Is CoreWeave the Silicon Valley Bank of 2026?

Remember Silicon Valley Bank? Its 2023 collapse triggered fears of a wave of US regional bank failures, which would stymie credit growth, cause a recession and force the Federal Reserve to cut interest rates. Whoops.

Instead, only a handful of small, similarly badly run banks imploded, economic growth remained robust, and the Fed actually raised rates that month — and twice more in the coming months.

The shift in expectations played out violently in markets. The 10-year Treasury yield round-tripped from over 4 per cent to a low of 3.3 per cent, before shooting back to nearly 5 per cent when investors realised the economy had entirely shrugged off the SVB mess. The KBW Bank Index lost as much as a third of its value, but ended the year almost flat.

Ajay Rajadhyaksha at Barclays reckons that something similar is now playing out with CoreWeave — investors are hysterically over-extrapolating from a single, almost unique case and wrongly questioning the entire AI investment spree.

Rajadhyaksha is specifically arguing that people like Chamath Palihapitiya and Elon Musk are misunderstanding CoreWeave’s convoluted balance sheet and seizing on the sell-off in its $2bn unsecured bonds.

Yikes— Elon Musk (@elonmusk) October 13, 2026

Yikes

This is silly, Rajadhyaksha argues, because they mostly just reflect how lowly these specific unsecured bonds rank in CoreWeave’s much larger $35bn debt stack — below the data centre operator’s “secured delayed-draw term loans”, its convertible bonds and billions of dollars' worth of lease obligations.

Basically, if CoreWeave gets into real trouble, these unsecured bonds will probably get wiped out, which makes them particularly vulnerable to even mild investor angst. Moreover, “even if the bonds might be rationally priced for their own recovery prospects they could still be being completely wrong as a signal about the viability of the entire AI CapEx cycle,” Rajadhyaksha writes.

Here’s the more detailed Barclays breakdown of CoreWeave’s balance sheet, with Alphaville’s emphasis in bold below:

CoreWeave funds itself in three tiers, and they have different risk profiles. The workhorse is secured DDTLs. The bulk of CoreWeave’s financing sits in ring-fenced special-purpose vehicles, each backed by specific GPU hardware, a customer contract, and the associated data center lease. When Meta signs a multi-year take-or-pay contract and that contract is assigned to an SPV, the bank is lending against Meta’s promise to pay.The progression of recent deals tells you the market is still open but getting more expensive. DDTL 4.0 in March – $8.5bn, backed by Meta, priced at SOFR+225bp, rated A3 by Moody’s. That’s investment grade. DDTL 5.0 in May – $3.1bn, SOFR+450bp. DDTL 5.5 in August – $2.6bn, SOFR+550bp. The spread is widening. But the market is open.The equity bridge is the convertibles. About $9bn issued over the past year, at coupons between 1.75% and 2.875%. These get less attractive as the stock falls – the conversion becomes more dilutive – but the existing paper is locked in. The September deal was upsized from $3bn to $3.7bn, which suggests there was demand. But after last week’s price moves, this channel is likely under pressure.We see the unsecured notes as the final piece of the puzzle. In a liquidation, there are tens of billions of dollars of claims ahead of it. A 15% yield on this paper tells you something about this paper. It doesn’t tell you the company is about to stop operations.If a company has a $35bn mortgage and much smaller credit card balance, that tells you something about the credit card. It doesn’t tell you that all the mortgage borrowers in that city are in trouble.

CoreWeave funds itself in three tiers, and they have different risk profiles. The workhorse is secured DDTLs. The bulk of CoreWeave’s financing sits in ring-fenced special-purpose vehicles, each backed by specific GPU hardware, a customer contract, and the associated data center lease. When Meta signs a multi-year take-or-pay contract and that contract is assigned to an SPV, the bank is lending against Meta’s promise to pay.

The progression of recent deals tells you the market is still open but getting more expensive. DDTL 4.0 in March – $8.5bn, backed by Meta, priced at SOFR+225bp, rated A3 by Moody’s. That’s investment grade. DDTL 5.0 in May – $3.1bn, SOFR+450bp. DDTL 5.5 in August – $2.6bn, SOFR+550bp. The spread is widening. But the market is open.

The equity bridge is the convertibles. About $9bn issued over the past year, at coupons between 1.75% and 2.875%. These get less attractive as the stock falls – the conversion becomes more dilutive – but the existing paper is locked in. The September deal was upsized from $3bn to $3.7bn, which suggests there was demand. But after last week’s price moves, this channel is likely under pressure.

We see the unsecured notes as the final piece of the puzzle. In a liquidation, there are tens of billions of dollars of claims ahead of it. A 15% yield on this paper tells you something about this paper. It doesn’t tell you the company is about to stop operations.

If a company has a $35bn mortgage and much smaller credit card balance, that tells you something about the credit card. It doesn’t tell you that all the mortgage borrowers in that city are in trouble.

Rajadhyaksha reckons that Coreweave’s order backlog is alone strong enough to finance its capex plans, and argues that markets simply “have a tendency to find the smallest, most volatile instrument in a complex system and treat it as the barometer for the whole thing”.

Even if CoreWeave has to halt all its investment plans, the impact on the broader AI build-out would be minimal, Barclays’ chair of global research argues:

CoreWeave’s 2026 CapEx guidance is $35-39bn. That sounds like a big number – and it is, for a single company. But set it against the total. The five largest hyperscalers – Microsoft, Meta, Alphabet, Amazon, and Oracle – are spending a combined $725-800bn on AI infrastructure this year, with estimates north of a trillion dollars for 2027. CoreWeave’s spend is roughly 4-5% of that total. The other neoclouds – Nebius, Crusoe, Lambda – are much smaller still, collectively 2%.So in the most extreme bear scenario – CoreWeave stops spending entirely – we lose about 4-5% of total AI capex. If its spend merely halves, which might be a more realistic downside, the impact is 2%. On a trillion-dollar-plus 2027 estimate, that’s $20bn. Not nothing. But not the cycle.The risk of course is contagion. The idea that if Coreweave is in trouble, that means a Google cannot finance its data center next. We think this is a very big stretch - extrapolating from a neocloud to one of the most profitable companies in the world. Moreover, much of hyperscaler spending is still financed by free cash flow. Microsoft, Meta, Alphabet, and Amazon are not uber-dependent on the unsecured bond market, even if they would like to – and will need to – issue more. The chain of logic the market is implicitly running toward – CoreWeave’s bonds at 15%, therefore AI infrastructure financing is broken, therefore the capex cycle is over – doesn’t hold up. CoreWeave’s unsecured bonds are a tiny slice of CoreWeave’s capital structure. CoreWeave’s capex is a tiny slice of total AI CapEx. And the hyperscalers doing the vast majority of the spending don’t finance it the way CoreWeave does.

CoreWeave’s 2026 CapEx guidance is $35-39bn. That sounds like a big number – and it is, for a single company. But set it against the total. The five largest hyperscalers – Microsoft, Meta, Alphabet, Amazon, and Oracle – are spending a combined $725-800bn on AI infrastructure this year, with estimates north of a trillion dollars for 2027. CoreWeave’s spend is roughly 4-5% of that total. The other neoclouds – Nebius, Crusoe, Lambda – are much smaller still, collectively 2%.

So in the most extreme bear scenario – CoreWeave stops spending entirely – we lose about 4-5% of total AI capex. If its spend merely halves, which might be a more realistic downside, the impact is 2%. On a trillion-dollar-plus 2027 estimate, that’s $20bn. Not nothing. But not the cycle.

The risk of course is contagion. The idea that if Coreweave is in trouble, that means a Google cannot finance its data center next. We think this is a very big stretch - extrapolating from a neocloud to one of the most profitable companies in the world. Moreover, much of hyperscaler spending is still financed by free cash flow. Microsoft, Meta, Alphabet, and Amazon are not uber-dependent on the unsecured bond market, even if they would like to – and will need to – issue more.

The chain of logic the market is implicitly running toward – CoreWeave’s bonds at 15%, therefore AI infrastructure financing is broken, therefore the capex cycle is over – doesn’t hold up. CoreWeave’s unsecured bonds are a tiny slice of CoreWeave’s capital structure. CoreWeave’s capex is a tiny slice of total AI CapEx. And the hyperscalers doing the vast majority of the spending don’t finance it the way CoreWeave does.

Alphaville has a lot of sympathy with this argument. But we think Rajadhyaksha is overselling it a little, for three reasons.

Firstly, given the dominant importance of the AI trade to virtually everything going on at the moment — even Goldman Sachs says the entire global stock market has virtually become “One Big Trade” — it is entirely natural that investors react skittishly to even oblique hints that it may be faltering.

Secondly, CoreWeave is not some small and obscure security, like the $KRE bank stock ETF that Rajadhyaksha uses as the example of the overdone US financials rout of 2023. CoreWeave is an exemplar of the AI build-out that is heavily backed by the likes of Nvidia, Jane Street and Magnetar Capital, and its biggest customers are reportedly Microsoft and OpenAI. Sure, it’s not Nvidia or Meta, or even Oracle, but it is deeply enmeshed in the AI ecosystem.

Thirdly, the reaction has hardly been that violent or broad. The Nasdaq 100 only fell 5.5 per cent last week, and has started this week in strong form. Oracle’s bond yields have also jumped, but they fell for the other hyperscalers thanks to the Treasury market regaining its footing. This is not the wholesale panicky repricing that we saw back in March 2023.

So what does Rajadhyaksha reckon would be a more worrisome harbinger for the AI capex fiesta? Any sign that the cost of computing is collapsing. But that just isn’t happening yet, he argues:

CoreWeave’s business model – and the business model of every AI infrastructure company – is a spread business. Buy GPUs, rent them out, pocket the difference between the rental yield and the cost of financing. That spread holds as long as demand for compute exceeds supply and rental rates stay high enough to cover debt service, depreciation, and operating costs.If compute prices collapse – if the market for GPU hours falls far enough that the economics of the buildout no longer work – then the bears are right, and CoreWeave’s bonds are the canary in the coal mine.But so far, that is just not happening. TSMC’s results tell you chip demand is rising. ASML’s backlog tells you fab capacity is constrained for years. The hyperscaler capex numbers – $725-800bn in 2026, with estimates of a trillion dollars or more in 2027 – tell us the buyers of compute are not pulling back. CoreWeave’s own backlog is $104bn of contracted, take-or-pay revenue.Unless the spot and forward price of compute drops drastically – and nothing in the supply-chain data supports that conclusion – we believe last week’s equity pullback in AI infrastructure should blow over. Quality should reassert itself, even if investors are more wary about the fringes and suspicious of the neoclouds. Just as the 2023 recession never arrived, we do not see the AI capex cycle breaking. What broke was a small, subordinated corner of one company’s capital structure. Investors, as they did in March 2023, should not mistake the part for the whole.

CoreWeave’s business model – and the business model of every AI infrastructure company – is a spread business. Buy GPUs, rent them out, pocket the difference between the rental yield and the cost of financing. That spread holds as long as demand for compute exceeds supply and rental rates stay high enough to cover debt service, depreciation, and operating costs.

If compute prices collapse – if the market for GPU hours falls far enough that the economics of the buildout no longer work – then the bears are right, and CoreWeave’s bonds are the canary in the coal mine.

But so far, that is just not happening. TSMC’s results tell you chip demand is rising. ASML’s backlog tells you fab capacity is constrained for years. The hyperscaler capex numbers – $725-800bn in 2026, with estimates of a trillion dollars or more in 2027 – tell us the buyers of compute are not pulling back. CoreWeave’s own backlog is $104bn of contracted, take-or-pay revenue.

Unless the spot and forward price of compute drops drastically – and nothing in the supply-chain data supports that conclusion – we believe last week’s equity pullback in AI infrastructure should blow over. Quality should reassert itself, even if investors are more wary about the fringes and suspicious of the neoclouds. Just as the 2023 recession never arrived, we do not see the AI capex cycle breaking. What broke was a small, subordinated corner of one company’s capital structure. Investors, as they did in March 2023, should not mistake the part for the whole.

So far, the market’s response this week suggests that investors broadly agree with Rajadhyaksha, that this was a squall and nothing more. But the tangled economics of compute might be worth keeping an eye on.