Analysis
I have spent the last two weeks reading rounds where the product is other people's GPUs, and I keep landing in the same place: this is not venture investing, and I would rather own the software running on top.
Look at what closed in five days, per Crunchbase's weekly tally:
- Crusoe -- $3B Series F at $30B (Atreides Management, Valor Equity Partners): AI cloud and data centers, Denver.
- Fluidstack -- $1.5B at $18B (Jane Street Capital): GPU infrastructure, New York.
- Gimlet Labs -- $300M Series B at $3B (a16z, Sapphire, Menlo): multi-silicon inference cloud, San Francisco.
“- Fluidstack -- $1.5B at $18B (Jane Street Capital): GPU infrastructure, New York.”
Roughly $4.8 billion in five days, all of it in the layer between the chips and the models, none of it in anything an end customer buys.
Jane Street leading an $18 billion round is the part everyone glossed over. That is a trading desk pricing a stream of contracted cash flows from Anthropic, Google and Meta. It is a credit underwrite. The returns look like credit returns: capped, dependent on counterparty behavior, sensitive to whether take-or-pay language survives a customer's own down round. Venture funds are participating in that trade at venture prices and calling it a growth investment.
Meanwhile the application layer is where the actual gross margins live and where nobody is bidding. Pulse tracked all three of these last week:
- Thyme Care -- $125M Series E at $2B (led by JPMorgan's Morgan Health): oncology care coordination, Nashville.
- HiddenLayer -- $100M Series B (Delta-v Capital): security for AI models and workflows, Austin.
- TabaPay -- $155M growth round (FTV Capital): money-movement infrastructure, Mountain View.
Companies with customers, contracts and pricing power, raising a combined fraction of one week's compute total.
My view is simple. Compute is a capacity business with a depreciation schedule and a small number of buyers who can dictate terms. Applications are a distribution business with many buyers and switching costs. In every prior infrastructure cycle -- fiber, cloud, mobile -- the capacity layer got overbuilt first and the durable equity value accrued to whoever owned the customer. I do not think this one is different, and I think the seed and Series A market is currently mispricing that.
Room for disagreement: The strongest counter is that this cycle's application layer has no moat. If model capability keeps improving, a vertical AI application is a thin wrapper that the model provider absorbs, while a data center with a signed ten-year contract and an interconnection queue position is genuinely scarce. That argument has real evidence behind it -- power and land are constrained in a way software is not, and several 2024-vintage AI application companies have already been flattened by a model release. If you believe capability improvement is the dominant force for the next three years, the compute trade is the rational one and I am early.
I would still rather be early on the layer that gets to charge the customer.
