Q2 2026: The Quarter That Built Its First Loop

Fable 5, the most capable AI model anyone had ever released to the public, was initially available for three days in June. I spent the first two of them telling myself I'd get to it tomorrow.

I have, interestingly, started to do this for a living. I run certain AI pilots at an ad-tech company, I now give conference talks and I built a fully-automated content system that auto-analyzes my own writing. And the biggest release of the quarter sat unopened for two days, because there is simply too much to keep up with. I finally sat down with it in what turned out to be its last twenty hours of availability. I did not know at the time they were the last twenty hours.

I gave it a somewhat menial assignment on purpose: an explainer page for defining MCPs instead of a typical deck. Interactive HTML. I asked Fable for “stunning visuals” and it came back with particle effects, animated graphs, charts you could grab and tweak, and a hero image of a hundred thousand points of light streaming through a glowing portal ring, every agent in the system entering through the same gate. I had asked for protocol documentation and I got what looked like the cold open of a sci-fi movie. I never learned to code. I learned music (which, fair enough, is its own technical language), but building this way feels like speaking through a translator who is somehow also composing and simultaneously playing the orchestra. I wanted to do more.

Then it was gone. The Commerce Department forced a suspension seventy-two hours after launch, the most powerful thing anyone had ever shipped to the public became a thing literally nobody could use, and my reaction, verbatim, was "oh crap." I spent the next two weeks pining after it like a high school crush. Now, it is back, albeit in a somewhat limited capacity. It is, and I want to be precise about this, freakin’ awesome. I still haven't gotten to most of what I wanted to do with it, and there is a limited time window for experimenting before it goes to eye-wateringly expensive token usage.

The Pace

Last quarter the models arrived every week or so. This quarter they arrived faster than I could write a LinkedIn post about the previous one.

Eight days flew by in April: Claude Opus 4.7 on the 16th, GPT-5.5 on the 23rd, and a day later DeepSeek V4, a near-frontier model you could download for free. Two American labs and a Chinese one, inside a single business week. Opus 4.8 followed in May. Fable 5 got its three days in June, and on June 26 the newest OpenAI model 5.6 Sol previewed to roughly twenty government-vetted organizations and nobody else. Two of the three leading American labs closed the quarter with their best models behind a rope.

Tap to zoom

The measuring sticks had a stranger quarter than the models did. The workhorse science benchmark everyone quotes, GPQA, is finished, clustered at 94 and 95% where the differences stop meaning anything. Humanity's Last Exam, the test built to be beyond current AI, is more than half answered on one leading evaluation, at 53.3%. FrontierMath, research-level mathematics, still stops everything at around half. It seems to be the one lock in the house the burglar can’t currently pick, and I find that oddly reassuring. But math is definitely “cooking”. And the industry's favorite coding benchmark got partly retired this quarter after researchers found the models had effectively seen the answers. We are generating capability faster than we can build honest rulers for it.

The rulers being built next don't look like the old tests at all.

METR, the group that measures how long a task an agent can finish on its own, says that length is doubling every three months for recent models. They admit their own suite can't measure past sixteen hours, a ceiling the labs' internal models have likely already reached. OpenAI's GDPval hands the models real work products from forty-four occupations, spreadsheets and briefs and diagrams, and grades them blind against human professionals. At launch, the best model won or tied on just under half. ARC-AGI-3, launched in March, is a set of games with no instructions at all: humans solve 100% of them, and frontier AI solved 0.51% at launch. The old benchmarks asked whether the model knows things. The new ones ask whether it can work. We've graded with the old ones for years, but pretty soon quoting them will sound like asking whether your fifteen-year-old can walk up a flight of stairs. When the toddler did it, it was a miracle. Now it's Tuesday.

The price of a token, the unit of everything these models do, has fallen something like 98% since 2022. Enterprise AI bills roughly tripled anyway. There's a name for this and it's a hundred and sixty years old: the Jevons paradox. Make the steam engine more efficient and the country burns more coal, not less, because cheap power finds new work to do. Cost per token keeps falling. The number of tokens keeps exploding, because a single agentic task, the kind that plans and acts instead of just answering, burns an estimated fifty to five hundred times the tokens of a chat message, and by some estimates around 72% of what companies actually spend on AI now sits outside the model bill entirely, in the connecting and retrying and monitoring around it. In ad speak: the impression got cheaper - the campaign got more expensive.

I can feel the paradox from inside my own subscription. Fable is the Ferrari now, and Opus, the flagship a quarter ago, is settling in as the workhorse. I have put a very modest amount of money into usage credits and I just watch it drain faster than I’ve ever seen when the Ferrari is out of the garage. Something gives eventually: the limits go up, or the bills do.

Tap to zoom

The Labor Force

In an earlier draft, this section was called ‘The Noise’, and that was lazy. The labor force isn't doing one thing that a headline can hold. It's doing at least four things at once, and they point in different directions.

Let’s start with the real cuts. In May, 40% of announced U.S. job cuts were attributed to AI, the highest share since anyone started counting. June cooled to 31% and AI still led every stated reason for the fourth month running. In the first half of 2026, AI was named in 101,743 cuts, nearly double all of last year. Oracle put its cuts in an SEC filing, under oath, in the careful language companies use when lawyers are watching. And the pain concentrates at the two ends of the seniority curve. Payroll data shows employment for 22-to-25-year-olds in the most AI-exposed jobs shrinking about 3.8% a year, down 16% relative to everyone else since gen-AI first arrived. And somewhere, mid-career, is an engineer who has done one thing superbly for twenty years, watching a model do most of it, with no second option to fall back on. That specific pain, the lifer with no fallback, doesn't show up in any index.

Some “AI job cuts” may be little more than a thin excuse. A lot of these cuts are COVID-era overhiring coming home to roost, because "the future made me do it" is an easier memo to write than "we over-hired and rates went up." Nvidia's chief executive called the AI-layoff framing lazy, and the math backs his skepticism: Meta's entire payroll is roughly a fifth of its AI infrastructure budget - you cannot cut your way to balancing a number that size by firing people. When researchers surveyed thousands of executives early this year, close to 90% reported no measurable productivity gain from AI at all. The trend is clear: AI shows up as the reason for cuts far faster than it shows up in any actual business results.

A third cohort is, by definition, almost impossible to accurately measure, because it's invisible: the hiring that doesn't happen. Meta grew revenue 33% last quarter on 1% headcount growth and still employs nine thousand fewer people than it did in 2022. Microsoft reported the same 228,000 employees two fiscal years running while adding thirty-seven billion dollars in revenue. Salesforce's CEO said it plainly on the May earnings call: "We're not hiring more engineers." Engineering has sat at about fifteen thousand people for two years. Amazon's CEO put it in writing a year ago: AI efficiency gains "will reduce our total corporate workforce." Software job postings sit about 69% below their 2022 peak. Indeed's own economist is careful with the causality and it's worth repeating her exact shape: AI didn't cause the tech-hiring bust, but it may be preventing the recovery. The researchers tracking young workers found the same mechanism, that the damage is making itself known through a lack of offers, not through pink slips. No company holds a press conference for the hundred thousand people they didn't hire. To be fair to the contested part: the New York Fed looked at job postings in May and found the slowdown isn't concentrated where the AI story says it should be, and started before ChatGPT. The freeze is real. How much of the blame should fall on AI is genuinely unsettled.

This last one is important to watch closely, because they are likely to surprise everyone in the long run: the refactorers. Companies that are hiring, using AI hard, and rebuilding how the organization works mid-operation, trying to grab as much ground as they can while heavier incumbents adjust. The ambition in that cohort talks along the lines of fifty-x and hundred-x. The record shows smaller numbers, but they are just as absurd.

• Lovable says it passed $500 million in annual revenue in June with 146 employees, roughly ten times the revenue per head of a normal software company, self-reported and unaudited, while its CEO publicly recruits the people Big Tech just laid off.
• Duolingo, in SEC filings, grew headcount through its AI-first reorg while its CEO claims four to five times the content output from the same team.
• Cursor went from $100 million to $3 billion (30x) in annualized revenue in sixteen months on a team of about three hundred, and then, two weeks before the quarter ended, SpaceX bought it for sixty billion dollars, days after its IPO - the biggest acquisition of a venture-backed startup on record.
• Palantir's CEO marks the boundary of the rhetoric, targeting ten times the revenue with fewer people.

Almost every number in this paragraph is a founder reporting his own score, so it's important to keep that in mind. But this market posture is real, and these companies are likely hiring people who were a) recently let go or b) looking to pivot as their own futures look less and less certain.

Tap to zoom

The Slop

Stop me if you’ve heard this one: this is like we’re in 1997. In the early internet in the dot-com era, most of these companies will die and the internet will carry on. It's tired because it's everywhere, and it breaks in the one place that matters. In 1997 the wild stuff was coming “someday”. Right now, the wild stuff is here, today, 100% absolutely guaranteed bananas, mind-blowing, available for a subscription fee, and most people are not touching it.

Roughly half of American adults have never used an AI chatbot ever. Not once. Only a quarter use one daily. Meanwhile the passive exposure is close to universal: 60% have had AI summaries put in front of their search results, roughly half of newly published web articles are primarily AI-generated by detector estimates, and 51% of spam email is now written by AI. Meta removed ten million impersonator profiles in six months. So the average person's relationship with this technology is: “I have rarely touched the tool and I’m swimming in its exhaust.”

And they've named the exhaust. "Slop" was the word of the year at Merriam-Webster, at the American Dialect Society, and at Australia's Macquarie Dictionary, all three, the same season. No pollster has asked "when you hear AI, do you picture slop?", so the dictionaries are the closest measurement we have, and their answer is yes. People hate spam. I hate spam. You hate spam. AI slop is spam that went to film school, and for most people, it is the main thing AI has personally done to their day. Three-quarters of Americans say it's important to be able to tell AI content from human content. Half aren't confident they can. Twenty percent of people worldwide trust an AI chatbot's answers about the news.

The media literacy is low, the curiosity is low, the passivity is high, and the distaste is high. And this is a completely rational response to the AI the average person has actually encountered.

The Inversion

In a 32-country Ipsos survey fielded this March and April, 85% of Chinese respondents said AI products and services offer more benefits than drawbacks. In America: 38%. Excitement about AI: 83% in China, 33% here. A separate 47-country study found 92% acceptance of AI in China against 54% in the U.S. Edelman's latest cut puts trust in AI at 87% there and 32% here. Every one of these studies carries the same asterisk: polling inside China means online panels, urban and educated and connected, answering inside a country where information can be tightly controlled. Hedge the results hard, but the gap is still forty to fifty points across three independent surveys.

So the country building half the frontier is the country that trusts it least, and like many things, I don’t think this is just a straight-forward simple answer.

I imagine three intermingling components. The first is that Americans aren't rejecting artificial intelligence so much as running out of institutions to trust it through. Congress polls at roughly 10% confidence, currently the lowest stat of the things Gallup measures. The Supreme Court fell below 40% approval last summer for the first time in a trend that goes back to 2000, when the same question polled at 62%. Confidence in higher education is down fifteen points over a decade. Nine of the country's biggest law firms pledged nearly a billion dollars in pro bono work to settle their way out of punitive executive orders. The rot predates any single administration, and I'm stating it as flatly as I can: the referee class has had a very bad few years, at the exact moment a technology arrived that needs referees. Ask about AI specifically and the numbers say it directly: 67% of Americans have little or no confidence in the government to regulate it, six in ten don't trust the companies building it, and in the freshest international polling, Americans trust their own government to regulate AI less than any other public surveyed trusts theirs. We are the lowest of anyone. Happy 4th of July!

Next, the messengers. The technology is arriving wrapped in the logos of companies Americans have spent a decade souring on, and the public reads "trust us with the most powerful technology ever built" differently when it comes from the people who brought them engagement=enragement optimization. The trust numbers tell the story plainly: Americans do not believe these megaliths have their best interest at heart.

Last, the slop! Most Americans meet AI as spam, scams, and brainrot feed content, and they've been told that's AI, so okay, that’s AI. Gross, I hate it. In China, the state is running an explicit national adoption campaign, AI Plus, targeting 70% penetration of AI agents and devices by 2027, and the public is introduced to AI as a national project. One population was introduced to this technology as a promise. The other was introduced to it as a plague.

Do China's numbers mean Chinese citizens are right to relax? I have no idea, and neither do the surveys. Internationally, only 27% of people trust China to regulate AI at all, and the 69% of Chinese respondents who say current safeguards are sufficient could be interpreted as either confidence or compliance. Meanwhile, the formal governance picture stayed jammed in both gears at once. The Pentagon's case against Anthropic sat unresolved at quarter's end, the EU postponed its own AI rulebook by sixteen months because the enforcement tooling wasn't ready, and the same government that pulled a model off the shelf in three days ordered its military, that same month, to move faster on AI. I'm laying those out matter-of-factly and walking away with my hands where everyone can see them.

Tap to zoom

The Grid

All of this runs on electricity. And this quarter, electricity graduated from a footnote to a constraint. Microsoft's cloud backlog, work it's been paid for and hasn't delivered, sits around $627 billion, nearly double a year earlier. The bottleneck isn't demand and it isn't even chips anymore. It's power. The four biggest tech companies plan around $700 billion in capital spending this year. The entire U.S. investor-owned utility industry, every wire and substation and power plant, comes to about $239 billion - Amazon alone plans to spend roughly what the whole utility industry spends!

Americans started paying for it on June 1, when the largest U.S. grid market's record capacity price took effect. Maryland's two big utilities alone project fourteen to sixteen dollars a month more per household, and the market's own monitor attributes roughly $21 billion of the past three capacity auctions' costs to data-center growth. In early May, the body that oversees the grid issued an alert it has used only about three times in its history, after data-center loads dropped a thousand megawatts offline in seconds. And somewhere in Memphis, a chatbot's exhaust turned into smog over a specific neighborhood: gas turbines running beside a residential community to power xAI’s supercomputer, the NAACP sued in April and the federal government intervened on the company's side in June, citing national security.

Okay, now the water question, because I walked into this quarter believing the data-center water panic was mostly a myth, and I made my skeptic fact-checker agents check whether I could still say that. The verdict: yes, mostly, with a slight limp. Data centers “drank” about 17 billion gallons of water in 2023, which sounds apocalyptic, but it is roughly 1/30th of what American golf courses pour on grass. 17 billion gallons is only about three and a half hours of the country's daily irrigation draw, and that 17 billion was for the whole year! Nationally, water is a rounding error. But, importantly, the electricity powering those data centers consumes about twelve times that much water upstream, and two-thirds of new data centers since 2022 sit in places already short on water. In one Georgia county, residents' wells failed as construction started next door, rates are rising 33%, and the county projects a water deficit by 2030. So: nationally a rounding error, locally very real and serious pain, and the national average is no comfort if you live next to a well that’s gone dry. Power is the bill that is coming due for all of this AI mania.

And here is the one place in this piece where I stop observing and argue. When a regular person's power bill goes up because of the data center down the block, while the county hands that data center an 85% property-tax abatement, the subsidy is flowing in the wrong direction. Virginia alone forfeited $1.6 billion in a single year to its data-center tax exemption. The towns are funding the companies. The companies should be funding the towns. Some shift has started over the past year, with more landing this very quarter: Ohio, Georgia, and Indiana pushed data centers onto minimum-pay tariffs, and in Indiana, Amazon, Google, and Microsoft agreed to pay them. Seven hyperscalers signed a White House pledge in March to fund their own grid upgrades "whether they use the electricity or not." Virginia signed the country's first per-kilowatt-hour data-center tax on June 30, the last day of the quarter. But it's all pledges and minimums, none of it audited, and the simple, visible thing remains undone: no company has yet put money directly back in the hands of the neighborhoods absorbing its buildout and said so out loud. It is the cheapest goodwill on the market, and it's just sitting there. Truly wild to me they aren’t doing more about this.

One more number which deserves a quick correction - when Senator Warren claimed bills near data centers had jumped 267%, a fact-check rated it mostly false. That was a wholesale figure, and the real five-year residential increase is closer to 42% nationally. Still enormous. Still real. The accurate number is bad enough, which is exactly why it's the one worth citing.

Tap to zoom

The Metabolization

The plumbing is arriving. It has not arrived. When I say plumbing, I mean interconnectedness, integrations, foundational layers that make things easier. Power grids. Water mains. Plumbing. The tendrils of it are beginning with things like plugins and connectors. There’s more to come.

The standards are starting to firm up: the A2A protocol that lets agents talk to each other hit its first stable version in April with 150-plus organizations behind it, and MCP, the one that lets agents reach their tools (the subject of my Fable explainer page), locked its next release in May. Then the control panels came in a rush: in roughly ten weeks, at least five major companies shipped platforms whose entire job is to orchestrate and govern fleets of AI agents, Google in April, Microsoft on the first of May, ServiceNow the same week, Salesforce and Amazon in June. My own corner of the industry went agentic in public too, autonomous buying agents in April, an agent across the biggest ad stack in May, agentic ad formats at Cannes in June, and, on June 11, my company Innovid's own agent layer for campaigns. I'm too close to that last one to hand you adjectives (Phenomenal! Superb! Unparalleled!). The honest version is that everyone, us included, is at the starting line.

What I can report from my own personal desk is where this section gets its name. I trust my own fleet. I outline with Fable, hand the plan to six or seven Opus and Sonnet agents, and they execute and self-test while I do something else. It is wild, it is very, very cool, and it is a little scary, and it is entirely too fast to expect the average person to absorb. Not because they’re dumb, because they’re busy and because they hate AI slop. That's the fastest speed: super-users and small, nimble companies, pinpoints of shock and electricity, metabolizing this in real time and mostly alone or in tight pockets.

The incumbents move at a different speed, and they have not metabolized much of anything yet. Gartner's first-ever hype cycle for agentic AI puts it at the peak of inflated expectations, with about 17% of organizations actually running agents. A May survey found 15% of data leaders felt ready for production agents while a reckless 41% were running them anyway. The same analysts expect most early orchestration projects to fail, and their reasons have nothing to do with the models: integration, governance, the boring work of making systems talk to each other. That’s what I’ve coined the 90% Problem, exactly: 90% of clients I talk to either don’t have agents and feel behind, or they do have agents and they don’t talk to anything else. Organizations at the scale of a national health insurer have not fully metabolized every possible benefit to agentic AI and they won't for a while. A body that size takes years to fully unpack what a two-hundred-person company digests in a quarter, which is exactly why the two-hundred-person companies can smell blood.

The public is slower still, and it isn't metabolizing AI at all. It’s marinating in the worst of it. Last quarter, I wrote that we had speed without trust. I'd sharpen that now, because the trust isn't merely lagging the plumbing; trust is falling while the plumbing goes in. Those are different problems. The first is an engineering schedule. The second is about relationships.

Tap to zoom

The First Loop

I owe you an ending, and I don't have a tidy synthesis. I’m merely chronicling this wild journey from inside the roller coaster car while it’s happening to all of us. Is the roller coaster metaphor tired? I’m keeping it anyway.

But let’s be specific about where we are on it. In the first quarter, we were climbing into the car, buzzing, pulling the bar down, full of anticipation and excitement. This quarter, we went down the first hill and through the first loop, and the thing about a first loop is that it recalibrates you: the speed stops being a theory and becomes something you are actively experiencing viscerally. We are moving at speeds we are not used to, through turns nobody called out in advance. We have no idea what the rest of the track looks like. There could be nine more drops and a hundred more corkscrews. The whole thing could fly off the railing at the end. Nobody in the car can see far enough ahead to tell you, and anyone selling you a track map is guessing at best.

It's thrilling. It's a little scary. I've stopped trying to pick one. The car is moving - see you at the next turn.