Are Citrini Research’s Predictions Correct?

When I read Citrini Research’s piece on Substack this past Sunday, the one titled “The 2028 Global Intelligence Crisis,” I felt something I have learned to distrust in myself. I felt the warm glow of confirmation.

I spent thirty years building technology companies. I watched the internet hollow out industries that seemed permanent. I watched mobile destroy businesses that survived the internet. I have been around long enough to recognize a genuine technological discontinuity when one is arriving, and I have been worried about this particular one for longer than I care to admit. So when a sophisticated macro analyst laid out, in precise and relentless detail, my first thought was finally the message is breaking through but instinct is a problem. Confirmation bias is the enemy of clear thinking, and nowhere is clear thinking more consequential than when you are trying to understand a threat that is still early enough to do something about.

So I made myself slow down. I read the piece again. And then I spent time thinking about the physical and structural constraints.

The core thesis of “The 2028 Global Intelligence Crisis” is sound and genuinely alarming. AI is not just another productivity tool. It is the first technology in history that improves at the very tasks that displaced workers would naturally redeploy to. Every previous disruption, the ATM, the internet, the smartphone, created new categories of human work even as it destroyed old ones. This one may not. When a competent AI agent can do the work of a product manager, a paralegal, or a financial analyst for a few hundred dollars a month, the assumption that underlies thirty years of mortgage underwriting, thirty years of consumer credit expansion, and the entire architecture of the white collar services economy quietly becomes fictional.

Citrini Research writes this as a future post mortem, dated June 2028, looking back at how the spiral unfolded. It is a clever device and the internal logic is tight. But the timeline is where I think the piece does too much work, and where my confirmation bias was leading me to accept conclusions I needed to question.

The physical world has brakes that the article underweights.

The electricity grid is the first one. The scenario requires AI inference to scale by something like ten to a hundred times over two years. The power does not not exist. What does not exist is the permission to use it. Permitting and transmission buildout in the United States run on three to seven year timescales, sometimes longer. Northern Virginia, the largest data center market in the world, is already seeing utility pushback and connection freezes. You cannot build a civilization-altering intelligence infrastructure faster than regulators approve the substation upgrades.

The semiconductor supply chain is an even harder constraint. TSMC’s Chairman stated publicly in late 2025 that advanced node capacity is roughly three times short of what customers plan to consume. The advanced packaging technology that lets AI chips function at their rated performance, a process called CoWoS, is fully allocated through 2026 and into 2027. High bandwidth memory, the component that feeds data to AI accelerators fast enough for them to be useful, is similarly booked out. You cannot train the next generation of models faster than you can manufacture the chips, and nobody is expanding capacity fast enough to close that gap on the article’s two-year timeline.

Water is the constraint nobody is talking about. AI data center cooling already competes with agriculture and municipalities in water stressed regions of the American Southwest. At the scale this scenario requires, that becomes a genuine geographic brake on where deployment can actually happen.

And then there is the autonomous vehicle assumption. The article treats the proliferation of self driving vehicles as simultaneous with the white collar disruption, compressing both into the same 2026 to 2028 window. Full AV deployment is a decades long regulatory project in most jurisdictions. The political economy of displacing four million truck drivers and two million ride share drivers in two years would generate a regulatory response that would slow things considerably.

Finally, the article slides somewhat quietly from “AI is very good at coding and cognitive tasks” to “AI has displaced general human cognitive labor.” That is a significant leap. Emotional intelligence, physical judgment, and the ability to navigate genuinely novel situations remain hard problems that current architectures have not solved. Also, read the article’s comments by user Peter about the current quality of the work in complex real world scenarios (net: there are still quality problems holding models back).

What the physical constraints tell me is not that the thesis is wrong. They tell me the timeline is probably stretched by three to five years. The destination Citrini Research describes may well be where we are heading. But the journey takes longer. And counterintuitively, that might be more dangerous, not less. A slower spiral gives institutions just enough time to convince themselves at each stage that the problem is contained, which is exactly what the article says happened in its fictional 2026 and 2027.

So with the constraints overlaid on the thesis, here is how I actually think this unfolds, year by year.

By early 2027, the white collar labor market is visibly softening but the headline unemployment number is not alarming. The real signal is in the composition of job openings, which are contracting fast in software, consulting, financial services back office, and entry level professional work while blue collar openings hold steady. The long tail of SaaS is being hollowed out. Consumer sentiment is deteriorating faster than the hard data reflects, because the people most exposed to disruption are high earners with savings buffers who maintain spending while quietly terrified. The financial markets are still in the euphoria phase. A very small cohort of capital owners is accumulating wealth at a historically unprecedented rate. The investment thesis here is still the infrastructure complex: chip makers, hyperscalers, utilities with contracted data center load, nuclear energy companies signing long term power agreements, and the physical materials that the buildout requires, copper, electrical switchgear, transformer manufacturers.

By early 2028, the semiconductor supply chain begins to loosen as new capacity comes online, and that is when capability improvements accelerate again after the Year 1 plateau. The disruption that was sector specific begins spreading into every white collar cost structure simultaneously. White collar unemployment breaks above six percent and the income mix of the unemployed shifts visibly toward higher earners. Consumer spending begins to crack. Housing stress appears in exactly the zip codes Citrini Research identifies. Financial stress concentrates in private credit, where PE backed software deals are finally being marked to something resembling reality. The investment rotation here is toward the physical constraints on the buildout: water infrastructure, electrical grid equipment, geographies with natural AI infrastructure advantages like the Pacific Northwest and Scandinavia. Defensive positions in consumer staples begin to make sense. Gold starts to look attractive as the fiscal implications become clearer.

By early 2029, the financial crisis mechanisms the article describes begin to activate, delayed by roughly a year from its timeline. Mortgage delinquencies in high income zip codes are undeniable. A recession is dated that has been obvious to anyone paying attention for six months. Private credit defaults are cascading. The gig economy that absorbed the first wave of displaced white collar workers is itself being disrupted as autonomous vehicles begin commercial deployment in the geographies that were proactively regulated. Capital preservation becomes the primary objective. Short duration government bonds and gold. Defense spending is rising globally as geopolitical tensions increase alongside domestic instability. Healthcare is relatively protected because demographic demand is inelastic and physical care remains genuinely hard to automate.

By early 2030, some version of the legislative response the article describes forces its way through because the alternative is political instability severe enough to threaten the incumbents of both parties. It will be smaller than needed and later than optimal, which is how American policy always works under stress. The equity market has drawn down substantially from its peak, and the trough is approaching or has passed. The long term investment thesis begins to clarify: AI infrastructure at reset valuations is the generational opportunity of this era, the equivalent of buying the railroads after the first speculative bubble collapsed. Physical scarcity plays, water rights, energy generation, critical minerals, have permanent structural tailwinds. Human skills that AI demonstrably cannot replicate, skilled trades, genuine craft, high touch healthcare, command meaningful premiums.

By early 2031, the new normal is taking shape. It looks like a society with a much wider gap between capital owners and everyone else, a significantly expanded government transfer system, and a permanently altered understanding of what work means and what it is worth. The companies and individuals who thrived share a common characteristic: they owned either the compute infrastructure itself, the physical constraints on that infrastructure, or the last genuinely scarce human skills, the ones requiring embodied judgment, authentic emotional attunement, or creative originality that emerges from lived experience rather than pattern matching on training data.

I left the United States for Portugal two years ago, partly for lifestyle reasons and partly because I had a nagging sense that the next decade in America was going to be harder than most people were prepared for. Reading “The 2028 Global Intelligence Crisis” did not make me feel vindicated. It made me feel the way you feel when a doctor talks about a diagnosis…

The canary, as Citrini Research notes, is still alive. We are in February 2026. The S&P is near all time highs. The spiral has not begun, or rather, it has begun but is not yet visible in the headline numbers.

The question I keep coming back to is not whether the thesis is right. It is whether the institutions designed for a world of scarce human intelligence can adapt fast enough to a world where that scarcity is dissolving. My experience across three decades of technology tells me that institutions adapt eventually, at great cost and with great delay. The people who do best are the ones who start adjusting before the headlines confirm what the data already shows.