Capital, Not Compute, is the Real AI Bottleneck
Satisfying AI’s ravenous appetite for power and compute has kicked off a once-in-a-generation infrastructure expansion that is also one of the most expensive in modern history. Nvidia CEO Jensen Huang has estimated a single gigawatt of capacity can cost as much as $50 billion. With McKinsey projecting data center demand could reach 156 GW by 2030, the total investment required could approach $7 trillion.
While the need for graphics processing units is immediate, the infrastructure needed to deploy them at scale can take years to finance and build. The result is a growing mismatch between appetite and access, with capital—not compute—emerging as the primary bottleneck in the AI build-out.
At a recent panel, The Information’s Anissa Gardizy sat down with three executives on the front lines of that shift to discuss what the future of AI financing looks like:
- Marc Boroditsky, chief revenue officer, Nebius
- Nick Robbins, vice president, corporate development, CoreWeave
- Charles Fisher, chief financial officer, Lambda
The Myths Shaping AI Finance
Charles Fisher is no stranger to capital-intensive infrastructure. Before joining Lambda, he spent years raising funds for the cable industry, another business defined by heavy up-front investment and long-term returns. On the surface, AI infrastructure—made up of large, fixed assets paired with predictable customer demand—should be similarly easy to finance.
“There ought to be a lot of similarities,” he said.
Lenders, however, don’t always see it that way. One reason is asset life. While cable networks were built to last decades, GPUs are estimated to depreciate over six years. Another is customer concentration, with many assuming demand is limited to a handful of hyperscalers.
Both assumptions, Fisher argued, are flawed.
Lambda, for example, has more than 10,000 customers across its public cloud business, representing roughly a third of its revenue.
“The customers behave a lot like a cable customer,” he said. “They are very sticky, with a lot of pricing power.”
The same goes for hardware longevity. According to CoreWeave’s Nick Robbins, older GPUs like Nvidia’s V100 and A100 continue to generate strong returns well beyond their expected lifespan.
“Everything about the pricing curve of these chips suggests they’re going to be in demand for a very, very long time,” he said.
If You Build It, Will They Come?
For AI cloud providers, customer contracts are often the foundation for unlocking capital.
“Capital markets will always have money to finance the build to honor investment-grade customer contracts,” said Robbins.
That dynamic is especially powerful when those contracts come from hyperscalers. Nebius’s multibillion-dollar deal with Meta Platforms, for example, includes a backstop that effectively guarantees demand for its infrastructure.
“The deal allows us to utilize Meta’s credit, because they’ve committed to purchase any GPUs we don’t sell,” said Nebius’ Marc Boroditsky.
Still, even with strong contracts in place, timing remains a challenge. Data centers can take years to build, while customer demand materializes much faster—creating a persistent mismatch.
“It can be less of a capital problem and more of a choreography problem,” said Fisher.
Boroditsky, meanwhile, said Nebius tackles the issue by running everything in parallel—securing sites, lining up demand and raising capital simultaneously.
“We’re generating demand as we build capacity, and as we’re filling those GPUs with revenue-producing customers, we’re supporting the next tranche of development,” he said.
Beyond the Hyperscalers
While hyperscalers dominate headlines—and often provide the easiest path to financing—they represent only part of the market.
“If you’re honest about where demand is coming from, it’s way beyond the hyperscalers,” said Boroditsky, pointing to fast-growing AI startups and enterprise adoption. Companies like Cursor and Harvey, he noted, represent early indicators of a broader shift as AI begins to reshape the software industry from the ground up.
“The real question is, how are you building for the rest of the market, and what’s the engine you’re putting in place to be there for developers all the way to the enterprise?” he said.
For Lambda, that long tail has been central from the start. Its roots in the developer community have also helped strengthen its relationship with Nvidia.
“Nvidia are happy doing business with Meta and Google, but they also want to do business with the next Anthropic and OpenAI—the next company you haven’t yet heard of,” said Fisher.