Year Four of the AI Revolution: The Real Winners May Emerge After the Highway Is Built

1. Where are we in the AI revolution?

It has been about three years and 11 months since ChatGPT launched on Nov. 30, 2022. The question has changed along the way. First, people asked what AI could do. Then companies competed over how many GPUs they could secure. Now the question is: after spending so much, how much is AI actually earning?

In Technological Revolutions and Financial Capital, the economist Carlota Perez divides technological revolutions into an installation period and a deployment period. Capital rushes into a new technology first and builds the infrastructure. Economic value expands later, as society and business put it to use.

The sequence of introduction, frenzy, turning point and deployment is a useful framework. It is not a formula with fixed dates or durations. In 2026, AI is in a phase where infrastructure installation and commercial rollout are happening at the same time. There is little basis to say the industry as a whole has already passed its turning point.

Perez's cycle, simplified for investors

1. Introduction: early demand
2. Frenzy: capacity expansion
3. Turning point: profitability tested
4. Deployment: productivity spreads

This is a simplified concept based on Perez's theory. It does not indicate actual years or durations.

2. Railroads and the internet succeeded. Did every investor?

Nineteenth-century American railroads cut logistics costs and linked a national market. But the railroad companies had to carry heavy construction costs and debt, and many went bankrupt or were restructured.

In the 1990s, telecom carriers poured huge sums into fiber-optic cable in anticipation of the internet's future. Those that overestimated demand and the speed of payback took heavy losses when the dot-com bubble burst. Yet search, e-commerce, video and the cloud grew on top of that network.

Mobile followed the same path. Once networks and smartphone operating systems spread after the iPhone, new services such as Uber and Instagram grew. The point is the same each time: a technology's success is different from the return on any single stock. Long-term investing does not mean holding every company for a long time. It means continually selecting the companies that can defend their competitive position.

3. The productivity J-curve: when costs and benefits are out of sync

Research on the productivity J-curve by Erik Brynjolfsson, Daniel Rock and Chad Syverson explains why the economic effects of a general-purpose technology arrive with a lag. Adopting AI takes more than buying GPUs. It requires complementary investment such as cleaning data, redesigning work, training staff and developing products. The costs show up first, while productivity gains may be measured later.

But the J-curve does not guarantee "three years of losses, then a profit explosion in year four." Nor is it empirical proof that AI as a whole has already passed its low point. Fast-adopting areas such as coding and customer support should be separated from areas that need time to validate, such as healthcare, finance and robotics.

The real question is not how many years have passed since launch. It is how much AI adoption is converting into actual customer spending and corporate profit.

The productivity J-curve (theoretical concept)

The chart plots effect against time. It dips during the period of complementary investment, then rises as the effects are realized. It is a conceptual illustration, not year-by-year data or a return forecast.

4. Why AI may spread faster than past technologies

The internet required a network to be built. Mobile required smartphones and wireless networks. AI can be distributed relatively quickly through smartphones, browsers, cloud services and business software that already exist. Stanford's AI Index 2026 counted generative AI reaching 53% of the surveyed population in about three years.

But adoption of a service is separate from profit for shareholders. Free users can grow, yet profitability may stay low if inference costs are high. Conversely, if paying enterprise customers sign long-term contracts, a business can be a good one even with relatively slower growth.

AI could be used in finance, healthcare, law, customer support, manufacturing, autonomous driving and robotics. Still, the global labor market should not be treated as an instant AI revenue market. The analysis has to account for which tasks can actually be automated, adoption rates, customers' willingness to pay, regulation and cost.

5. Was the worry about too much spending wrong?

The absolute scale should not be downplayed. Major AI infrastructure spending plans for 2026 are reported to exceed $700 billion. But companies' total capital expenditure includes non-AI businesses, so that sum should not be described as entirely AI-specific.

Comparisons can also mislead. Dividing global corporate investment by U.S. GDP, or comparing 19th-century railroad construction costs with today's Big Tech capex on different scopes, creates an illusion. An accurate historical comparison needs the same country, the same investment scope, GDP for the same period and conversion to real prices. That is why figures of 3% to 6% and 1.2% to 1.6% should not be treated as confirmed comparisons.

Companies with large operating cash flow, such as Microsoft, Alphabet, Amazon and Meta, are better able to endure than the speculative telecom carriers of the past. That does not mean the whole industry is built on cash. Debt-reliant AI infrastructure companies such as CoreWeave are far more sensitive to interest rates and funding conditions. In fact, CoreWeave's large debt load was at the center of the market jitters in October 2026.

In the end, investors should look at how the spending is financed as much as how large it is. If capex grows faster than revenue and depreciation accumulates, even a company with excellent technology can come under cash-flow pressure.

6. Who makes money once the AI highway opens?

For Nvidia and AMD, what matters is demand for AI accelerators, product competitiveness and total cost per system. Ecosystem lock-in such as CUDA helps defend profitability, but customers' expanding in-house chips and a slowing growth rate of orders are risks.

For Micron, SK hynix and Samsung Electronics, what matters is not only HBM shipments but also the pace of supply expansion, conventional DRAM prices, yields and customer concentration. Even if AI demand grows, product prices can fall if supply grows faster.

TSMC benefits from manufacturing advanced chips for many designers. But it cannot avoid capital spending, advanced-packaging bottlenecks, utilization rates and the cost of moving to new process nodes.

For Microsoft, Alphabet and Amazon, the test is how quickly cloud AI revenue growth catches up with the burden of capex and depreciation. For Meta, the key is whether AI improves ad efficiency and user engagement enough to offset the investment.

Software companies can benefit from falling inference costs. But the same savings apply to their competitors. Without paying-customer retention, net revenue retention and operating margin improvement, the spread of the technology alone cannot justify high share prices.

Tesla and robotics companies hold a long-term option on AI's physical spread. But self-driving and humanoids need to be assessed separately on technology demonstrations, regulatory approval, production scale-up and unit economics.

Conclusion

A short-term pullback does not mean the AI revolution is over. At the same time, the fact that the revolution continues does not mean every AI stock is worth buying at current prices. The histories of railroads, the internet and mobile support both statements.

The highway will likely keep getting longer. But what investors need to keep checking is not construction progress. It is the road's demand, its construction cost, how it is financed and the tolls actually collected. A long-term investor's edge is not unshakable conviction. It is that facts become easier to verify as time passes.

This analysis is for informational purposes and is not a recommendation to buy or sell any security.

Insight Times Editorial Desk