“The average Roman didn’t know their empire had collapsed,” the TikTok read.
The market has officially flipped the calendar to the second half of 2026 — and if you thought the first six months were dramatic, buckle up. Investors are rotating, leadership is changing hands, and the trades everyone loved six months ago are suddenly getting a lot less attention. The biggest opportunities rarely come when everyone is celebrating. They show up when the market is busy, ignoring what’s happening right in front of it. (Source: TikTok)
What caught our eye, though, was the image.
On March 26, 2024, the Francis Scott Key Bridge in Baltimore collapsed after the 100,000-ton cargo ship Dali lost power, propulsion and steering, and crashed into one of the bridge’s main support piers. The catastrophic impact caused the bridge to lose structural equilibrium and fall into the Patapsco River, killing six construction workers on the roadway.
The rest of TikTok goes on to describe all the things we see happening – 14-hour waits at “emergency rooms”, taxes and fees rising with no tangible benefits, public hygiene and transportation failing, cities giving out the same advice to citizens as they do in war zones.
“The West is turning into the third world,” reads the cautionary last frame, “you may not see it yet, but once you do, you can’t unsee it.”
The contrast of a rising stock market and AI euphoria versus a middlin’ real economy and horrid government finances could not be more pronounced.
Today, we’re giving the last of our 2026 forecasts a mid-year update. Let’s begin.
🧟 The Intelligence Bubble and Creatures It Has Spawned
Three years ago, when OpenAI released ChatGPT to the general public, artificial intelligence escaped the lab and walked into the kitchen.
Researchers had spent decades predicting that artificial intelligence and machine learning would reshape the way people work, write, code, research and think, but most of that language remained safely abstract until ordinary people could type a question into a box and receive something useful in return.
ChatGPT made the abstraction usable. People asked for code, summaries, travel plans, sales copy, legal outlines, spreadsheets, poems and memos, then wondered within the same afternoon whether the thing was a toy, a tool, a colleague or a rather polite burglar casing the joint.
While investors were debating whether the AI boom was real, the world kept building it anyway. Data centers expanded. Chip demand surged. Companies raced to integrate AI into everything from software to energy systems. The next phase of the AI story won’t be about finding out whether the trend survives — it’ll be the hard task of identifying the companies positioned to capture the biggest wave of growth. (Source: First Principles and Butterfly Effects, Substack)
Investors did what investors always do when the future appears in a browser window. They reached for their checkbooks, first for OpenAI, then Anthropic and the other companies building frontier models that looked like the new operating layer for knowledge work. The “Magnificent Seven” went along for the ride. Then, SpaceX (SPCX) took investors for another one.
That was the first visible bubble. The model bubble came first because the models were easiest to understand, and the second bubble followed because the magic trick needed chips.
🧠 The Chip Bubble
Large language models do not run on optimism. They run on specialized semiconductors, especially graphics processing units (GPUs), and demand for those chips rocketed as cloud providers, software platforms, hyperscalers and AI developers raced to secure capacity.
Supply became constrained, inventory had to be allocated, customers asked for more than suppliers could deliver, order books filled and revenues surged. The semiconductor companies became the shovel makers in the new gold rush, except the shovels cost billions, require specialized factories, consume enormous power and arrive with lead times that could make a railroad baron tap his foot.
As end users adopted AI, cloud capacity became a limitation. The market moved from models to chips, then from chips to the physical infrastructure required to host and serve demand.
Data centers became the third bubble. Land, wiring, cooling systems, electrical equipment, technicians, fiber, backup power, memory, specialized networking, permits, transmission capacity and local tolerance all became part of the trade because the AI future may be weightless in a product demo, but it is heavy in a county zoning meeting.
🏗️ The Infrastructure Bubble
The infrastructure bubble pushed the AI trade into the traditional economy. Construction firms, electrical contractors, power companies, fiber suppliers, memory manufacturers, component makers, testing firms and service technicians found themselves pulled into the same demand vortex.
At first, manufacturers hesitated because production capacity is not a software feature that can be toggled on and off from a dashboard. It requires buildings, machinery, labor, financing and time.
Then the orders kept coming, and quarter after quarter of rising demand has a way of dissolving prudence. Boards that had once worried about overbuilding began worrying about missing the train, and every company in the supply chain saw the same curve and made the same inference: whatever could be produced would be absorbed.
That is how bubbles recruit serious people. Once power became the limiting factor, capital moved toward nuclear energy, especially small modular reactors; quantum computing entered the conversation as the next computational frontier; humanoid robots and robotaxis became symbols of the future of work; and orbital data centers began appearing in serious discussions because land, power, water and local opposition had become real constraints.
💸 The Funding Bubble
None of this would have been possible without capital.
Six years ago, the Federal Reserve unleashed trillions of dollars through quantitative easing and near-zero interest rates as the economy threatened to buckle amid the pandemic, and that intervention helped prevent depression while also leaving consequences in inflation, asset prices and housing.
After the 2008 financial crisis, regulators pushed risk out of the banking system by raising capital requirements and limiting the risks banks could take. Risk did not disappear; it migrated into private credit, where funds provide higher-yielding loans and financing opportunities that are often unavailable in public markets.
Unlike public markets, private credit does not trade daily, so it lacks continuous price discovery. Values can appear more stable than equivalent publicly traded assets even when the underlying economics are changing, which is a handy feature right up until it becomes a blindfold.
Over time, a sizeable portion of the excess liquidity generated by the Fed found its way into private markets in search of yield. Pension funds, insurance companies, family offices, endowments and banks participated in the migration, turning what began as a market for well-heeled investors into a much larger pool of capital financing projects, either directly or indirectly, tied to the AI boom.
The pandemic created the ultimate investing sugar rush: trillions in stimulus, cheap money and a market willing to fund almost any idea with a catchy enough pitch. But sugar highs don’t last forever. The second half of 2026 could be the great sorting period — where real businesses rise, and the “growth stories” built on endless funding finally face reality. (Source: AdvisorPerspectives)
That is the final bubble, and perhaps the least visible one: the funding bubble. Mature, dividend-paying businesses with significant cash flow were often ignored as capital gravitated toward narratives promising transformational growth.
The cash-flow split inside the AI complex now shows where the money is landing. Nvidia (NVDA), Micron Technology (MU), Broadcom (AVGO) and Applied Materials (AMAT) are expected to generate a record $430 billion in combined free cash flow over the next 12 months, more than triple what they generated just two years ago.
At the same time, Amazon (AMZN), Alphabet (GOOGL), Meta Platforms (META), Microsoft (MSFT) and Oracle (ORCL) are projected to turn combined free cash flow negative for the first time on record, a stunning reversal from their roughly $260 billion positive free cash flow peak in 2024. Their AI-related capital spending is estimated to reach about $1.8 trillion across 2026 and 2027.
That is the Age of Intelligence in one balance-sheet snapshot. The chipmakers are becoming cash machines, while the AI giants burn record amounts of capital to build infrastructure that may one day justify the spending.
🫧 The Bubble Cluster
These are not separate bubbles. The economics of model companies depend on continued growth in AI usage, the economics of semiconductor expansion depend on continued demand from model developers and cloud providers, and the economics of data centers depend on continued growth in inference, training and enterprise adoption.
The case for new power generation depends on continued data-center growth. The case for robotics depends on continued improvement in AI, while the case for quantum computing often rests on future computational requirements that have not yet fully materialized.
The Age of Intelligence is less like one bubble and more like a cluster of partially fused bubbles, each drawing support from assumptions embedded in the others. A model company needs chips, the chipmaker needs hyperscaler demand, the hyperscaler needs enterprise adoption, the data center needs power, the power project needs long-term load growth, the private credit lender needs the borrower’s valuation to hold, and the investor needs the public market to accept the next story at a higher price.
The whole arrangement looks elegant while capital is abundant. It becomes more interesting when someone asks what the “intelligence” is good for… and actually costs.
🪶 The Recurring Dotcom Memory
The dotcom crash was not caused by the internet failing. The internet had enormous potential to transform commerce, media, communications and business, and the bulls were ultimately right about the direction.
They were wrong about cost, timing and which companies would survive long enough to enjoy the future they predicted.
In March 2000, the realization emerged that execution would be harder, slower and more expensive than investors had assumed.
Capital had flooded into internet businesses, data infrastructure, logistics networks and digital models, while far less attention was paid to the cost structure required to serve those opportunities profitably. The simple question was widely ignored: if delivering a product or service online costs more than the existing model, how does the new model win, and how do the companies actually make money?
Customers like superior products, but they rarely abandon cheaper alternatives unless the value proposition is strong enough. Many early internet companies discovered that acquiring customers, fulfilling orders and supporting operations at scale were far more expensive than their business plans suggested.
The opportunity was real, but it took longer to realize and required more investment than investors expected. Companies that survived, such as Amazon, spent the following decade building infrastructure, attacking costs and creating operating leverage, helped by the ultra-low interest-rate environment that followed the dotcom bust and later the global financial crisis.
AI now faces a similar test. The question is whether today’s valuations account for the time, capital and execution required to make that future economically sustainable.
🧮 Tokens Meet Budgets
For several years, companies encouraged experimentation. Teams deployed copilots, built agents and generated enormous volumes of tokens in pursuit of productivity gains, while usage became a proxy for adoption, adoption became a proxy for value, and value became a proxy for eventual profits.
Now the budget people have entered the room. Companies from Uber to Walmart are clamping down on token usage after AI consumption threatened to blow through annual corporate IT budgets in a single quarter without showing significant return on investment.
Current AI economics must account for model training, inference costs, chips, memory, data centers, networking, cooling, electricity, water, software integration, security, compliance, consulting, workflow redesign and human supervision.
Even before grid strain and water consumption are priced honestly, the cost structure remains far from where it needs to be for broad enterprise deployment at attractive returns.
A crash in such a scenario would not occur because AI does not work. It would come because customers begin asking whether the economics work.
🌐 Industrial Scarcity, Network Abundance
AI sits between two economic logics.
An industrial economy builds wealth by manufacturing physical goods, where value comes from scarcity, while a network economy builds wealth by connecting people, software and data, where value comes from abundance.
This distinction matters because AI has the demand profile of a network economy and the cost structure of an industrial economy. The models improve with scale, data and usage, and the applications spread through software, promising abundance, personalization, lower marginal costs and access to capabilities once limited to specialists.
The supply chain is physical. Chips must be fabricated, data centers must be built, power must be generated, fiber must be laid, cooling systems must operate, land must be acquired, communities must tolerate the infrastructure and capital must be raised.
The prompt looks weightless, and the response arrives in seconds. Behind it sits a chain of capital, electricity, water, silicon, copper, concrete, code and debt.
Government can make that harder. A network economy thrives on seamless connections and rapid scale, while governments often impose digital borders, data localization, ex-ante regulation, platform restrictions and legacy rules designed for older industrial businesses.
The more fragmented the network becomes, the harder it is for platforms to scale. The more expensive the physical layer becomes, the harder it is for AI to deliver cheap abundance.
🏛 The Fiscal Wall Behind the Bubble
The funding question does not stop with private credit because Washington has its own financing problem, and it is moving closer.
The political winds threaten to put Democratic Socialists in charge of the public purse. If you haven’t looked at the policy platform they released over the weekend, it’s worth a look for the entertainment value:
Here are a few gems:
The rise of democratic socialism has gone from a fringe conversation to a force shaping the political debate. For investors, the question isn’t about picking a side — it’s about understanding the potential impact on industries, regulation and capital flows. Because history has shown that political movements eventually find their way into markets. (Source: Google Gemini)
The linchpin of the current welfare state – Social Security and Medicare/Medicaid – is all we need to see where government control of essential services leads.
Social Security’s trustees say the main trust fund will be empty by late 2032, at which point incoming payroll taxes cover only 78% of scheduled benefits, implying an automatic 22% cut for retirees.
Jason Fichtner, former chief economist of the Social Security Administration, recently warned that the real deadline may arrive earlier because the bond market will move before the trust fund runs dry. He told CNBC that the bond market may look at Washington and say, in effect, that Congress has 12 months to get its act together because it will soon need another $600 billion or more per year.
Fichtner and economist Veronique de Rugy estimate that filling the Social Security gap would require roughly $600 billion in new borrowing in the first year, rising toward $700 billion annually by 2036. That money comes from the same savings pool that funds mortgages, car loans and business investment, so when the world’s largest borrower demands another $600 billion per year, the price of money changes for everyone.
Markets are forward-looking. A 10-year Treasury bought today matures after the trust fund runs dry, so investors do not need to wait until 2032 to ask whether Congress will reform Social Security, cut benefits, raise taxes or borrow the difference.
Nine months into fiscal year 2026, the federal deficit has already reached $1.4 trillion, according to the Congressional Budget Office. This is happening without a major crisis draining the Treasury, while the economy is still growing and unemployment remains low.
That matters for the AI bubble because capital is not infinite. The same system trying to finance AI models, chips, data centers, private credit, power infrastructure, housing, deficits and future entitlement borrowing must also persuade lenders that the dollar claims they hold will be honored in real terms.
The Age of Intelligence has arrived to a world in debt.
🔮 The Second-Half Forecast
The Grey Swan forecast for the second half of 2026 is that the AI bubble moves from the imagination phase into the economics phase.
The first phase rewarded possibility, the second phase rewarded capacity, and the next phase will examine cost, return on investment, customer willingness to pay, token budgets, electricity demand, water use, financing structure, credit quality and supply-chain overbuild.
The most vulnerable companies will be the ones whose valuations assume demand curves with no ceiling, margins with no gravity and capital markets with no memory. The free-cash-flow split deserves close attention because the market may continue rewarding shovel makers while it begins questioning whether the miners can earn enough to justify the digging.
If enterprise customers clamp down on token usage and demand clearer returns, the bullwhip effect can move back through chips, memory, fiber, data centers, power and private credit.
Back in the bust of 2000, the internet winners emerged after the dotcom crash, and the railroad winners emerged after waves of overbuilding and capital destruction, so AI may reshape the economy while still punishing investors who paid too much too early.
The market often assumes innovation is the hard part, but innovation is usually the beginning. The harder work is reducing costs, improving processes, scaling operations and translating possibility into profitability.
In 2026, the costs, energy demands, infrastructure requirements and questions around return on investment are visible. The Fed does not have the same cushion it had after the dotcom bust, because inflation and consumer strain limit how easily policymakers can cut rates or flood the system with liquidity without consequences.
The gap between the story and reality will close more quickly this time.
~ Addison
P.S. AI may change the world and still disappoint investors who paid for perfection in advance. The railroads changed America. The internet changed civilization. Both destroyed plenty of capital before the durable winners became obvious. The Age of Intelligence now has to prove that intelligence can pay its own electric bill.
Meanwhile, the barbarians are already inside the gate. In this week’s Grey Swan Live!,we’re going to take a look at what happens to a prosperous country after a century of socialist policies. And to what lengths they have to go to extricate themselves from the nightmare.
Argentina is running one of the biggest economic experiments in the world right now — and the results could have implications far beyond its borders.
That’s why this Thursday at 2 p.m. ET, Joel Bowman will join Grey Swan Live! from Buenos Aires to break down Javier Milei’s dramatic reforms, the battle over free markets versus socialism and what investors should understand about the future of global capital flows.
