Why AI Vendor Lock-In Is Already Here:
Mozilla's CTO

Raffi Krikorian argues that most enterprises have already handed control of their AI to big platform companies. Krikorian is the CTO of Mozilla, which develops the Firefox browser; earlier, he led the infrastructure group at Twitter and the self-driving car division at Uber. His framing: "Most enterprises are renters, not owners." With their AI reached through API calls to those companies, "you're basically turning over your destiny over to a different system that you might be making a call to."

A Zapier survey found most enterprises cannot switch AI providers

Krikorian's evidence is a survey that Zapier conducted among enterprises on whether they could switch away from their AI provider. "Something like 85, 86% of them thought they could, but when they actually tried to do it, only about 30% of them could actually pull it off." His reading: "I don't think they willingly walked into that situation. In fact, they believed they could walk out of it, but it turns out to be quite hard to go do for a variety of different reasons."

He compared it with the cloud era, when buyers used tools such as Terraform as an abstraction layer so they could move between providers: "When AWS shows up one day and it's 'I want to give you a huge bill this month,' we can be 'Well, fine, we'll pay it, but next month we're moving,' and then that starts a negotiation."

Three things the buyer does not fully control

The model comes first: providers "can just change model behavior at any given moment. If they're having capacity issues, they could shuttle you off to a different model that might behave differently." For code written against that model, "it's not under your control; it's under control of the provider."

Cost comes second. Krikorian had expected token prices to fall. "They kind of have, but token usage has skyrocketed," he said, and he thinks that could be partly because "new models and new harnesses just need more tokens, which is in the benefit of the people I'm buying from." Usage comes third: "Once I set off my initial prompt, all the different things that happen in the harness could just exponentially grow, and that is also a little bit out of my control."

He cited Uber, where "post-IPO, the prices changed. All the subsidization that we did of Gen Z just basically disappeared, and the real true pricing showed up." Krikorian expects the $200 monthly plans to follow the same path: "I think that price either floats up significantly or we go into this per-API model."

Krikorian records every coding prompt to run his own model evaluations

Krikorian built a tool called Morph, on GitHub, that is "basically Git, but for prompt-based workflows." It records "what's the prompt, what's the harness, what's the model, when I ran it, all the settings and stuff like that," so a real workload can be replayed against a new model. He planned to do that over the weekend with the Opus model that had just come out, to see "if I had this model six months ago, this is the direction this piece of code would've gone into."

Public benchmarks "can be tuned for. You can tune yourself to be able to pass a particular benchmark." What has been lost, he said, is "doing our own evaluations on the workloads that we really care about."

An engineer told him his bosses allow only frontier models because they believe those will probably get the right answer the first time. Krikorian's reply: "It's a good belief, but where's the data behind it?"

Use frontier models first and scale with a model mix

Krikorian said "a responsible organization should get to" a hybrid of rented and locally hosted models, but timing matters. His team moved much of Twitter's code base from Ruby on Rails to the JVM around 2010 and 2011. He gives AI buyers the advice he gave startup engineers then: "Back in the beginning stages, you should be doing the thing you can do the fastest, and you can learn the most from. And in those cases, it probably is just use the frontier models."

Krikorian runs Qwen 30B on his laptop as his usual first resort, and may call out to Claude when it gets stuck: "It's stuck. Help me out here for a second." The case for local hosting goes beyond cost: "I want my data to be protected by architecture, not legal handshakes."

Write permissions are the open problem for agents

The question Krikorian keeps getting asked is "how do you manage write versus read permissions?" Whether actions are idempotent, "so if an agent does it 3 times I don't accidentally bill myself $1,000 each time." Asked whether companies should cap the number of agents, he offered an analogy: "The IT team is slowly becoming the HR team for agents effectively." On who should care about token costs, he said everyone, then named the CFO in partnership with the CIO or CTO, "potentially even at some point a board level conversation."

Krikorian's newsletter mapped an open AI stack from bare metal up

"It turns out almost all the layers are covered. Like you could do this today. I mean, the map is kind of red and screaming at us when it comes to enterprise readiness, but all the pieces exist," Krikorian said of the catalog he built in his newsletter, Owners Not Renters, mapping open source projects from bare metal to an OpenAI-compatible API. The gaps are compatibility, ease of use, and hardening. Mozilla is working on the bundle, "something someone can just pip install, apt-get install, whatever, onto their server, and it's up and running." His proof point is Pinterest: "Q3 of last year, Pinterest deployed open models instead of closed, and saved something in the order of $10 million that quarter alone by just making that switch." Krikorian said: "I think open source can get us there pretty fast and pretty well if we focus on it."

Watch the full conversation with Raffi Krikorian and read the complete transcript on the episode page.

CXOTalk prepared this article with AI assistance from the verbatim transcript of episode 920. Quotations are unedited from the broadcast.