly every other lab and even invests in many of them, now faces a credible scenario where the second-largest frontier lab runs a growing share of its compute on a competitor's accelerators. The financing is not just a balance-sheet maneuver; it is a multi-year hardware allegiance, and it tilts the long-running Nvidia-versus-custom-silicon contest in Google's favor at the precise moment that contest matters most.

Hidden Insight: The Real Risk Is Hidden in the Lease, Not the Loan

Everyone is staring at the $36 billion debt figure, but the more dangerous variable is the depreciation curve on the chips themselves. The entire structure assumes TPUs hold enough value over the lease term that the financing vehicle can be repaid even if Anthropic stumbles. Yet AI accelerators are improving so fast that today's flagship silicon can lose much of its competitive value in two to three years. If Google ships a TPU generation that is twice as efficient, the leased hardware underpinning this deal becomes a stranded asset, and the lenders' collateral evaporates faster than the loan amortizes. The math only works if either the chips stay competitive or Anthropic stays solvent, and ideally both.

The bear case, however, is straightforward and worth stating plainly: this is vendor-adjacent financing dressed as independent credit. Google designs the chips, Broadcom co-designs and backstops the senior tranches, Anthropic is the tenant, and Google is also a major equity investor in Anthropic, having committed sums reported in the tens of billions. The circularity is uncomfortable. A chipmaker guaranteeing the debt that buys its own chips, leased to a company it partly owns, is the kind of structure that looks brilliant in a boom and incestuous in a bust. Critics argue this is exactly how vendor financing inflated the dot-com telecom bubble, where suppliers lent customers the money to buy supplier equipment and booked the resulting purchases as revenue.

There is a second, quieter insight. Off-balance-sheet treatment is a feature for Anthropic today and a liability for public investors tomorrow. When the company files its S-1, sophisticated analysts will reconstruct the true economic leverage regardless of where the debt formally sits, the same way they learned to read airline and shipping balance sheets through operating leases. The accounting can move the liability off the page, but it cannot move the cash obligation. Anthropic will owe lease payments whether Claude's revenue grows or stalls, and those payments are fixed costs in a business whose pricing power is anything but guaranteed in a market where DeepSeek and others keep cutting prices.

The deepest signal is about who now controls the AI economy. The frontier was supposed to belong to researchers and the labs that employ them. Increasingly it belongs to whoever can finance the compute, and that is a small club of asset managers, chip designers, and cloud providers. When a model lab needs Blackstone's appetite to afford its own training runs, the balance of power has quietly shifted from the people who build the models to the people who build the financing. That dependency does not show up in any benchmark, but it shapes every strategic decision the labs will make for the next five years, from which chips they buy to how aggressively they can price.

What to Watch Next

In the next 30 days, watch the final terms and the order book. If Apollo and Blackstone have to widen the spread to fill the deal, it signals that even sophisticated credit investors are nervous about AI hardware as collateral. A tight, oversubscribed close at favorable rates would confirm that compute has cleared the bar to become genuinely financeable. Watch too whether other labs, OpenAI in particular, rush to copy the structure now that the template exists and the rating logic has been road-tested by the two biggest names in private credit.

Over 90 days, the key marker is Anthropic's IPO disclosure. The S-1 will reveal how the company presents these lease obligations and whether regulators or auditors push back on the off-balance-sheet framing. Pay attention to the discussion of compute commitments and minimum payments, buried deep in the filing, because that is where the true leverage lives. The gap between Anthropic's reported debt and its real fixed obligations will tell investors how much of the trillion-dollar valuation rests on operating performance and how much rests on accounting presentation.

On a 180-day horizon, the question is whether this becomes the standard financing rail for AI or a one-off enabled by Broadcom's backstop. If insurers and pension funds get comfortable holding AI infrastructure debt, the buildout accelerates and a multi-hundred-billion-dollar lending market emerges almost overnight. If the first stress, a model that disappoints or a chip generation that obsoletes the collateral early, spooks lenders, the spigot closes fast. The next two quarters of TPU and GPU performance data will quietly decide which path the capital takes, and with it the pace of the entire AI buildout.

AI compute just became an asset class, and the people who finance the chips now have more leverage over the frontier than the people who design the models.

---