$700 Billion in AI Capex. Adoption Wide but Shallow. The Bottleneck Moved.

Benedict Evans published his Spring 2026 AI deck this week, and the headline number does most of the work. Per the deck summary, big tech is on track for roughly $700 billion in AI capital expenditures in 2026. The framing inside boardrooms, also per the deck, is that “under-investing is seen as the bigger risk.” That sentence is the consensus position of an entire industry. It is also the most important thing to read carefully.

Because the deck goes on to say two more things that, taken together, change what the $700 billion actually means.

First, foundation models are commoditizing fast. Second, value is shifting up the stack to applications, agents, and workflows. And the most quoted line, repeated by the TLDR newsletter summary of the deck: “Adoption is wide, but shallow. Deep integration is still rare outside of tech and finance.”

Read those three statements as one paragraph. The substrate is getting cheaper. The value is moving to integration. And integration is not happening. That is not a bullish picture of capex efficiency. It is a picture of capacity outrunning the organizations that are supposed to absorb it.

What “Wide but Shallow” Actually Describes

Wide adoption is easy to measure. License counts. Seat counts. Pilot counts. By those numbers, AI adoption is essentially universal at the top of the enterprise market. Every Fortune 500 has Copilot or ChatGPT Enterprise or Claude for Work somewhere. Every consulting firm has decks with adoption percentages above 80%.

Deep integration is harder to measure, which is why people stop trying. Deep integration means the work has been redesigned around the tool. The org chart has shifted. The review process has changed. The success metrics on someone’s quarterly scorecard are now downstream of an agent decision. The legal contract template references AI-generated content as a recognized category, not as an exception. Auditors know what to ask for.

Almost none of that has happened. The deck’s claim that deep integration is “still rare outside of tech and finance” is the polite version. The honest version is that most enterprises have AI in their hands and have not yet figured out how to put it into their bones.

This matters because the value Evans says is shifting up the stack, into applications and workflows, can only be captured by organizations that have done the deep work. If the foundation model layer commoditizes, the differentiated returns live in the integration layer. And the integration layer is empty for most of the market.

That is the macro picture in one sentence: substrate capacity outran organizational capacity to integrate it deeply. The capex is real. The returns require a second build that almost nobody has started.

Why the $700B Is Not the Problem

There is a temptation, reading these numbers, to call the $700B a bubble. That is the wrong frame.

The capital expenditure is going somewhere productive at the infrastructure layer. Data centers get built. Power contracts get signed. Chips ship. Energy capacity comes online. The substrate is being laid down. In ten years it will be useful regardless of which specific model providers survive the commoditization.

The problem is not the spend. The problem is the mismatch between spend velocity and absorption velocity. The capex curve is steep. The integration curve is flat. And the gap between them is being filled, today, with optimism and slide decks.

This is not a unique pattern. It is the same shape as the dotcom buildout in 2000, the broadband buildout in 2003, the cloud buildout in 2010. Substrate gets built ahead of demand, demand catches up later, the second wave of returns goes to whoever absorbed the substrate fastest. The companies that did the second build won. The companies that bought the magazine subscription did not.

The question facing every board in 2026 is whether they are in the first group or the second group. And the honest answer, for most, is that they have not started the second build. They have bought tools. They have not redesigned work.

Where the Bottleneck Actually Sits

If foundation models are commoditizing and value is moving to applications, agents, and workflows, then the binding constraint on enterprise AI returns is no longer compute. It is the capacity of organizations to govern the integration of AI into their actual work.

That phrase, capacity-to-govern, is doing real work in this sentence. It is not the same as risk management. It is not the same as compliance. It is the organizational ability to make decisions about where AI fits, who owns the outputs, what the new operating model looks like, and how the human-and-agent workforce will be measured. It is the design work that turns substrate into productivity.

The reason this matters now, and not eighteen months ago, is that the substrate has caught up. Models are good enough. Tooling is good enough. APIs are stable enough. The technical excuses for shallow integration have largely evaporated. What remains is the organizational work, and organizational work compounds slowly when it has been neglected.

We have written about three pieces of this before. The organizational debt of AI covered the BCG finding that 70% of AI implementation hurdles are people and process. The 81,000 people governance demand showed the Anthropic data on how fast enterprise demand for governance roles has grown. Governance and the adoption mandate addressed the tension between executive mandates to use AI and the mental model gap that prevents teams from using it well.

What Evans’s deck adds to that picture is the macro number. $700 billion in substrate spend, paired with shallow integration, names the problem at the size where boards have to take it seriously. It is no longer a complaint from change management consultants. It is the dollar number on the other side of the integration deficit.

What to Do With This

If you are an executive reading the deck summary and trying to translate it into your own organization, three things are worth doing in the next ninety days.

Audit your integration depth, not your adoption breadth. Pick three workflows where your organization has deployed AI. For each, write down what has actually changed in how the work gets done, who owns the output, and how success is measured. If the honest answer is that the workflow looks the same and someone just types into a prompt box now, you have wide adoption and zero integration. That is the population the Evans deck is describing.

Name the second build. The first build was procurement. The second build is integration. Treat them as separate programs with separate leaders, separate budgets, and separate timelines. Procurement is mostly done. Integration has barely started. Conflating them is how organizations spend another year confusing license counts for transformation.

Stop calling it a technology investment. The $700 billion at the industry level is a technology investment. Your spend, inside your organization, mostly is not. Inside your walls, the binding constraint is organizational design, not model capability. Budget accordingly. If your AI program has more spend on licenses than on operating model redesign, the program is mis-shaped for the moment we are actually in.

The Evans deck is, in the end, a polite warning to the people writing checks. The substrate will be there. The returns will not be automatic. The companies that capture the value Evans says is moving up the stack will be the ones that did the integration work while everyone else was buying seats.

That work is governance work. And in 2026, the bottleneck is not how much we can spend. It is how much we can absorb.

---

This analysis synthesizes AI Eats the World, Spring 2026 (Benedict Evans, Spring 2026).

Victorino Group helps boards and executive teams close the gap between AI capex and AI capacity-to-govern. Let’s talk.