On February 12, Anthropic closed a $30 billion Series G. Post-money valuation: $380 billion. That makes it the second-largest private tech financing ever, behind only OpenAI's $40 billion raise in April 2025. GIC and Coatue led the round. D.E. Shaw Ventures, Dragoneer, Founders Fund, ICONIQ and MGX co-led. Microsoft and Nvidia contributed portions of investment commitments they had already announced.
The company says run-rate revenue hit $14 billion. Claude Code, the coding tool, supposedly generates $2.5 billion annualised. Both figures were stated by Anthropic to investors and reported by Reuters and Axios. They have not been audited.
"Run-rate" means you take your best recent month, multiply by twelve, and call it annual revenue. If there was a usage spike around the Opus 4.6 launch or a big enterprise contract kicking in, that number could be painting a rosier picture than what Anthropic will actually collect across 2026. It is a projection wearing a trenchcoat pretending to be a fact. Even at face value, $380 billion on $14 billion run-rate works out to about 27 times sales. Microsoft and Alphabet trade at roughly 9 to 10 times their actual, audited, GAAP revenue. Whether 27x is reasonable depends entirely on whether growth continues at this rate, and since Anthropic is private, there's no way to verify the trajectory from outside.
What we do know: The Information reported in February 2025 that Anthropic burned $5.6 billion in cash in 2024, with revenue that year likely in the range of $850 million to $1.2 billion based on its reported annualised run-rate of $1 billion at the end of 2024. That is a company spending roughly five dollars for every dollar it earns. Anthropic projected it would nearly halve its cash burn in 2025 and has told investors it expects to stop burning cash by 2027 and break even by 2028. Those are pitchbook projections designed to get people to hand over $30 billion. To be fair, it worked.
The Infrastructure Loop
Here is the part that should bother people more than it does.
Anthropic has committed to buying large volumes of compute from Microsoft Azure, Amazon Web Services and Nvidia hardware. Ed Zitron's reporting at Where's Your Ed At has documented that Anthropic spent more than 100 per cent of its estimated revenue on Amazon Web Services through September 2025 alone. If that pattern holds even roughly for the new round, the majority of the $30 billion is flowing straight back to infrastructure providers. The headline figure wildly overstates how much fresh capital actually enters the system.
Nvidia invests in Anthropic. Anthropic spends the money on Nvidia chips. That spending shows up as Nvidia revenue, which supports Nvidia's stock price, which supports Nvidia's ability to invest more. Microsoft invests in Anthropic. Anthropic spends on Azure. Microsoft books the revenue. The Register called it tossing money onto a furnace. This is not fraud. It is not even unusual for venture-backed industries. But every player in this ecosystem is deeply dependent on every other player continuing to spend. Pull one thread and the whole thing goes.
The Debt Everywhere
This is where it stops being a story about one company's funding round and starts being a story about systemic risk.
Building AI data centres costs money that even the richest companies on earth don't want to pay out of pocket. So they're borrowing it. From everywhere. Corporate bonds, junk debt, private credit, securitised loan pools. Bloomberg reported that AI-related financing has become "all-consuming for debt markets." AI-related investments now account for roughly 30 per cent of net issuance in the US dollar-denominated investment-grade corporate bond market.
The scale is staggering. In 2025 alone, AI-related companies tapped debt markets for at least $200 billion, and that's likely an undercount since many deals are private. Data centre financing nearly doubled to $182 billion. Meta sold $30 billion in public bonds in a single offering in October 2025, drawing $125 billion in investor demand. Alphabet raised nearly $32 billion in under 24 hours last week, including a 100-year bond, the first century bond from a tech firm since Motorola in 1997. That came one week after Alphabet announced it would nearly double its capex to $185 billion.
Morgan Stanley estimates hyperscalers will borrow around $400 billion in 2026, more than double 2025 levels, and projects they will spend about $3 trillion on data centre infrastructure through 2028, with roughly half financed through debt. JPMorgan projects the tech sector may need $1.5 trillion in new debt over the next few years.
And here's where it gets properly concerning. Data centre loans are being pooled into tranches and sold to bond investors as commercial mortgage-backed securities and asset-backed securities. If the phrase "securitised loans being sliced into tranches" sounds familiar, there is a reason. The structure is the same one used in the mortgage-backed securities that contributed to the 2008 financial crisis: package loans together, slice them by risk profile, sell the slices to investors, and assume the underlying cash flows will keep coming. The asset class is different (data centres instead of houses), but the mechanic is identical. Loans bundled, risk dispersed, everyone assumes someone else is doing the due diligence.
The tenants of these data centres are often AI companies that are themselves unprofitable. The AI companies are funded by venture capital. The venture capital comes partly from the same hyperscalers building the data centres. Everyone in this chain is either borrowing money or spending someone else's. The whole thing only works if AI revenue eventually materialises at a scale that justifies $3 trillion in buildout costs. This debt is not sitting in some isolated corner of the financial system. It is spreading across corporate bond portfolios, structured credit products, private credit funds and convertible bonds. It is sitting in pension funds and retirement accounts belonging to people who probably have no idea they're exposed to the question of whether AI coding tools will generate enough revenue to pay the electricity bill at a data centre in Texas.
CoreWeave: The Canary
If you want to know what financial distress looks like in the AI infrastructure space, look at CoreWeave.
CoreWeave went public in March 2025 at $40 per share, peaked near $180, and has since fallen over 60 per cent. Their balance sheet is genuinely alarming: $14 billion in total debt, with over 50 per cent tied to GPUs. Interest rates on their loans run between 9 and 15 per cent. In Q3 2025, interest expense hit $310.6 million against GAAP operating income of just $51.9 million. That is a coverage ratio of 0.17x. For the first nine months of 2025, interest expense hit $841.4 million, up 297 per cent year-over-year. Credit default swap spreads have more than doubled since October, which means bond investors are actively pricing in higher default risk.
They've already had to amend their loan covenants as of December 31, 2025, pushing back the dates when lenders can test whether the company meets its financial targets. They reduced their minimum liquidity requirement. They gave themselves unlimited "equity cures" for covenant failures through October 2026. That is the language of a company negotiating with its creditors to stay alive.
The really precarious bit: they can't stop spending. CoreWeave's entire model is to borrow money, buy Nvidia GPUs, put them in data centres, and rent compute capacity under long-term contracts. They exited Q3 2025 with $55.6 billion in revenue backlog. To fulfil those contracts, they need to keep building. If they stop buying GPUs, they can't deliver on the backlog. If the backlog evaporates, the story that justifies the borrowing collapses. If the borrowing collapses, they can't service the existing debt. So they're planning to "more than double" 2025 capex in 2026, implying north of $24 billion. They're spending $2.35 to $2.77 for every dollar of revenue they earn.
Read that again. A company whose operating income covers a sixth of its interest payments is planning to spend over $24 billion this year. Not because they're flush with cash, but because stopping would kill them faster than continuing. They're funding it with new debt facilities, Nvidia's $2 billion equity injection, a $2.7 billion convertible note issued in December 2025, and whatever else they can find. Every new obligation pays for the last one.
Meanwhile, GPUs already purchased are sitting in warehouses. Depreciating. While interest on the debt used to buy them keeps accruing. Revenue that was supposed to come in Q4 2025 got pushed to 2026 because third-party data centre developers couldn't deliver "powered shells" on time. Every day those chips sit idle, CoreWeave bleeds money. And the loans themselves are collateralised by GPUs whose rental rates have been falling. The collateral backing the debt is shrinking in value while the debt stays the same size. This is what a financial treadmill looks like. The company needs continuous access to capital markets to survive. Not to grow. To survive. If CoreWeave hits a wall, it won't just be their problem. They're a significant data centre tenant, a major Nvidia customer, and their debt is spread across lenders, private credit funds and convertible bond holders. A default would send shockwaves through AI infrastructure financing.
Has the Bubble Popped?
Not yet. But the early signals are here. J.P. Morgan estimates the software sector has shed roughly $2 trillion in market cap from its October 2025 peak, the largest non-recessionary 12-month drawdown in over 30 years. Apollo cut its direct lending funds' software exposure almost by half in 2025, spooked by AI disruption risk. Oracle is sitting on over $100 billion in debt and reportedly delayed data centre deliveries for OpenAI from 2027 to 2028, which they deny but Bloomberg reported citing people familiar with the projects. Morgan Stanley suggested investors short the stock. And a 100-year Google bond now exists, which is the kind of financial creativity that turns up when normal financing isn't cutting it anymore.
What hasn't happened yet: no major hyperscaler has cut capex. OpenAI and Anthropic can still raise money. No high-profile data centre loan has defaulted. No AI company IPO has forced real numbers into the open and tanked on contact with reality.
The triggers to watch, in rough order of likelihood: CoreWeave restructuring or defaulting, which feels like a matter of when rather than if given 0.17x interest coverage and active covenant renegotiation. A mid-tier AI startup failing to raise its next round, signalling that investor appetite has limits (several alternative asset managers fell 3 to 11 per cent in a single session on concerns that software weakness would cause credit problems). Data centre loan delinquencies showing up first in private credit and then in securitised products. A hyperscaler cutting capex, which Bank of America's chief investment strategist identified as the event that would flip the narrative. And an AI company IPO, which Anthropic is reportedly preparing for. The moment one of these companies has to file an S-1 and publish audited financials, the gap between the pitch and the reality becomes public. OpenAI projects it will burn through $115 billion in cash between now and 2029 according to Wall Street Journal reporting of internal documents. Those numbers in an SEC filing will read very differently than they do in a pitch deck.
None of these individually would be catastrophic. But they don't need to happen individually. The whole structure is interdependent. CoreWeave struggles, lenders get nervous about data centre debt, financing costs rise for new projects, construction slows, Nvidia sells fewer chips, hyperscaler investments in AI companies look less justified, the next funding round gets harder, the AI company can't pay its data centre lease, more debt goes bad. Each domino tips the next.
I don't know if this thing collapses in six months or holds together for three more years. But a system where every participant is either borrowing money or spending someone else's, where the debt is being securitised and spread across the financial system, and where the whole thing depends on revenue that largely doesn't exist yet, is fragile. And fragile systems don't send warning letters before they break.
This analysis draws on SEC filings, company announcements, and reporting from The Information, Reuters, Axios, Bloomberg, CNBC, The Register, Fortune, and specialist financial analysis from Where's Your Ed At (Ed Zitron), Level Headed Investing, and TradingKey. CoreWeave financial data from 10-Q filings and earnings call transcripts. Debt market figures from BBN Times, Portfolio Adviser, and bank research notes from Morgan Stanley, J.P. Morgan and Bank of America. Anthropic revenue and loss data from The Information, Sacra, and company disclosures. All links verified as of February 16, 2026.