Over the New Year, I had one of those moments where a belief you thought you held gets confirmed in an unsettling way. Being chronically online made me feel behind with everyone vibecoding, agentmaxxing, shipping with copilots, while I was still figuring out how to deploy on Vercel.
Then I went offline, spent time with family, with friends of family, with functioning adults who work in real industries spanning Accountants, Engineers, Managers, Consultants in their late twenties to early fifties.
Surprisingly, none of them actively use AI. They don’t use it daily or for real work regularly, a few had tried ChatGPT when it went viral, one mentioned the Ghibli image thing. That was it.
These are people with graduate degrees and real careers and zero reason to be dismissive of new tools. People who should be adopting AI but for some reason, that exposure started and stopped at viral moments of curiosity. They never used it for a single meaningful brainstorming prompt, nor redesigned a workflow around AI. They describe AI as “interesting” the same way they would describe a new restaurant that they’d try eventually.
I thought I was behind the curve. I realized I was just online too much.
Offline, the median knowledge worker is still operating in 2019.
That’s the paradox of this moment: AI is experiencing a capital-markets bubble but is behaviorally under-adopted. Both measure different layers of the same phenomenon.
The Spread
The gap between what’s been invested and what’s been adopted is the most important number in tech right now. Nobody is tracking it carefully.
Here’s the framework. Call it the Reality Gap:
Reality Gap = Upstream Enthusiasm - Downstream Penetration
Upstream refers to the capital allocation level that the markets have priced in. These include Hyperscaler capex, VC funding flows, GPU orders, data center buildouts, AI-exposed multiple expansion. This is the source of the headlines of SNDK 10x-ing, NVDA annual revenue going from $60B to over $200B in 2 years, Memory prices skyrocketing… (ie. where the bubble perception is).
Downstream refers to the the behavioral layer of what humans actually do at work. Daily AI tool usage for real deliverables, enterprise seat activation rates, workflow redesign, junior hiring trends. This is where "still early" lives.
When upstream enthusiasm dramatically outruns downstream penetration, you get both camps being right simultaneously. The bubble critics are correct about valuations but there is also a lot of headroom when it comes to adoption.
What Markets Priced Upstream
These numbers are genuinely hard to digest. Hyperscaler capex across big tech like Amazon, Microsoft, Google, Meta, and Oracle hit $256B in 2024, up 63% YoY. Projected to reach $443B in 2025 (+73% YoY) and $602B in 2026. Roughly 75% of that is going to the AI buildout: GPUs, servers, data centers built for training and inference.
To fund this, hyperscalers now spend almost 80% of their operating cash flow on capex (up from 40% in 2024), with firms like Amazon going as far as 94%. They raised over $100B in debt this year alone which is more than triple the average of the past 9 years. Meta and Oracle issued $75B in bonds and loans in just two months (Sep - Oct 2025) to fund AI buildouts. Hyperscalers plan to add $2T in AI-related assets to their balance sheets by 2030. At 20% annual depreciation on AI assets, that’s $400B in annual depreciation expense (more than their combined profits in 2025).
This essentially that captures what this actually means:
“The reason RAM has become four times more expensive is that a huge amount of RAM that has not yet been produced was purchased with non-existent money to be installed in GPUs that also have not yet been produced, in order to place them in data centers that have not yet been built, powered by infrastructure that may never appear, to satisfy demand that does not actually exist.” - @jatinkrmalik
This is defensive capex where if Hyperscaler 1 is doing it, Hyperscaler 2 can’t be seen not doing it. Spend gets pulled forward in anticipation of demand that may be mis-timed by years. It’s an arms race where everyone has to play and nobody knows whether the war is coming.
VC funding is equally concentrated with AI startups capturing 52.7% of all global venture capital in 2025 which is the first time AI has taken majority share. In Q3 2025, AI absorbed 62.7% of US venture dollars. OpenAI and Anthropic alone accounted for 14% of all global venture investments, and over 2,000 AI-related deals in Q2 2025 alone.
What Humans Actually Do Downstream
Now for the other side.
In BCG’s AI at Work survey of over 10K workers across 11 countries found that only 51% of frontline employees (BCG terms them as white-collar non-managerial roles) frequently use AI in their daily work, which is a decrease from 52% in 2024.
Interestingly, higher AI adoption correlates with higher fear of job loss, which actually rhymes with the thesis whereby the people closest to the technology are most aware of what it can replace while the offline majority isn’t worried because they haven’t crossed that literacy gap yet.
The most useful data point to consider here however is what’s happening wiht Microsoft's own Copilot AI. Microsoft 365 has >450M paying subscribers, but Copilot’s monthly active paying users sit at a reported 15M (~3.3%) 2 years after launch. What I intend to point out here isn’t a Microsoft-specific criticism, but rather the description of what “adoption” actually means across enterprise AI beacuse seats sold =/= seats used. MAU numbers can easily be inflated with a low threshold for what is defined as an “active user” since i can accidentally open Gemini on my Samsung phone when I actually intended to search Google.
Why Adoption Is Slow
(And no, it’s not what you think)
The standard explanation for slow adoption is that people are lazy or resistant to change, which isn’t quite right. Structurally there seems to be 3 things that drastically reduce the level of true “people-replacing” capabilities of AI usage:
AI literacy isn’t binary, there is a knowledge curve to it. The gap between “I tried chatgpt bro” vs “I use AI daily to cut research time from 12 hours to 2 hours” is enormous, and frankly not visible from the outside. Crossing it requires redesigning workflows, prompt engineering and context-setting knowledge , and iterative refinement skills that take (a lot of) time to develop. I believe most people, including many young professionals who should theoretically be most motivated, haven’t crossed this chasm. The analogy I keep coming back to is:
AI is the new Excel. Excel has been standard since the 90s, yet a meaningful chunk of the workforce still can’t build a functional pivot table. AI has a steeper learning curve than Excel, a less standardized (open-ended chat) interface, and a significantly larger variance in output quality based on how you use it.
Institutional friction and stigma, which is severely underappreciated. Even capable individual users often don’t use AI openly beacuse of a “cheating” stigma in many workplaces. There is an underlying fear of looking replaceable and being seen as cutting corners, all while organizations haven’t redesigned processes around AI. Then there are regulated industries (finance, law, medicine) with active restrictions so AI gets capped at the individual level but never compound across the organization.
BCG found that only about 25% of frontline employees say they receive strong leadership support for AI adoption. When leaders demonstrate strong support, the share of employees who feel positive about AI rises from 15% to 55%. The bottleneck to stigma is actually mostly just permission and cultural cover to use it at work.
Lastly, the friction to change workflows. From an economic pov, change happens when the pain of staying exceeds the cost of switching. Most knowledge workers haven’t hit that threshold yet as their current tools work well enough. The forcing function hasn’t personally affected them yet.
The Mirror
What makes this more interesting than “normies don’t get it” is that the Twitter-native crowd which is supposed to be a more motivated, chrnically online, more tool-obsessed cohort that should be furthest ahead, is also failing to compound.
As Ryan Watkins put it:
“Dude I have 10 agents running while I sleep. No one is prepared for AGI in 2 years man.”
“So what are you building?”
“Bro all my smartest friends are vibe coding until 3am every night. It’s all about agency. Intelligence is a commodity man.”
“So what are you building?”
“Do you even study exponentials? Have you seen the latest METR chart? You’re going to be stuck in the permanent underclass bro.”
“So what are you building?”
“I promise you I’m 10x more productive bro! You just don’t understand! Please bro just…. I know you use this stuff everyday too, but you must not be prompting it right! Please broo…”
The “10x more productive” claim without any shipped work. In relation to my New Year observation, its a complete mirror image where one end of the distribution under-adopts entirely, the other over-signals without building anything. Both failure modes confirm the same gap that actual, compounding, output-generating AI use is rare across the whole distribution, not just the offline majority.
If even the most motivated, risk-native, tool-obsessed cohort hasn’t figured out how to reliably compound with AI, what does that tell us about the timeline for everyone else? To me it suggests the adoption gap is going to be wider than consensus assumes.
The Gap Doesn't Close Smoothly
Tne adoption gap among individuals exist but is largely irrelevant to the overall outcome because the displacement mechanism bypasses the individual level entirely. Perhaps this is portrayed more clearly in Citrini’s piece which I’m sure many have read.
The core question it asks, “what if AI bullishness keeps being right... and what if that’s actually bearish?” This mechanism is worth considering.
In this scenario, the adoption gap doesn’t close because individuals finally get it. It is forcefully closed because CFOs start cutting headcount to fund AI tooling with institutional adoption growing at a pace where workers (especially juniors) can’t match.
The Flashing Signal
The leading indicator for this dynamic is entry-level hiring.
What I intiially expected to see (and what I’m watching now) isn’t massive unemployment announcements. Its likely something quieter with fewer junior, entry-level roles hired, slower replacements when someone leaves, flatter/leaner teams shipping the same output, budgets shifting from headcount to AI tools.
A Stanford working paper from 2025 documented a 13% relative employment drop for 22-25 year-olds in AI-exposed occupations (customer service, accounting, software development) since 2022. This isn’t as severe among more experienced workers in the same fields.
The one category worth stress-testing this against is software. The BLS projects +15% employment growth for developers through 2033. There’s a Jevons paradox argument here, and loosely quoting CampbellRamble:
If AI makes code 10x cheaper to produce, does the world want 10x less code or 100x more?
Every company that couldn’t afford a dev team now can. Demand for what software does is arguably infinite. So software might restructure rather than shrink where senior engineers are fine, but junior CRUD grinders not.
But Jevons doesn’t save the other categories in that Stanford paper as there’s no latent infinite demand for more customer service tickets or accounting reconciliations. These are the jobs that don’t expand when they get cheaper. They just disappear.
Where This Lands
The whole point of writing this is to address the bubble vs early debate by pointing out that both camps are measuring different things. The bubble is evident in valuations, capex intensity, and VC concentration. But we’re also early in low daily usage rates and in the sheer number of knowledge workers who have never crossed the AI literacy chasm.
The AI buildout likely continues regardless of job displacement because capex is defensive and the Hyperscalers pouring billions into this cannot afford to be wrong. The first leg ran through chip designers and GPU manufacturers, the second through fabs, HBM, and memory. The next logical stops are energy infrastructure powering the clusters and the interconnects pulling GPU servers together at scale.
The more interesting short leg pointed out in Citrini’s piece is friction-based consumer businesses, but that trade is obvious enough to be crowded and wrong in its timing. The cleaner expression of this thesis is replicable mid-market SaaS where competent internal teams can replicate million-dollar annual contracts in weeks. Others like Asana (-61% past year), Monday.coms (-73% past year), ie workflow tools whose moats were mostly switching costs gets hit from the labor displacement thesis (fewer knowledge workers, fewer seats).
The people who compound AI literacy now will have a durable edge when the downstream catches up to the upstream. If you're entering the workforce in the next few years (which is uncomfortably close to Citrini's scenario window), I believe the best path to take isn’t necessarily to “find a job AI can’t do”, but rahter to become the person who makes a small team output what a large team used to.
The setup isn’t particularly optimistic, but what needs to be done on the individual-level is clear, and so is the trade: long the infrastructure that has to get built regardless, short the software that gets commoditized by the same buildout.
DISCLAIMER: The information, opinions, and projections contained in this report are provided exclusively for discussion and educational purposes. They do not constitute investment, legal, accounting, tax, or other professional advice and should not be relied upon to make, or refrain from making, any investment decision. You should perform your own independent analysis (including consultation with appropriately qualified advisers) before acting on any information presented herein.
All quantitative data were gathered from publicly available sources. Wherever possible those figures were cross-checked against offical copmany or government reports. Nonetheless, historical metrics can be restated and dashboards may contain errors. No representation or warranty, express or implied, is given as to the accuracy, completeness, or correctness of the data as of either the pull date or the date you are reading this report.
References
https://www.microsoft.com/en-us/investor/events/fy-2025/earnings-fy-2025-q4
https://www.statista.com/chart/35046/capital-expenditure-of-meta-alphabet-amazon-and-microsoft
https://news.crunchbase.com/ai/big-funding-trends-charts-eoy-2025/
https://www.bcg.com/publications/2025/ai-at-work-momentum-builds-but-gaps-remain
https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai
https://web-assets.bcg.com/fd/0d/bcc5dfae4cbaa08c718b95b16cf5/ai-at-work-2025-slideshow-june-2025-edit-02.pdf
https://plusai.com/blog/how-consulting-firms-use-ai
https://aifundingtracker.com/q2-2025-ai-funding-report/
A caveat to this "When leaders demonstrate strong support, the share of employees who feel positive about AI rises from 15% to 55%." is that leaders walk the line between being aspirationally supportive and actually supportive because of the experiences they themselves are having. In other words, there's a difference between "you should be using AI" and "I'm using AI, you should too". I should get off the Internet too sometimes