Two of Silicon Valley’s biggest venture firms are now competing over compute, not only over deals. The Information reports that Andreessen Horowitz and General Catalyst are racing to secure the AI chips their portfolio startups need. Getting GPU access to founders is turning into a core part of what a VC sells.
⚡ The Quick Read
- What’s happening: a16z and General Catalyst are working to lock down chip supply for the startups they back, according to The Information.
- Why it matters: For AI startups, compute is often the hardest input to get. A VC that can supply it has an edge when founders choose whose term sheet to sign.
- The shift: Venture firms used to bring money, advice and introductions. The top firms are now adding infrastructure to that list.
- Who’s exposed: Smaller funds that can’t buy capacity in bulk could lose competitive AI rounds to firms that can.
🧠 How We Got Here
This didn’t come out of nowhere. In 2024 The Information reported that a16z was building a stockpile of thousands of Nvidia H100 GPUs, with plans to grow past 20,000. The firm later named the program Oxygen. It rents that capacity to portfolio companies so they can train and run models without fighting for scraps at market rates.
The reasoning was simple. a16z partner Anjney Midha said at the time that big cloud providers were favoring their largest customers, so startups got pushed to the back of the line. If you’re a seed-stage lab burning cash on a training run, that kind of delay can kill the company.
Not everyone agreed with the approach. Sequoia argued back then that the GPU crunch had peaked and that rising production would eventually flood the market. That hasn’t happened yet. Demand from frontier labs, hyperscalers and enterprise inference has kept supply tight, and the race The Information describes suggests the top firms still see compute as scarce and strategic.
🔍 Why This Is Significant
What stands out is General Catalyst joining in. When one firm does something, it’s an experiment. When two of the largest firms compete for the same resource, it’s turning into standard practice.
This changes the venture business in a few concrete ways:
- Capital is no longer the main thing VCs compete on. Money is plentiful for hot AI companies. Guaranteed compute isn’t.
- VCs are taking on infrastructure risk. Buying or reserving chips means betting billions of dollars on hardware that loses value fast as each new Nvidia generation ships.
- Portfolio lock-in gets stronger. If your training cluster comes from your lead investor, switching becomes harder and that investor gains more influence.
- Deal terms could shift. Earlier coverage of a16z’s program suggested compute might be traded for equity. That could become a new negotiating lever.
📉 The Risks Nobody Should Ignore
This strategy has real downsides. GPUs depreciate quickly, and a firm that overbuys right before supply catches up ends up holding expensive, aging hardware. Running compute also isn’t a VC’s core skill. Managing clusters, uptime and allocation between portfolio companies takes operations work that doesn’t fit the usual fund model.
There’s a fairness question too. When a VC decides which of its startups gets limited chips, it’s effectively picking winners inside its own portfolio. Founders should ask how those calls get made before they sign.
🎯 What Founders and Operators Should Do
- Ask about compute in every AI fundraise. It’s now a fair question for any lead investor: what access to GPUs comes with this check?
- Read the terms closely. Find out whether access is priced at cost, at a discount or in exchange for equity, and what happens if you leave.
- Don’t depend on one source. Investor-provided compute is useful, but keep relationships with cloud providers and specialized GPU clouds too.
- Watch smaller funds’ response. Expect pooled compute deals, partnerships with neoclouds and more creative ways to compete.
🔮 What Comes Next
If this race keeps growing, the line between venture firms and cloud providers will blur further. The firms with the deepest pockets may start to look partly like infrastructure companies. That would push the industry toward fewer, bigger players at the top.
The main question is whether chip scarcity lasts. If supply finally catches up, these programs become a costly perk. If it doesn’t, compute access could become the most valuable thing a VC brings to an AI startup. Full details are in The Information’s original report.