The AI shift in tech, by the numbers — and why E-E-A-T matters more than ever

I’ve spent more than eight years building software for a living — distributed systems at AWS and Microsoft, then production AI at two startups I co-founded. Nothing in that stretch has changed the job as fast as the last three years. But “AI is changing everything” is a claim anyone can make, and a lot of people are making it without evidence. So this post does the opposite: every number below comes from a public, linked source, and where I add my own view, I’ll say so.

1. The money: an infrastructure build-out without precedent

Start with where the money goes, because capital is the most honest signal in tech. Epoch AI combined the SEC filings of Alphabet, Amazon, Meta, Microsoft and Oracle — cash spent on property and equipment plus new finance leases. Spending was flat from 2022 to 2023, then took off. By 2025 it reached roughly $448 billion, and 2026 is on track for about $770 billion. Epoch’s summary: hyperscaler capex has quadrupled since GPT-4 was released.

The rest of the market is moving with it. Stanford’s 2026 AI Index puts global corporate AI investment at $581.7 billion in 2025, up 130%, with US private AI investment at $285.9 billion — 23 times China’s $12.4 billion. Whatever you think of the valuations, the servers are real, and they are being built on the assumption that demand keeps compounding.

2. Capability got cheap — and much better

The spending would be hard to justify if the product weren’t improving this quickly. The 2025 AI Index found that the cost of running a model at GPT-3.5’s level fell more than 280-fold between November 2022 and October 2024. The 2026 edition tracks the capability side: on SWE-bench Verified, a benchmark built from real GitHub issues, top scores went from 60% to nearly 100% in a single year, and agents on OSWorld — real tasks on a real computer — jumped from about 12% to roughly 66% success.

3. Developers: adoption up, trust down

The most interesting chart of the year isn’t about models at all. In Stack Overflow’s developer survey, the share of developers using or planning to use AI tools rose from 76% in 2024 to 84% in 2025. Over the same year, favorable sentiment fell from 72% to 60%, and the share who trust the accuracy of AI output fell from 43% to 33%. In 2025, 46% said they actively distrust it.

My read: this isn’t a contradiction, it’s maturity. When a tool is new, people judge it by its best demo. Once it’s in daily use, they judge it by its worst Tuesday. The work hasn’t disappeared; it has moved from typing code to specifying intent and reviewing output. That shift has a cost, too — the AI Index reports that employment among software developers aged 22–25 has fallen nearly 20% since 2024, which makes entry-level engineering the first white-collar category with a measurable contraction.

4. Search: the click is shrinking

If you publish anything on the web, this is the shift that reaches you directly. Pew Research Center analyzed 68,879 real Google searches from 900 US adults. When an AI summary appeared, people clicked a traditional result in 8% of visits, versus 15% when there was none. They clicked a link inside the summary just 1% of the time, and they were more likely to end their browsing session entirely: 26% versus 16%.

The takeaway isn’t “SEO is dead.” It’s that ranking is no longer the finish line. The page that wins is the one an AI system chooses to cite, and the one a reader still wants to open after reading the summary. Generic content loses on both counts: a model can already produce it, so there is nothing left to cite and nothing left to click for.

5. Why E-E-A-T is the answer, not a buzzword

E-E-A-T comes from Google’s Search Quality Rater Guidelines: Experience, Expertise, Authoritativeness and Trustworthiness. Google added the first E — Experience — in December 2022, weeks after ChatGPT launched, and it describes Trust as the most important of the four. Google has also said it judges content by quality, not by how it was produced. Put those together and the logic is clear: when fluent text costs nothing, the scarce ingredients are the ones a model can’t fake — first-hand experience, a verifiable author, and claims you can check.

Here is how this post tries to practice what it describes:

• Experience — first-hand observations are marked as mine and kept separate from the data.
• Expertise — it’s written by someone who builds production AI systems, in the field being discussed.
• Authoritativeness — every number links to its primary source: Stanford HAI, Epoch AI, Stack Overflow, Pew Research.
• Trustworthiness — publish and update dates are visible, projections are labeled as projections, and each chart ships with its data table.

What I’d do with this, as a builder

• Design for intelligence that keeps getting cheaper. Features that look too expensive to run today may cost a fraction as much in eighteen months.
• Invest in verification, not just generation. The trust gap is where the value is: evals, tests and review loops are now core product work.
• Write from what only you have seen. Data you gathered, systems you shipped, mistakes you made — that is what gets cited and clicked.
• Make trust machine-readable. Named authors, dates, sources and structured data help both people and AI systems decide whether to rely on you.

The AI shift is real, measurable and still accelerating. But the lesson in the data isn’t that people matter less. It’s that the parts of the work only people can do — judgment, taste, first-hand experience and accountability — are now the parts that carry the most weight.