Stop Thinking of AI as Software. It's Infrastructure.

Here's the mistake I see everywhere, from boardrooms to analyst reports: we're treating AI like software. Evaluating it like software. Budgeting for it like software. But AI isn't software. Its infrastructure is more analogous to electricity or railroads than to SaaS applications, and the two kinds of asset do not reward the same decisions.

The capital involved is what makes the category error expensive. The four largest hyperscalers spent roughly $350 billion on AI infrastructure in 2025, more than the inflation-adjusted cost of the entire Apollo program. When I put a figure like that in front of board members, the first question is almost always: "Is this a bubble?" The answer depends on which layer of the AI economy you're examining. But the bubble question, while important, is the wrong starting point.

The better question: what kind of economic phenomenon are we actually witnessing? Answer that one and the investment horizon moves, the risk model moves, the vendor relationship moves, and so does the place competitive advantage actually lives. This isn't semantics. It's a reframe with a balance sheet attached.

The Historical Parallel We're Missing

In 1935, only 11% of American farms had electricity. By the late 1950s, that number was 97%. Not because private markets suddenly discovered rural profitability, but because the government decided electricity had become essential infrastructure and created the Rural Electrification Administration to make it happen.

The pattern is instructive: massive private investment in generation and distribution, followed by consolidation as weaker players failed or were absorbed, and then regulation once it became clear who the dominant players were.

AI is following the same trajectory. This isn't a scrappy startup revolution. It's a second-wave infrastructure build driven by giants with fortress balance sheets. Microsoft, Google, Amazon, and Meta aren't gambling on AI. They're building the equivalent of power stations and transmission networks, and they can afford to wait a decade for returns.

Here's what most people miss about electrification: the companies that won big weren't the ones building power plants. General Electric didn't become an industrial titan by generating electricity. It became a titan by figuring out what to do with electricity. Light bulbs. Motors. Refrigerators. The resistance welder that made Ford's assembly line possible.

The infrastructure providers captured their share. But the transformative value, the value that reshaped entire industries, went to those who figured out the applications.

I can hear the objection: AI moves too fast for this analogy to hold. The electricity infrastructure was stable. Once you built a power plant, it worked the same way for decades. AI capabilities double every eighteen months while costs plummet. How can infrastructure consolidate when the technology keeps shifting beneath it?

That's the strongest case against my thesis. But I'd argue the speed of change actually reinforces infrastructure concentration rather than preventing it. Only companies with massive capital reserves can continuously rebuild capacity at the pace AI demands. The hyperscalers aren't winning despite the rapid change. They're winning because of it. Everyone else gets exhausted.

The tempo is different. The structural dynamics are the same.

Why Software Thinking Fails

The last twenty years of enterprise technology taught us to ignore infrastructure. Virtualization, containers, serverless computing, the whole trajectory of cloud was about abstraction. Making lower layers invisible. We got comfortable treating compute as a utility that just worked.

AI workloads reverse this trend entirely.

Performance at scale depends directly on the hardware underneath. A single AI search query uses 10 times more electricity than a traditional internet search, according to the Electric Power Research Institute. Token costs vary by orders of magnitude depending on which model you're calling, how you're calling it, and what infrastructure sits beneath it. The abstraction layer is leaky, and the leaks are expensive.

If you're budgeting for AI the way you budget for SaaS, predictable per-seat licensing, minimal infrastructure considerations, and easy vendor switching, you're setting yourself up for cost blowouts and disappointing performance.

More importantly, if you're strategizing about AI the way you strategize about software, you're asking the wrong questions. "Which vendor should we choose?" is a software question. "What capabilities should we build versus consume, given that we're becoming dependent on utility-like infrastructure?" is an infrastructure question. They lead to very different answers.

Where Value Concentrates

Start with Nvidia. They don't just sell chips. They control CUDA, the software layer that makes those chips valuable. Developers build on CUDA; switching costs compound; Nvidia's moat deepens with every application written for their ecosystem. The hardware is valuable, but the stack integration is what makes them dominant.

Google sees this clearly. They're not content to offer cloud AI services. They build their own TPUs, train their own foundation models, and control the orchestration layer through Vertex AI. Apple is building custom silicon specifically to run AI workloads on-device, capturing value from chip to interface. These companies aren't picking one layer and excelling at it. They're integrating vertically, and each layer reinforces the others.

This is the pickaxe economics of AI. During the California Gold Rush, the merchants selling tools made more reliable fortunes than most miners. Nvidia is the pickaxe seller, capturing value whether OpenAI, Anthropic, or Google ultimately wins the model race. At current valuations, the market has already priced this in. Nvidia trades at multiples that assume continued dominance.

So where's the asymmetrical opportunity? Not in infrastructure, that value is already claimed and priced. The opportunity is in the application layer, where domain expertise meets AI orchestration.

Here's what took me years to see: for decades, software value concentrated at the user interface. The screen where humans interact with systems. AI inverts this. When agents can execute multi-step workflows autonomously, value migrates to the orchestration layer, the logic that coordinates models, retrieves context, validates outputs, and triggers actions. The UI becomes a monitoring dashboard rather than the main event.

Companies still pouring resources into interface excellence while ignoring orchestration are optimizing for a world that's already shifting beneath them.

The Enterprise Implication

Electricity didn't just show up and declare victory. It needed tools. Light bulbs made it visible. Motors made it powerful. Refrigerators made it indispensable. These inventions, not the power plants, transformed daily life and created new industries.

For those of us in financial services, the lesson is that you don't need to build power stations. You need to figure out what your light bulbs and motors are.

I've watched colleagues at tier-1 banks agonize over which foundation model to select, running elaborate benchmarks, debating GPT-4 versus Claude versus Gemini as if this were the strategic decision. It isn't. The models are commoditizing rapidly. Eighteen months from now, the performance gap between leading models will be negligible for most enterprise use cases.

The strategic decisions are different: Which processes can AI fundamentally transform? What proprietary data makes those transformations defensible? What governance structures let us move quickly without regulatory disaster?

I learned this the hard way. Early in my AI work, I watched a well-funded initiative spend nine months evaluating models, detailed benchmarks, proof-of-concept after proof-of-concept, and endless vendor presentations. By the time they selected a model, two of the three finalists had released newer versions that invalidated the benchmarks. Meanwhile, a smaller team with a fraction of the budget had already deployed a working solution. They'd picked a "good enough" model in week two and spent the remaining time on integration, workflow design, and change management. Guess which initiative delivered value?

The model selection was noise, and the orchestration was the signal.

Consider trade surveillance, an area I know well. Today, compliance teams review thousands of alerts daily, the vast majority false positives. Analysts spend hours on each case, pulling data from multiple systems, reconstructing transaction sequences, and documenting their reasoning. It's expensive, slow, and soul-crushing for talented people who joined compliance to catch bad actors, not to click through endless false alarms.

The "light bulb" here isn't the AI model. Any decent LLM can summarize a trading pattern. The light bulb is the orchestration layer: the system that triages alerts based on risk indicators, automatically enriches cases with relevant context from disparate data sources, generates draft narratives for analyst review, and learns from analyst feedback to improve future triage. The model is a commodity input. The workflow transformation is the competitive advantage.

JPMorgan's COiN platform offers a public proof point. By applying AI to commercial loan agreement analysis, they eliminated 360,000 hours of annual legal work. The value wasn't in which model they used, it was in the application layer: understanding which documents needed analysis, what fields mattered, how to validate outputs, and how to integrate results into existing workflows. That's the light bulb.

Banks that figure this out will run compliance functions that are simultaneously more effective and less expensive. Banks that keep debating which model to license will still be drowning in false positives five years from now.

This is why I keep returning to the same thesis: governance enables trust, enables speed. The organizations that build robust AI governance frameworks first won't be slowed down by compliance. They'll be freed to move faster than competitors, still negotiating each use case from scratch with legal and risk teams.

What This Means for Your Strategy

If AI is infrastructure rather than software, three strategic implications follow:

Stop asking "build or buy?", ask "what must we own?" Most enterprises are debating whether to build or buy AI capabilities. Wrong question. The right question: what's the minimum viable stack we must own to avoid permanent disadvantage? For most, the answer is orchestration and data pipeline, the layers where your domain expertise translates into a defensible advantage. Everything else can be rented. You wouldn't build your own power station, but you'd invest heavily in the machinery that makes electricity useful, same logic.

Some vendor relationships need board-level scrutiny , or exit. Your hyperscaler relationship isn't a software procurement exercise you revisit every contract cycle. It's a utility dependency. Sticky, consequential, and increasingly hard to unwind. If you're uncomfortable with that concentration risk, the time to diversify or exit is now , before you're locked in by data gravity, integration depth, and switching costs that compound quarterly. Ask your board: if our primary cloud provider doubled prices tomorrow, what's our realistic alternative? If the answer is "we'd pay it," you've already lost negotiating leverage.

Governance isn't a gate. It's an operating system. In a software paradigm, governance reviews happen at procurement. In an infrastructure paradigm, governance must be embedded in how you operate daily. Organizations that treat AI governance as a one-time approval will move slowly and anxiously, re-litigating every deployment. Those that build governance into their operating model, establish clear policies, pre-approve use-case patterns, and embed risk assessment will move fast with confidence. The gap between these two approaches will become the gap between market leaders and laggards.

Where I'm Placing My Bets

If the infrastructure framing is correct, several things should happen over the next 18-24 months. Here's where I'm placing my bets:

Model companies will consolidate or get absorbed. The standalone model company, selling wholesale capability without owning infrastructure or applications, occupies an unstable position. When your suppliers (hyperscalers) and your customers (enterprises) can both build competitive models, your moat erodes quickly. Expect acquisitions, acqui-hires, and a few high-profile down-rounds. The wholesale model layer is not a sustainable business.

Hyperscaler concentration will trigger regulatory attention. When three or four companies control the infrastructure layer of a general-purpose technology, regulators eventually notice. I expect early moves toward AI infrastructure regulation within two years, starting with compute capacity disclosure requirements and data center energy reporting, mirroring the requirements utilities must meet for reporting generation capacity. This won't be the model-level AI Act provisions. It will be utility-style oversight of the companies controlling compute.

Enterprise winners will emerge from the application layer. The organizations that pull ahead won't be those with the biggest AI budgets or the most aggressive experimentation programs. There'll be those who identified high-value applications early, built proprietary orchestration capabilities, and established governance frameworks that let them deploy with confidence. Their advantage will compound while competitors remain stuck in perpetual experimentation, forever evaluating and never shipping.

"AI strategy" will become as quaint as "electricity strategy." Within five years, organizations still treating AI as a distinct initiative will look as dated as those running "digital transformation programs" in 2025. AI will become infrastructure, assumed, embedded, unremarkable. The question won't be "what's your AI strategy?" but "how have you transformed your operations?" We're not there yet. But the trajectory is clear, and the organizations that see it early will define what the industry looks like on the other side.

The Question That Matters

The bubble debate will continue. Nvidia's multiples may correct. Model companies will consolidate. But the infrastructure build will proceed because the underlying economics, once you see them clearly, make sense.

The question isn't whether AI is a bubble.

It's whether you're building power stations when you should be building light bulbs.

For most of us in financial services, the power stations are someone else's business. Our job is to identify the workflows where AI transforms operations and build them before competitors do.

What's your light bulb? For me, it's the orchestration layer in compliance, the workflows that turn AI capability into operational reality. I'd like to know what it is for you.