There's a number floating around right now that should make you uncomfortable. The Big Four tech companies have budgeted $650 billion in AI capital expenditure for 2026 alone. That's more than the entire global telecom industry spends annually. Roughly two-thirds of what oil and gas invests worldwide.
Benedict Evans flagged this in his Spring 2026 deck, and it crystallizes something I've been thinking about for months: the supply side of AI is scaling at a pace that makes the demand side look almost comically unprepared.
When you ask most enterprise leaders what they're actually building with AI, the answers are thin. Summarize documents. Draft emails faster. Generate marketing copy variations. We have an industry pouring unprecedented capital into infrastructure while the organizations meant to use it remain stuck in first gear.
Every technology wave follows the same deployment pattern:
- Absorb: doing old things slightly better with the new technology
- Innovate: building things only possible with the new tech
- Disrupt: redefining the question entirely
Most organizations are deep in absorb mode. And history suggests many will stay there for a long time. After two decades of cloud computing, only about 30% of enterprise workloads have migrated to public cloud. Platform shifts don't happen on conference keynote timelines.
But here's what I find more interesting than the diagnosis: what actually moves a company from absorb to innovate? After spending the last six months deep in enterprise AI implementations across product and marketing organizations, I think there are exactly three levers that matter. And none of them are about picking the right model.
📊 Lever 1: Build Your Own Context Layer
LLMs are commoditizing fast. Benchmarks are saturated. Weekly model releases blur together. OpenAI, Google, and Anthropic perform similarly on standardized tests.
There are no obvious moats at the model layer.
This is actually liberating. If swapping from GPT-4 to Claude to Gemini is increasingly trivial, then your competitive advantage was never going to live in the model. It lives in what sits above the model: the context layer.
The context layer is the organizational knowledge that makes any model useful inside your specific walls:
- Decision history
- Project context
- Institutional memory
- Workflow patterns
Without it, every AI interaction starts from zero.
I'm working with a product organization right now where the entire AI strategy turned out to be a context engineering problem. They had six AI tools approved for use. Adoption was scattered. The tools worked fine in isolation but produced generic output because they knew nothing about the org's design system, product priorities, or how decisions actually get made. The breakthrough wasn't better tools. It was building a context system that feeds organizational knowledge into whatever tool a team member happens to use.
When models are commodity, context is your moat. The companies investing in their own context infrastructure are decoupling their fate from any single vendor. That's a fundamentally different strategic posture than "we're a GPT shop" or "we're an Anthropic shop."
🔧 Lever 2: Redesign the Process, Not the Task
When you automate existing tasks, you get a one-time efficiency bump. A 20% speed improvement on email drafting. A 30% reduction in first-draft time. These numbers look fine in a quarterly review.
They do not compound.
The ROI that actually matters lives in process redesign: rethinking workflows end-to-end rather than accelerating individual steps. This is the bridge from absorb to innovate.
Think about how a marketing operations team typically runs a campaign. Brief gets written in one tool. Media plans live in another. Performance data sits in a third. The campaign manager is the integration layer, copying, translating, reconciling across systems.
The absorb-mode response is to use AI to draft briefs faster or summarize performance reports. Fine. Maybe you save a few hours per campaign.
The innovate-mode response is to redesign the campaign lifecycle itself. What if the brief, the media plan, and the performance data lived in one system with a shared campaign ID? What if AI agents could synthesize insights across all three, not because they're smarter, but because the architecture eliminated the translation overhead?
That's not automating a task. That's building a capability that didn't exist before.
The distinction matters:
- Task automation has a ceiling. You hit it fast.
- Process redesign has a compounding curve. Each cycle generates data that makes the next one smarter.
That's the difference between a 20% speed bump and a structural advantage.
💡 Lever 3: Expand What People Can Do
There's a concept from 1865 that keeps proving itself relevant. William Jevons noticed that more efficient steam engines didn't reduce coal consumption. They increased it, because efficiency made new applications economically viable.
AI works the same way. Every executive I talk to frames it as a headcount conversation: "Can we do the same work with fewer people?" But the teams actually getting results are asking a different question: "What can our existing people do that wasn't possible before?"
I built a playbook for designers at a large product organization recently. The assumption going in was that AI would make designers faster at mocking up screens and iterating on components.
The actual outcome was different. Designers started prototyping interactive experiences they previously would have specced and handed off to engineering. They weren't doing the old thing faster. They were doing a new thing entirely: building functional prototypes with real design system components, testing scenarios with users, arriving at engineering handoff with validated decisions instead of static mockups.
The headcount didn't change. The role boundary did.
- Prototypes that would have taken an engineering sprint were tested in a day
- The team explored 3-4x more design directions per feature
- Decision quality went up because more options got evaluated before build
This pattern keeps showing up. PMs generating project context from organizational data instead of spending three days in stakeholder interviews. Marketing ops teams running insight synthesis in hours instead of weeks. AI didn't reduce the need for these roles. It expanded what they could accomplish.
What This Means
The companies making real progress share three characteristics: they're building context layers that compound, they're redesigning processes rather than automating tasks, and they're expanding role boundaries rather than shrinking headcount.
None of these require picking the right model. None of them require $650 billion in capex. They require something harder: the willingness to rethink how work actually gets done, and the patience to build infrastructure that compounds rather than chasing the next benchmark.
The interesting question isn't whether AI will transform enterprises. It's whether most organizations will ever get past absorb mode. The ones that do will have built something their competitors can't easily replicate. Not better AI, but better context, better processes, and people who can do more.
What are you seeing in your organization? Still absorbing, or starting to build toward something different?