Key Takeaways

  • Silicon Valley consensus expects 90% of token volume to run through open weights, but ClickHouse CEO Aaron Katz predicts a clean 50/50 enterprise split.
  • Enterprise software historically split evenly between open-source tools and proprietary vendors, and AI architecture will follow the exact same historical boundary.
  • The limiting factor for open models is not raw performance; it is the total lack of third-party indemnification for output inference, particularly from models originating in China.
  • ClickHouse allows internal teams to use open models for tasks like code reviews, but bars untrusted open weights from runtime production code.

The 90% Token Myth

Every venture capitalist on your timeline claims open weights will eat the software market. The thesis sounds clean: open models run cheaper, catch up to closed models within months, and let developers own their weights. They project that 90% of all future tokens will execute on open models.

Katz runs ClickHouse, which powers real-time analytics for thousands of organizations and scaled past $350 million in ARR. He thinks the 90% prediction ignores how enterprise legal teams actually work.

“Well, the easy answer is 50/50 in the same way that what percentage of enterprise software today is open source versus proprietary,” Katz explains. “I think it's a pretty even distribution.”

Enterprise software reached equilibrium a decade ago. Companies run Linux and PostgreSQL alongside closed SaaS platforms. They use open tools where internal control matters and buy proprietary software when they need a vendor on the hook for service level agreements. AI will settle into the exact same pattern.

The Indemnification Wall

The real barrier to open-weight adoption is legal liability. When a company deploys an AI model to generate customer-facing code or process enterprise data, the board cares about IP contamination and lawsuits.

Frontier labs like OpenAI, Microsoft, and Google sell more than API access. They sell legal protection. If a proprietary model generates copyrighted code or triggers an IP lawsuit, the frontier provider covers the legal defense.

Open-weight models, especially models released by Chinese labs, offer zero commercial indemnification. If that model outputs code derived from an unlicensed repository, the liability sits entirely on your balance sheet.

“I don't because I think it especially in the enterprise they want provisions and protections that potentially openweight models especially those that come out of China cannot provide around indemnification for example and output inference,” Katz says. “I think when you need the legal protection that most enterprises do you're going to want to work with one of the frontier lab providers. I think there's too much security concern around some of these openweight models.”

Even inside ClickHouse, engineering leadership maintains strict boundaries. “We have our own concern internally, right? Because you worry about the output from that code generation. You worry about third party indemnification if you were to consume code that's being derived from another repository.” ClickHouse permits open models for internal developer code reviews, but isolates them from production environments that touch proprietary IP.

What to Do With This

Audit your company's AI stack before your next enterprise customer security review. Categorize every model deployment into two buckets: internal developer assistance versus production customer runtime. If you use open-weight models to generate code or process customer data in production, meet with your legal counsel this week to quantify your exposure to copyright and output liability without vendor indemnification.