Summary of How CoreWeave Sees the Market for Compute Right Now

Overview of How CoreWeave Sees the Market for Compute Right Now

In this Odd Lots episode from Bloomberg, Joe Weisenthal and Tracy Alloway speak with CoreWeave co-founder and Chief Development Officer Brandon McBee about the rapidly evolving AI compute market. The conversation focuses on whether demand for inference is still accelerating, how CoreWeave’s customer base is changing, why NVIDIA remains the dominant hardware choice, and whether compute could eventually become a tradable commodity with a futures market.

Key Takeaways

• AI compute demand is still extremely strong.

  • CoreWeave says it is not seeing a meaningful pullback in inference demand.
  • McBee argues that the market is still in the phase of broadening adoption, especially beyond AI labs and hyperscalers into enterprise and financial services.

• Enterprise adoption is expanding the customer base.

  • CoreWeave says its client mix has diversified significantly.
  • It now serves:
    • hyperscalers,
    • major AI labs,
    • and a growing enterprise segment.
  • McBee cited 9 of the top 10 AI labs as customers and said the company has 10+ billion-dollar clients.

• Model routing matters for cost efficiency.

  • The hosts and guest discussed the idea that not every task needs the most advanced frontier model.
  • A major future efficiency gain may come from routing workloads to the cheapest model that can do the job well.

• NVIDIA remains the default infrastructure choice.

  • CoreWeave says customers continue to overwhelmingly demand NVIDIA-based compute for both training and inference.
  • The company sees NVIDIA’s CUDA ecosystem and operational reliability as key reasons for that dominance.

• Compute is not yet a true commodity.

  • McBee’s central argument against a near-term compute futures market is that GPU compute is not fungible enough.
  • Performance can vary depending on:
    • the data center setup,
    • the operator’s software stack,
    • rack configuration,
    • cooling and power delivery,
    • and uptime/failure management.
  • In his view, this makes it hard to standardize compute the way markets do with commodities like gold or oil.

• The biggest bottleneck has shifted.

  • Three years ago the issue was getting GPUs.
  • Now CoreWeave says the limiting factor is more often “powered shells” — data centers with power, cooling, and supporting infrastructure ready to receive GPUs.
  • Other constraints include transformers, backup power, electricians, and broader construction/supply-chain limits.

What CoreWeave Thinks Is Changing in AI

Inference Is Driving the Current Wave

McBee says the recent AI boom is being powered less by experimentation and more by real usage at scale, especially inference. He believes the market has moved beyond novelty use cases like simple prompts or poems, and toward practical, business-critical deployment.

Longer-Term Contracts Are Becoming the Norm

CoreWeave says customers are asking for longer-duration access to infrastructure:

• contracts moved from roughly 3 years,
• to 4 years,
• and now in some cases 5 years.

That shift suggests customers want stability and continuous access to high-end infrastructure rather than short-term bursts.

Infrastructure Lifespan May Be Longer Than Expected

A key point from the episode: the market may have been too aggressive in assuming today’s GPUs will become obsolete quickly.

• McBee argued that hardware like H100s and even A100s may remain useful for years.
• Why? Because different workloads can be served by different model sizes and different levels of infrastructure.
• That extends the useful life of existing compute assets and weakens the idea that GPU hardware rapidly becomes worthless.

Hardware and Architecture Discussion

NVIDIA Generation Cycle

The conversation touched on the progression of NVIDIA hardware:

• Hopper
• Blackwell
• Vera Rubin (the next architecture in NVIDIA’s roadmap)

McBee described Blackwell as a major change in deployment format, moving from older 42U racks to larger, more complex, often liquid-cooled systems. Vera Rubin is expected to continue that progression.

Custom Silicon vs. NVIDIA

The hosts also asked about custom chips from companies like Microsoft and whether future AI systems might optimize around proprietary hardware.

CoreWeave’s view:

• it can run multiple types of silicon,
• but customers are still overwhelmingly requesting NVIDIA,
• especially because NVIDIA’s ecosystem is proven, efficient, and operationally mature.

Financing and Capital Markets

CoreWeave emphasized that its business model is now much more financeable than it was in earlier years.

What Improved

• The company says it has raised over $21 billion in financing year to date.
• Its cost of capital has improved as it has built a track record of delivering infrastructure and billable GPU hours.
• Investors are becoming more comfortable with the model because demand is visible and contracts are long-dated.

How CoreWeave Finances Itself

McBee described a structure where:

• customer contracts are paired with infrastructure assets,
• financing is organized through asset-level vehicles,
• revenue from contracts pays down the debt tied to the infrastructure.

This helps convert uncertain future demand into financeable cash flow.

Will There Be a Compute Futures Market?

This was one of the most interesting parts of the episode.

The Case for a Market

The hosts noted that if compute became more standardized, a futures market could:

• improve hedging,
• lower financing costs,
• and create new trading opportunities.

CoreWeave’s Skepticism

McBee’s main objection:

• GPU compute is not yet fungible enough to trade like a commodity.
• The same nominal GPU can perform very differently depending on how and where it is deployed.
• Until operations become more standardized, a liquid futures market may be difficult to sustain.

Notable Quotes and Ideas

• “What people need is the right model for the right prompt.”

  • This underpins the idea of model routing and cost optimization.

• “Inference demand remains unrelenting.”

  • CoreWeave says there is no sign of a demand slowdown.

• “The bottleneck is the powered shell.”

  • A concise description of the real infrastructure constraint now facing AI buildouts.

• “Compute is not fungible yet.”

  • The core reason CoreWeave thinks a commodity market for GPU compute is still premature.

Bottom Line

CoreWeave’s view of the market is bullish but nuanced: AI demand is still accelerating, enterprise adoption is widening, and infrastructure remains scarce. However, the industry’s challenge has shifted from simply buying GPUs to building and operating highly complex systems efficiently. That operational complexity, CoreWeave argues, is exactly why compute is not yet a clean commodity — and why NVIDIA-backed infrastructure still dominates the market.