DNF Opinion: AI Efficiency Gains Are Not That Efficient

It is incredible to think that Chat-GPT launched only three and a half years ago in November 2022. When Open AI (then a non-profit, now gearing up for a very much for-profit IPO) launched its chatbot, the term artificial intelligence became mainstream and people were amazed at the LLM’s ability to answer important queries (“Build a $10 million drop shipping business, make no mistakes”) in a conversational manner.

The principal AI enthusiasts (then, and now) were the techies in Silicon Valley which emerged from their poorly lit dungeons to tout the technology’s ability to change the world. Indeed, no industry was beyond the reach of your friendly and very affirming chatbot. Altman, Musk and co. were all over the show stating that AI would redefine how we work, solve medical mysteries, push the frontier of physics (Hawking’s theory of everything, which attempts to unify gravity with Albert’s theory of relativity is so close to being solved I can feel it!) and change the entire world. Fast forward to today, and we have more chatbots (Grok, Gemini, Claude, etc.), but we are still waiting for the world to change. While these models’ capabilities have certainly improved significantly (Claude can now build a three statement financial model in Excel in 10 minutes, make sure to check its assumptions though, my AI driven portfolio is down 80% YTD), the large scale changes promised are yet to reflect for the average user.

Other than junior bankers failing to land an internship and AI experts trading in the market like Premier League players, the average person has little more to show for the technology than Grok summarizing tweets and generating non-consensual nude photos of your favorite celebrity (Sydney Sweeney had the right idea by simply taking off her shirt in every second scene, no ground-breaking technology needed).

Central to the problem is the difference in cost-per-user-added (CPUA). When Zuckerberg launched Facebook in 2004, he spent money on building the platform and software, after which every user added hit the revenue line but left the operating expenditure line largely untouched. This CPUA is reflected in most internet/dot-com companies. They were free to spend cash on R&D and grow the bottom line by simply adding more users.

With AI, this is not the case. With each “Hallo GPT!”, the massive data centers which run these models use electricity and puts wear and tear on the hardware used to generate a response (chips also need replacing every year, at least according to Mr. Huang). As the models scaled in ability, so did the computation per query. While the techies (not a derogatory term by the way) were living in fantasy land, dreaming of finally building a robot girlfriend, the laws of computing remained unchanged. More complex queries require more complex computations require more electricity. Open AI’s o3 model was claimed to cost upwards of $1,000 per query for the most intensive modes (this is now an outdated model, but proves the point). The key takeaway here is that while the private market can slap another multi-billion dollar valuation on a start-up with AI in its mission statement, the reality is that the models will continue to be constrained by the availability of electricity and the associated cost.

Let’s look at a few real world examples:

Microsoft CEO Satya Nadella revealed earlier this year that the company now writes up to 30% of its code using generative AI (Microsoft’s own AI product, Co-Pilot, might be the single worst piece of software I have ever seen, luckily the have invested $13 billion in Open AI, not-for-profit my @$$). Shocking, then, that on 14 May the company sent thousands of its own engineers a message that cut against the use of AI. The tech giant cut most internal Claude Code licenses with a deadline of 30 June 2026. Interestingly, the 30 June deadline coincides with the last day of Microsoft’s fiscal year, a good time to reduce costs. Claude Code launched inside Microsoft only 5 months ago in December 2025. Token based pricing is standard for frontier AI APIs. If this is getting too technical for you – here is a laymen’s summary: when Microsoft engineers put on The Boys season finale while their laptop runs Claude Code to finish their big project, Microsoft has to pay enormous amounts of money to Anthropic (who built Claude) for the model to finish their work for them. Might be a good time for Microsoft to call those junior software engineers they didn’t hire.

Uber had a similar experience. Between increasing their own share of the driver’s revenue each year and asking the Uber Eats’ driver to shake your pizza box before delivery, they too ventured into the AI efficiency promise. Claude Code was deployed to 5,000 engineers and monthly usage rates climbed to 84-95% by April 2026. Per-engineer costs reached between $500 and $2,000 per month. The result? Uber burned through its entire $3.4 billion 2026 AI budget in four months. Time to switch off The Boys, boys, and return to the good old keyboard and mouse.

While Uber and Microsoft bore the brunt of the attention recently, these are not isolated cases. The first wave of AI adoption was driven by the promise of what they could do, the second wave seems to be shaped by what they cost.

For Anthropic, which is raising capital at a $900 billion valuation (!), this is concerning. Enterprise adoption is part of the pitch. But if a company with Microsoft’s balance sheet and cashflow profile cannot pay for its usage, it sure seems unlikely that many other market participants can.

So, where does this go? Two options: either increase the availability of electricity by building out generation capacity (nuclear power is the best bet, clean and consistent, but takes 5+ years to build, also, ask the NIMBY crowd how they feel about uranium-235 undergoing fission near their home town’s new government sponsored data center, finally, avoid asking Iran about their nuclear ambitions entirely) or nerf the models. Weaker models cost less, but can do less, again failing to deliver on the promise. So, we are at an inflection point – save for significant efficiency gains in the chips running these models (get Nvidia on the phone), it seems unlikely that the large scale, world changing adoption of AI is coming soon.

I would encourage you to avoid the opinions of so called AI experts in what the future of the technology holds. These individuals, generally, have significant financial incentive to market the technology as ground breaking. While I by no means want to imply that the technology is not going to have a significant impact on various industries, I simply do not buy the “AI will change the world” narrative. At least not in the short term. AI is a frontier technology and should be treated as a potential catalyst for changing how we work, learn and live. But until Claude stops assuming a 9% terminal growth rate in my DCFs, I simply don’t believe that the technology is what the techies claim it to be.

(This article, by the way, was written entirely by Claude)

(Kidding, I asked it to, but my tokens ran out and I only get access back in 4 hours)

(DNF is written by a human , now and always. Although that’s exactly what Claude would say?)