Apple Cut Some Engineers' Daily AI Budget to $200 From $300, Bloomberg Reports. Revolut Says Self-Hosting Cut Its AI Coding Bill About 90 Percent.

On Wednesday, November 4, 2026, Bloomberg reported that Apple has started putting its own employees on a tighter AI budget. In October, according to Bloomberg's sources, Apple quietly lowered the daily spending cap for some engineers using AI models to $200 from $300. More cuts are planned.

A day earlier, The Register published a very different answer to the same problem. Revolut, the British fintech, says that in less than a year it cut its AI coding inference costs by roughly 90 percent. It did not do it with a cap. It changed the default model, and then it ran that model itself.

Put the two side by side and you get a clear picture of how big companies are learning to pay for AI coding. One is rationing tokens. The other is making them cheaper.

What Bloomberg Reported About Apple

• Engineers: some daily caps fell to $200 from $300 in October. Further cuts could take some engineering teams to as little as $100 a day.
• Everyone else: many non-engineering groups already had a $100 daily limit, and Apple has considered going as low as $25 for some teams.
• Workarounds: some teams are being told to use cheaper models or models that run on the device. Employees who need more can ask senior leadership for an exception.
• Tools: employees have access to several services, including Anthropic's Claude, OpenAI's ChatGPT and Google's Gemini. 9to5Mac notes it is not clear which tools the new limits cover.
• Savings: Bloomberg's sources said "few workers were using their full $300 daily allowances," so it is unclear how much Apple will actually save.

The figures come from people familiar with the changes, not from Apple. Apple has not commented publicly; TechCrunch said it had asked. MacRumors and 9to5Mac tie the cuts to a broader push for leaner operations under John Ternus, who became CEO on September 1.

This is not Apple spending less on research. Its annual report, filed with the SEC on Tuesday, November 3, shows research and development expense of $45.77 billion for fiscal 2026, up about 32 percent from $34.55 billion a year earlier. The squeeze is on the per-person AI tab, not on the R&D budget.

What Revolut Says It Did

David Abitbol, Revolut's ops and infrastructure lead, laid out the numbers in a talk at AI Engineer NYC on Tuesday, October 13. The Register wrote it up on November 3. All of the figures below are Revolut's own.

• Default model: self-hosted open-weights GLM from Chinese developer Z.ai, first GLM-5.1 and now GLM-5.3. Before that, Revolut deployed Gemini in early 2025 and Claude Code in early 2026.
• Hardware: 56 Nvidia B200 GPUs serve GLM-5.3 to roughly 2,000 engineers.
• Cost: about $0.30 per million input and output tokens for self-hosted GLM-5.1, which The Register says came to about a tenth of the cost of Codex or Claude Code. That is Revolut's cost on its own hardware, not a price anyone else can buy.
• Speed: a target of under 3 seconds to the first token, and about 200 tokens per second of interactive throughput.
• Limits: GLM is text-only, so a router sends image work to hosted models such as Gemini.

Here is the detail that matters most for Apple. Revolut started where Apple is now. "We were seeing engineers spend $1,000 a month, $10,000 a month with not so much ROI, and so when we launched it we launched it with strict budget limits," Abitbol said, per The Register. The Register says those limits are what led the company to investigate self-hosting.

Once tokens were cheap, the use changed. The biggest gains, Abitbol said, came from letting models handle debugging and Jira and Confluence tickets; the biggest bottleneck, he noted, was not the code. "When you have unlimited tokens, you're free to do that," he said. Hosted frontier models did not disappear. In his words, "the frontier is the exception."

He was blunt about latency, too. "As soon as you exceed 5 seconds, my Slack inbox starts to blow up."

Two Ways to Ration Tokens

Some quick arithmetic shows how far apart these are. Over a 250-day working year, a $300 daily cap is up to $75,000 per engineer, $200 is $50,000 and $100 is $25,000. Those are ceilings, not spend: BloombergThose are ceilings, not spend, since Bloomberg says few people hit them.apos;s sources said few workers used the full $300. At Revolut's reported $0.30 per million tokens, $200 would buy about 667 million tokens. Revolut is spreading about 36 engineers across each of its 56 GPUs, by our math.

Our Take

A per-day cap is the bluntest tool a finance team has. It is easy to announce and easy to enforce. But if few engineers were reaching $300, cutting the ceiling to $200 mostly trims the outliers. That is why I think the most important line in Bloomberg's report is the quieter one: some teams are being told to use cheaper models or on-device AI. That is a routing decision, and routing is where the money is.

Revolut shows what that looks like when you commit to it. It tried strict budgets first, then stopped asking 2,000 engineers to use less AI. It changed what the default request costs, and kept the expensive frontier tools for the cases that need them. Its 90 percent figure is self-reported and depends on its own hardware and tuning, so I would not treat it as a benchmark for anyone else. The direction is still hard to argue with.

The awkward part is for the labs. Apple is one of the richest companies in the world, and its R&D expense grew by about $11.2 billion in fiscal 2026. If even Apple is putting a daily dollar number on its employees' AI use, enterprise coding revenue is not going to grow just because seats keep getting added. On June 27, we wrote that the tokenmaxxing era was ending. Four months later, the bigger risk for hosted model vendors is not caps. It is the Revolut playbook: the cheap open model becomes the default, and the frontier model becomes the exception.

My read: Apple's cap is a signal to employees. Revolut's router is a signal to vendors. The second one is the one Anthropic, OpenAI and Google should be worried about.

Three Signposts for the Next 60 Days

First, whether Apple's planned cuts to $100 for some engineering teams happen, and whether Bloomberg or Apple says which tools they cover. Second, whether other large employers disclose per-person AI caps or, like Revolut, a self-hosted default model. Third, whether the hosted coding vendors answer with cheaper default tiers or volume pricing aimed at the companies now rationing them.

Sources: Bloomberg, MacRumors, 9to5Mac, TechCrunch, Apple R&D expense in SEC filings (XBRL), The Register, AI Engineer NYC talk page, Runtime Wire.