Facing Down the "AI Is Useless" Argument, Again
MIT recently published a report titled "The GenAI Divide: State of AI in Business 2025," claiming that 95% of companies see zero return from generative AI. The finding immediately triggered a selloff in US tech stocks. But if you actually read the report, it isn't nearly as sensational as the headlines suggest — it's a fairly measured piece about when companies succeed or fail at deploying AI. Honestly, reports like this have been coming out constantly for the past three years; it's just that the market currently feels a bit bubbly, so any hint of trouble immediately triggers a selloff.
Beyond this sociological research, another source of recent jitters is that the major AI labs themselves seem to be in a consolidation phase. When GPT-5 finally arrived after endless anticipation, it was a serious letdown for anyone who assumed it would basically be AGI. And while GPT's agent capabilities were the headline feature, in practice there still seem to be a lot of unresolved problems before it's truly usable.
This has brought the "AI investment can't pay off" crowd back out in force, mocking the whole enterprise once again — the oft-cited take from economist Daron Acemoglu, that "AI will only add 0.5% to GDP over the next decade," has resurfaced yet again. (Worth noting: even the optimistic estimates, like the IMF's, put that 0.5% figure on a per-year basis.)
I think the argument that AI investment can't pay off self-destructs the moment you look at reality. AI has already reshaped the very industry that created it — software engineering itself. Coding AI agents basically exploded from last year into this one. A recent article making the rounds argues that US software engineers are struggling to find work, precisely because entry-level jobs have already been taken over by coding agents.
The Engineer Layoff Wave in the US Reflects Accelerating AI Substitution
Historically, software development costs were hard to compress, because no matter how you slice it, you still needed engineers writing code line by line — and those people needed years of training just to get there. If AI can replace even 30% of that work, you're talking about astronomical savings in headcount and cost — and highly paid headcount at that. So coding is exactly the kind of domain where AI agents are pouring in their firepower.
The immediate blast radius of AI agents is generally expected to hit entry-level workers first, and that's exactly what's playing out in software engineering. The cost structure for a senior engineer using AI to boost their own output is completely different from that of a junior engineer being "helped" by AI — one is a direct force multiplier on someone already effective, the other adds internal communication overhead on top, so 1 plus 1 doesn't equal 2. As a result, junior CS grads are starting to find it genuinely hard to land jobs, because companies increasingly prefer to replace junior headcount with AI agents outright. So the claim that AI is a bubble or a hoax simply doesn't hold up on the ground in software engineering — try telling that to all the unemployed CS grads.
Compared with progress on foundation models, the new trend this year is that AI agents have become the main thing worth watching, because the underlying foundation-model layer — with its concentrated compute requirements — has become essentially impossible for small teams to compete in. A lot of teams have pivoted toward building agents instead.
AI's Value Far Exceeds the "Zero Effect" Narrative — This Revolution Is Just Getting Started
But even this direction isn't easy, because the big labs will happily jump in and compete there too. The wave of M&A and consolidation among AI coding startups in June and July is clear evidence of exactly this trend. The two areas where AI agents currently have the biggest impact and the highest valuations are coding and law — both fields built on document-heavy, line-by-line work that requires years of specialized training, where the cost-cutting incentive is huge and companies have strong motivation to automate. For industries like these, the value AI creates is happening right now, in real time — it can't be waved away with a line like "95% zero effect."
In foundation models, the landscape right now is roughly: Google owns enterprise customers, OpenAI owns consumers, and Anthropic owns engineers. The pecking order among the big foundation-model labs is more or less settled at this point, which is part of why they all seem to be moving at a slower pace, working through their priority feature lists at a more measured cadence than before. One major reason the pace of foundational AI progress has slowed is that it now demands ever more compute. DeepSeek's struggles, for instance, are a visible real-world case of hitting the compute wall — the reality is that the compute threshold for foundation models is extremely high, and no clever technical trick can dodge it entirely.
Right now, only Google has managed to build a genuinely effective alternative stack outside the Nvidia ecosystem — though even that only works within certain limits. In nearly every other case, everyone still runs into the software moat created by Nvidia chips and Nvidia's development environment. There have been plenty of attempts to route around Nvidia, but most have failed commercially — to the point where some of these efforts almost feel like proof-of-concept demonstrations of exactly which paths don't work around Nvidia. So for the foreseeable near term, Nvidia's performance is unlikely to falter.
Longer term, though, whether small AI agents will actually require heavy compute is still very much an open debate. A lot of small AI agent startups pitch investors on the idea that they can run on smaller chips with less compute — and if these agents don't need to do much inference, that claim might hold up. If that kind of application matures, a non-Nvidia path could well open up eventually — but it will take a while.
Saying that 95% of companies haven't used generative AI well seems basically accurate. But concluding from that alone that AI has no substance, and that AI company valuations are universally inflated, strikes me as jumping to conclusions far too quickly. This AI revolution has really only just begun. Stock prices may rise and fall, but AI adoption is genuinely, substantively spreading and being deployed at scale — we're still at a very early stage of this story.
台股20年風雲錄──從市值50大消長看台灣產業變遷與投資趨勢
限貸緊箍咒 房市三重壓力來襲！第7波信用管制效應連環套 消費者、建商兩頭受創
吳嘉昭接任台塑集團總裁 王文淵淡出經營，未來王家人誰說了算？