J.P. Morgan 54th Annual Global Technology Conference, May 19, 2026
CoreWeave co-founder and Chief Development Officer Brannin McBee took the stage at J.P. Morgan's technology conference to an unusually crowded room, a fitting metaphor for a company that has gone from a niche GPU cloud operator to one of the most closely watched infrastructure businesses in the world. In roughly 30 minutes, McBee covered new financing milestones, a pointed response to the Blackstone-Google TPU deal, a surprisingly large financial services exposure in its backlog, and a blunt message to investors who think building AI infrastructure is simply a matter of signing power leases.
DDTL 5.0 Closes With Record Demand — and a Structural First
The most concrete piece of news from the session was the official close of DDTL 5.0, a $3.1 billion to $3.5 billion publicly syndicated term loan facility that attracted $19 billion in demand — what McBee described as "the largest ever TLB demand book." The facility was priced at SOFR plus 450 basis points at roughly 70% loan-to-cost, covering non-investment-grade counterparties including OpenAI and Cohere. More structurally significant is that this is the first time CoreWeave has syndicated one of these asset-level instruments into a publicly traded security, a meaningful evolution in how the company accesses capital markets. "When you have $20 billion worth of demand for these credit instruments showing up in the market, that's the reality," McBee said, pushing back on any narrative that credit appetite for AI infrastructure has softened.
The prior milestone, DDTL 4.0, was CoreWeave's first investment-grade rated instrument and introduced an ABS-style financing unlock that takes the facility to 104% LTC during the revenue phase — meaning the asset essentially finances itself and then some. McBee noted that for investment-grade offtake, the company is achieving 90% to 100% LTCs, compared to the 70% range for non-investment-grade. He framed the entire financing architecture around a two-tier structure: AssetCo, where individual GPU clusters and their long-term take-or-pay contracts sit, and ParentCo, which handles the residual capital needs through converts, high yield, and equity. The explicit goal is to make AssetCo increasingly self-funding, reducing the parent's need to tap capital markets. "AssetCo is kicking off net proceeds to parent," McBee said. "Parent will get a larger and larger stream of clean net proceeds from AssetCo that it just turns around and back down into AssetCo."
The cost of capital trajectory is striking. McBee noted that two to three years ago, the first DDTLs against Microsoft offtake were priced at SOFR plus 850. The equivalent transaction today, all else equal, would price at SOFR plus 200 to 225. That 625 basis point compression is entirely attributable to execution track record, not market conditions.
The Blackstone-Google TPU Deal Is a Non-Event for CoreWeave's Core Business
McBee addressed the widely covered Blackstone-Google TPU cloud deal directly and without much concern. His argument was straightforward: CoreWeave's clients come to it explicitly for NVIDIA GPUs, not TPUs, and the two architectures are not interchangeable. "The pipeline isn't full of clients saying, here's my TPU pricing, can you match that on GPU? No one asks that." He also pointed out that both Blackstone and Google remain active CoreWeave partners — Blackstone has participated in every financing transaction the company has done and will participate in the DDTL 5.0 close, while Google is described as a multibillion-dollar GPU client. McBee's framing was that the deal is simply another demand signal in an already demand-saturated market, not a competitive threat.
Inference Is Now Materially Over Half of Power Draw — and the Mix Is Shifting Fast
On the Q1 earnings call, CoreWeave disclosed that inference workloads now account for materially more than 50% of power consumption on its platform. McBee added important texture here. He attributed the shift partly to Hopper generation hardware maturing past its initial training phase, and partly to net-new clients — particularly in financial services — who are arriving on the platform for inference only, with no interest in training foundation models. The practical implication is that the same physical infrastructure is being used for a widening range of workloads, often within hours of each other. "They can be using the exact same infrastructure for training of next-gen foundation model across hundreds of thousands of GPUs, and then the next hour, they're running inference on it." There is no dedicated inference buildout planned, as clients are not requesting it and latency sensitivity does not yet demand it.
The inference trend also has a favorable implication for GPU asset life. Older Hopper and Ampere generation chips — the very assets that critics worry will depreciate quickly — are seeing rising spot prices and contract renewal demand. McBee noted that CoreWeave's original plan was to let those older GPUs roll into on-demand pools after their initial contract periods, but clients are instead signing new one-to-three-year contracts at full utilization. "It's really hard to say no to 100% utilization rates and firm economics for a multiyear period." He was careful not to pre-announce any depreciation policy change — the company uses a six-year life in line with peers — but suggested that assumption may prove conservative given actual market behavior.
Financial Services Is $10 Billion of Backlog — A Number the Market Is Missing
Perhaps the most underappreciated data point McBee dropped was the scale of financial services in CoreWeave's $100 billion backlog. "Financial services, in my notes, they're over $10 billion of our backlog today. I don't think many people in this room expect that our backlog has that much financial services associated with it." This cohort is described as a heavy inference user running primarily on Hopper, Blackwell, and even Ampere. It also serves as a concrete counter to the persistent investor question about whether enterprise demand is real — at $10 billion, it is real, it is contracted, and it is growing.
More broadly, McBee said CoreWeave added more than double its previous record for new logos in Q4, with the incremental growth coming almost entirely from enterprise. These are eight- and nine-figure deals, not the ten- and eleven-figure commitments that drive headline backlog numbers, but the volume and acceleration are meaningful. Enterprise clients are signing four-to-six-year contracts at the same margin profile and structure as AI Labs and hyperscalers, and McBee expects them to systematically underestimate their own demand — just as the company's earlier AI Lab clients did. "That's the exact same cadence we saw with our AI Lab clients, our hyperscale clients when we were really growing within those sectors."
Margin Path to Year-End: Mechanical, Not Speculative
McBee was direct in reaffirming CoreWeave's guidance that Q1 was the margin trough and that the company will exit the year at a low double-digit pro forma operating margin, with a long-term target of 25% to 30%. His confidence rests on visibility into specific deployments coming online in Q2 and Q3, all of which are tied to existing contractual commitments with known economics. "We know the cost of the infrastructure for all these deployments, and thus we know what the margin profile is for them." He framed the earlier margin compression as a timing artifact — investment precedes revenue in infrastructure — and noted that the Q4 concentration of last year's deployments was an anomaly that front-loaded cost without yet showing the revenue benefit.
The One Insight McBee Wants the Market to Internalize
Asked what investors sitting in the same room a year from now will appreciate that they do not appreciate today, McBee returned to a theme he described as one he "hammers on a lot": signed power does not equal revenue. "There is an oversimplification in the market of, well, this company has 500 megawatts of signed power. That must mean they're going to be able to easily translate that to 500 megawatts of GPU-associated revenue. It's just not the case." The bottlenecks he identified as most acute today are not electricity availability but powered shell capacity — the ability to actually consume electricity at the rack level — constrained by electricians, transformers, and backup battery supply chains that were not built to scale at the pace of AI demand. His expectation is that supply-demand balance in this market does not arrive before the end of the decade.
CoreWeave's answer to that constraint is its operational track record across 43-plus sites and what McBee described as a proprietary software stack — Mission Control — that bridges the gap between contracted power and billable GPU hours. That execution capability, not the hardware itself, is what he argues gives the company preferential access to NVIDIA's GPU supply and a structurally lower cost of capital than any peer attempting the same model without the same history. Whether that competitive advantage proves durable as more capital enters the space is the question this conference room did not fully resolve, but on the evidence presented, CoreWeave is operating well ahead of the field on the metrics that matter most.