CoreWeave co-founder Brannin McBee recently argued that compute still does not behave like a true commodity because it is not fully fungible across providers. His comments, made on Bloomberg's OddLots podcast, focused on the idea that one GPU hour is not always equivalent to another, even when the hardware appears similar on paper.
McBee's core point was that compute varies in ways that affect performance, utilization, and reliability. He pointed to differences in goodput and model-flop utilization, saying those operational factors mean two seemingly identical workloads can produce different results depending on where they run. In his view, that lack of interchangeability makes compute harder to treat like gold, oil, or other standardized commodities.
The argument matters because CoreWeave has built part of its business around specialized AI infrastructure, and the question of whether compute can be traded more like a commodity has implications for pricing, margins, and valuation. If compute were widely seen as interchangeable, customers might bargain more aggressively on price and investors could view the business more like a utility. If it remains differentiated by performance and service quality, operators may retain more pricing power.
Still, the case against commoditization is not as simple as saying non-fungible assets cannot become commodities. The source analysis notes that many established commodity markets operate with imperfect interchangeability. Natural gas, for example, is priced against benchmark hubs, while local delivery points trade at premiums or discounts. Power markets also vary by location, delivery timing, and reliability. Those markets are deeply developed despite differences in the underlying product.
That comparison suggests compute could follow a similar path. Instead of being perfectly uniform, a future market might build around a standard reference product and then price the differences separately. In that model, buyers and sellers would still trade a benchmark, but quality, topology, software environment, scale, and service-level terms would all affect the final price.
McBee's comments also highlight the lack of a clear convergence mechanism in compute. In gas and electricity markets, infrastructure such as pipelines, storage, transmission, and delivery nodes helps prices converge around a benchmark. Compute does not yet have an equivalent system. The market is fragmented by clusters, queues, security rules, hardware environments, and data location. Those differences make it difficult to create a single, universally accepted reference price.
The analysis also frames McBee's comments as part of a broader investor debate around CoreWeave. The company has been able to raise large amounts of capital, and its non-commodity positioning supports a valuation narrative based on contracted revenue, operator expertise, and long useful lives for expensive GPUs. A commodity framing, by contrast, would imply more aggressive depreciation and weaker margins.
From that perspective, McBee is not dismissing commodity logic altogether. Instead, he is emphasizing that compute still lacks the standardization needed to trade cleanly at a single price. Even so, the source argues that the right lesson may be less about whether compute can ever become a commodity and more about how its spreads, service levels, and performance differences could eventually be priced.
For now, the debate remains open. Compute may not yet be fungible enough to trade like gold, but the market structures that made gas and power tradable suggest that imperfect products can still support sophisticated financial markets.