Can CoreWeave Control the GPU Market?
Financialization of the GPU market seems inevitable. CoreWeave's executive team comes from commodities trading. Will they position themselves to control this market?
Snowflake vs Glencore: Two Very Different Futures
In a recent post, I argued that CoreWeave has no moat.
I compared it unfavorably to Snowflake—a company that owns no hardware yet commands one of the deepest moats in SaaS. Snowflake's edge doesn't come from controlling infrastructure. It comes from abstraction: a sticky software layer, developer entrenchment, and a control plane that monetizes workflows instead of raw compute.
CoreWeave, I wrote, owns GPUs—but not gravity. It sells access, not context. It leases data centers, relies on third-party chips, and serves customers who can leave as soon as the market normalizes.
That critique still holds. But I now see a different path for CoreWeave. One that doesn’t compete on abstraction at all. A path that plays not in SaaS but in commodities.
Two Frames, Two Futures
To understand the strategic tension here, you have to see CoreWeave through two irreconcilable paradigms.
🛠️ Frame 1: Snowflake vs CoreWeave — The Abstraction Thesis
Power flows from developer entrenchment
Moats come from control planes, not physical infrastructure
Switching costs and platform stickiness drive defensibility
By this logic, CoreWeave is exposed. It lacks an SDK. It lacks a native developer platform. It’s a logistics business dressed as a platform.
💎 Frame 2: Glencore vs CoreWeave — The Commodities Thesis
Power flows from controlling scarce infrastructure
Moats come from timing, liquidity, and capex scale
Levers include futures contracts, regional arbitrage, and structured offtake agreements
In this frame, CoreWeave is not a SaaS company. It’s a capital allocator for the AI age. Think: Exxon with an API. Or Glencore for GPU cycles.
Why These Frames Clash
The Snowflake frame asks: Can CoreWeave build a sticky abstraction layer that customers won’t leave?
The Glencore frame asks: Can CoreWeave become the market-maker for GPU access—across sectors, geographies, and sovereigns?
These aren’t two answers to one question. They’re answers to two entirely different games.
This is not contradiction. It’s divergence.
The Glencore Playbook, Applied to GPUs
Glencore is one of the largest commodity traders in the world. It rarely manufactures or extracts directly. Instead, it arbitrages supply, structures complex offtake agreements, exploits regional mismatches, and creates liquidity where none existed. Its power comes from timing, logistics, and information asymmetry—not ownership.
Now imagine that playbook applied to AI compute.
GPU blocks sold years in advance.
Private bilateral deals with sovereign buyers.
Spot market inventory held off-market to drive volatility pricing.
This isn’t hypothetical:
One-year private leases for H100s have traded at $4–5 per watt-hour—higher than some solar PPAs
AI compute demand surpassed the market value of global copper in 2024
Hyperscalers already use long-term structured commitments to manage capacity
The preconditions for financialization are already here.
But today’s market is dominated by opaque, illiquid, non-transferable forwards. The hyperscalers have built multibillion-dollar forward markets for GPU access—bespoke deals that serve their internal risk hedging, but do not establish price discovery, liquidity, or accessibility for the broader market.
A true futures market would change this. It would:
Standardize contracts across chip classes and time horizons
Enable hedging for smaller players, startups, and sovereigns
Introduce transparency, depth, and liquidity to what is now an insider’s game
In effect, it would turn compute into a real commodity.
What might these contracts look like?
1-Year H100 Futures (US-East Delivery): Standardized contracts for 1,000 GPU-hours of H100 compute, deliverable in Q2 2026 at a designated co-lo facility in Virginia. Priced per GPU-hour in USD. Tradable on a centralized exchange with daily margining.
Inference Capacity Strip (Rolling): A bundled contract offering sustained inference throughput (in TFLOPS) across three regions—US-West, EU-Central, APAC—over six months. Settled monthly in cash based on indexed spot GPU-hour pricing.
Sovereign Reserve Option: An options-style contract where a buyer (e.g., a national lab or defense agency) pays a premium for the right, but not the obligation, to secure 100,000 GPU-hours within a defined window at a pre-set strike price, protecting against price spikes.
These are not merely financial instruments. They are hedging tools, capacity guarantees, and allocation frameworks, all of which enable smaller players to compete, sovereigns to plan, and capital markets to engage.
The capital allocation impact is profound. With standardized futures:
Infrastructure builders can secure financing against predictable future revenue, just as solar developers do with PPAs
Investors can underwrite build-outs based on credible forward curves, enabling GPU infrastructure to be financed like energy projects
Equipment lessors and sovereign buyers can use futures as collateral, enhancing credit access for high-capex initiatives
Second-order effects emerge quickly.
A yield curve for compute unlocks structured products (e.g., AI income funds)
Insurers can create coverage products against GPU delivery risk, slippage, or regional outages
AI-native ETFs could track long or short GPU capacity pricing, allowing institutional investors to express directional bets on AI demand
Credit markets could price lending for startups based on their hedged compute exposure, much like energy developers with fixed offtake agreements
In other words: compute liquidity doesn’t just lubricate AI growth—it transforms it into an investable asset class. They are hedging tools, capacity guarantees, and allocation frameworks, all of which enable smaller players to compete, sovereigns to plan, and capital markets to engage.
So Where Is CoreWeave?
📉 In the Short Term: Exposed
Microsoft/OpenAI is its largest customer and existential risk
Gross margins are brittle
It lacks a software abstraction layer to create switching costs
The moment GPU prices normalize, its value proposition evaporates
📈 In the Long Term: A Commodity Market Maker
If compute becomes a tradable industrial commodity, abstraction-first platforms will lose to infrastructure liquidity.
This is where CoreWeave’s background gives it an edge: it emerged from crypto mining—an industry defined by capex scale, market volatility, and energy arbitrage. Before its founders played in crypto, they were commodities traders. It already speaks the language of risk spreads and power pricing.
Its path to a moat lies not in abstraction, but in becoming the liquidity layer for intelligence capacity.
The Transition Window
But to get there, it has to survive the valley.
This is the most dangerous stretch in CoreWeave’s life:
Hyperscalers are vertically integrating
Nvidia is abstracting at the software layer (via NIMs)
Spot GPU prices are volatile and could collapse
The abstraction layer players (like Hugging Face or Mosaic) are rising fast
In this window, capital missteps can kill. So can overexposure to a single buyer. To cross this chasm, CoreWeave needs to:
Diversify offtake partners geographically and sectorally
Build long-term structured contracts with sovereigns and industrial AI consumers
Begin financializing compute blocks like energy futures or LNG
The Endgame
If CoreWeave makes it across, it won’t be the Snowflake of compute.
It will be something rarer and more powerful:
The CME of GPU futures
The ICE of AI liquidity
The Glencore of cognition
And that’s a very different kind of moat.