CoreWeave says compute isn't a commodity. He's right, and he's selling.

CoreWeave co-founder Brannin McBee went on Bloomberg’s OddLots podcast and argued that AI compute cannot commoditize because it is not fungible. An H100-hour at CoreWeave, he said, is not the same as an H100-hour anywhere else: different goodput, different model-flop utilization, different operational reality under the same nameplate. No fungibility, no commodity.

The claim is accurate. It is also the keystone of CoreWeave’s entire equity story, and the two facts are related.

Writing in his “Buy the Rumor; Sell the News” newsletter on June 11, Dave Friedman takes the argument apart with unusual precision. Friedman points out that McBee spent years trading natural gas and power before founding CoreWeave, two markets defined by locational basis and firmness differentials. Power at one grid node is worth a different amount than power at another, and the congestion between them is itself a traded product. Gas has Henry Hub plus an entire complex of basis swaps pricing every delivery point against it. Both markets are among the deepest derivatives markets in the world, and they got there through non-fungible, location-specific physical assets. Commodity market design has never required sameness. It requires a standardized reference plus a basis you price separately.

McBee knows this. Friedman’s point is that when a former power trader tells you locational differentials prove there can be no market, he is not describing the commodity. He is pricing it, and revealing where the spread still hides.

The spread hides in the hard tier. Large training runs requiring specific high-bandwidth interconnects, reliable failure recovery, and tight operational SLAs cannot simply move to the cheapest available cluster. Switching costs are real, performance requirements are exacting, and the operational complexity of running at scale on an unfamiliar stack introduces risk that customers price into their vendor decisions. That is where CoreWeave’s moat lives, and it is a genuine moat.

The spread collapses in the easy tier. Batch inference, lower-priority jobs, and workloads with flexible scheduling can run on almost any provider with a compatible GPU class. DeepSeek’s efficiency gains, which have driven down the compute cost of inference significantly, push more workloads into this fungible tier. Oracle’s decision to commit roughly $40 billion in annual capex to AI infrastructure assumes a sustained demand for contracted, premium-tier compute. If DeepSeek-class efficiency continues to widen the fungible tier, a larger share of demand migrates toward commodity pricing, and the contracted, premium-tier market that underwrites Oracle’s capital plan and CoreWeave’s valuation becomes a smaller fraction of the total.

The same logic applies to the subscription-versus-API margin question. Frontier labs that sell API access at usage-based pricing are operating in the fungible tier, where customers will reprice on every new model release. Subscription products with deep workflow integration sit closer to the hard tier, where switching costs hold the spread open. CoreWeave’s $21 billion in capital raised year to date reflects an investor bet that the hard tier stays large enough to sustain a premium-operator business at scale.

Friedman’s framing resolves the apparent contradiction in McBee’s position. McBee is not denying that a compute market exists. He is specifying what the basis looks like: physical and spec fungibility (GPU class, DGX reference architecture, cooling) is largely settled; operational fungibility (goodput, failure recovery, software stack) and contractual fungibility (duration, SLA, priority, delivery window) are not. The non-settled layers are the spread. McBee is arguing that CoreWeave prices those layers better than its competitors, which is a market argument, not an anti-market argument.

The honest read: compute is partially commoditizing, and the pace of that commoditization in each tier is the central question for every AI infrastructure investment made in the next two years. Teams building on contracted GPU capacity should stress-test how much of their workload genuinely requires hard-tier performance, because every job that can move to the fungible tier will, and the price gap between the two tiers will only widen.

Source: Dave Friedman, “Buy the Rumor; Sell the News” newsletter (davefriedman.substack.com), published June 11, 2026.