Roula Khalaf, Editor of the FT, selects her favourite stories in this weekly newsletter.
Well, there it is:
CoreWeave is planning to slash the size of its initial public offering and bring in Nvidia as an anchor investor, another sign of wavering investor demand for artificial intelligence infrastructure on Wall Street.
The cloud computing provider will formally set the price of its shares later on Thursday and is expecting to pare back its offering to around $1.5bn, according to people close to the matter.
CoreWeave had initially targeted raising $4bn and dropped that figure to $2.7bn when it began a roadshow to generate interest for its shares last week.
No official details yet on float pricing and structure, including the size of Nvidia’s anchor, so things could still change. CNBC reported earlier that Nvidia’s fresh backing of CoreWeave would involve a $250mn order, presumably in addition to the $320mn of server time it agreed to buy in April 2023.
We write at length elsewhere about how reliant CoreWeave is on Nvidia, its sole GPU supplier, 5.97 per cent shareholder and second-biggest customer. We also mention how hard it’s been for CoreWeave’s team of 14 IPO advisers to convince the buy-side that its debt-burdened business model is sustainable.
On the one hand, bringing in Nvidia to shore up the IPO might be seen to deepen their relationship and guarantee early drops on new hardware that could provide a competitive advantage. On the other hand, it won’t make worries about concentration go away.
In an anonymous poll seen by FT Alphaville, RBC Capital Markets asked hedge-fund and long-only clients: “Does CoreWeave have a sustainable moat?” Ninety per cent voted no.
Here are a few of the clients’ explanations as to why:
Their moat is priority access to GPUs – that’s it
Capital/relationships are the barrier and won’t last
Near-term they have capacity which is needed but longer-term, no. Anyone can buy GPUs, put them into a cluster, and sell the capacity to larger players. Competing with the hyperscalers with deeper pocket books who are also doing this whole nother thing.
The business model is predicated on the scarcity of NVDA chips. If, and when, the market loosens a bit or a competing chip manufacturer ramps up, the need for their “conduit” business model will be less needed.
Equipment rental business with cost of capital being the only LT advantage . . .
In answer to “What is the least attractive financial aspect of CoreWeave’s financials?”, more than half of RBC survey respondents said “customer concentration” (meaning Microsoft and Nvidia). Respondents’ reasoning included:
CoreWeave’s largest customer [Microsoft] is publicly telling investors it no longer has any need for CoreWeave and will build its own datacenters from here on out
If this is truly ‘overflow’ capacity for MSFT, then this is a tough model to invest [in]
I don’t like that I would be investing in OpenAI by proxy and that feels like an investment that is a function of Sam Altman’s ability to raise capital, first MSFT, now SoftBank, then… Saudi Arabia? Further and further out the risk curve.
And under “What do you think investors are missing about the CoreWeave story?”, one money manager wrote:
NVDA is 15% of revs which they levered up to buy billions ($) of GPUs. Why does NVDA need to pay somebody to access their own GPUs? It is a gimmick to create competitive tension for GPUs outside of the hyperscalers to give NVDA pricing leverage. As big as CRWV is, it looks small relative to Stargate scale. Move downmarket will require customers hand-holding and features akin to a hyperscaler. That will be tough. In the meantime, the banks are racing to get the deal done and their bank loans refinanced with bonds before this story meets reality.
Official IPO pricing is due after the US closing bell, but it looks a lot like reality is already catching up.
Imagine a caravan maker. It sells caravans to a caravan park that only buys one type of caravan. The caravan park leases much of its land from another caravan park. The first caravan park has two big customers. One of the big customers is the caravan maker. The other big customer is the caravan maker’s biggest customer. The biggest customer of the second caravan park is the first caravan park.
Sorry, not caravans. GPUs.
As Tabby Kinder and Rob Smith wrote last week for MainFT:
CoreWeave . . . which leases computing capacity to tech groups building artificial intelligence models, is gearing up for the largest stock market debut of the year.
This week it revealed it was seeking to raise as much as $2.7bn in the share sale, valuing the business at $32bn. As the New Jersey-based group prepares to start an investor roadshow, it is attracting scrutiny for its huge debt burden, borrowing at high interest rates, and forthcoming maturities on billions of dollars of loans.
CoreWeave started out in 2017 as the side hustle of some traders at Hudson Ridge Asset Management, a defunct gas futures hedge fund. First it was an Ethereum miner that pivoted during the 2019 crypto crash to pay-per-hour 3D video rendering. The phrase “machine learning and AI” was added to CoreWeave’s blurb in November 2022, the same month OpenAI launched, and soon grew to consume the whole. Shortly after CoreWeave’s Series C funding round in May 2024, its website title changed from “The GPU Cloud” to “The AI Hyperscaler”.
MainFT’s coverage focuses on the Blackstone and Magnetar Capital-backed company’s $8bn of debt. The crux of the story is the WeWork-style mismatch between its assets and liabilities, along with some apparent carelessness around debt covenants:
CoreWeave . . . violated several key terms of a $7.6bn loan last year, triggering a series of so-called technical defaults.
[The company] disclosed in the exhibits to its IPO document that it had to ask its biggest lender Blackstone to amend the terms of the loan and “waive” these defaults in December.
While CoreWeave did not miss any payments under the loan facility, it made a slew of serious administrative errors, which stemmed from beginning to use the financing to expand into western Europe. This clashed with key terms that in effect restricted the debt’s collateral to the US.
But with CoreWeave due to price its IPO later today, there’s plenty more in the S-1 filing that deserves attention. Here’s a quick tour of other notable items.
The Nvidia thing:
Tim Bradshaw last year asked CoreWeave CEO Michael Intrator about the company’s reliance on Nvidia, its 5.97 per cent shareholder, key supplier and key customer. Intrator . . .
. . . batted off questions about whether prospective investors were concerned about backing a business that had raised capital from Nvidia, only to spend a significant portion of those funds on that company’s products.
“It’s such a crap narrative,” he said. “Nvidia invested $100mn. We’ve [raised] $12bn in debt and equity. It’s an inconsequential amount of money in the relative scale of the amount of infrastructure we’re buying.”
Crap narrative it may be, but let’s take a look at what the S-1 says about customer concentration:
We recognized an aggregate of approximately 77 per cent of our revenue from our top two customers for the year ended December 31, 2024.
And . . .
Our largest customer accounted for 16%, 35%, and 62% of our revenue for the years ended December 31, 2022, 2023, and 2024, respectively.
CoreWeave’s revenue was $1.9bn in 2024. Sixty-two per cent of $1.9bn is $1.18bn. That squares with its Microsoft Master Services Agreement (our bold):
In February 2023, we entered into a Master Services Agreement (the “Microsoft Master Services Agreement”) with Microsoft, pursuant to which we provide Microsoft with access to our infrastructure and platform services through fulfillment of reserved capacity orders submitted to us by Microsoft and as may be amended upon our and Microsoft’s mutual agreement. We have recognized revenue of $81 million and $1.2 billion for the years ended December 31, 2023 and 2024, respectively, pursuant to the Microsoft Master Services Agreement.
That leaves 15 per cent of $1.9bn, or $285mn. That’s not far off the 20-month number CoreWeave gives for its Nvidia contract:
In April 2023, we entered into a Master Services Agreement (the “Master Services Agreement”) with NVIDIA, a beneficial owner of more than 5% of our outstanding capital stock, pursuant to which we provide NVIDIA with our infrastructure and platform services through fulfillment of order forms submitted to us by NVIDIA. As of December 31, 2024, NVIDIA has paid us an aggregate of approximately $320 million pursuant to the Master Services Agreement and related order forms.
And since . . .
None of our other customers represented 10% or more of our revenue for the year ended December 31, 2024.
. . . it seems fair to conclude that CoreWeave’s second-biggest customer in 2024 was Nvidia — which makes the following line feel a bit incestuous:
[O]ur current customers have contractually specified our use of NVIDIA GPUs.
Much more on Nvidia later, but first . . .
The Core Scientific thing:
“We relentlessly and creatively explore additional opportunities to add power capacity”, says CoreWeave’s S-1. Nowhere is that better demonstrated than with Core Scientific.
Core Scientific is a publicly traded crypto miner that collapsed into bankruptcy protection in 2022 alongside its main customer, Celsius Network, whose founder/CEO Alexander Mashinsky last year pleaded guilty to fraud and market manipulation. (Core Scientific’s co-founder, the Viper Room nightclub co-owner and Fatburger promoter Darin Feinstein, stepped down as group co-chair in 2023.)
CoreWeave last year tried to buy Core Scientific. After its takeover proposal was rejected, CoreWeave announced several contracts to rent and modify Core Scientific’s rack space. At the 2024 year-end, Core Scientific’s data centres accounted for “more than 500MW” of CoreWeave’s approximately 1300MW of total capacity (72 per cent of which was not yet switched on).
Core Scientific, in a 2024 results presentation, says its contracts with CoreWeave last 12 years.
Core Scientific’s disclosures also reveal who funds the conversion. A footnote to the above graphic says CoreWeave is paying Core Scientific “up to $1.5mn per HPC [high-performance computing] MW of data centre build-out costs” to a value of about $750mn. In exchange, CoreWeave gets an up-to-50 per cent rebate on its hosting costs. A follow-on deal involves CoreWeave funding $104mn of capex for no hosting rebate; more on that below.
It’s hard to shake the impression that CoreWeave is sinking a lot of capital and wearing most of the risk. Core Scientific’s data centres will still be there long after CoreWeave’s chips are fried.
Yet the market gives Core Scientific an enterprise value of approximately 10 times EBIT — a deep discount to conventional real estate investment trusts, which probably reflects some uncertainty about its anchor tenant.
Meanwhile, CoreWeave’s syndicate of 14 IPO advisers had reportedly been aiming for approximately 15 times forward EBIT. So even if recent speculation proves accurate that the price range has moved down by around 20 per cent, it still looks pretty punchy.
The depreciation thing:
GPUs are quickly depreciating assets. Not only do they burn out, they’re constantly being superseded by new models. Massed Compute estimates value loss of 20 to 30 per cent a year. The investment case for an AI data centre hinges on the rental market growing fast enough to cover sunk costs before their hardware is obsolete.
Here’s how CoreWeave’s S-1 estimates the useful life of its property and equipment:
Technology equipment: 6 years
Software: 3-6 years
Data center equipment: 8-12 years
Furniture, fixtures, and other assets: 3-5 years
Six years in AI is an eternity. Nvidia’s server-grade V100 GPU cost around $10,000 in 2019 and can now be picked up for a few hundred dollars. It’s already four generations behind the times, with another generation due to arrive next year.
CoreWeave’s rapid expansion last year makes it a big bet on Hopper, Nvidia’s last-but-one architecture, which debuted in 2022. The S-1 doesn’t give a detailed breakdown of assets but says a majority of its GPUs use Hopper.
Some content could not load. Check your internet connection or browser settings.
Its expansion has been mirrored across the data centre industry, which has moved inside a year from a critical shortage of AI compute to a glut.
“I start to see the beginning of some kind of bubble,” Alibaba chair Joe Tsai told a conference this week. “I start to get worried when people are building data centres on spec. There are a number of people coming up, funds coming out, to raise billions or millions of capital.”
What’s probably happening is that hyperscalers are no longer compelled to sprint and establish competitive moats by training the biggest models, so don’t need to rent as much emergency capacity, while smaller operators are waiting to see where things land before committing funds. Here’s what Goldman Sachs (a CoreWeave IPO lead underwriter) says in a note dated March 24 that downgraded ratings and forecasts for several Taiwanese AI server makers:
‘Training’ server will remain the growth driver given the increasing need for computing power to upgrade advanced AI models, but the volume ramp up is slower than we previously expected due to the combined reasons of product transitioning and uncertainties of demand and supply. As the GPU platform is transiting to next generation in 2H25, shipment can potentially slow during the transition period. Uncertainties remain in production ramp up, given the complexity of full rack systems and there remains debates on the demand for intense computing power after the release of more efficient AI models like DeepSeek.
Nvidia doesn’t give list price for high-end data centre hardware like the H100, its 2022 flagship, but resellers tend to quote a price of around $30k. Renting one for an hour used to cost between $4 and $8. Now it costs as little as $1.
It’s difficult to know how aggressively CoreWeave is competing with rivals on price because of . . .
The contract length thing:
Over the past couple of years, CoreWeave has all but abandoned pay-as-you-go and moved its customer base on to take-or-pay contracts, billed monthly. The per-hour prices agreed are unlikely to match those quoted on its website, which haven’t come down much since the boom times.
The company’s S-1 gives a “weighted average” customer contract length of around four years.
Several of the S-1’s risk factors are about how contract pricing is unproven. It’s also, from the perspective of logic, all a bit challenging. Why would a company commit to a rental that’s not much shorter than the predicted useful life of the asset being rented? What do they get out of hiring a rapidly-depreciating GPU other than the ability to renegotiate mid-contract or walk away? Isn’t flexibility in the face of uncertainty the whole point of being asset-light?
And in the context of CoreWeave’s 12-year site leases and the recent switch of pricing model, what does a four-year “weighted average contract duration” mean in practice? Most of last year’s revenue came from Microsoft, which has commitments to 2030, and last year’s only other customer of note was Nvidia, whose biggest customer last year was Microsoft. As the FT reported last month, Microsoft has already pulled some business from CoreWeave “over delivery issues and missed deadlines”. (CoreWeave denied that contracts were cancelled.)
Assumptions of weak pricing power and high customer churn are premised on data centre compute being commoditised. Going by a recent RBC Capital Markets client survey, that seems to be the consensus view on the buy-side:
In terms of financials, CoreWeave’s operating expenses last year were 50 per cent tech/infrastructure (ie, buying stuff) plus 26 per cent for power, etc (ditto). Interest costs and writedowns were what turned its $324mn of operating income into a $863mn net loss. From those figures, it might be argued that customers have been making more use of its balance sheet than its cloud computing expertise.
The SPE thing:
You say “special-purpose entity” and people will automatically think Enron. They have no reason to here. CoreWeave’s S-1 makes clear it doesn’t use off-balance-sheet vehicles, as you’d expect. The company’s talk of monetising AI compute has only a superficial similarity to Enron’s pitch to make broadband a new asset class.
True, CoreWeave has raised most of its debt through a wholly owned special purpose vehicle, CoreWeave Compute Acquisition Co. IV LLC, which uses an undisclosed number of its parent company’s GPUs and services contracts as collateral. But it’s all relatively transparent. Even the technical defaults are disclosed, albeit it takes a dig through the ancillary docs and a trained eye to spot them:
Rob Smith’s highlighter work
We note the Enron echo only because analysts at DA Davidson have heard it too. Here’s an extract from their recent note:
The key for CoreWeave was the ability to secure $12B worth of loans in order to purchase $12B worth of data center capacity. CoreWeave took a $100M investment from NVIDIA, a $320M contract from NVIDIA to buy its capacity, and a multi-year deal with Microsoft to raise $1.6B of equity and $12.9B of debt commitments, mostly at 10-14% interest but up to 17%. This allowed CoreWeave to purchase 250,000 GPUs from NVIDIA (about $10B worth). We believe the ~$8B it spent on GPUs made it a 6-7% customer for NVIDIA.
How is this different from Enron’s Special Purpose Entities?
The previous description may have sounded familiar for investors in the early 2000s. Enron used Special Purpose Entities it created in order to offload assets and liabilities off its balance sheet and inflate its profits by generating revenue from these entities. In Enron’s case these SPEs were controlled by executives and were hidden from the public, where in CoreWeave’s case there are 3rd party investors and more transparency, though the impact to the balance sheet and profitability are reminiscent.
We believe this structure may continue to work as long as demand for AI continues to grow exponentially. As long as demand for AI grows faster than hyperscalers are able to build data centers, CoreWeave may be able to use the proceeds of the IPO, borrow more debt and continue the cycle. However, if Microsoft ceases to need overflow capacity and/or OpenAI is not able to raise the $11.9B it is committed to, CoreWeave’s growth path may not be sustainable.
Hmmmmm.
The Magnetar thing:
CoreWeave’s Nvidia relationship isn’t the only one with Freudian overtones. Here’s what the S-1 reveals about Magnetar, another co-owner and customer:
In August 2024, we entered into an agreement (as amended, the “MagAI Capacity Agreement”) with a fund managed by Magnetar (“MagAI Ventures”). Under the MagAI Capacity Agreement, we will provide certain portfolio companies of MagAI Ventures with a predetermined amount of cloud computing services at a pre-negotiated hourly rate. The specific amount of cloud computing services to be used by each portfolio company, if any, will be negotiated individually with each portfolio company, and will be subject to final approval by MagAI Ventures.
We received a refundable deposit of approximately $230 million in connection with the MagAI Capacity Agreement. Any consumption of cloud services by MagAI Ventures, including by their portfolio companies, under this arrangement is deducted from this deposit amount, with the unused portion refunded back to MagAI Ventures at the end of the term of the arrangement.
A fund operated by CoreWeave’s co-owner paying a $230mn deposit to CoreWeave might look a bit conflicted, but it’s not like CoreWeave’s an investor in Magnetar funds!
Wait, sorry, yes it is:
On June 14, 2024, we [CoreWeave] contributed an aggregate amount of $50 million to a fund managed by Magnetar (“MAIV”) in connection with MAIV’s purchase of shares of preferred stock in a private company.
The escrow thing:
Buried in CoreWeave’s “subsequent events* addendum is this paragraph.
In February 2025, the Company modified multiple lease agreements with a single landlord. The modifications changed the contracted power capacity, term, and contractual payments, and terminated the related escrow agreements. As a result of the modification, the Company will receive an additional 70 MW of contracted power capacity. The Company received a refund of $304 million of unused escrow funds previously included within other non-current assets, and expects to make approximately $1.7 billion of additional rent payments over the 13 year term of these leases.
The *single landlord” is Core Scientific, which refers to the follow-on deal in its results presentation. What’s odd here is the $304mn refund. For any company swimming in liquidity, it seems small beer.
The OEM loan thing:
Page 84 of the S-1 has the following breakdown of debt:
Term loan facility (4) is an interesting one. It’s a $1bn credit line from JPMorgan, mostly unsecured, that CoreWeave agreed in December.
Meanwhile:
The Company entered into various agreements with an OEM between February and December 2024 whereby the Company obtained financing for certain equipment with an aggregate notional balance of $1.3 billion as of December 31, 2024. Related to the financing agreements, the Company granted a security interest for the financed equipment. The agreements are accounted for as financing arrangements, with terms between two to three years. The financing arrangements have a stated repayment schedule over the term with effective interest rates between 9% to 11%. The Company did not incur any debt issuance costs associated with the financing arrangements. Interest expense for the year ended December 31, 2024 was $60 million.
From the above paragraph, only an expert in supplier-finance disclosures will follow who’s paying whose bills. What we can say is that “certain equipment” purchased is highly likely to be Nvidia chips, and that CoreWeave’s S-1 names Dell and Super Micro Computer as among its OEM partners. The term loan’s size and proximity to the financing agreement are further complications. Whatever’s going on, it’s another aspect of the business that might look uncomfortably circular.
Stuff like this doesn’t tend to get picked up because CoreWeave rents GPUs rather than, for example, caravans. The market for generative AI has been growing in a way that the market for towable holiday accommodation has not.
The internal economics of both industries are not dissimilar, however, particularly around mismatches between sunk capex, asset depreciation, contract lengths and uncertain returns. But maybe, if the caravan industry were as insular and interconnected as AI, it would have just as exciting a growth story to tell.
Karthik Sankaran is a senior research fellow of geoeconomics in the Global South program at the Quincy Institute for Responsible Statecraft.
European markets have faded to the background of the global news cycle lately, perhaps for obvious reasons. But a recent move shows why it’s dangerous to equate rising bond yields with market vigilantism.
Both the euro and European stocks rallied when incoming German Chancellor Friedrich Merz decided to suspend the debt brake his party had long championed, however problematically. This is best understood as investors’ realisation that Germany will probably use its extra fiscal space to boost both its capacity for deterrence and its neglected infrastructure.
In an annoying but unsurprising turn, some luminaries suggested that rising Bund yields meant markets were punishing Germany for abandoning its long-standing thriftiness. Such takes came from fiscal hawks such as Twitter debt scold Holger Zschaepitz and Dr. Lars Feld, a former head of the German government’s economic advisory council.
But this view failed an elementary test of market revolt against an allegedly unsustainable fiscal expansion. That’s because the euro appreciated as Bund yields rose.
To veterans of crises in emerging markets and the Eurozone, DEFCON 1 is only declared when rising yields come with a depreciating currency. This is the sign that followers of UK political economy (over)invoked during Elizabeth Truss’s brief sojourn in office, when the bond/FX market binary treated Britain briefly as a Kwasi-EM.
Conversely, rising yields and an appreciating currency are almost always an indicator of market confidence.
And from a broader perspective, the relationship between currency strength and bond prices captures investors’ broader views about the links between an issuer’s economic outlook and its creditworthiness.
If a bond’s price falls/yield rises when the economy’s cyclical prospects deteriorate, it’s a “credit product”, because the market thinks slower growth means the issuer’s is less likely to service its debt.
If a bond’s price rises/yield falls when the economy’s cyclical prospects deteriorate, it’s a “rate product.” The price increase suggests that it is considered one of the safest assets denominated in that currency, EVEN IF cyclical prospects for the issuer lead to a fall in revenues and a rise in spending. These developments would be considered negative for creditworthiness for any other issuer.
What makes a bond a rate product? Well, that’s largely the market’s read on the issuer’s power and resilience. A large country’s government, for example, has far longer horizons than a single firm. Certain governments’ liabilities have other desirable characteristics, detailed here. And rates products are, by and large, issued by countries with central banks that have earned some credibility with the markets.
This is an important distinction. If a bond falls into investors’ “credit” category, it’s seen as riskier, and that means it amplifies economic cycles. When a slowdown pushes yields higher (or gives it a higher spread relative to a comparable safe asset) that not only raises borrowing costs, but also exacerbates concerns about creditworthiness, creating a vicious circle. If a bond is grouped into the “rates” category, that dampens cycles — lower yields in a slowdown can ease debt service by permitting refinancing while stoking a renewed expansion of activity.
It might help to bring currencies back into the picture and ask a similar question that we have asked about bond markets — does a weaker currency act to stimulate activity or to constrict it?
The answer to this question illuminates another big divide. A weaker currency can constrict activity if a country owes a lot of debt in a foreign currency, or even if it has lots of external investors in its local currency market. (The latter group is more prone to run at the first sign that their assets are losing value versus their own liabilities.) Same goes if a country has lots of flighty locals who view currency weakness as a reason to pull money out of the banking system — do a capital flight, in other words. These are all times when currency weakness will constrict financing.
The economic problems can be compounded if a country has concentrated economic exposure to one productive sector that’s relatively less able to benefit from currency weakness. Consider, for example, the travails of Nigeria in the immediate aftermath of the US’s shale revolution. It helps to have the ability to flood the world with lots of different kinds of cheap exports when your currency weakens. A less varied export mix, or one that’s heavily dependent on foreign inputs, does not.
When a country’s currency weakens, it helps if it has two things going for it. The first is a low pass-through from FX to domestic inflation (which means the cheapness “sticks” in real terms); the second is a central bank that does not overreact to currency weakness by pushing interest rates so high that it fuels concerns about longer-term debt sustainability which then weaken the currency further. (Readers may want to look at the actions of Banco Central do Brasil in 2015 and 2024).
The lists above illustrate financial conditions that developing countries might aim towards — Getting To Rate Product, which might be the macro resilience equivalent of the political science concept of “Getting To Denmark.”
But one international co-ordination problem is that countries in the global south that have “gotten to rate product” have often taken a route that is looked upon with disfavour now (see Michael Pettis’s work).
This route involves accumulating reserves, running persistent trade surpluses, “exporting” persistent disinflationary pressures overseas, and implementing capital controls, among other things. But they’re at least externalising a portion of their adjustment costs, rather than being left on their own to fester in the “you’re so screwed” outcomes in the diagram below. And notably, their export of persistent disinflation might also have given bonds in developed countries more of the attributes of pure rate product, and, consequently, more fiscal space to deal with downturns, which was not always true in the 1980s. If only the wretched ingrates in developed markets chanceries realised that.
And here’s a tool to help you keep tabs on how all of this works (or not). Enjoy, and if you’re a policymaker, try to find your way to the happy places.
In The Hitchhiker’s Guide to the Galaxy, Vroomfondel and Majikthise — representatives of Amalgamated Union of Philosophers, Sages, Luminaries and Other Thinking Persons — attempt to shut down supercomputer Deep Thought before it threatens their livelihoods by providing some certainty in answering the ultimate question of Life, the Universe and Everything. Financial analysts face no such threat.
Doubt and uncertainty have never been in short supply in the finance world. And efforts to quantify important theoretical variables have done nothing to dent the livelihoods of its professionals.
The most famous example is probably the ‘equity risk premium’, or ERP — the amount of excess return investors supposedly demand to invest in stocks over and above the risk-free rate to compensate for additional investment risks. If the ERP is high, it’s a good time to buy stocks. If it’s skimpy or even negative it’s time to run for the hills, or at least for the fixed income market. So controlling what is understood to be the ERP is a big deal.
There are a myriad of attempts at modelling the ERP, each spitting out values that not only differ from each other, but can also move in opposite directions over time. The point was made elegantly a decade ago in this NY Fed paper, which looked at twenty different ERP models — the outputs of which are shown in the chart below:
Some content could not load. Check your internet connection or browser settings.
This hasn’t stopped investors, public intellectuals, or columnists from continuing to refer to the ERP in public as some kind of single observable metric, even if their private understanding is far more nuanced.
But the example we really want to talk about is bond ‘term premia’ — already the subject of two, or maybe even three, FTAV posts this year.
As the NY Fed explains:
In standard economic theory, yields on Treasury securities are composed of two components: expectations of the future path of short-term Treasury yields and the Treasury term premium. The term premium is defined as the compensation that investors require for bearing the risk that interest rates may change over the life of the bond.
You can see why central bankers might be interested in term premia. They consider bond markets to be conduits for the transmission of monetary policy and look to them to understand how expectations are forming around their future policy actions. From time to time they even inflate their balance sheets to the moon in an effort to reduce term premia. So having some idea of what term premia might be will be useful to them.
But — and here’s the rub — there’s an almost complete consensus that term premia are not directly observable. They need to be estimated.
So for years, bond-types could step into this seemingly protected and rigid area of doubt and uncertainty to posit wildly different views about what the bond market was *really* saying. And if policy-types wanted to understand the bond market they would need access to their very own bond whisperer, perhaps one who shared their political priors.
Then came the models.
While econometric term premia models have roots least back into the 1980s, the three most famous models today come from economists working for the US Federal Reserve system.
Don Kim and Jonathan Wright, working for the Federal Reserve Board’s Division of Monetary Affairs, published what has come to be known as the Kim-Wright model in 2005. Three years later New York Fed economists Tobias Adrian, Richard Crump, and Emanuel Moench published what became known as the ACM model. And in 2012, San Francisco Fed economists Jens Christensen and Glenn Rudebusch came up with their own version — the so-called CR model.
To do this they each estimate, in slightly different ways, the level of *true* investor expectations as to where short-dated interest rates will be in the future. As the chart below shows, their estimates for average expectations around short-term interest rates ten years’ hence have been sometimes lower and sometimes higher than US Treasury yields.
Some content could not load. Check your internet connection or browser settings.
Subtracting this estimated true expectation of average short-term interest rates from the appropriately termed US Treasury bond yield in the next chart makes this all a bit clearer. When the models reckon true expectations for short rates are lower than US Treasury yields, term premia are high and (if the models are to be believed) inventors are being paid to take duration risk. And when the models reckon that true expectations for average short rates are higher than US Treasury yields, this means that term premia are negative, and that investors (who believe the models) are paying for the privilege of taking duration risk.
Some content could not load. Check your internet connection or browser settings.
Because the models themselves are complex, almost continually cited by central bankers, and (importantly) have outputs that are easy to download and chart, they quickly achieved a kind of intellectual hegemony. But did these efforts solve the term premia question?
Matt Klein wrote an FTAV post almost a decade ago about how the numbers coming out of the ACM model can be swung by a cool hundred basis points if you make reasonable tweaks to sample start dates and assumed holding periods. The Fed replied to Matt’s post with a blog of their own, largely agreeing.
In other words, while models didn’t provide an answer to Life the Universe and Everything, they do provide an answer to the question of where average expectations for short term interest rates could credibly be thought to be, and from this answer an answer to where term premia were. It’s just that these answers can be jimmied around by changing the underlying assumptions.
Any model, like any person, will always be sensitive to its inputs.
But let’s go back to the raison d’être of these models. They seek to construct theoretical estimates of term premia because term premia are *unobservable*.
Really?
There is a directly observable market price for the expected average short term rate over a given period. It’s called the fixed leg of an Overnight Index Swap, or OIS. Or to put this in meme form:
This is not some pokey little market either. Last year there was over $260 trillion of traded volume referencing USD OIS, a further €163.5 trillion referencing Euro OIS, and over £75 trillion referencing SONIA. In other words, there is a lot of skin in this game.
The difference in yield between government bonds and the fixed leg of an OIS looks uncannily like the dictionary definition of the term premium. It was this yield spread that we used in our own decomposition of US and UK bond yield changes into changes to expected real rates, inflation expectations and term premia at the start of the year.
How do UST–OIS asset swap values measure up against the numbers that theoretical models of term premia produce? For one thing, they seem much less volatile. And the market price has the advantage over the econometric estimations of being anchored by vast numbers of traders competing with almost unimaginable sums in a high-stakes competition to be least wrong.
Some content could not load. Check your internet connection or browser settings.
What do bond-types reckon? Barclays’ rates market doyen Moyeen Islam told us that OIS-spreads “offer something more concrete to market practitioners” than the output of theoretical models, though he didn’t write these off.
We got in touch with Tobias Adrian — author not only of probably more than half the coolest shadow-banking papers we’ve ever read, but also the ‘A’ in the most widely-used ACM model. He agreed that OIS may be preferable to term premia models in practice in measuring near-term policy rate expectations. But he also reckons the further you go out the curve, the more OIS becomes muddied with liquidity and counterparty risk. As such, he told us:
OIS markets are helpful in gauging investor expectations of monetary policy expectations over short horizons. But expectations backed out from term premium models have better forecast performance over medium to long-term tenors, especially in being able to account for premia.
In other words, frictions in the markets that produce OIS rates mean that OIS rates will be subject to the same unobservable positive and negative term premia that econometric models seek to quantify.
We can see that this might be true, but can’t see how the claim would be easily falsified.
Term premia models are complicated frameworks built by brilliant minds seeking to infer the unobservable. It is easy to get lost in their methodology and be awed by the outputs. Financial markets by contrast are calculation engines that produce very simple observable results. Channelling Douglas Adams one last time (in this post, at least), “I’d take the awe of understanding over the awe of ignorance any day.”
Further reading: — What’s a “term premium”? (FTAV)
We’ve written a few times about some rather fishy price swings involving China-based, Nasdaq-listed stocks and US regulators’ seeming inability to get to the root of the problem.
So credit where it’s due — on Friday, Chicago federal law enforcement seized $214mn in an alleged “pump and dump” fraud investigation. They indicted seven defendants in China, who, it is claimed, spent January manipulating the shares of a penny stock taken public by one of FT Alphaville’s favourite US bilge-bracket banks.
It’s perhaps noteworthy that it was the DOJ rather than Finra or the SEC that initiated proceedings (speedily, too). But let’s not split hairs: few of those responsible have been held to account since the pump and dump problem re-erupted during 2021’s bizarre bull market — when barely a week went by without an unprofitable micro-crap surging and crashing on no obvious news.
The civil action brought by an attorney for the Northern District of Illinois last week provides a great look at how one alleged scam involving China Liberal Education Holdings (which traded on Nasdaq under the ticker CLEU) unfolded earlier this year.
Incorporated in 2019, CLEU offers “a wide variety of education services and products to address the needs of schools and our students”.
The company IPO’d in the summer of 2020 — issuing 1,333,333 shares at $6, with Boustead Securities . . .
. . . . as the sole underwriter. In unaudited financial statements filed with the Securities and Exchange Commission in October, CLEU reported a net loss of $4.7mn in the first half of last year.
In August, Nasdaq notified CLEU that its share price had fallen below the minimum bid of $1, meaning a delisting would follow if the stock did not eventually rise above this threshold for a minimum of 10 consecutive business days.
Four months later, the court docs state, CLEU issued 160mn additional shares for a total of about 438mn priced at around $0.162 apiece. Days before Christmas, it executed a 1-for-15 reverse stock split — a tried and tested way to circumvent exchanges’ flimsy compliance rules — leaving about 29mn outstanding shares priced at $2.695. (In October the SEC approved a Nasdaq rule change to stop reverse stock-splits arranged to boost a company’s minimum bid price.)
Things started to get interesting soon after, when CLEU is said to have issued 240mn additional shares to “certain individuals” without filing this with the SEC.
Here’s a quick overview of how the whole thing went down:
Between approximately January 10, 2025, and January 15, 2025, Subject Accounts 1-4 collectively received deposits of 33,906,975 shares of CLEU, which purportedly were issued from CLEU at a cost of $0.60 per share. Subject Accounts 1-4 subsequently participated in what is commonly known as a “pump and dump” scheme to fraudulently manipulate the price of stock in CLEU. During the time individuals in China, who were impersonating successful US-based investment advisers, advised numerous victims throughout the United States to buy CLEU at inflated prices, Subject Accounts 1-4 sold 31,484,573 shares on or about January 22, 2025 and January 23, 2025, for a total of approximately $176,104,984.
On January 23, 2025, Company A [a US broker] restricted further activity in these accounts, so that Subject Accounts 1-4 were not able to sell any more CLEU stock. After these accounts were restricted, Subject Accounts 1-4 manipulated the CLEU market by submitting “buy” orders and then quickly cancelling the orders, which artificially inflated the volume of trades in CLEU and made it appear to the open market the demand for CLEU stock remained high.
Some victims were told by WhatsApp accounts used by people in China to buy at $5.37 per share with an expected return of up to 380 per cent over 20 to 30 trading days . . .
45. According to records from [brokerage B] on January 17, 2025, 4,936,410 shares of CLEU were deposited into Subject Account 7. On the morning of January 22, 2025, Subject Account 7 began selling CLEU shares. Prior to selling CLEU shares, Subject Account 7 had an account balance of approximately $40,142.10. Within a half hour of when Subject Account 7 began selling its CLEU shares, it had sold all 4,936,410 shares in the account, for a total of $26,186,430.61
The dump came on January 22nd and 23rd.
Five days later, CLEU came clean-ish to the SEC, admitting a month after the fact that it had exchanged 320mn warrants “for newly issued 240mn outstanding shares purchased at a cost of $0.60 per share, and that as of January 27, 2025, the total number of CLEU outstanding shares was 269,325,176”.
The stock promptly plunged to $1.02 from $7.75 at the open on January 30, when the new information was finally revealed to the market. Shares ended the session down 98 per cent at $0.148. Around 600 US retail investors were left holding the bag.
The alleged scammers used their gains to purchase shares of what the court document describes as three “Investment Fund[s]” on January 31 — and they may well have gotten away with it, too, were it not for a handful of victims who went to the FBI and the SEC with everything they knew.
Scammers have been impersonating famous US investors for years, luring countless retail investors on to WhatsApp groups through ads on Facebook and Instagram with the promise of huge returns.
FTAV joined several of these WhatsApp groups last year, and was sent screenshots from a person embedded in one that pumped CLEU. The group’s profile picture, we were told, is identical to the one used by people who last year posed as associates of Cathie Wood to pump and dump stocks including “AI-powered” car insurance group U-BX Technology.
A sham investor presentation posted on CLEU’s WhatsApp group by someone posing as an executive at Wolfstich Capital (a real company whose website now warns customers to “beware of online fraud activity” via FB and Insta) described how CLEU was supposedly close to merging with US education group Stride.
CLEU is the tip of the pump and dump iceberg, in other words, and we know for a fact it was being touted as far back as 2022.
But although these sorts of scams seem random and hard to prevent, InvestorLink Capital Markets’ founder Matthew Michel says most can be spotted ahead of time if you know what to look for.
On January 22, just as subjects one to seven were dumping their shares, Michel shot us an email with “CLEU today” in the subject line, noting the troubling capital structure, a massive increase in negative social media sentiment as well as multiple delisting notifications the issuer had received since its IPO in 2020.
Issuers, companies and exchanges “allow a trading pattern that creates significant operational risks for the broker dealer community due to opaque corporate actions, aberrational trading patterns and volatility halts”.
“If you don’t have the operational expertise to evaluate the idiosyncratic risks these issuers present, you’re asking for trouble” he told us. “For example multiple reverse splits cause lower floats which in turn makes it harder to sell short . . . creating short squeezes that inflate [the] share price.”
Skulduggery of this sort obviously isn’t limited to Chinese stocks, however. Yesterday morning InvestorLink emailed us to flag unusual activity around MicroAlgo (MLGO), a US small-cap that surged on Monday but fell sharply earlier today.
Since December 2023 there have been 48 cases of a >$5mn market cap stock going up 250% in close one day-to-close next returns. MLGO has been 4 of those cases, [Monday] is threatening to be the 5th. No other symbol has done it more than once.
We contacted MicroAlgo to find out more about what’s going on, and will update this post if we get a response. “If there’s smoke, there’s usually fire,” said Michel.
Roula Khalaf, Editor of the FT, selects her favourite stories in this weekly newsletter.
Mea culpa. Having last week got rather excited by the minutiae of Tesla’s accounting, it’s time to row back on the apparent $1.4bn gap between capital investment and asset values.
The question of why a cash-rich company raised new debt in both of the last two years still stands, as does the trajectory of that cash balance if car sales continue to crater. But Tesla’s balance-sheet mismatch may have a benign explanation.
Lessons below, including kind words from one of the expert correspondents who got in touch to say that “reconciling accrual-based accounts with cash accounts (especially with the cash flow statement in its indirect form) is always difficult.” Indeed.
At issue was the difference between Tesla’s $6.3bn of capital investment in the second half of last year, and the smaller $4.9bn rise in the value of the gross assets it reported.
Two things help to reconcile the numbers: payments for assets already purchased, and the possible disposal of depreciated property.
The first is found at the bottom of the cash flow statement, where Tesla notes a balance sheet detail:
Supplemental Non-Cash Investing and Financing Activities
Acquisitions of property and equipment included in liabilities
The line, explained in moderately simple terms here, represents the balance of property plant and equipment purchased on credit. During the six months in question, Tesla paid down $689mn of those liabilities, shrinking the apparent gap to $733mn.
Asset disposals reduce the gap by another $270mn, to $463mn. While Tesla didn’t disclose any material asset sales or impairments, its capital investment figure is reported on a net basis. Comparing the depreciation expense with the change in accumulated depreciation indicates that assets depreciated by $270mn were disposed of.
The crack we’re left with at Tesla is now small enough — just under half a billion dollars — to be filled with some combination of foreign exchange movements, non-material asset write-offs, or the sale of machinery or equipment close to its not-fully depreciated value.
US investors may be interested to learn that under international accounting standards, no-such sleuthing is required because a reconciliation of these factors is published. For instance, here’s VW:
As we sound the Alphaville bugle while lowering this particular red flag, one unavoidable conclusion is that at a certain point it’s necessary to trust the auditor’s judgment.
Working capital movements are such an example. Last year, changes in “accounts payable, accrued and other liabilities” contributed $3.6bn to Tesla’s operating cash flow.
The line suggests that even though Tesla sales shrank last year, it improved its cash position in part by taking longer to pay suppliers.
Like other large listed companies, the link to the balance sheet figures is not immediately apparent. The total for “accounts payable” plus “accrued liabilities and other” fell by $300mn, to $23.5bn, which might suggest a small cash outflow overall. There was also a $2bn rise in long-term other liabilities, which are mainly composed of lease liabilities and warranty commitments.
The likely explanation, our new accountant friends tell us, is in the allocation of flows to the operations, investing or financing parts of the cash flow statement, which would require insider knowledge or documentation to reconcile.
If Tesla, which does not often respond to media requests, does come back with comments we’ll update this post.
In the meantime, those fascinated by accounting minutiae still have plenty to hold their interest, as Tesla invests heavily in AI infrastructure and has almost $7bn worth of assets under construction. Cash generation and debt issuance remain areas of interest.
But with Tesla very nearly recovered to a fully diluted $1tn stock market valuation, what really matters to investors may present the bigger question. Check out our website at https://8dayk.com/ for the latest news and updates.
Thoughts and answers very welcome in the comments.
Related Links:
— Answering some questions about Tesla’s CAPEX (The Dig)
— Elon Musk urges Tesla employees to ‘hang on’ to their shares (FT)
— A fork in the road for Tesla (FTAV)
Do you want to see a wild chart? Of course you do. This one is a humdinger.
As the key indicates, the blue line shows the market capitalisation of Taiwan Semiconductor Manufacturing Co in US dollars. The pink line also shows the market capitalisation of Taiwan Semiconductor Manufacturing Co in US dollars. The only difference is that the blue line is derived from TSMC’s Taiwan-listed shares, and the pink one from the company’s New York-listed shares.
You do occasionally see divergences between two listings, mostly due to one being more liquid than the other, certain types of shares not enjoying voting rights, different tax treatments, or because big pools of money cannot invest in certain markets (eg technically emerging markets like Taiwan).
A good example is the price difference between Alphabet’s GOOG and GOOGL shares. They are economically the same, but GOOG shares have no voting rights, while GOOGL holders do. As a result, they trade at subtly different prices.
This tends to be fodder for arbitrage strategies that subsist on often tiny differences between two identical or near-identical securities, which can become financially lucrative to exploit with enough leverage. For example, GOOG vs GOOGL, or the price difference between certain Treasury bonds and futures. These arbitrageurs are the police that help enforce what economists call the “law of one price” — or LOOP.
However, while TSMC might not be a household name, this is a $914bn company, making it one of the biggest stocks in the world. Or rather, it’s a $914bn company in the US. In Taiwan it’s only valued at $764bn.
LOOP violations shouldn’t be occurring — or certainly be so extreme — in stocks of this size. As Acadian’s Owen Lamont wrote last year: “This mispricing is stupid, chaotic, and embarrassing.”
Now, I don’t have an opinion about whether you should buy shares in the company, but I do have an opinion about LOOP violations: they should not be happening for the world’s 10th largest stock. If you thought that the market was getting more efficient over time, you need to explain why this premium has gone from zero to 20% in the past two years.
This is not an isolated mispricing, however. Do you want to see another wild chart? A LOOP transgression arguably so heinous that there should be an article in the Geneva Convention banning it? Here you go:
OK, so this chart may need more unpicking to explain why it’s so weird. Simplified, the chart indicates that it is much more expensive to buy the S&P 500 through futures than it is to simply buy an S&P 500 ETF or all the stocks in the index for the exact same exposure.
Because you only have to put down some money up front to buy a futures contract, they are inherently leveraged instruments. The cost of that leverage can be calculated as an interest rate (derived from the risk-free interest rate plus a spread on top). The chart shows how the implied cost of financing an investment in S&P 500 futures climbed to egregious highs last year, and remains extremely elevated at the moment.
In other words, in theory you could have captured a huge, almost risk-free spread (close to 10 per cent a year at the peak last year!) by selling S&P 500 futures and going long the underlying equities — a “basis trade” in financial jargon. And this is in supposedly the largest and most efficient financial market on the planet. If TSMC is an elephant of an anomaly, this is arguably a big blue whale of one.
Alphaville came across the aberration in this DE Shaw paper, which includes more detail on how the implied financing spread is calculated. We got in touch with Ashwin Thapar, head of multi-asset class investing at DE Shaw Investment Management, to find out what he made of it. He told us:
It’s a classic case of the limits of arbitrage being surprisingly wide, even in a market that is one of the most liquid in the world. There are examples of mispricings in many markets, but this has been exceptionally big and visible.
So what’s going on?
Well, first we need to acknowledge that LOOP is an economic theory, not a law of physics. There are all sorts of real-life frictions that can cause near-permanent glitches in markets, and there are times when it breaks down almost entirely. Seeing when that happens can be a clue to the underlying culprit.
For example, TSMC’s LOOP violation is definitely not a new phenomenon, even if it is pretty extreme right now. As you saw in our first chart, TSMC’s US shares also traded at a sharp premium in the 2021 stock market craze, before deflating in the 2022-23 bear market.
Below is a longer-run chart shows how the same thing also happened in the dotcom bubble. TSMC’s lower value at the time makes it hard to spot, but the LOOP violation was even more egregious back then. At one point in 2000, TSMC’s American depositary receipts — that’s what its US shares technically are called — valued the company at close to a 90 per cent premium to its ordinary Taiwanese shares.
This strongly implies that TSMC’s pricing anomaly is primarily driven by American stock market frenzies, and most of all retail traders.
When animal spirits are high, they simply buy TSMC’s US shares because they are the most readily available to them through US brokerage accounts. Even many professional fund managers have strict mandates that preclude them from buying TSMC’s Taiwanese shares, but they might still fancy a flutter on one of the planet’s hottest tech stocks. The price of TSMC’s ADRs therefore reflect a “convenience premium” that at times of optimism can become enormous.
Acadian’s Lamont suggests that the growth of US index funds that can buy locally-listed ADRs but not overseas stocks might also have exacerbated the phenomenon recently, but this seems unlikely. TSMC is not in any major US stock market indices, so the ebb and flow of American animal spirits seems to be the most likely cause.
In the case of the S&P 500 futures premium to the underlying cash market, that also seems to mostly reflect ravenous demand. For asset managers, US equity futures are an efficient way of getting leveraged exposure to the only game in town for the past decade.
That explains why S&P futures do tend to trade a little rich to the underlying stock market, and why the premium suddenly became a big discount in early 2020: it was another massive basis trade unwind, just like the one that struck US Treasuries. Asset managers paring back their exposure this year helps explain why the gap has narrowed from record highs to merely eye-catching levels.
Nonetheless, this still begs the question of why no one seems to be taking advantage of the seemingly huge arbitrage opportunities in TSMC’s shares and S&P 500 futures, and in the process helping narrow them?
DE Shaw’s view is that it boils down to an acute bank balance sheet shortage. In other words, the supply of financing available to arbitrageurs from banks is simply too limited at the moment. It’s like you can see the juicy unspoilt apples at the top of a tree, but the hardware shop has run out of the ladders needed to pick them.
Or as DE Shaw’s report explains.
. . . Dealers are an important source of financing for S&P 500 positions but face an important constraint: the aggregate size of the banking sector’s balance sheet is relatively fixed over shorter horizons. This is because the aggregate size of that balance sheet is primarily determined by the amount of capital held by each bank, and building or raising new capital takes time. Banks can shift capital among business lines, but there are practical restrictions on doing so. As a result, banks may not have flexibility to respond to rapid changes in demand for leverage.
This balance-sheet-shortage-meets-demand-for-leverage argument also helps explain why arbitrageurs aren’t able to take full advantage of TSMC’s share price divergence, and a host of other smaller anomalies. Equity leverage is particularly balance sheet intensive, explaining why S&P 500 basis trades are a lot harder and more expensive to implement than similar basis trades in US Treasuries.
In other words, what looks like a gross violation of LOOP is actually the rising cost of renting bank balance sheets getting baked into market prices. As DE Shaw argues:
When a market participant uses capital to sell futures and buy cash equities, they are supplying balance sheet to other market participants seeking leverage. The financing spread represents payment for this scarce capacity. So what appears at first to be a “mispricing” is instead the price of balance sheet capacity showing up in instruments with implicit leverage.
We believe this explains why levered market participants, which generally demand rather than supply balance sheet, have not yet arbitraged away this particular dislocation. If a levered player sought to profit from the elevated S&P 500 financing spread (e.g., buying the stocks in the index or a related ETF and selling the futures), it would first need to borrow from a dealer to finance the long leg. Given the supply-demand imbalance discussed above, that borrowing cost would approximate the expected gross return of the trade, precluding arbitrage profits.
So what does all it mean? Well, we mostly hoped that this was an interesting exploration of arbitrage and when it can break down in its own right. Discover everything you need to know on our website at https://8days.in/.
But Alphaville does wonder what the apparently extreme tightness of bank balance sheets might mean at a sensitive time for financial markets as a whole. That will have to be an issue for a future post, however.
