Cisco gave a 27-year-old unlimited AI tokens to rewrite 98% of its old code. What are AI tokens, what does that cost, and who checks the work?

Cisco's president, Jeetu Patel, has given a 27-year-old engineer unlimited AI tokens and a goal that sounds unreal. The goal is to rewrite 98 percent of Cisco's legacy code within 24 months, with a team of two or three people. "I've given him unlimited tokens because he's going to refactor code," Patel told Fortune at a small roundtable dinner. "So I'm like, you can go, you can go nuts." The engineer is Padmanabhan Krishnamurthy, known as Paddy. He joined Cisco when it bought his translation startup, EzDubs, in 2025. Patel says he found him "buried away" five levels below him in the company. He has four weeks to prove the idea works, and no result has been reported yet.

The headline made it sound like a story about Gen Z getting a big break. I read every report I could find, Cisco's own filings and about 450 comments on Hacker News and Lemmy, and I think the headline points at the wrong thing. The unlimited tokens are the cheap part. Even a wild AI bill would be a tiny slice of Cisco's research budget. The expensive part, the part nobody is talking about, is who checks code that runs a big piece of the internet once a machine has rewritten it.

What are AI tokens?

Tokens are how AI companies measure and charge for work. A token is a small piece of text, a whole word, part of a word or a few characters. Anthropic's own guide says a token on its older models is about 3.5 English characters. Code works out to roughly 7 to 12 tokens per line, depending on the language.

Every time an AI model reads your code, it uses input tokens. Every line it writes back uses output tokens, which cost more. So "unlimited tokens" means Krishnamurthy can let AI read and rewrite as much code, as many times over, as he wants, and Cisco pays the bill without a cap. One note, because people search for this too: AI tokens have nothing to do with crypto tokens.

How much do AI tokens cost?

Here are the official list prices today for the top coding models, per million tokens:

• Anthropic Claude Opus 5.5: $4 for input and $20 for output
• OpenAI GPT-6 Astra: $10 for input and $50 for output
• Google Gemini 4 Argon: $2 for input and $10 for output, doubling to $4 and $20 on 1 December 2026

For a normal developer, that adds up slowly. Anthropic says companies using Claude Code spend about $150 to $250 per developer per month on average. Heavy users go far beyond that. In March, The New York Times found one Claude Code user with a bill of more than $150,000 in a single month, and Nvidia's Jensen Huang floated giving engineers about half their base pay again in tokens. In Silicon Valley, a token budget is turning into a job perk, the way free lunch used to be.

What could "unlimited tokens" cost Cisco?

Nobody has said which code Krishnamurthy is rewriting, which AI tool he uses or how many lines are involved, so here is our own rough estimate. Treat it as a sense of scale, not a fact.

Cisco's own blog said in 2021 that IOS XE, the software that runs many of its switches and routers, had 190 million lines of Cisco code. With outside libraries, it was more than 300 million. And that is one product family. The best public cost example I found is a case study InfoQ wrote up in September 2026. AI agents refactored a C codebase of 300,000 lines in three weeks for about $4,000 in tokens.

If you scale that up to 190 million lines, you get about $2.5 million. If Krishnamurthy and two teammates each spent at the $150,000-a-month pace the Times reported for all 24 months, the bill would be about $10.8 million. Cisco spent $9.56 billion on research and development last year. Even the wild version is about one tenth of one percent of one year's R&D.

So "go nuts" is a cheap thing to say. The cost that matters is the people needed to read, test and stand behind the result.

What is legacy code, and why does Cisco want to rewrite it?

Legacy code is old code that still runs something important. It was written years ago, often by people who have since left, and usually in older languages. Nobody wants to touch it, because nobody fully remembers why it works.

Nobody has said exactly which code is in the plan, but Cisco's oldest code is the software inside its network gear, much of it written in C and C++. Those languages let a coder make memory mistakes that hackers can use. After Fortune's story, Patel replied on X, as India Today reported. He called this a "2 to 3 person team" that catalyzes the work, and said the core teams then "own the final product". He also said it could help move code to memory-safe languages, which prevent that whole type of bug.

That reason is real. The US security agencies CISA and the FBI have said that software makers whose products run critical infrastructure, and who use memory-unsafe languages like C and C++, should publish a plan to fix that by 1 January 2026, starting with code that faces the network. The advice is voluntary, but it is clear. CISA's list of bugs that hackers are actively using has 102 Cisco entries, and 20 of them were added in 2026 alone, against 8 in all of 2025. Most of those are login and web flaws rather than memory bugs, so a new language alone would not fix them. One added in July 2026 was an 18-year-old bug in old Cisco IOS software. Old code stays in use, and old code stays a target.

Who checks code that an AI wrote?

Fortune's story never answers this question. Nobody has said how "98 percent" will be measured, how the rewritten code will be tested or who signs off.

Network software is hard to test. Former Cisco engineers on Hacker News were blunt. One who worked on IOS XR, Cisco's software for big routers, said the code is far more complex than outsiders think. It has to handle hardware quirks, strict timing and odd behaviour between different makers' gear. The same engineer called the project a "deluded vanity project". The InfoQ case study that gave us the cost number had a special test setup. It replayed the old code and the new code and compared every result after every change. InfoQ warned that most legacy systems have no such test to check against, which is why rewriting them is risky.

Lemmy users put it in a line that stuck with me: rewriting code is easy, but knowing why a strange workaround from 1998 is there is the real job. AI rewrites tend to clean those up, because they look pointless.

There is a warning from Amazon too. In March 2026, the FT saw an internal Amazon document about a run of outages with a "high blast radius". It named "GenAI-assisted changes" as one cause. One outage broke parts of Amazon's shopping site for six hours. The FT also reported that changes made with AI help would now need a senior engineer's sign-off. Amazon pushed back. It said only one incident involved AI tools and none involved AI-written code. It also said the meeting was routine and no new approval rule was added.

The good news is that the other side has real voices too. A security researcher wrote on Hacker News that he used AI to study a core router's code for handling internet routes, found nothing exploitable, but saw huge effort spent on memory safety by hand, which a safer language would handle on its own. He is, in his words, on team rewrite. I am closer to him than to the sceptics on the goal. My worry is the review, not the rewrite.

Is this good news for young engineers?

Patel tells it that way. He says companies need young people to "challenge the status quo", he treats interns as reverse mentors, and he wants more people like Krishnamurthy. That is a better attitude than most bosses have.

Still, one hand-picked person is not a hiring plan. Krishnamurthy is not a new graduate. He is a former startup co-founder who came in through an acquisition. And look at what he has been asked to do: automate the maintenance work that many experienced engineers at Cisco do today. Fortune's story does not say what happens to them.

The same company cut fewer than 4,000 jobs in May. The memo came out on the same day it reported record quarterly revenue, and it said Cisco was putting money into "our employees' use of AI". Cisco's yearly filings show its workforce fell from about 86,200 to about 82,400. Earlier this year Patel told Euronews about a new way of working. He said it lets Cisco turn a team of eight people into three people plus five AI agents, with triple the output.

The wider numbers point the same way. SignalFire found that new-graduate hiring at big tech is down about 65 percent from 2019. A Stanford study, updated in August 2026, looked at people aged 22 to 25 in the jobs most open to AI. Their employment is 19 percent below where it would otherwise be, mostly because companies hire fewer of them.

What does this mean for India?

More than for most countries. Patel said at Davos in January 2026 that India has Cisco's largest group of staff outside the US. Back in 2005, Cisco called its Bengaluru centre its biggest place for building products outside America.

The bigger link is Indian IT services. Keeping old code running and moving it to new systems is the bread and butter of TCS, Infosys, Wipro and HCLTech. In April, HCLTech's chief executive said the kinds of IT work most open to AI, like keeping old apps running, could shrink 3 to 5 percent a year for a few years. He put the hit to HCLTech's own revenue, which he called "AI deflation", at 2 to 3 percent a year. The top five Indian IT firms shrank by 7,389 people in the year to March 2026. TCS had made only 25,000 fresher offers by April and said more would depend on demand, though hiring picked up again from July to September. If two or three people with AI can redo a giant's old code, that is the work, and the fresher seats, that shrink.

In India the story spread fast as a win for an Indian-origin engineer, and abroad a lot of replies to it were plainly racist about Patel's name. Neither of those tells you whether the plan will work.

"This frees engineers to do better work"

That is the strongest case for Patel, and it deserves a fair hearing. If AI handles the boring rewrite, engineers can build new things. Patel has said AI will create more jobs than it destroys. A move to memory-safe code could also remove a whole type of bug that has hurt Cisco customers for decades.

I believe the memory-safety part. I do not believe the jobs part follows. AI can create jobs in the world while the jobs left at Cisco go to fewer, older reviewers rather than new graduates. Cisco's own record this year is fewer people, record revenue, and a finance chief celebrating the best earnings per employee in 30 years. The work is moving from writing code to checking it, and checking is a job for experienced people.

What Jeetu Patel should tell us

Patel turned a dinner comment into a story that travelled far beyond Cisco. He should now answer the questions it raised. Which code is in the 98 percent, and how will it be measured? How will it be tested on real hardware before it reaches customers? Who signs off? What happens to the engineers who maintain that code today? And will Cisco publish the four-week result, good or bad?

Cisco could prove me wrong. If it shows that AI moved its network software to memory-safe code with no new outages, that would be one of the most useful results any company shares this year. If it cannot say how the work is checked, then "go nuts" is the wrong order to give for software that runs other people's networks.

My prediction

By 31 October 2027, Cisco will announce another round of job cuts, and the announcement will again point to investment in AI. If I am wrong, I will say so here.

Cisco cut jobs in February 2024, August 2024 and May 2026, and each time it talked about AI and growth. Its president's own model turns eight people into three plus five agents. Its finance chief is measuring success in earnings per employee. A two-person rewrite of the company's old code is the kind of result that gets used to justify the next round.

Questions people are asking

What are AI tokens? Small pieces of text, a word, part of a word or a few characters, that AI models read and write. AI companies charge by the million tokens, with output tokens costing more than input.

How much do AI tokens cost? It depends on the model. Claude Opus 5.5 costs $4 per million input tokens and $20 per million output tokens. A typical developer using Claude Code spends about $150 to $250 a month, while one heavy user ran a bill above $150,000 in a month.

What does "unlimited AI tokens" mean? That the company pays for as much AI use as the person wants, with no cap. At Cisco's size, even a huge bill is small next to its yearly R&D spending.

What is legacy code? Old code that still runs important systems, often written years ago in older languages by people who have left. It is risky to change because nobody fully remembers why every part is there.

Who is Padmanabhan Krishnamurthy? He is a 27-year-old Cisco engineer. He co-founded EzDubs, a live translation startup that Cisco bought in 2025. Cisco's president gave him unlimited AI tokens to try to rewrite 98 percent of Cisco's legacy code in 24 months.

Did Cisco have layoffs in 2026? Yes. In May 2026 Cisco said it would cut fewer than 4,000 jobs, under 5 percent of staff, while investing in AI, and its workforce fell to about 82,400 by July.

Will AI replace software engineers? Not all of them, but the work is changing from writing code to checking it, and hiring of new graduates has fallen sharply. Experienced reviewers are in demand, entry-level coders less so.