More and more interesting perspectives over just how productive AI is making engineers are coming to the fore.
Hock Tan, CEO of Broadcom — one of the world’s largest semiconductor and infrastructure software companies — has offered a concrete articulation yet of what AI-driven productivity actually looks like in practice. Speaking about Broadcom’s use of Claude Opus, Tan laid out a compelling case for why the return on investment in AI tooling, once engineers truly learn to use it, is difficult to argue against.
On the broader picture, Tan acknowledged that Broadcom is perhaps not as far along as some peers in terms of using AI. But the directional logic, he argued, is clear. A big part of what’s happening is that when engineers start using a tool like Opus 4.7, they go through a learning process — even in terms of how to use it. You can’t get a tool to be very productive at the start. But as you keep using it consistently, and use it long enough, you get very good at it, and then the tool becomes much more productive.
And as productivity compounds, the economics become striking. Think about it this way, Tan said: you can get one great, very senior engineer to produce an application design in one week — something that might otherwise take ten engineers, each earning $300,000 a year, three months to produce the same thing. The return on investment is still pretty compelling.
That framing — one engineer, one week, versus ten engineers, three months — is not a theoretical projection. It is Tan’s read of what is already happening at the frontier of AI-assisted engineering, conditional on teams that have put in the time to get genuinely good with these tools.
Tan also pushed back on any suggestion that Broadcom is throttling its AI usage. The question of whether a company is pulling back, he said, is really a question of return on investment and which application you’re using. And it’s also a question of being trained well enough to use the tools better and better over time.
The implications of Tan’s remarks extend well beyond Broadcom. What he is describing is a learning curve dynamic: AI tools are not plug-and-play productivity multipliers from day one, but they compound significantly as engineers develop fluency with them. That framing matters because it shifts the conversation from “does AI make engineers more productive?” — which is now largely settled — to “how quickly can teams climb the curve, and what does the ceiling look like?”
The broader industry is already grappling with those questions. Anthropic’s own internal data shows its engineers are writing eight times more code than they were 18 months ago, with the acceleration tracking almost directly against its own model releases. Salesforce stopped hiring software engineers entirely in 2025, citing productivity gains of over 30% from AI — and by mid-year, CEO Marc Benioff was reporting that AI was doing between 30 and 50 percent of the work across the company. NVIDIA’s Jensen Huang has made a near-identical productivity-expansion argument: when a single engineer can generate dramatically more output, the economic incentive is to hire more of them, not fewer.
Tan’s version of that argument is more grounded in specific numbers — and the numbers are arresting. A 10x reduction in headcount and a 12x compression in timeline, at a fraction of the salary cost, represent a fundamental shift in how software gets built and what it costs. The caveat, and it is a meaningful one, is that the productivity gains Tan describes are not automatic — they accrue to engineers who invest the time to get genuinely skilled with these tools. That learning curve may, in the near term, be one of the most important competitive differentiators between engineering teams.