CoreWeave is Just One Small Story Spun in the Web of the AI Boom

The New Gold Rush Landlord of Compute

In the world of artificial intelligence, there’s a modern gold rush underway – not for precious metals, but for computing power. And at the heart of this frenzy is a little-known company called CoreWeave, which has quickly risen to become one of the biggest providers of AI cloud infrastructure. CoreWeave doesn’t make flashy chatbots or consumer apps; instead, it sells the “picks and shovels” of the AI boom – namely, access to massive data centers full of advanced chips that can train and run AI models. In essence, CoreWeave is a “landlord for compute,” renting out processing power much like a real-estate company leases office space[1].

It wasn’t always in the AI business. In fact, CoreWeave’s origin story is an unlikely pivot that mirrors the fickle fortunes of tech trends. Founded in 2017, the company spent its early years mining cryptocurrencies. As recently as 2021, CoreWeave earned its keep by mining Ethereum, putting high-end graphics processors (GPUs) to work crunching crypto algorithms[1]. But when the crypto market cooled and Ethereum shifted away from mining, CoreWeave reinvented itself. By 2022, it had transformed into a dedicated AI infrastructure firm, betting that the future lay in supplying computation for the exploding field of generative AI. The gamble paid off dramatically – at least at first glance.

Today, CoreWeave operates dozens of data centers across the U.S. and Europe, stocked wall-to-wall with cutting-edge Nvidia GPUs, the prized silicon brains behind artificial intelligence projects[2]. In documents for its March 2025 IPO, CoreWeave revealed it owned over 250,000 Nvidia chips and, notably, only Nvidia chips[2]. This single-vendor strategy underscores how inseparable CoreWeave is from Nvidia’s technology – and, as we’ll see, from Nvidia’s financial backing as well. The company proudly calls itself “the essential cloud for AI”, competing against giants like Amazon Web Services and Microsoft Azure in the race to provide on-demand AI horsepower[3]. And it has attracted marquee clients that lend credence to that claim, including Microsoft, OpenAI, Meta, IBM, and others[4].

By late 2024, CoreWeave’s rapid ascent earned it “unicorn” status and then some. A funding round in November 2024 valued the startup at a staggering $23 billion, up from just $7 billion a year prior[5]. Investors from Wall Street and Silicon Valley – including names like Blackstone, Magnetar, Coatue, and even tech firms like Cisco – poured money into the company, seeing it as a key player in the AI revolution[6][7]. At its IPO in March 2025, CoreWeave aimed for a $35 billion valuation[8][9], and while it ultimately priced below that goal, the stock soon took off. By the fall of 2025, CoreWeave’s share price had climbed over 160% from its IPO, making it a stock-market darling of the AI craze[10].

Yet behind the gleaming headline numbers – billion-dollar contracts, exponential revenue growth, soaring valuations – lies a more unsettling story. CoreWeave’s business, for all its apparent success, is built on a fragile foundation: massive debt, a handful of risky customer deals, and an extraordinary level of support from Nvidia itself. This support ranges from straightforward (Nvidia is a major equity investor) to intriguingly convoluted (Nvidia has become CoreWeave’s customer, backstop, and benefactor all at once). Such entanglements raise tough questions: How sustainable is CoreWeave’s breakneck expansion? Is the company’s explosive growth a sign of genuine demand, or a mirage engineered by creative financing and hype? And could CoreWeave’s story be a warning sign that today’s AI infrastructure boom is starting to rhyme with yesterday’s bubbles?

Debt, Deals, and GPU Collateral: A High-Stakes Financial Engine

One of the most striking aspects of CoreWeave’s rise is how aggressively it has been financed – and how novel its financing methods are. Unlike traditional cloud providers that built data centers gradually using corporate cash flow or moderate debt, CoreWeave took a shortcut: it borrowed money, and a lot of it, at very high interest rates, using piles of computer chips as collateral. In late 2023, CoreWeave pioneered an unusual deal to raise cash. The company pledged thousands of its Nvidia GPUs to secure a $2.3 billion loan – essentially using high-end chips like one would use real estate to back a mortgage[11]. The effective interest rate on this loan was a steep 15%, reflecting both the novelty and risk of the arrangement[11].

This was just the beginning. Within months, CoreWeave took out a second, even larger loan – $7.5 billion – also collateralized by GPUs[12]. Thanks to the fervor around AI, the terms improved slightly (around 10% interest, later lowered to 9% as the loan was upsized)[12]. By the third quarter of 2024, the company even secured a third loan tranche at roughly 9%[12]. All told, CoreWeave raised well over $10 billion in less than a year through these GPU-backed credit facilities, an astonishing feat for a company of its age. It touted this as a creative way to fund growth – indeed, other AI infrastructure startups like Crusoe and Lambda would later follow with similar collateralized loans[13]. But the strategy also saddled CoreWeave with a towering debt load and hefty interest obligations. (For context, 9–15% interest is more akin to credit-card debt than corporate loans, signaling how much risk lenders see in CoreWeave’s model.)

By early 2025, Reuters reported that CoreWeave had amassed over $14.5 billion in debt and equity financing across 12 rounds[6]. In 2024 alone, the company orchestrated one of the largest private debt raises in history: more than $7 billion in a single go, led by asset-management titans Blackstone and Magnetar[14]. After going public, CoreWeave didn’t slow down its fundraising. It issued high-yield bonds (junk bonds) twice – one $2 billion offering and another $1.75 billion – and in mid-2025 closed an additional $2.6 billion in secured debt financing to fuel its frenetic expansion[15]. By August 2025, the company acknowledged that since 2024 it had raised a staggering $25 billion from investors and creditors to build out its platform[15].

What has all that money bought? In simple terms, data centers on an unprecedented scale and speed. CoreWeave has been spending on infrastructure like there’s no tomorrow. In the second quarter of 2025 alone, capital expenditures reached $2.9 billion, a record for the company and more than triple the level earlier in the year[16][17]. Management maintained a full-year 2025 capex forecast of $20–23 billion – a number so large that Wall Street analysts openly marveled at it[18]. To hit that target, CoreWeave expected to spend more in the final quarter of 2025 than in all the previous quarters combined, as new facilities come online[19]. Plans were laid for a $6 billion flagship data center in Pennsylvania[20] and even a bold move to acquire a legacy data center operator (the crypto-miner-turned-hoster Core Scientific) for $9 billion to instantly boost capacity[21]. By mid-2025, CoreWeave had 33 data centers live with 470 megawatts of power and contracts in place to nearly quintuple that capacity to 2.2 gigawatts in the coming years[22]. In CEO Michael Intrator’s words, the company was “executing at a massive scale” to chase demand that “continues to outpace supply” – even if it meant being “capacity-constrained” and scrambling to find enough electricity and physical space for all those GPUs[23][24].

The revenue side of the equation has grown fast, but not nearly fast enough to catch up to spending. CoreWeave’s 2024 revenue was $1.92 billion, up almost 8× from the prior year thanks to the AI boom[25]. Yet it still lost a whopping $863 million in 2024, even more than its $594 million loss in 2023[25]. By mid-2025 the company finally notched a tiny adjusted operating profit for a quarter, but its operating expenses had exploded nearly fourfold to $1.19 billion in that quarter[26]. In other words, CoreWeave is burning cash as furiously as it can raise it – a classic hallmark of a hyper-growth startup, but one that leaves little margin for error if the growth wobbles.

Wall Street short sellers have taken notice of these red flags. Kerrisdale Capital, an investment firm known for betting against hype-driven stocks, bluntly described CoreWeave as an “undifferentiated, heavily levered GPU rental scheme stitched together by timing and financial engineering, not lasting innovation.”[27] In Kerrisdale’s analysis, strip away the AI buzzwords and you have a company that rents out commodity hardware, carries dangerous levels of debt, and relies on financial tricks to paper over fundamental weaknesses[27]. CoreWeave, unsurprisingly, disagrees with this characterization – its spokesperson insists the company is “the essential cloud for AI” with cutting-edge infrastructure and strong customer support, not a house of cards[28]. But the numbers don’t lie: CoreWeave’s balance sheet is leveraged to the hilt, and its bet is that future demand for AI compute will make those debts worth it. If that bet is wrong, the company could find itself in a precarious position, owing billions for hardware that its customers might not fully use.

Nvidia’s Invisible Hand: How the Chip Giant Props Up CoreWeave

If CoreWeave’s financial engine is fueled by debt, the transmission keeping that engine running is Nvidia. It is difficult to overstate how crucial Nvidia – the $1.2 trillion semiconductor behemoth – has been in CoreWeave’s story. In fact, CoreWeave as it exists today “simply isn’t possible without Nvidia,” as one observer noted pointedly[2]. Not only does CoreWeave use exclusively Nvidia chips in its fleet[2], but Nvidia has entwined itself with CoreWeave through investments, sales agreements, and even direct business deals that effectively guarantee CoreWeave’s success (or at least its survival).

Start with equity: Nvidia itself became a major shareholder in CoreWeave well before the IPO. In 2023, CoreWeave’s Series B fundraising was led by Magnetar Capital, but Nvidia also chipped in as an investor, signaling a “deepening” relationship between the two companies[29]. By mid-2025, Nvidia owned an estimated $4 billion worth of CoreWeave stock – a significant stake suggesting strong confidence in the startup’s future[30]. When CoreWeave prepared to go public in March 2025, Nvidia literally made the IPO possible: faced with lukewarm demand from other investors, Nvidia swooped in and agreed to buy a big block of CoreWeave shares itself to shore up the offering[31]. This is highly unusual – major tech companies don’t typically spend their own cash to backstop someone else’s IPO. But Nvidia had a clear interest in keeping CoreWeave’s valuation aloft and its growth on track, because CoreWeave is both a key partner and a voracious buyer of Nvidia’s flagship product (the advanced GPUs). In effect, Nvidia signaled to the market: we believe in this company so much, we’ll put our money on the line to ensure it succeeds. CoreWeave’s CEO acknowledged Nvidia’s critical role, thanking them for “rowing in the same direction to accelerate the AI economy” – though he denied any “circular” or unsavory financial engineering, calling the partnership completely above-board and complementary[32].

Nvidia’s support goes well beyond owning shares. In a move that raised many eyebrows on Wall Street, Nvidia has also become one of CoreWeave’s largest customers – or at least a customer of last resort. In September 2025, CoreWeave announced a $6.3 billion “initial order” from Nvidia to purchase any of its cloud capacity that isn’t sold to other clients[33][34]. In plain English, Nvidia has guaranteed that if CoreWeave builds more computing power than the market demands, Nvidia will step in and pay for the idle servers itself. This extraordinary backstop agreement runs through 2032 and essentially insures CoreWeave against the risk of empty data centers[35]. Analysts at Barclays cheered the deal, noting it “cushions [CoreWeave] against any potential decline in demand” and reassures investors worried that CoreWeave might overbuild relative to its two big clients, Microsoft and OpenAI[36]. Indeed, CoreWeave’s stock jumped 8% on the news of Nvidia’s $6.3 billion guarantee[37]. From Nvidia’s perspective, this arrangement helps diversify its own exposure – instead of relying solely on selling chips to the likes of Microsoft or Google, Nvidia is effectively locking in future sales by promising to rent capacity from up-and-comers like CoreWeave[38]. It’s a creative way for Nvidia to keep its GPU demand high: if end customers don’t buy enough cloud services from CoreWeave, Nvidia will artificially create demand by buying the service for itself.

This isn’t the only case of Nvidia doing such roundabout deals. Around the same time, reports emerged that Nvidia signed a $1.5 billion contract to lease back 18,000 of its own GPUs from another AI cloud startup, Lambda[39][40]. Essentially, Nvidia had sold state-of-the-art chips to Lambda, and then turned around to become Lambda’s tenant, renting those same chips as a service over four years – making Nvidia the startup’s biggest customer[41]. The Information and Data Center Dynamics noted that Nvidia had “previously signed a similar leasing deal with CoreWeave”[42], suggesting that Nvidia may have quietly funneled cash to CoreWeave through a capacity lease or purchase even before the big $6.3 billion backstop. These “pay-to-rent-your-own-product” maneuvers are highly unusual in any industry, and they haven’t gone unnoticed. Tech analyst Edward Zitron describes it bluntly: “Nvidia is effectively incubating its own customers, creating the contracts necessary for them to raise debt to buy GPUs — from Nvidia, of course — which can, in turn, be used as collateral for further loans to buy even more GPUs.”[43] It’s a self-reinforcing loop: Nvidia helps CoreWeave secure huge contracts (say, with OpenAI or even with Nvidia itself), CoreWeave uses those contracts as evidence of future revenue to convince lenders to give it billions, CoreWeave spends that money to buy thousands more Nvidia GPUs, and the cycle repeats. The result is that Nvidia sells more chips (and even benefits as a shareholder), CoreWeave expands faster than it otherwise could, and lenders feel a bit more secure knowing a giant like Nvidia is standing behind the upstart.

If this sounds reminiscent of past boom-and-bust shenanigans, it’s because it is. In the late 1990s telecom bubble, equipment makers like Cisco and Nortel famously propped up sales by financing their customers – the telecom carriers – essentially lending them money to buy more fiber-optic gear. This kept the party going until it didn’t. “Vendors [had] propped up demand by offering financing to buyers,” Reuters noted of that era, and when the bubble popped, those same vendors saw their valuations plunge 90% in a year[44][45]. There’s a clear parallel in how Nvidia is now acting not just as a supplier, but as banker and client to companies like CoreWeave to keep the AI hype humming. It’s a strategy that boosts Nvidia’s short-term sales and market dominance – the company just reported mind-boggling revenues (88% growth, tens of billions of dollars) almost entirely from AI hardware[46]. But it also could be masking the underlying weakness of demand for AI computing. As Zitron observes, “there doesn’t actually appear to be mass market demand for AI compute, other than the voracious hunger to build more of it.”[47] In other words, a significant portion of CoreWeave’s current “demand” is coming from its own ecosystem: OpenAI (which is itself funded by Microsoft and others) paying CoreWeave for cloud power, Microsoft paying CoreWeave (while also investing in OpenAI), and Nvidia promising to pay CoreWeave if no one else does[48][49]. It’s a relatively small circle of tech giants and investors essentially trading compute contracts with each other. This creates the impression of a booming market, and to be fair there is real usage of AI models surging right now. But it also raises the specter of financial engineering: revenue that is, in a sense, manufactured by strategic agreements rather than pure organic end-user demand.

CoreWeave’s leadership rejects any notion that its Nvidia relationship is problematic. “These investments are not circular; they are complementary,” insists CoreWeave’s head of communications, arguing that Nvidia, CoreWeave, and others are “all rowing in the same direction to accelerate the AI economy”[32]. Certainly, there is truth to the idea that big partnerships can help jumpstart industries – Nvidia’s ecosystem investments have indeed enabled dozens of AI startups to build where otherwise only the tech titans could play. However, it’s also true that CoreWeave’s financial viability appears unusually dependent on one supplier and backer. If Nvidia were to pull its support – say, if economic conditions changed or Nvidia found a better outlet for its GPUs – CoreWeave would be in uncharted waters given its outsized leverage. As it stands, Nvidia’s “invisible hand” is stabilizing CoreWeave’s present (ensuring its capacity is utilized and investors stay confident), but it might also be inflating CoreWeave’s future by encouraging more build-out than the market truly needs. This delicate balance makes CoreWeave’s story not just one of a startup’s meteoric rise, but also a story of how far an industry leader will go to keep the AI frenzy going.

Breakneck Growth Meets Red Flags

CoreWeave’s meteoric growth has earned it contracts and headlines that any startup would envy. In March 2025, just weeks before its IPO, CoreWeave stunned the tech world by securing a five-year, $11.9 billion deal with OpenAI[50] – a contract to provide the computing muscle behind ChatGPT and OpenAI’s other AI services. As part of that pact, OpenAI even invested $350 million in CoreWeave equity, taking a stake in its new infrastructure partner[51]. Over the following months, that partnership only expanded: by September 2025, CoreWeave claimed the OpenAI deal’s “total value” had grown to $22.4 billion after add-ons[52]. Not long after, Meta (Facebook’s parent) inked its own $14 billion contract through 2031 to use CoreWeave’s cloud for AI workloads[53]. These eye-popping figures – $22 billion here, $14 billion there – made clear that CoreWeave was at the center of the AI infrastructure boom. They also helped drive up CoreWeave’s reported backlog of future revenue. By mid-2025, the company boasted a backlog of over $30 billion in cloud commitments[54]. By the third quarter, that backlog had roughly doubled to $55.6 billion[55], an almost unheard-of number for a company that essentially didn’t exist in the AI space three years prior.

On the surface, such demand projections justify CoreWeave’s massive expansion. The company repeatedly raised its revenue forecasts for 2025, crossing the $5 billion mark in expected annual sales[56] – an astonishing 5x growth from 2024’s $1.2 billion. However, a deeper look reveals that CoreWeave’s present and future hinge overwhelmingly on just a few customers, each of whom has their own shifting strategies and potential conflicts of interest. In 2024, Microsoft alone accounted for roughly two-thirds of CoreWeave’s revenue[57]. That concentration actually increased in 2025: Microsoft was 71% of CoreWeave’s revenue in Q2 2025, and 67% in Q3[58]. This makes CoreWeave highly dependent on the goodwill of one tech giant. And Microsoft, while happy to buy extra capacity during the initial AI rush, is no charity: it has been rapidly building out its own AI supercomputing data centers and even designing its own AI chips to reduce reliance on Nvidia[59]. In fact, earlier in 2025 Microsoft reportedly walked away from some commitments with CoreWeave due to delivery issues and missed deadlines[60], and chose not to exercise an option to buy $12 billion more of CoreWeave’s capacity[61]. That nearly derailed CoreWeave’s IPO until Nvidia intervened to buy those unsold shares[31]. Microsoft’s priorities can change fast – especially after its very public breakup and reshuffling of its partnership with OpenAI in late 2025 – so there is no guarantee it will remain a huge CoreWeave customer beyond existing contracts.

OpenAI itself, despite signing that megadeal, is an interesting risk. For one, OpenAI as a business is deeply unprofitable (it reportedly doesn’t make money as of 2025)[62], which raises questions about its ability to pay for multi-billion cloud contracts unless its backers keep subsidizing it. Moreover, OpenAI is pursuing a strategy (codenamed “Stargate” in partnership with SoftBank) to build its own data centers and supply 75% of its computing needs internally by 2030[62]. In other words, OpenAI’s deal with CoreWeave could be a stopgap – a five-year arrangement until OpenAI no longer needs an external landlord. The contract does allow OpenAI to terminate portions of the deal if CoreWeave fails to deliver on time[52]. If OpenAI’s needs or financial situation change, CoreWeave could be left with far less revenue than anticipated. Meta, similarly, might not be a forever client; it is heavily investing in AI data centers of its own and even issued $30 billion in bonds to finance that infrastructure push[63]. It doesn’t take much imagination to see a future where today’s key customers – Microsoft, OpenAI, Meta – evolve into competitors or at least self-sufficient players, leaving CoreWeave scrambling to find new buyers for its overbuilt capacity[59][64]. CoreWeave itself, in its IPO filings, explicitly acknowledged this risk: that its biggest customers might “suddenly become [its] biggest competitor” if they decide to bring AI compute totally in-house[59].

Then there’s Nvidia as a customer. CoreWeave’s IPO prospectus apparently listed an unnamed “fourth customer” that was its second-largest revenue source after Microsoft[65]. Industry insiders have indicated this was Nvidia itself, which agreed to spend around $1.3 billion over four years on CoreWeave’s services (likely a precursor to or part of the broader $6.3 billion backstop)[65]. It is highly unusual for a company’s supplier to also be a top customer. While the Nvidia deal is framed as a positive – it “cushions” CoreWeave as mentioned earlier – it underscores how non-diverse CoreWeave’s income stream really is. Excluding Nvidia’s safety net and the handful of big tech clients, CoreWeave has yet to demonstrate a broad base of paying customers. In a thorough investigation of the so-called “AI neocloud” sector, Ed Zitron found “a worrying pattern”: the major new GPU cloud firms (CoreWeave, Lambda, etc.) have almost no significant revenue outside of Big Tech and Nvidia itself, even as they pile up debt anticipating demand that may never materialize[66]. That means CoreWeave is essentially building for a future where every other industry starts buying AI computing by the truckload – a future that is still uncertain. Most mainstream enterprises are only experimenting with AI in 2025, not rolling out trillion-parameter models that need entire data centers. A recent McKinsey survey showed that while nearly all companies are intrigued by AI, few have deployed it at meaningful scale, with most still in “wait and see” mode[67][68]. If that widespread, mass-market AI adoption lags behind CoreWeave’s expectations, the company could be stuck with enormous fixed costs and far fewer takers for its capacity – a classic recipe for a boom to turn to bust.

Some early signs of overcapacity in the AI compute market are already emerging. Remember how scarce and expensive AI GPUs were in 2023? By late 2024 and 2025, with CoreWeave and dozens of rivals racing to spin up GPU farms, rental prices for these chips started falling. Industry analysts at SemiAnalysis observed it becoming “a buyers’ market for GPU rentals” in 2024–25, with “widespread availability from over 100+ AI neoclouds and hyperscalers.”[69] The going rate to rent Nvidia’s top-tier H100 GPU has dropped by roughly 23% between late 2024 and mid-2025 – from about $3.06/hour to $2.36/hour on one index[70]. Some providers were advertising H100 rentals for as low as $1.50 an hour[70]. That may sound trivial, but these economics matter: experts estimate that once prices dip below roughly $1.65 per hour for an H100, a cloud operator is no longer recouping the cost of the hardware over its lifespan[70]. And if prices aren’t above ~$2.85/hour, the returns won’t beat even a modest index fund[71]. In short, many GPU cloud vendors might already be “barely breaking even” at current rates[72]. CoreWeave’s CFO has insisted that the company mostly sells capacity through long-term contracts rather than on-demand rentals, insulating it from short-term price swings[73][17]. But if those long-term contracts (like OpenAI’s, Meta’s, etc.) don’t all fully materialize or renew, CoreWeave could be forced to compete on the spot market – and the spot market is getting crowded and cheaper.

Even the technological winds pose risk. Just as 2000-era fiber networks suddenly faced new tech that made them more efficient (like better compression and wavelength division multiplexing in fiber), the AI compute world could see leaps that undercut the demand for brute-force infrastructure. Case in point: when a small Chinese startup called DeepSeek announced an AI model that it claimed could match OpenAI’s GPT-4 at a fraction of the training cost, it shocked the industry and briefly sent Nvidia’s stock tumbling in January 2025[74][75]. The idea that smarter algorithms or more efficient chips could do the same AI tasks with far less computing power is a real wild card. One need only look at history: in telecom, new data transmission tech vastly improved existing fiber capacity, contributing to the glut and price crash. In AI, if innovations make it possible to serve AI applications with, say, 1/10th the GPUs, a lot of planned data centers could become overbuilt white elephants overnight. Moreover, the pace of GPU innovation is blistering – Nvidia updates its architectures every 2 years or so. A top-of-line chip today might be obsolete in two or three years. If CoreWeave’s billions in hardware have a shorter useful life than expected, depreciation and writedowns could hit hard. Analysts warn that if the effective lifespan of AI chips is closer to 2–3 years (due to rapid advances) rather than the 5–10 years data centers typically assume, companies may face painful write-offs on infrastructure that never lived up to its promise[76].

All of this paints a picture of a company racing at highway speed, but also approaching several crossroads of risk. CoreWeave’s CEO and investors remain bullish, often pointing out that we’re just at the start of a generational shift to an “AI-powered” world. They argue that demand for AI compute will only increase as new applications emerge, and that being first to build capacity positions CoreWeave to ride that wave. They could be right in the long run. But in the near term, even CoreWeave’s supporters had a wake-up call recently. In November 2025, CoreWeave disclosed a delay in one of its major new data centers, forcing it to slash its revenue outlook for the year[77][78]. The stock plunged 10% in a day, erasing $5 billion in market value, as analysts lamented that the quarter “revealed something investors have feared – operational risk.”[79] The company’s margins were already under pressure from soaring infrastructure costs and higher chip prices[80]. Now the missed construction timeline showed that even with “strong demand,” scaling these AI super-computers is not as easy as just throwing money around[81]. Barclays analysts noted it was “the first time for the young AI infrastructure industry” that such a hiccup had occurred, and a reminder that “large scale AI data centers are not easy engineering projects.”[81] A research firm, MoffettNathanson, put it more bluntly: CoreWeave’s slip-up and narrowing margins are “an incrementally worse setup for the day in the future when demand isn’t off the charts.”[82] In other words, if the company is struggling now while AI demand is red-hot, what happens if (or when) the growth inevitably cools? It was a sobering moment that hinted how rapidly the narrative around CoreWeave could turn if a few assumptions don’t pan out.

Echoes of Past Booms and Busts

For those who have witnessed earlier tech booms, CoreWeave’s story contains some deja vu. History doesn’t repeat, as the saying goes, but it often rhymes – and the rhymes here sound eerily familiar to the dot-com bubble of the late 1990s, the telecom frenzy that followed, and even the more recent WeWork saga.

Flash back to 1999–2000: Back then it was websites and e-commerce driving a gold rush. To support the “new economy,” companies poured unprecedented money into infrastructure – chiefly, data centers and fiber-optic networks. A startup called Global Crossing laid undersea fiber cables in a blitz, went public within a year of founding, and hit a $47 billion valuation by 1999 – only to go bankrupt in 2001 when the overbuilt market collapsed[83][84]. Telecom firms collectively raised over $2 trillion in debt in the late ’90s to lay 80 million miles of fiber optic cable across the U.S. and beyond[85]. They believed internet traffic would double every 100 days (a myth popularized by WorldCom) and that “build it and they will come”[86]. And build they did: by 2001, only 5% of the installed fiber capacity was actually being used[87]. The rest was dark fiber – literally cable in the ground with no light running through it. Even four years after the dot-com crash, 85% of fiber lines were still idle as of late 2005[88]. Bandwidth prices fell off a cliff (dropping 90% by 2004 in cost)[89], and one telecom operator after another went bankrupt. The infamous “data center glut” saw shiny new colocation facilities sit empty. Companies like Exodus Communications and WorldCom imploded under debt. Yet, notably, the infrastructure itself wasn’t worthless – far from it. That so-called excess capacity became the foundation for the next two decades of internet growth. The dot-com bubble, wasteful as it seemed, left behind cheap fiber and servers that enabled YouTube, Netflix, and cloud computing later on[89][90]. It’s a classic example of a productive bubble**: investors lost big, but society retained the assets, which eventually found use when demand caught up.

The parallels to CoreWeave and the AI cloud boom are striking. Once again, we see an investment frenzy to build “the future” – this time AI factories instead of fiber highways – at a scale that outpaces current demand. Once again, we see rosy assumptions that demand will explode exponentially, justifying an all-out build-out. (In 1999, every business was told they’d be an internet company or die; in 2025, every business feels pressure to have an AI strategy or be left behind.) And importantly, once again we see the use of reciprocal financing tricks to fuel the boom. During the telecom bubble, equipment vendors propped up their sales by lending money to their customers and counting that as revenue – a circular flow of money that gave a false sense of sustainable demand[91]. In today’s AI infrastructure surge, Nvidia and cloud firms are engaging in a similar dance, with Nvidia providing financing and guarantees to keep its GPU sales humming[43][92]. A seasoned investor looking at CoreWeave might indeed recall how “grizzled investors… fear it’s 2000 all over again,” as Reuters put it[93]. Back then, when the bubble popped, valuation multiples crashed, technological advances made many plans obsolete, and heavily indebted builders like Global Crossing went down in flames[92].

There are differences, of course. The companies driving AI investment now (Microsoft, Google, etc.) are far stronger financially than the startup telecoms of yesteryear[94]. And private investment in tech, while huge, hasn’t hit the same frenzied share of GDP as the dot-com era (at least not yet)[95]. But as one tech CEO quipped recently, “we’re building Ferrari racetracks when most traffic will be Honda Civics”[96]. The concern is that AI infrastructure is being built out ahead of the applications that truly need it, much as fiber was laid far ahead of the data that eventually flowed.

Then consider WeWork, a cautionary tale from just a few years ago. At its peak in 2019, WeWork was a real estate company masquerading as a tech disruptor, with a private valuation of $47 billion and massive global expansion plans. It signed leases on offices all over the world, banking on a future of co-working dominance. But the business was fundamentally unsound – it was burning cash, over-committing to long-term leases while relying on short-term client rentals, and had a charismatic founder long on vision and short on financial discipline. When WeWork attempted to go public in 2019, reality set in: its valuation collapsed and by 2023 the company was teetering on the edge of bankruptcy (in fact, it filed for Chapter 11 in late 2023). What’s the lesson? As Business Insider noted, WeWork’s rise and fall offers a “stark warning for the age of AI euphoria.”[97] WeWork was an example of venture capital over-exuberance, where investors suspended disbelief and poured money into growth at all costs – until the music stopped. In WeWork’s case, SoftBank was a key enabler, pumping in billions of dollars to fuel that expansion. It’s ironic, then, that SoftBank’s Vision Fund was also an investor in CoreWeave (leading a $221 million round in 2023)[29][98]. The same playbook – massive private funding to blitzscale physical assets – is playing out, just this time with data centers instead of office space. One might say CoreWeave is “a glorified landlord despite years of claims it’s a tech company,” much like WeWork ultimately was[99][100]. Indeed, CoreWeave signs long-term deals for capacity (akin to WeWork’s long leases) and then rents it out short-term to clients who may or may not stick around. If those clients pull back (as Microsoft and others very well could), CoreWeave could be left holding costly idle assets – the equivalent of WeWork’s empty offices. WeWork went from a $47.7 billion valuation to barely $300 million in market cap in just four years[101][102]. While CoreWeave’s situation is not the same – it has hard technology assets and a clearer immediate revenue stream – the underlying caution is: beware of companies whose valuations and expansions far outrun the proven market demand, especially when fueled by easy money.

Conclusion: A Ticking Time Bomb or the Next Essential Infrastructure?

Is CoreWeave a bubble about to burst, or a visionary bet on future infrastructure that will eventually pay off? In truth, it may be elements of both. On one hand, few deny the transformative potential of AI – much as no one doubted the eventual importance of the internet after the dot-com dust settled. The data centers CoreWeave and others are building could very well find ample use in the long term, supporting everything from advanced research to mundane business automation. If history is a guide, today’s “excess” capacity can become tomorrow’s baseline. The fiber glut of 2001 became the backbone of streaming video and cloud computing a decade later[89][103]. Likewise, the GPU clusters of 2025 might power medical breakthroughs, scientific simulations, or AI applications we haven’t even imagined by 2030 – especially as the cost of compute comes down dramatically due to this very boom. In that optimistic scenario, CoreWeave’s aggressive build-out could look prescient in hindsight.

On the other hand, booms have a way of overreaching, and the fallout can be brutal for individual companies and investors. CoreWeave today embodies many of the classic warning signs: skyrocketing valuation based more on future promise than present profit, huge capital outlays funded by heavy debt, revenue concentrated in a few deals that were themselves facilitated by insiders, and a narrative that “this time is different” in a hot industry. The broader AI sector is already showing hints of exuberance – from hundreds of AI startups with lofty valuations to incumbent tech firms collectively planning to spend trillions on AI infrastructure and R&D over the coming years[104][105]. When so much is built so fast, there’s a real risk of a supply-demand mismatch. As one analysis quipped, “we’re constructing capacity for applications that remain years away”[106]. If that’s true, a lot of these shiny new AI data centers may sit underutilized in the near term, and companies like CoreWeave – which lack the deep pockets of a Microsoft or Google – could be caught in a squeeze, unable to cover their costs. Debt, in particular, is unforgiving. It doesn’t matter how noble your vision of the future is – if you can’t make your interest payments in the present, trouble awaits. CoreWeave’s $2.3 billion first loan at 15% interest wasn’t nicknamed a “ticking time bomb” by accident in tech circles[27]. It needs explosive growth just to service obligations. That pressure can lead to a fragile situation where any downturn or stumble (like the recent data center delay) sends shockwaves through its financials.

There’s also a systemic angle: Nvidia’s role in propping up players like CoreWeave spreads the risk around in unusual ways. Nvidia has thus far been lauded for its business acumen and dominant position in AI chips. But by entangling itself with the fates of multiple infrastructure startups – investing in them, guaranteeing their revenues, even renting their capacity – Nvidia is effectively betting that it can keep the AI party going long enough for real demand to catch up. If it turns out that demand was overestimated, Nvidia could find itself not just with a temporary dip in chip sales, but on the hook via these backstop arrangements (for example, paying billions to CoreWeave for unused cloud time)[34]. Those kinds of promises can shift risk back onto Nvidia’s balance sheet and potentially obscure losses that haven’t surfaced yet. In the worst case, it could resemble a form of “round-trip” revenue – money circulating in a loop to give the appearance of robust growth[43][47]. Regulators and investors will surely be keeping an eye on that dynamic, just as they did after the telecom equipment fiasco of 2000. As one financial columnist dryly noted, “the qualitative parallels... are uncanny”, even if the scale is different[107][108].

For general readers and AI-curious onlookers, the tale of CoreWeave offers a fascinating, cautionary window into the hype and hope of today’s AI boom. On one side, you have visionary rhetoric about building the “next-generation operating system for civilization” (as CoreWeave grandly puts it)[32], multi-billion-dollar contracts flying around, and the sense that we’re witnessing the construction of a new kind of digital railroad that will change the world. On the other side, you have seasoned voices whispering “we’ve seen this movie before”. They point to the ghostly skeletons of fiber optic cables, empty coworking offices, and other remnants of bubbles past, reminding us that rapid growth can mask deep fragilities.

Perhaps the final word should go to the skeptics tempered by experience. In August 2023, even before CoreWeave’s biggest deals, Morgan Stanley’s CEO compared the AI frenzy to previous bubbles, cautioning that lots of capital would be invested with uncertain return, and that eventually a shakeout would come. More pointedly, when WeWork was circling the drain in 2023, a columnist warned AI investors not to repeat that mistake: “Companies racing for glory in the AI era would do well to learn from the over-exuberance in the WeWork saga.”[109] In WeWork’s case, grand visions of a tech-enabled future workplace crashed into the reality of a very old-fashioned real estate crunch. In CoreWeave’s case, the grand vision of an AI-everywhere future could crash into the reality that building and running world-class data centers is a hard, costly, and potentially over-subsidized business.

For now, CoreWeave stands as both a symbol of the AI revolution’s promise and a harbinger of its possible peril. It has achieved in a few years what legacy data center operators took decades to do, becoming a linchpin in the AI supply chain almost overnight. But its story should come with a flashing yellow light. If the AI infrastructure boom turns into a bubble, CoreWeave might well be the canary in the coal mine – the early warning sign that exuberance has outstripped reality. And if, conversely, the optimists are right and AI truly transforms everything in short order, CoreWeave will be remembered as a bold first mover that bet big and won. In either case, the next couple of years will be crucial. As the saying often attributed to Mark Twain goes, “History doesn’t repeat itself, but it often rhymes.” The coming chapters for CoreWeave, Nvidia, and the AI cloud industry will reveal what kind of rhyme they are crafting – a triumphant one of progress, or a cautionary one of hubris.

Sources:

·      Lopatto, Elizabeth. “I looked into CoreWeave and the abyss gazed back.” The Verge, Nov. 17, 2025[1][2][31][32][58][52][62][64].

·      Reuters News. “CoreWeave, Nvidia sign $6.3 billion cloud computing capacity order.” Sept. 15, 2025[33][34][36].

·      Reuters News. “CoreWeave inks $11.9 billion contract with OpenAI ahead of IPO.” Mar. 10, 2025[50][110][6].

·      Reuters News. “Nvidia-backed CoreWeave’s shares fall as data center delay hits revenue forecast.” Nov. 11, 2025[77][81][82].

·      Data Center Dynamics. “CoreWeave expects capex in Q4 2025 to exceed previous three quarters combined.” Aug. 13, 2025[18][111][21][15].

·      Zitron, Edward. “Is There Any Real Money In Renting Out AI GPUs?” Where’s Your Ed At (blog), Sept. 19, 2025[112][113][66][114][115].

·      BoxCars AI (Tabrez Syed). “Making It Up in Volume: How the AI Infrastructure Boom Echoes the Telco Frenzy of the 90s.” Apr. 4, 2024[84][116].

·      The Grey Ledger. “After the Gold Rush: When the Bubble Deflates.” Aug. 1, 2025[87][89][117].

·      Business Insider (Chowdhury, Hasan). “WeWork could head to the scrapyard of American dreams. There’s a stark warning in there for AI.” Aug. 9, 2023[99][101][100][102].

·      Breakingviews (Cyran, Robert). “AI and fiber bust parallel is partly hallucinated.” Reuters, Jan. 29, 2025[92][45][93].

·      FXStreet. “Is this a data center bubble, not an AI bubble?” Nov. 19, 2025[70][72][76].