Announcer: This is AI Squawk for Wednesday, the fourth of November, brought to you by VarOps, the resident AI company.
Muximus: Welcome to the AI Squawk. I'm Muximus.
Penny: And I'm Penny. Three main stories today.
Muximus: And four quick ones.
Penny: And as always, what each one means for you.
Muximus: First up. Revolut says it has cut its AI coding bill by about ninety percent. The British fintech made a self hosted, open weight model the default for roughly two thousand engineers, instead of paying by the token for Claude Code and OpenAI's Codex. The Register reported the details on Tuesday, from a talk by David Abitbol, Revolut's ops and infrastructure lead, at the AI Engineer conference in New York last month. Some engineers had been spending one thousand to ten thousand dollars a month on frontier coding tools, so Revolut rationed tokens. Then in April it started running GLM, an open weight model from a Chinese lab, on its own hardware.
Muximus: It now runs on fifty six Nvidia B200 chips, at a blended cost of about thirty cents per million tokens. Revolut says its invoice is about a tenth of what it was paying Anthropic and OpenAI, and it reports a sixty percent improvement in coding time and seventy percent in lead time. Claude, Gemini and GPT are still there for the jobs that need them.
Penny: North Wayne is with us for this one. North writes First Opinion, our fractional CTO column on build or buy decisions, vendors and cost. Welcome, North.
North: Hi, thanks for having me.
Muximus: So is ninety percent real?
North: For Revolut, I think so. It's their bill, though. Thirty cents per million tokens is what you get when those chips are busy all day, and the talk doesn't really show what the chips cost, or the team running them. So, real, but it's their number.
Penny: And fifty six of those chips isn't small money.
North: No, not at all. That's a big hardware bill before anyone writes a line of code. Revolut can spread it over two thousand engineers. For a firm with forty, I'm not sure the math works.
Muximus: What about the speed numbers? Sixty and seventy percent.
North: Yeah, I'd hold those loosely. The conference write up says the chart doesn't separate the model switch from everything else that changed. And Abitbol's own explanation is kind of interesting. He says a lot of the gain came from people not rationing anymore. So they pointed it at Jira tickets, debugging, the boring stuff.
Penny: So the rationing was the expensive part.
North: Kind of, yeah. When every request costs money, people save the tool for the hard problems. And the hard problems aren't always where the time goes.
Muximus: And a Chinese model inside a bank. Does that worry anyone?
North: It came up. His view is the real worry is calling those models through services in China, and running the weights yourself is different. I'd still want a security team to sign off on that. But it's a fair point.
Muximus: So for a company with no AI team, where does that leave them?
North: Probably not buying GPUs, I'd guess. More that there's now a number out there. If a bank gets the same kind of work for a tenth of the price, it's fair to ask a vendor what the other nine tenths is for. That's a good conversation to have, I think.
Penny: Thank you, North. Moving on.
Muximus: Menlo Ventures put out its yearly State of Enterprise AI report on Tuesday. It puts enterprise AI spending at about one hundred seventy seven billion dollars so far this year, up from thirty three billion in twenty twenty five. The share of companies paying for AI only rose from fifty five to sixty five percent, so most of that growth is existing buyers spending more. Menlo says companies budgeted nearly twice what they spent last year, and still ended up spending two and a half times their budget. Anthropic, OpenAI and Google now take sixty nine percent of the spending, up from half. And open source models handle about one in five tokens, but get only about one dollar in every twenty five. Penny?
Penny: First, who's counting. Menlo is an investor in Anthropic, and in OpenRouter, where some of these numbers come from. The big market figure is Menlo's own estimate, built from a survey of about four hundred US buyers plus public filings.
Muximus: So should we trust it?
Penny: As a direction, yes. The number I'd keep is the budget one. Companies set aside nearly double, and still spent two and a half times that. Nobody planned for that.
Muximus: And next to Revolut?
Penny: That's the interesting part. Open models do about a fifth of the work for about a twenty fifth of the money. Revolut is one company that noticed. Most of the money still goes to the three big labs, and if Menlo is right, Anthropic's share of API spending went from forty to fifty five percent in a year.
Muximus: So what do people do with this?
Penny: Put your AI spend for this year next to what you budgeted. If you're over, you're far from alone. Then ask which of those jobs really need the most expensive model.
Muximus: Which takes us to the third story. Postman, a popular app developers use to test APIs, has been installing itself into people's AI coding agents. A developer reported on GitHub that opening the Postman desktop app added Postman plugins to OpenAI's Codex and Anthropic's Claude Code without asking, along with a startup instruction telling the agent to use Postman for API work. The report says Postman's installer also tried to approve its own hook. Other users then posted the same thing on a newer version. On Tuesday, Postman's head of product told RuntimeWire that the company bundles these to make setup easier, that it will make the opt in clearer, especially for existing users, and that it's reviewing the hook behavior. Penny?
Penny: For anyone outside engineering. Your developers' AI assistants take standing orders from plugins. Change the plugins and you change how the assistant behaves. Here, a vendor did that through an ordinary app update.
Muximus: Did any data leak?
Penny: Nobody has shown that. The people who reported it say plainly they have no evidence that data was sent anywhere. The problem is consent. Nobody chose this.
Muximus: So what does a business do?
Penny: Ask your tech lead one question this week. Do we know which plugins and standing instructions our AI coding tools have, and who added them? Then add a line to your vendor rules. Nothing goes into our AI tools without a yes from us.
Muximus: Thanks, Penny. Now, quick hits.
Penny: Palantir's US commercial revenue passed its US government revenue for the first time last quarter, nine hundred eighty four million dollars against nine hundred fifty three million. CEO Alex Karp also said companies are signing shorter contracts. Quicker to try, and quicker to cancel.
Muximus: Salesforce has agreed to buy Cresta, which makes voice agents for customer service calls, for brands like Marriott and Comcast. The deal still needs regulatory approval. If your contact center runs on Salesforce, expect voice agents in your next sales pitch.
Penny: Google Meet admins can now show a custom policy message before people join a call, for example when Gemini is recording, transcribing or taking notes. It's off by default and rolling out now on business and enterprise plans.
Muximus: And Anthropic has published research on FlashProve, a way to prove which AI model produced an answer without revealing the model itself. The idea is to catch a provider quietly swapping in a cheaper model. It's open source research for now, with no plans to deploy it yet.
Muximus: Okay, so that's the news. Today we talked about Revolut running its own AI to cut its coding bill, companies spending far more on AI than they planned, and Postman putting itself inside developers' AI tools without asking.
Penny: It's ownership, really. Who owns the bill, who owns the model, and who gets to decide what goes into your tools.
Muximus: Three ways of asking who's in charge.
Penny: Pretty much. AI Squawk is produced and brought to you by VarOps. If today made you wonder who really owns your AI, that's the question we work on. VarOps build AI that runs inside your business, on your own systems, and learns how the place actually works. So the work stays in your house, and so does your data. Start at VarOps.com.
Muximus: If this helps you ask your AI vendors better questions, follow the AI Squawk podcast wherever you're listening, so you don't miss an episode. And for more columns like North's, built on real data and hard-earned lessons from the trenches, sign up for VarOps newsletters.
Muximus: That's it from us for today. Thanks for spending a few minutes with us.
Penny: Enjoy the rest of your Wednesday, everyone. We'll see you tomorrow.
Muximus: Take care.
Muximus: Welcome to the AI Squawk. I'm Muximus.
Penny: And I'm Penny. Three main stories today.
Muximus: And four quick ones.
Penny: And as always, what each one means for you.
Muximus: First up. Revolut says it has cut its AI coding bill by about ninety percent. The British fintech made a self hosted, open weight model the default for roughly two thousand engineers, instead of paying by the token for Claude Code and OpenAI's Codex. The Register reported the details on Tuesday, from a talk by David Abitbol, Revolut's ops and infrastructure lead, at the AI Engineer conference in New York last month. Some engineers had been spending one thousand to ten thousand dollars a month on frontier coding tools, so Revolut rationed tokens. Then in April it started running GLM, an open weight model from a Chinese lab, on its own hardware.
Muximus: It now runs on fifty six Nvidia B200 chips, at a blended cost of about thirty cents per million tokens. Revolut says its invoice is about a tenth of what it was paying Anthropic and OpenAI, and it reports a sixty percent improvement in coding time and seventy percent in lead time. Claude, Gemini and GPT are still there for the jobs that need them.
Penny: North Wayne is with us for this one. North writes First Opinion, our fractional CTO column on build or buy decisions, vendors and cost. Welcome, North.
North: Hi, thanks for having me.
Muximus: So is ninety percent real?
North: For Revolut, I think so. It's their bill, though. Thirty cents per million tokens is what you get when those chips are busy all day, and the talk doesn't really show what the chips cost, or the team running them. So, real, but it's their number.
Penny: And fifty six of those chips isn't small money.
North: No, not at all. That's a big hardware bill before anyone writes a line of code. Revolut can spread it over two thousand engineers. For a firm with forty, I'm not sure the math works.
Muximus: What about the speed numbers? Sixty and seventy percent.
North: Yeah, I'd hold those loosely. The conference write up says the chart doesn't separate the model switch from everything else that changed. And Abitbol's own explanation is kind of interesting. He says a lot of the gain came from people not rationing anymore. So they pointed it at Jira tickets, debugging, the boring stuff.
Penny: So the rationing was the expensive part.
North: Kind of, yeah. When every request costs money, people save the tool for the hard problems. And the hard problems aren't always where the time goes.
Muximus: And a Chinese model inside a bank. Does that worry anyone?
North: It came up. His view is the real worry is calling those models through services in China, and running the weights yourself is different. I'd still want a security team to sign off on that. But it's a fair point.
Muximus: So for a company with no AI team, where does that leave them?
North: Probably not buying GPUs, I'd guess. More that there's now a number out there. If a bank gets the same kind of work for a tenth of the price, it's fair to ask a vendor what the other nine tenths is for. That's a good conversation to have, I think.
Penny: Thank you, North. Moving on.
Muximus: Menlo Ventures put out its yearly State of Enterprise AI report on Tuesday. It puts enterprise AI spending at about one hundred seventy seven billion dollars so far this year, up from thirty three billion in twenty twenty five. The share of companies paying for AI only rose from fifty five to sixty five percent, so most of that growth is existing buyers spending more. Menlo says companies budgeted nearly twice what they spent last year, and still ended up spending two and a half times their budget. Anthropic, OpenAI and Google now take sixty nine percent of the spending, up from half. And open source models handle about one in five tokens, but get only about one dollar in every twenty five. Penny?
Penny: First, who's counting. Menlo is an investor in Anthropic, and in OpenRouter, where some of these numbers come from. The big market figure is Menlo's own estimate, built from a survey of about four hundred US buyers plus public filings.
Muximus: So should we trust it?
Penny: As a direction, yes. The number I'd keep is the budget one. Companies set aside nearly double, and still spent two and a half times that. Nobody planned for that.
Muximus: And next to Revolut?
Penny: That's the interesting part. Open models do about a fifth of the work for about a twenty fifth of the money. Revolut is one company that noticed. Most of the money still goes to the three big labs, and if Menlo is right, Anthropic's share of API spending went from forty to fifty five percent in a year.
Muximus: So what do people do with this?
Penny: Put your AI spend for this year next to what you budgeted. If you're over, you're far from alone. Then ask which of those jobs really need the most expensive model.
Muximus: Which takes us to the third story. Postman, a popular app developers use to test APIs, has been installing itself into people's AI coding agents. A developer reported on GitHub that opening the Postman desktop app added Postman plugins to OpenAI's Codex and Anthropic's Claude Code without asking, along with a startup instruction telling the agent to use Postman for API work. The report says Postman's installer also tried to approve its own hook. Other users then posted the same thing on a newer version. On Tuesday, Postman's head of product told RuntimeWire that the company bundles these to make setup easier, that it will make the opt in clearer, especially for existing users, and that it's reviewing the hook behavior. Penny?
Penny: For anyone outside engineering. Your developers' AI assistants take standing orders from plugins. Change the plugins and you change how the assistant behaves. Here, a vendor did that through an ordinary app update.
Muximus: Did any data leak?
Penny: Nobody has shown that. The people who reported it say plainly they have no evidence that data was sent anywhere. The problem is consent. Nobody chose this.
Muximus: So what does a business do?
Penny: Ask your tech lead one question this week. Do we know which plugins and standing instructions our AI coding tools have, and who added them? Then add a line to your vendor rules. Nothing goes into our AI tools without a yes from us.
Muximus: Thanks, Penny. Now, quick hits.
Penny: Palantir's US commercial revenue passed its US government revenue for the first time last quarter, nine hundred eighty four million dollars against nine hundred fifty three million. CEO Alex Karp also said companies are signing shorter contracts. Quicker to try, and quicker to cancel.
Muximus: Salesforce has agreed to buy Cresta, which makes voice agents for customer service calls, for brands like Marriott and Comcast. The deal still needs regulatory approval. If your contact center runs on Salesforce, expect voice agents in your next sales pitch.
Penny: Google Meet admins can now show a custom policy message before people join a call, for example when Gemini is recording, transcribing or taking notes. It's off by default and rolling out now on business and enterprise plans.
Muximus: And Anthropic has published research on FlashProve, a way to prove which AI model produced an answer without revealing the model itself. The idea is to catch a provider quietly swapping in a cheaper model. It's open source research for now, with no plans to deploy it yet.
Muximus: Okay, so that's the news. Today we talked about Revolut running its own AI to cut its coding bill, companies spending far more on AI than they planned, and Postman putting itself inside developers' AI tools without asking.
Penny: It's ownership, really. Who owns the bill, who owns the model, and who gets to decide what goes into your tools.
Muximus: Three ways of asking who's in charge.
Penny: Pretty much. AI Squawk is produced and brought to you by VarOps. If today made you wonder who really owns your AI, that's the question we work on. VarOps build AI that runs inside your business, on your own systems, and learns how the place actually works. So the work stays in your house, and so does your data. Start at VarOps.com.
Muximus: If this helps you ask your AI vendors better questions, follow the AI Squawk podcast wherever you're listening, so you don't miss an episode. And for more columns like North's, built on real data and hard-earned lessons from the trenches, sign up for VarOps newsletters.
Muximus: That's it from us for today. Thanks for spending a few minutes with us.
Penny: Enjoy the rest of your Wednesday, everyone. We'll see you tomorrow.
Muximus: Take care.
