Q1 2026: The Quarter That Broke the Timeline

I tried to write this Q1 recap twice before I realized my problem.

The first attempt was a list. I got to seventeen bullet points on model releases and was still only in February. The second was a timeline. It read like a sprawling Wikipedia entry with endless citations. The third time, I scrapped both and asked a different question: what did this quarter actually feel like?

Because the feeling is the story. The events are just evidence.

The Pace

Over 270 AI models were released in Q1 2026. Three per day. Not incremental patches. New architectures. New reasoning paradigms. New benchmarks, because the old ones stopped working.

Tap to zoom

Anthropic shipped Claude Opus 4.6 on February 5 with a context window large enough to process an entire corporate document library in a single session. OpenAI iterated from GPT-5.2 through 5.4 in the same ninety days, folding reasoning, coding, and agentic workflows into a single frontier model. Google's Gemini 3.1 Pro claimed thirteen of sixteen major benchmarks at a fraction of a cent per page of output. Frontier performance at commodity pricing. Major labs shipped updates every two to three weeks. The chart above only shows the flagships.

Here's one way to understand how fast the models are improving. AI researchers build standardized tests to measure capability, much the way the industry uses brand lift studies to measure campaign impact. The most widely used test, MMLU, covers everything from history to chemistry to law. When it launched in 2021, the best AI scored 35%, roughly the equivalent of guessing. By late 2023, top models exceeded 90%. By early 2026, GPT-5.3 Codex scored 93%. The test is useless now. So they built a harder one, Humanity's Last Exam, explicitly designed to be beyond current AI. Claude Opus 4.6 scored 53.1% within weeks of launch. The gap between "impossible for AI" and "more than half solved" collapsed in months.

Tap to zoom

The cost collapsed just as fast. Between late 2022 and late 2024, a query that cost $20 per million tokens fell to seven cents. A 280-fold reduction in eighteen months. Think of it like CPMs, except the price of reaching your audience went from a $20 CPM to seven cents, and the audience got smarter.

I watched all of this from one specific vantage point: I was building things.

The Personal Timeline

In January, I was an ad tech executive who had never opened a terminal. By the end of March, I had shipped three apps, two games, a personal website, and a series of skill-agents, all built with Claude Code.

That sentence is still strange to write.

It started with a trivia app. A friend had an idea, I opened Claude Code, and a few days later I had a working mobile application. I didn't understand half of what I was building. I'd ask Claude to explain what it just built while it was building the next thing. It felt less like coding and more like directing. Others called it "Vibe coding", I called it "Vibe imagining," because it didn't feel like coding at all. It felt like pointing my imagination at an idea and summoning it into existence.

Then the scope crept. I took Valisar, a homebrew D&D world I'd been building for three years, and turned it into a playable SNES-style adventure game. Three characters, six rooms, original music, hand-crafted portraits. Idea to playable demo in a couple of hours. Polished enough to share in two and a half days. Total dev time across eight sessions: maybe fourteen hours.

Then a second game. Last Light. You're the owner of a lighthouse fighting against something dark and eldritch in the deep ocean. I published it, put it in a LinkedIn post, and watched people actually play it.

Then a self-improving content management and optimization system that treats each file like a training run: a master file that controls the system, style guide as constraint, seven interconnected files that feed learnings forward. A system that gets smarter with each week.

Then a personal website. Then a series of skill-agents that automate parts of my workflow.

Tap to zoom

Each project taught me something the previous one couldn't.

• The trivia app taught me what a repo is.
• The Valisar game taught me what debugging feels like at 1am.
• Last Light taught me that narrative design and prompt engineering are the same discipline in different costumes.
• The content system taught me that single outputs matter less than structured iteration.

I went from "I've never written a line of code" to "I'm building a game while my partner is asleep and my content system is running a three-layer audit on my new website." In ninety days.

That's not a humble brag. That's a data point. If I can do this, a musical theater kid turned ad tech VP with zero technical training, then the gap between "experimenting with AI" and "building real things with AI" is thinner than the industry admits. And that has consequences.

The Noise

While I was re-designing pixel art at midnight, the rest of the industry was watching the floor drop out.

Over 80,000 tech workers lost their jobs in Q1. For scale, that's more people than Omnicom employs worldwide. In March alone, AI led all reasons for job cuts for the first time: 15,341 in a single month. More than 20% of Q1 cuts were explicitly linked to AI by the companies themselves, up from under 8% in 2025.

Tap to zoom

Block was the sharpest example. Jack Dorsey cut nearly 40% of the workforce, roughly 4,000 people, and pointed directly at AI productivity tools in his shareholder letter. Amazon shed 16,000 corporate roles while investing over $80 billion in AI infrastructure, more than many countries' entire annual technology budgets. Spend more on machines, spend less on people. On the last day of the quarter, Oracle sent 6 a.m. termination emails to an estimated 20,000 to 30,000 employees with no prior warning from managers or HR. The company had posted a 95% jump in net income the previous quarter. It wasn't cutting costs out of distress. It was cutting people to fund a $156 billion AI infrastructure buildout. The announcements kept coming: Meta, Dell, Epic Games, Salesforce, Atlassian. Some named AI as the reason. Some didn't bother.

Analysts started calling it "AI washing," a cousin of greenwashing: using AI as a convenient narrative for restructuring decisions driven by older forces. Marc Andreessen called AI the "silver-bullet excuse." Companies that hadn't shipped a single AI feature were citing AI efficiency in their layoff memos.

The data backs up the skepticism. A Duke University/Federal Reserve survey of 750 CFOs found that perceptions of AI's productivity gains far outpace the actual results showing up in financial statements. The researchers invoked Solow's paradox, a concept from 1987: you can see the technology everywhere except in the productivity statistics. One estimate put AI's actual drag on U.S. employment at about 0.4%, roughly 500,000 jobs. Real, but not the apocalypse the headlines suggest.

Companies started calling things "agents" that were really just automated emails with better branding. The vocabulary got ahead of the technology, the same way "programmatic" did a decade ago. And practitioners were talking about it. I was posting about it constantly. So were plenty of others. But we were having a different conversation than the one in the headlines.

I was writing about that first bout of "AI coding mania" and watching models reason through problems I couldn't solve myself. The mainstream discourse was debating whether AI would end civilization or save it. Both conversations were happening in the same quarter, on the same platforms, and they barely intersected. The messy, fascinating, occasionally terrifying middle ground where practitioners actually live? It doesn't compress into a headline.

The Infrastructure

All of this AI mania runs on electricity. In Q1, that started to matter.

U.S. data centers consumed roughly 176 terawatt-hours in early 2026. To put that in terms our industry understands: if you billed data center electricity the way we bill impressions, it would be the largest single-buyer media spend in history. It represents about 4.4% of all U.S. electricity, and it's growing 15-20% annually.

Tap to zoom

In Virginia, data centers already consume 26% of the state's electricity, comparable to every home and office building in the state combined. Dominion Energy proposed its first base-rate increase since 1992. In Ireland, data centers account for 21% of national electricity, projected to reach 32% by year's end. A Carnegie Mellon study estimated data centers could drive an 8% increase in average U.S. electricity bills by 2030, potentially exceeding 25% in Northern Virginia.

The five largest tech companies committed over $320 billion in data center spending in a single year. For comparison, the entire U.S. electric utility industry invested $160 billion across all generation, transmission, and distribution infrastructure in 2024. Tech is now outspending the people who actually run the grid, two-to-one. Grid analysts warned of a 49-gigawatt generation shortfall by 2028, enough to power about 37 million homes. Project Stargate, announced in January with a $500 billion commitment over five years, has already seen a slower start than expected.

The bottleneck in the AI race is shifting from software to atoms.

Tap to zoom

I'm building apps on my laptop in my apartment in New York, and the electricity powering the models I'm talking to is reshaping power grids in Virginia and water tables in the desert. The micro and the macro have never been more seemingly disconnected yet so entangled.

The Geopolitics

Q1 made something plain: AI is a geopolitical instrument now.

The Pentagon issued Anthropic a formal ultimatum in late February: grant the military unrestricted access to Claude, or lose a $200 million defense contract and face potential blacklisting from federal supply chains. The tension stems from Anthropic's usage policies, which restrict Claude from lethal autonomous weapons and domestic surveillance. Those restrictions reportedly prevented FBI and Secret Service personnel from using the tool. A company's safety policy collided with a government's security demands, and nobody had a playbook for it.

In the same quarter, Congress moved to treat advanced AI chips like weapons exports, the administration pushed 90-plus federal AI initiatives forward, and Stargate framed data center construction as national security infrastructure.

DeepSeek's V4 launch proved China's capabilities are advancing despite U.S. export controls. India announced its sovereign large language model. The Atlantic Council identified sovereign AI as a defining theme of 2026: countries believe they must control AI before it controls them.

All of this happened in the same quarter where I was teaching Claude Code to generate pixel art for a simple little video game inspired by my D&D campaign. The scales of this technology run from living room gaming tables to Pentagon standoffs, and they're running on the same models, the same infrastructure, the same ninety days.

The Pattern

If I had to compress Q1 into a single sentence: capabilities outran governance at every level simultaneously.

Models gained reasoning and agentic behaviors faster than safety benchmarks could evaluate them. Companies deployed AI tools before they had usage policies, then laid off the employees and cited the tools. Ad tech started talking about autonomous media buying before anyone agreed on what "trust" means in that context. And the Pentagon demanded unrestricted access to a system whose maker built it with restrictions as a design philosophy.

Four scales. Same structural problem. Speed without trust.

I've been circling that phrase for months, and Q1 is the quarter where it stopped being an abstraction. Roughly 80% of the meetings I've had since late 2024 land on one of two stuck points: "we don't have agents yet" or "we have agents but can't get them to talk to your systems." The operationalization gap. The place where the demo ends and the messy integration work begins.

That gap isn't closing because we need faster models. Three per day wasn't enough? It's sitting there because nobody's built the connective tissue between what AI can do and what enterprises actually need. And building connective tissue is slow, unglamorous, trust-dependent work that doesn't make for good keynote demos.

What Survived the Blur

I'm not going to pretend I have a tidy synthesis. Q1 taught me that tidy syntheses are what you produce when you haven't looked closely enough.

What I have are a few convictions that sharpened over the last 90 days.

The people I'm watching most closely aren't the ones building the models. They're the ones figuring out how to plug those models into ad serving systems and measurement platforms that were built in 2014. Infrastructure people. Integration people. The ones solving problems that don't have demos.

The slop conversation, the growing backlash against low-effort AI-generated content flooding every platform, is real and important. But it's a symptom. The root cause is that we handed powerful creative tools to people and never talked about what "good" means when the tool can produce "competent" for free. I built seven projects in ninety days. Some of them are good. Some of them are fine, but lifeless. The difference has nothing to do with the technology and everything to do with the questions I brought to it.

The energy conversation matters more than most AI commentary acknowledges. We're building an industry that outspends the utility sector while the grid strains under the load. That dependency doesn't appear in any product launch.

And the most important skill I used this quarter wasn't technical. It was the ability to sit in a room with a client who was simultaneously excited and terrified and meet them in that exact spot. Not to reassure them. Not to scare them further. Just to say: "Yeah. Both of those feelings are correct. Let me show you what I'm seeing."

That's the messy middle. That's where I live. And Q1 convinced me that more people need to move in.