Artificial intelligence (AI) infrastructure is about to get a lot easier to finance. Nvidia has teamed up with some of the world's largest asset managers and financial institutions, including Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR, to mobilise more than $500bn of third-party capital for AI infrastructure. The implication is profound: capital may no longer be the industry’s main constraint.

One prominent beneficiary of this flood of capital is CoreWeave, one of the world’s leading specialised AI cloud infrastructure providers. Its multibillion-dollar business model is elegantly simple: it buys chips, lights them up and rents out the compute.

Telecom operators have been doing variants of this for decades. Why, then, is CoreWeave valued as an AI growth company while many telcos are valued like mature utilities?

AI is rewriting the rules of infrastructure, shifting economic value by redefining what is scarce. Every technology cycle creates a new bottleneck. In the internet era, it was distribution. In the cloud era, it was software platforms. The AI era brings an entirely new structural argument.

The real question is less about whether telcos can “beat” the neoclouds outright, and more about identifying what remains genuinely scarce in an AI-driven economy. As capital pours into the stack, the market’s major archetypes are placing very different bets on where the next structural bottlenecks will form.

Neoclouds such as CoreWeave are betting that compute is the scarce asset – or, more precisely, high-density computing clusters that can be deployed at scale. They are buying chips to build specialised, hyperconnected physical environments in which tens of thousands of graphics processing units (GPUs) work together as a single supercomputer. Their thesis rests on the conviction that compute will stay scarce for longer than most people expect. The current supply squeeze, combined with CoreWeave’s contracted backlog of more than $104bn, appears to vindicate that aggressive strategy.

But buying GPUs is only part of the equation, and the neocloud bet has other points of failure. All the chips in the world are useless if neoclouds can’t light them up. Nor should anyone underestimate the risk that compute shifts from scarce asset to cheap commodity. If that happens, the neoclouds’ expensive infrastructure would lose its pricing power.

Hyperscalers such as Amazon Web Services (AWS), Microsoft Azure and Google Cloud are betting instead on ecosystem gravity. Their thesis is that enterprise AI cannot exist in a vacuum. Because enterprises already run their core workloads and keep sensitive data in the public cloud, they will naturally deploy AI where that data lives. Model training, moreover, is only a fraction of what enterprises need – the rest is a trusted, all-in-one blanket of data platforms, developer tools and security compliance.

That ecosystem cuts both ways. Hyperscalers face a growing backlash from CIOs wary of costly vendor lock-in, and retrofitting legacy datacentres for high-density compute is far less efficient than building pure-play AI factories from scratch.

Telecom operators, meanwhile, are making a third bet. Not on AI models, nor on enterprise software, but on the physical infrastructure the tech world once took for granted: secure power supplies, distributed edge facilities, extensive fibre networks and physical proximity to governments and enterprises.

The examples vary, but the direction is consistent. Singtel’s Nxera is investing in AI-ready infrastructure across Southeast Asia. SK Telecom is combining AI compute with sovereign AI ambitions. In the Middle East, e& is positioning itself around trusted and sovereign AI services. Each is making a slightly different bet, but all assume the bottleneck lies below the application layer.

Telcos have made similar predictions before, though. A decade ago, the industry argued that distributed infrastructure, edge computing and network proximity would become decisive, high-margin cloud battlegrounds. Instead, enterprise workloads gravitated relentlessly towards ever more centralised hyperscaler platforms. And today, those same tech giants are securing land, power and fibre at an unprecedented, multibillion-dollar scale. If capital ultimately solves the world’s infrastructure constraints, power and location may prove to be temporary bottlenecks rather than enduring moats.

The uncomfortable possibility is that telcos have not discovered a new scarcity. They have simply rediscovered assets they already own – and assumed the market will value those assets more highly than it does today.

The AI infrastructure race is not simply a contest between telcos, neoclouds and hyperscalers. It is a contest between competing beliefs about what remains scarce in an AI-driven economy. Neoclouds believe it is compute. Hyperscalers believe it is ecosystem. Telcos increasingly believe it is physical infrastructure.

The winners will not be the companies that build the most infrastructure. They will be the ones that correctly work out which assets are structurally scarce for the long haul – and which only looked scarce because nobody had enough capital to challenge them.

Edwin Lin is principal consultant at Omdia, part of Informa TechTarget.

• CoreWeave connects AI coding tools to infrastructure intelligence

  

  By: Kinza Yasar

• Why every telco's AI strategy looks different
• With Nvidia GPUs, CoreWeave hits architectural milestone

  

  By: Shane Snider

• Zayo, Nvidia team to address AI network capacity shortage

  

  By: Joe O’Halloran