Someone's finally building the Blue Book for AI chips — and CME wants to build futures on it.
The Summary
- Silicon Data raised $30.5M to build pricing and performance benchmarks for AI compute, positioning itself as an "independent referee" for the GPU market
- CME Group is planning to launch GPU futures contracts tied to Silicon Data's benchmarks, creating the first derivatives market for AI compute
- The company tracks real-time depreciation curves for older chips, giving buyers actual price discovery in a market that's been flying blind
The Signal
The AI compute market has a used car problem. Nobody knows what anything is actually worth. Silicon Data just raised $30.5M to fix that by becoming the Kelley Blue Book for GPUs.
CEO Carmen Li says the company is building "independent referee" benchmarks for both pricing and performance across the AI compute stack. That means tracking what H100s actually trade for today, not what Nvidia says MSRP should be. It means measuring real inference speed, not spec sheet promises. And it means publishing depreciation curves so buyers can see how fast their $30,000 GPU investment turns into a $15,000 doorstop.
The real signal here is CME's involvement. CME Group doesn't build futures markets for fun. They build them when there's genuine price risk that sophisticated players want to hedge. If CME is planning GPU futures tied to Silicon Data benchmarks, that tells you two things. First, the compute market is big enough and volatile enough that hedging demand exists. Second, someone credible finally built benchmarks that can serve as settlement prices for real contracts.
"CME's planned GPU futures tied to Silicon Data benchmarks mark the first derivatives market for AI compute."
This is infrastructure for the agent economy. Right now, if you're building an AI company, you're guessing at compute costs six months out. You're signing cloud contracts with pricing that might look stupid in three months when the next chip generation drops. You're buying hardware with no idea what the resale market looks like. Silicon Data turns that guessing into data. CME turns that data into hedges.
The depreciation tracking is the sleeper feature. Older chips don't just get slower, they get less useful for frontier model training. But they might still be perfectly fine for inference, for fine-tuning, for mid-tier workloads. Silicon Data's benchmarks show the actual value curve, not just the obsolescence narrative. That creates secondary markets. It creates arbitrage opportunities. It means the GPU you bought for training might have a second life you can actually price.
Key implications for the compute stack:
- Price discovery turns compute from magic to commodity
- Depreciation data unlocks secondary markets for older hardware
- Futures contracts mean builders can hedge their biggest cost
The Implication
Watch what happens when CME launches those futures. The moment you can hedge GPU prices, the moment you can short compute capacity you think is overvalued, you get real price discovery. That changes planning for every AI company. It changes how VCs model compute costs in their projections. It changes whether you build on cloud or buy hardware.
Silicon Data isn't just benchmarking chips. They're building the financial rails for the compute economy. If they get this right, "independent referee" undersells it. They're building the pricing engine for Web4 infrastructure.