The smartest technology analyst you’ve probably never heard of just published a presentation that reframes the entire AI investment thesis — and if you’re still thinking about this as a software story, you’re already behind.

Benedict Evans — former Andreessen Horowitz partner, and one of the clearest thinkers in technology and capital — laid out a case that the AI boom is less like the software revolution and more like the railroad era. The four largest tech companies alone are expected to spend roughly $700 billion on infrastructure in 2026. That’s nearly triple the annual capital spending of the entire global telecom industry. And just like railroads, the companies building the tracks may matter far more than the ones buying the tickets.

That realization points to a specific set of companies — in power generation, grid infrastructure, semiconductors, data centers, and enterprise software — that are positioned to get paid no matter which AI model “wins.” The picks-and-shovels playbook has worked through every major technology transition in history. Here’s why it may work again now, and exactly where to look.

One of the most useful documents I’ve read on artificial intelligence this year doesn’t predict what’s going to happen. It explains what questions investors should be asking.

The presentation is called AI Eats the World, and it was put together by Benedict Evans.

If you’re not familiar with Evans, he spent years as a partner at Andreessen Horowitz after a career in equity research and telecommunications strategy. Unlike many technology commentators, Evans approaches technology as both an investor and an economist. He spends very little time making grand predictions and a great deal of time thinking about incentives, business models, capital allocation, and how value actually gets captured.

That last point is important.

The technology industry has a long history of creating enormous value while simultaneously destroying shareholder capital. Railroads changed America. Airlines changed the world. Telecommunications networks connected the globe. Investors who owned the wrong companies during those revolutions often discovered that being right about the future and making money are two very different things.

That may end up being the most important lesson from the current AI boom.

Right now, everyone is obsessed with artificial intelligence. Every earnings call mentions it. Every venture capitalist is funding it. Every corporate executive is trying to explain how it fits into their business model. Every stock promoter on social media has discovered that adding the letters “AI” to a company description is apparently worth an extra 20% on the share price.

Evans takes a step back and asks a simple question: What if AI is not primarily a software story? What if it’s a capital spending story?

The numbers are staggering. Microsoft, Amazon, Alphabet, and Meta are expected to spend roughly $700 billion on capital expenditures in 2026. The four largest technology companies are planning to spend nearly three-quarters of a trillion dollars in a single year building infrastructure — more than double what they were spending only a few years ago.

For perspective, Evans notes that global telecommunications capital spending runs roughly $300 billion annually. Global oil and gas capital spending is around $1 trillion. Artificial intelligence has become one of the largest infrastructure construction projects in human history.

Everyone focuses on Nvidia (NVDA) because the stock has become the poster child of the AI era. Evans points out that Nvidia can’t get enough capacity from Taiwan Semiconductor Manufacturing (TSM) fast enough to satisfy demand. Semiconductor manufacturers are scrambling. Memory suppliers are scrambling. Data center developers are scrambling. Electric utilities are scrambling. Construction companies are scrambling.

That observation leads to the first investment conclusion: if AI really does transform the economy, the biggest winners may not be software companies. They may be the businesses selling picks, shovels, and electricity.

The Power Story Everyone Is Still Underestimating

Every AI query requires electricity. Every inference requires electricity. Every data center requires electricity. The more you study the AI boom, the clearer it becomes that power generation and transmission are among the most overlooked investment themes in the market.

Constellation Energy (NYSE:CEG) is probably the purest publicly traded AI power story. The company operates the largest fleet of nuclear power plants in the United States, producing roughly 10% of all carbon-free electricity generated in the country. AI data centers require enormous amounts of reliable baseload power. Wind and solar can contribute, but hyperscale operators need electricity every second of every day. Nuclear solves that problem.

Recent agreements between data center operators and nuclear facilities demonstrate that large technology companies are increasingly willing to pay premium prices for reliable power. Constellation owns an asset base that would be nearly impossible to replicate today, given regulatory hurdles and construction costs. If AI demand continues growing at anything close to current projections, Constellation could find itself in the enviable position of owning exactly what everybody suddenly needs more of.

Vistra (NYSE:VST) offers broader exposure to power demand growth. The company owns natural gas, nuclear, solar, and battery assets across key markets, including Texas, which remains one of the most important data center markets in North America. While Constellation is primarily a nuclear story, Vistra provides a diversified way to participate in rising electricity demand. As AI workloads expand, power demand growth in ERCOT could remain significantly above historical trends.

NRG Energy (NYSE:NRG) has historically been viewed as a cyclical power producer, but the AI buildout changes the equation. Data centers are becoming anchor customers capable of consuming enormous amounts of electricity over very long periods. The market may still be pricing NRG as a traditional utility while the underlying economics shift toward supplying one of the fastest-growing categories of industrial demand in the world.

The Bottleneck Nobody Is Talking About

One of Evans’ most important observations is that the bottleneck isn’t just power generation. It’s transmission. You can build all the generation capacity you want — if you can’t move electricity from where it’s produced to where it’s needed, the entire system breaks down.

Quanta Services (NYSE:PWR) may be one of the highest-quality infrastructure businesses in America. The company designs, builds, upgrades, and maintains electrical transmission and distribution systems. If the United States is going to spend hundreds of billions upgrading the grid to support AI, data centers, electrification, and industrial reshoring, Quanta is likely to be one of the largest beneficiaries. It’s not glamorous. Nobody talks about it at cocktail parties. But the company occupies a critical position in a massive, long-term spending cycle.

MYR Group (NYSE:MYRG) is a smaller version of the same basic theme — a company that builds and maintains electrical infrastructure and transmission systems. Unlike many AI-adjacent stocks, MYR trades at valuations that occasionally become reasonable during market pullbacks, making it worth attention for investors looking for a second-tier beneficiary of the AI power buildout.

Data Centers Are the New Office Buildings

Evans highlights another data point that deserves more attention: data center construction spending has now surpassed office construction spending in the United States. For decades, office buildings were the physical infrastructure of the knowledge economy. Today, the knowledge economy is building server farms instead.

That shift is bullish for data center REITs.

Digital Realty (NYSE:DLR) owns more than 300 data centers globally and serves many of the world’s largest technology companies. What makes it attractive is that it’s not a bet on which AI model wins. Whether OpenAI, Google, Meta, Amazon, or some startup becomes dominant, they all need data center capacity. It’s a classic toll-road investment — and as a REIT, Digital Realty also provides current income, making it more accessible for income-oriented investors than most AI stocks.

Equinix (NYSE:EQIX) is arguably the highest-quality data center operator in the world. The company has built a global interconnection ecosystem with powerful switching costs: customers locate equipment within Equinix facilities because that’s where their partners, suppliers, customers, and network providers already are. That network effect makes Equinix considerably harder to replicate than a simple warehouse full of servers.

The Commodity Risk Nobody Wants to Admit

Here is where Evans makes his most controversial argument, and it’s one worth taking seriously.

Large language models are increasingly starting to look similar. Performance differences exist, but they appear to be narrowing. There are no obvious network effects. There are no obvious switching costs. There is no guarantee that today’s AI leaders will maintain pricing power.

That should sound familiar. Telecommunications companies spent trillions building networks. Consumers captured most of the benefit. Application developers captured much of the profit. Network operators often earned mediocre returns.

If AI models become commodities, the real winners will be the businesses built on top of those models. History provides a useful guide. The internet created enormous fortunes — and most of those fortunes were not made by fiber-optic cable manufacturers. They were made by Amazon, Google, Netflix, Meta, and thousands of software businesses that used the infrastructure to solve specific customer problems.

Evans believes the same dynamic may play out in AI. The real opportunities may emerge in vertical software, workflow automation, healthcare administration, legal technology, customer service automation, and business process management.

ServiceNow (NYSE:NOW) may be one of the strongest AI application stories available today. The company already sits at the center of enterprise workflows. AI allows it to automate more tasks and increase customer value without requiring customers to overhaul their entire technology stack.

Salesforce (NYSE:CRM) has spent years building customer relationship software. AI allows it to enhance those workflows with automation and predictive capabilities. The company’s enormous installed base provides an advantage that most AI startups cannot match.

Palantir Technologies (NYSE:PLTR) may be the most direct beneficiary of enterprise AI adoption. Unlike companies focused on generic models, Palantir helps organizations apply AI to specific operational decisions — a distinction that may become increasingly important as enterprises move beyond experimentation and into production deployments.

The Consultants and the Equipment Makers

One chart in the presentation is worth pausing on: despite all the excitement surrounding ChatGPT, only a small percentage of users actually pay for the service. Many people use it. Far fewer depend on it.

Technology adoption typically follows a predictable path. People experiment. Businesses test. Consultants get hired. Pilot programs proliferate. Years later, mission-critical applications emerge. We are still early in that process — and that creates real opportunity.

Every company in America wants to implement AI. Most have absolutely no idea how to do it. That gap creates significant opportunity for consultants and systems integrators. Companies such as Accenture (NYSE:ACN), Infosys (NYSE:INFY), and Cognizant Technology Solutions (NYSE:CTSH) may end up capturing a significant portion of the AI gold rush simply because somebody has to help businesses deploy these systems. Implementing technology has historically been harder than inventing it.

The semiconductor equipment companies are equally compelling for the same structural reason. Applied Materials (NYSE:AMAT) is one of the largest suppliers of equipment used in chip manufacturing. Regardless of whether Nvidia, Advanced Micro Devices (NYSE:AMD), Broadcom (NYSE:AVGO), or some future competitor dominates AI chips, somebody has to buy Applied Materials equipment to manufacture them.

Lam Research (NYSE:LRCX) specializes in etching and deposition technologies essential to advanced chip manufacturing. As chips become more complex and require more processing steps, Lam’s equipment becomes increasingly critical. KLA Corporation (NYSE:KLAC) provides inspection and process control systems — and the more sophisticated semiconductors become, the more important quality control becomes. KLA is easy to overlook because inspection equipment isn’t exciting, but it occupies a critical position in the semiconductor ecosystem and consistently generates impressive profitability.

The Only Honest Conclusion

The final lesson from Evans’ presentation may be the most valuable: nobody knows how this ends. Nobody.

Every platform shift looks obvious in hindsight. Personal computers, the internet, smartphones — each seemed inevitable after the fact. At the time, each transition was filled with failed business models, bankruptcies, false starts, and wildly incorrect predictions. Artificial intelligence will be no different. Some of today’s stars will disappear. Some of today’s forgotten companies will become tomorrow’s giants.

As investors, the job isn’t to predict the future with certainty. It’s to identify situations where the payoff is attractive if you’re right and the downside is manageable if you’re wrong.

That’s why power infrastructure, data centers, semiconductor equipment, and software businesses with real customers and real cash flow keep coming back as the most defensible positions. The AI future may be uncertain. The need for electricity, computing capacity, and business productivity improvements is not.

Evans closes his presentation by suggesting that every AI question ultimately has one of two answers: “Nobody knows,” or “What happened the last time everything changed?”

For investors, that may be exactly the right framework. History doesn’t repeat itself — but it generally rhymes. The smartest move isn’t trying to predict the exact shape of the future. It’s owning the businesses most likely to get paid regardless of how the future unfolds. When uncertainty is high, owning the businesses selling picks and shovels has proven more reliable than guessing which prospector strikes gold. That approach has worked through every major technology transition in financial history. There’s little reason to think the AI era will be any different.