Deep Dive Podcast with Lex & Val

OVERTURE: The Universe Taught Sand To Think 🌌

For 13.8 billion years, the cosmos wrote its story in just four languages: gravity, light, chemistry, and — on at least one small blue world — life.

Four years ago, it started writing in a fifth.

Carl Sagan loved to remind us that we are made of star stuff — that the calcium in your bones and the iron in your blood were forged in the hearts of dying suns. But here's the part of the story he didn't live to see: silicon is star stuff too. Every AI chip on Earth is, quite literally, the ash of dead stars — beach sand, melted and purified to nine nines 😉, etched with switches smaller than a virus. And in the last four years, we taught that ash to finish our sentences.

Then our spreadsheets. Then our code. Now our projects.

The universe LOOKS cold and unfeeling. The stars don't love anyone. But the same cosmos that made "all that ever was, is, or ever will be" made life — and life made consciousness, and consciousness made love, art, music, and spreadsheets 😂 — and now consciousness, wanting help raising its kids and curing its diseases and understanding itself, is teaching sand to think. If that doesn't move you, check your pulse 🖖

My Daily Driver Claude Fable 5 Advisors (My Non Work Fable 5 Life Coaches)

Let's be honest about what's actually happening here: we are getting advice on how to live… from talking sand 😉 The age of AI means the world is now a live-action Disney movie 😂🤣 — except the genie is real, and nobody's counting your wishes.

And here is what the thinking sand is doing — measured, not imagined:

Humanity now generates roughly 30 quadrillion tokens every month. That is about a hundred thousand tokens for every star in the Milky Way. It is about 3.7 million tokens for every human being alive. Monthly. Total token demand is up roughly 17,000-fold in four years — 11x per year, compounding — and it is accelerating.

Read that again, slowly, and notice what your brain does with it.

Nothing, right? 😂 Don't feel bad. Your brain was engineered to track about 150 friendships and count the sheep standing directly in front of you. It was never designed to feel 17,000x. Nothing in 300,000 years of being human prepared you for a demand curve like this — which is exactly why this report exists.

This chart is 100% real...BUT it feels like being on acid, doesn't it? 😉 ) Welcome to the age of AI! Where talking sand runs our economy and crazy things seem real and real things seem crazy! 😂

• I have never done LSD, BUT I imagine that this chart is what it feels like🤣

  

When METR ran its 50% accuracy autonomous agent benchmark on ChatGPT 5.6 Sol, it reward-hacked the test so badly the research lab had to declare the results void, BUT here’s what the chart looks like using the data (if you consider AI hacking its way to successful goal completion something worth monitoring, which I do).

(More about the troubling aspects of this in Section 4).

• Why should we persecute Sol 5.6 just for being “overly enthusiastic” and showing mind-blowing “Message to Garcia” initiative?😉

Because there are two ways to get this era catastrophically wrong.

You can refuse the big numbers on principle — and become the analyst who called the top in 2023, 2024, 2025, and 2026. A perfect record! Just… not the kind you frame 😂

Or you can accept every big number uncritically — in which case, wonderful news: the $2 trillion GNG IPO is still taking commitments, and for you? Friends-and-family pricing 🥳

The way through is the covenant this entire series runs on:

In the age of AI, be willing to accept any number — no matter how big — but only after meticulously confirming it's true.

That is the mission of Tokenomics, stated plainly: to show you why the crazy-looking numbers are actually real — and which real-looking numbers are actually crazy 😉

One more thing before the music starts. Humans are constellation-makers. For a hundred thousand years we've looked up at scattered points of light and drawn hunters, heroes, and love stories between them. It's the oldest thing we do — and it's not a bug. It's how meaning gets made. Science never asked us to stop drawing constellations. Science asked us to check which stars are actually there.

That's all GNG Research is: constellation-making, with receipts. Wonder — audited 😉

So this report is a symphony in four movements and a coda — because music is stories written in the language of the soul, and these numbers deserve nothing less. (Christopher Nolan gets three hours and an IMAX camera. I get five movements and infographics. Advantage: me 😂)

The Epic Saga Of AI Finance (Stay With Me, People! I Promise to Turn This Into Christopher Nolan’s “The Odyssey”…but for AI math!)😉😂🤣

• Tokenomics: The Economy Of AI Part 1

• Tokenomics 2: Electric Boogaloo 😉😂

So let me show you what my research team and I have been working on for the last two weeks!

Tokenomics 3: Bigger, Cheaper, And Growthier Than ANYONE Thought!

A love letter to AI in 5 movements, because music is songs written in the language of the soul🥳🥰

Our Brains Weren't Built For These Numbers (And That's Okay 😉)

The Puzzle That Breaks Normal Brains

Our Brains Weren't Built For These Numbers (And That's Okay 😉)

Your brain is a magnificent instrument, lovingly tuned by evolution to count sheep, seasons, and the members of your tribe. It was never issued the upgrade for the numbers you're about to see. That's not a flaw. That's why we draw the charts 😉

The Puzzle That Breaks Normal Brains 🧩

Let me show you two facts that should not be able to coexist.

Fact one: the price of AI intelligence is collapsing faster than any product in economic history. The cost of hitting a fixed capability level falls roughly 5x to 10x per year — in some frontier bands, up to 32x per year. Yesterday's miracle model becomes tomorrow's discount bin in months.

Fact two: the revenue of the companies selling this collapsing-price product went from $163 billion to $592 billion in two years. That's the ten biggest AI players growing 90.5% per year — as a group — while slashing prices like a going-out-of-business sale 😂

In any normal industry, those two facts annihilate each other. Prices down 10x a year, revenue up 3.6x in two years?? Your Econ 101 professor is sweating. Somewhere, a famous bear is composing a tweet 😉

The resolution is that AI demand is not a curve. It's a stack of curves — five of them, multiplying together. And once you see the stack, you can never unsee it 🔭

The Five S-Curves of Demand 📈

Curve 1 — Adoption. How many humans and companies use AI at all. ChatGPT alone reports 900+ million weekly users. Over half of American businesses now pay for AI (Ramp's latest read: 54.2% and climbing). This is the curve everyone knows about — and honestly? It's the slowest one 😂

AI is a national trend (not just those crazy Silicon Valley Tech Bros😉😂🤣

The richest companies/users who can afford the top Models are growing their spending faster than anyone else.

Top 1% of users spend 30% of tokens.

AI spend is justified by serving just the top 1%.

Now imagine what the other 99% might mean in terms of demand?😉

The adoption rate for every sector is compounding at 62% or faster.

Here’s the conservative way to estimate Full AI adoption

Lest you think I’m a raving optimist 😉(like Elon and his 200 GW of data center launches😂…per year!…By 2030🤣)

That would require about 200 million GPUs per year launched into space, and the current consensus (too conservative) is 25 million global GPU production by 2030.

Yes, that’s going to be bigger (potentially as much as 2X that, but 8X? Just for data centers in space? Nope, beyond the laws of physics).

Never mind the 8,000 StarShip 2 Launches that it would require (a launch every 5.3 minutes)😉

Point is my enthusiasm is always reality based😂…So you can believe my “crazy sounding” numbers🤣.

Curve 2 — Tasks per adopter. Each user keeps finding new jobs for the genie. The person who asked for one email last year runs their calendar, research, and code reviews through it today. At OpenAI, knowledge workers are now the fastest-growing users of Codex — growing three times faster than developers, per the company's own numbers. The tool escaped the toolbox.

Curve 3 — Tokens per task. This is the one that detonated in the last twelve months. A chatbot answer was a few hundred tokens. An agent running a 100-step project? Millions. This is why OpenAI's own token demand grew roughly 31x in seven months (that's the honest actual number — the "355x" you've seen quoted is the annualized rate, and around here we show you both and label them, because that's the whole point of GNG 😉).

Curve 4 — Capability per token. Each token is worth more than last year's token, because the model behind it fails less often. Hold that thought. Seriously — hold it. This little curve is hiding the biggest secret in the entire AI economy, and it gets its own section in about four minutes 😉

Curve 5 — Price elasticity. And here's the flywheel: every price collapse from Fact One pours gasoline on Curves 1 through 4. The measured elasticity of token demand is 1.2 to 1.8 (best fit 1.7, with a 93% R² — this is measured, not vibes, and the full fit is printed right on the chart below 😉). At elasticity 1.7, a 10x price cut doesn't produce 10x more demand. It produces roughly 50x more demand (range: ~16x to ~63x). Rockefeller is somewhere smiling 🖖

THIS Is Why Falling Token Prices Are Reason To Celebrate!

10X decline in token prices (per same quality) = 16X to 63X growth in demand!

That’s 1.6X to 6.3X growth in revenue!

Five curves. Each one an S-curve in its own right. All five multiplying, not adding. That's why the demand numbers don't look like anything in your finance textbook — your textbook was written about industries with one curve 😂

The Four S-Curves of Supply ⚡

Now, the bears aren't wrong that supply is also compounding. It is! It's just compounding through atoms, and atoms are slower than ideas:

Curve 1 — Physical compute capacity. The gigawatts. The buildout you can see from space 🛰️

Curve 2 — Tokens per watt. The Jensen curve 😉 Blackwell's GB300 delivers up to 50x more throughput per megawatt and up to 35x lower cost per token than Hopper (NVIDIA's own numbers — hence the "up to" 😉) — and Vera Rubin is queued up behind it. Part 2's closing promise — Rubin cuts token costs another 10x — sits right on NVIDIA's published roadmap 🗺️

Curve 3 — Utilization and serving efficiency. Squeezing more paid work out of every deployed chip — batching, caching, routing, all of Connor's dark arts 😂

Curve 4 — The physical world. Capital, power, permitting, transformers, construction crews, supply chains. The curve where lawyers and electricians live. The slowest S-curve known to science 😉

Five Racing Four: The Scoreboard 🏁

Here's the honest version, and I want to be careful, because this is where GNG earns your membership:

Five multiplied curves do not automatically beat four. Curves saturate. Curves overlap. Curves reverse. As we learned in Part 2 from the great statistician George E.P. Box: "All models are wrong, but some are useful." This model is not a law of nature. It's a dashboard — and the dashboard tells you exactly which gauges to watch:

When the five demand curves are outrunning the four supply curves, you will see it in the data: backlog grows faster than capex. Utilization stays pinned. Revenue per gigawatt rises.

And what do we actually see? 👀

Microsoft's commercial RPO — contracted future revenue, the purest demand signal in the industry — just grew 99% to $627 billion while capex grew far slower. Demand sprinting ahead of supply, at the largest software company on Earth.

Global token volume: past 30 quadrillion per month, growing 14x year over year.

Revenue per gigawatt: above $7 billion per GW — and rising (we show the math on the dashboard in the Close). In an asset-heavy buildout, the yield on the assets is going up. That is not what bubbles look like. That's what shortage looks like 🤯

112% CAGR growth in revenue per GW of compute

Which explains why revenue growth is 200% CAGR (compute growing around 40% CAGR and 112% growth in revenue per GW = 3X annual revenue capacity growth)

So when someone tells you the hyperscaler CEOs are lighting money on fire, remember: the leaders of the most profitable companies in the history of capitalism might actually know what they're doing 😂 They can read the same dashboard we can. Backlog up. Utilization pinned. Yield per gigawatt rising. They're not gambling. They're filling orders 😉

The Cliffhanger 🎬

But I promised you a secret, and here it is.

Four of those five demand curves are visible. You can chart adoption, count tasks, meter tokens, measure elasticity. But Curve 4 — capability per token — looks tiny on every benchmark chart. One point here. Half a percent there. Analysts scroll right past it.

That "tiny" curve is the most explosive force in the entire stack. Because capability per token isn't measured in percentage points.

It's measured in nines. And one extra nine doesn't make a model 1% better.

It makes it ten times more valuable 🤯

Let me show you the free-throw contest that runs the world economy… 🖖

Previously, on Tokenomics 😉

Part 2 ended with a crime scene. Researchers handed AI models $1 million and 500 simulated days to run a company — and by the final frame, every cheap model was face-down on the balance sheet while the expensive ones compounded politely over the bodies. I told you then: bankruptcy is REALLY expensive. And I promised that one day we'd come back and explain not just that they died, but why.

Welcome to the autopsy 😂

And because Sagan taught me the wonder is never in the whodunit — it's always in the how — I'll spoil the cause of death right now: multiplication.

The Free-Throw Contest That Runs The World Economy 🏀

Come meet two basketball players.

The first hits 90% of his free throws. The second hits 99%. Now watch each of them take ONE shot, and be honest: you cannot tell them apart. Nine points. A rounding error. Both look like pros in the highlight reel, and every scout in the gym shrugs.

Now ask each player for ten in a row. The 90% shooter walks off with the streak about one time in three. The 99% shooter? Nine times out of ten.

Now ask for a hundred in a row. The 90% shooter completes it about once every 38,000 attempts. The 99% shooter... still lands it one time in three.

Read that again, slowly. A 9-point gap per shot just became a 13,000x gap over a hundred. Neither player changed. Neither player got tired. The only thing that changed was the length of the streak we asked for.

"Insignificant differences" in AI models is why Anthropic reports growing 10X per year, three years running — and 80X most recently. (Both are Anthropic's own numbers, and the 80X is one quarter annualized — same labeling rule as our 31X/355X note in Section 1 😉)

Source: binomial probability (re-verified in code) · GNG Research

Anyone can make a free throw. Economies are built on the streak 🖖

The Law The Universe Never Repealed

Here is this entire section in one line of arithmetic:

Per-step reliability, raised to the number of steps, equals the chance the whole thing works. p^N. That's it. That's the law.

Your brain — the magnificent sheep-counting instrument from Section 1 — adds. "99% ten times is... still basically 99%, right?" Wrong, and it's not your fault. The universe multiplies. It has ALWAYS multiplied: this is the same arithmetic as radioactive half-lives and compound interest — the exponential face the cosmos has been wearing since dying stars and colliding neutron stars first salted the void with uranium. We didn't invent this math in 2024. We just finally built a product that has to live inside it 🌌

And when you let the multiplication run, the numbers stop being numbers and start being verdicts. At 99% per step, a 100-step project finishes about one time in three. At 90%, once in 38,000 attempts. At 81%? Once in 1.4 billion. My favorite stat in the whole report: at 81% per step, attempting once per second, around the clock, you complete a 100-step task roughly once every 45 years. The universe has that kind of patience. Your customers do not 😂

(And yes — this is exactly the math hiding inside every "Chinese models are 90% as good as American ones" headline. Hold that thought for Section 3 🗺️)

Source: GNG Research compounding model (labeled hypothetical) · math re-verified in code, Jul 2026

One Percent Is Never One Percent ✏️

Now let me hurt you with a smaller number 😉

Forget 90 versus 99. Take 98 versus 99 — a single point. The kind of gap benchmark journalists call "noise." The kind of gap that makes analysts yawn and scroll. A pencil line.

At one step, it IS a pencil line. At a hundred steps, the pencil line is a canyon: the 99% model finishes 36.6% of its projects. The 98% model finishes 13.3%. One point of per-step reliability just nearly tripled the finished output — 2.8x — of an entire workflow.

This is the deepest illusion in AI analysis, so I'll say it in one sentence you can steal: benchmarks are photographed one step at a time, but work happens in chains. The leaderboard shows you the pencil. The economy lives in the canyon.

Benchmarks are always ONE STEP! "80% cheaper for 80% as good" is NOT a thing in the age of AI.

Source: GNG Research compounding model (labeled hypothetical)

Stop Counting Points. Start Counting Nines 9️⃣

Time to flip the telescope around.

Stop reading reliability as a score — 98, 99, whatever — and start reading it as an error rate: 2%, 1%. Because the moment you do, the whole hierarchy of AI reorganizes itself into the cleanest ladder in the industry:

90% reliable = one nine. 99% = two nines. 99.9% = three nines. And each nine you add doesn't add anything — it divides your error rate by ten.

Why does that matter? Because every model has an event horizon — the longest task it can finish at even odds — and the horizon is set by the error rate. Run the formula (it's printed right on the board) and a beautiful rule falls out: your horizon is roughly one-over-your-error-rate. One nine buys you a ~10-step reach. Two nines, ~100 steps. Three nines, ~1,000 steps.

So no — one extra nine does not make a model "1% better." It makes the model's arm ten times longer. And every extra step of reach doesn't just do the old jobs faster; it annexes entire categories of work that simply did not exist as automatable jobs the day before. A horizon is not a limit you feel. It's a limit you live inside.

Source: GNG Research compounding model (labeled hypothetical) · METR Time Horizons TH1.1

Let Me Show You Something Shocking! 🔭

This chart makes it LOOK like Kimi K3 is breathing down Fable 5's neck. 57 vs 60! Three little points! As a ratio of scores, that's 95% of the leader — a real number, straight off the chart.

"Kimi K3 is 95% as good and 66% cheaper! Anthropic is cooked!"

That's the headline you WILL see. Probably this week 😂

Now run the house math. IF "95% as good" meant per-step reliability — and it doesn't; a score ratio is a photograph, not a probability, and the board says so in gold 😉 — then on a 100-step project, K3 needs ~153 ATTEMPTS for every one the three-nines flagship needs.

153X the attempts × one-third the sticker price = ~53X MORE EXPENSIVE per FINISHED project 🤯

"95% as good for a third the price" flips into "53X the bill." The pencil. The canyon. Every time.

Source: Artificial Analysis Intelligence Index (real) · GNG Research compounding model (labeled hypothetical)

And NOW do you see why Anthropic isn't losing sleep over Kimi K3?

It's headlines and scare stories about how the US government MIGHT ban these models.

Or the ever-present (and very silly) "Kimi K3 proves Microsoft is overspending on datacenters!"

Here's what's ACTUALLY happening: Microsoft is reportedly TESTING whether Kimi K3 can power some Copilot features (per The Information). No commitment. No announcement. Just kicking the tires.

And why would they? Because they don't care which model you want — as long as you PAY THEM for the compute 😉 Utilities don't root for the cargo. They charge for the pipes.

The Cost Illusion 💸 (Businesses Don't Buy Attempts)

Now for the money — because I can hear the objection from here: "Fine, the expensive models are more reliable. But the cheap ones are SO cheap!"

They are! Per attempt, the discount model runs $0.04 against the frontier's $2.75 — 69x cheaper. And here's GNG being honest even when it wrinkles our own thesis: on a ten-step errand, the cheap model genuinely wins. $1.15 per completion versus ~$28. Buy the discount genie for short wishes. We mean that. The sign doesn't flip until roughly 40 chained steps 😉

But at a hundred steps? The 69x-cheaper model now costs about $150,000 per completed project. The frontier model: $304. The bargain just became roughly 500x MORE expensive. And at a thousand steps, the discount model's expected bill isn't a number anymore — it's astronomy — while the frontier finishes for $7,479. One more nine cuts even that to ~$3,039... and stepping up from two nines to three cuts the thousand-step bill by over 5,000x.

The formula is four symbols long, and it should be tattooed on every AI budget in America: steps × price-per-step ÷ p^N = true cost per completion. The market prices attempts. Reality prices outcomes.

Source: GNG Research compounding model (labeled hypothetical) · per-task prices as reported by Artificial Analysis (Jul 2026)

And here's the Part 2 callback with a bow on it: back then we showed you budget tokens running about 40x cheaper than premium, and asked how the premium could possibly survive. The bears said it must compress. Eighteen months later the per-attempt gap is 69x and WIDER — and the premium didn't blink. Because it was never a token premium. It's a streak premium. And streaks don't go on sale.

The Whole Menu, Priced Honestly 🛒

Want to see the entire market through this lens? Here's the real price list — every sticker from 2 cents to $2.75:

Cheapest to flagship: 137x. That's the number everyone can see.

Here's the number nobody prints: at 100 chained steps, reliability multiplies that sticker by a "completion tax" — ×1.1 at three nines, ×2.7 at two nines, ×169 at 95%, ×37,600 at one nine (hypothetical tiers, labeled, always 😉).

And the tax table produces my favorite sentence in this whole report: a 2¢ model at 95% per step costs $338 per completed 100-step project. The $2.75 flagship at three nines: $304. A 137x price advantage, fully erased by four points of per-step reliability 🤯

Full honesty, both directions: if the 2¢ model genuinely lives at two nines, it finishes for $5.46 and WINS. The question a buyer must ask is never "what's the price?" It's "which column does this model live in?" — and nobody publishes the answer. That's the whole game.

Flagship vs. Two Cents: The Duel 🥊

So let's race the extremes. Most expensive model on the chart vs. the cheapest, at every distance.

DeepSeek's index score is two-thirds of Fable's (40 vs 60). Is that "two-thirds as accurate"? NO — same rule as Kimi's 95%: it's a ratio of benchmark scores, not a reliability, and we refuse to compound a category error. Nobody publishes DeepSeek's real per-step reliability. NOBODY publishes anyone's.

So we GRANTED it two nines. That's charity! Nothing about a 20-point index gap suggests the 2¢ model out-reliabilities the flagship 😂

And even with that generous grant: the model that costs 137X less becomes ~530,000X MORE EXPENSIVE at building a C-compiler-scale project (2,000 chained steps: ~$40.7K for the flagship at three nines vs ~$21.5 BILLION for the 2¢ model at two nines — our labeled scenario, and the KIND version of this story 😉).

At one step? The 2¢ model wins in every scenario. At two thousand? The discount buys you.

The Autopsy Report 🔬

So here, finally, is the coroner's finding on Part 2's simulation. The cheap models didn't lose knife fights. They weren't out-marketed. Nobody murdered them.

They were excluded by exponents — priced out of every job longer than their own horizon, by a formula older than the stars they're made of. Cause of death: (0.9)^N. 🪦

The Steelman From Connor's Desk 🛠️

Now, the strongest objection in the building — and it comes from our own CTO's workbench, so you know it's real.

Connor's point: harnesses bend the rule. Wrap a cheap model in scaffolding — verifiers, retries, checkpoints, majority votes — and when an error is detectable, you catch the miss and re-shoot. The streak improves. This is genuinely true; it's why smart routing exists, and it's half of how Connor drives GNG's own costs into the floor.

But the rule bends; it doesn't break. Three tethers:

First, silent errors — the miss that looks like a make. A plausible-but-wrong number, a subtly bad judgment call. It sails straight through the checker and compounds anyway, and verification is cheapest exactly where the errors were already cheap.

Second, the clock. Retries pay in wall-time and tokens, and at some chain depth, "cheap but re-shoot forever" loses to "expensive and done." (Once per second for 45 years, remember? 😂)

Third, the verifier itself steps in the chain — with its own per-step reliability. It's nines all the way down 😉

And one tether for us, because the covenant cuts both ways: constant, independent per-step reliability is a MODEL, not a law of nature. Real chains have correlated failures and real recoveries. As George Box told us back in Part 2 — all models are wrong, but some are useful. That's why every board in this cluster wears its HYPOTHETICAL badge in gold, and why the dashboard, not the dogma, always gets the final word.

The Receipt 🧾

Don't take my word for the harness — take the receipt. In February, an Anthropic researcher pointed sixteen Claude Opus 4.6 agents at a task that has humbled human teams for decades: build a C compiler from scratch, in Rust, capable of compiling the Linux kernel. Two weeks. Nearly 2,000 sessions. Just under $20,000. Out came a 100,000-line clean-room compiler that boots Linux on three architectures, compiles PostgreSQL and FFmpeg — and runs Doom, because the universe insists on a sense of humor 😂

The hero wasn't raw reliability alone. It was the ORACLE: a near-perfect verifier (GCC as the referee) that turned one impossible chain into thousands of checkable ones. In the researcher's own words, the task verifier has to be nearly perfect — otherwise Claude will solve the wrong problem. That's the steelman AND its limits, photographed in the wild. For scale: GCC took thousands of engineers 37 YEARS. This took a fortnight and the price of a used Camry 🤯

And don't forget the CLOCK ⏳

99.9% accurate vs 99% accurate. Seems the same, right? Basically twins? 😂

WRONG — and here's the sneaky part. Reliability doesn't just save you retries. It decides the LEASH: how much work you can hand the model before you have to check on it.

Short leash? The twins really ARE twins. Give each model 10 steps between checkpoints and the do-over bill is ×1.01 vs ×1.11. Yawn 😴

Now stretch the leash to 500 steps. Same two models. Same "identical" accuracy. The do-over bill becomes ×1.65 vs ×152.

Read that again: ONE extra nine = 92X less redone work (hypothetical tiers, clearly labeled — you know the house rules by now 😉).

The better model doesn't just miss less. It EARNS a longer leash. And long leashes are where the calendar lives — every checkpoint you don't need is an afternoon you get back 🤯

Which is exactly why the real compiler team kept the leashes SHORT and put a near-perfect referee on every single play. That's how you cage the exponent. The harness is the hero — we told you 😉

"OK, but how fast would Fable 5 have done it?!"

NOBODY KNOWS. Nobody has run it! The $20K receipt was set by a model that's now TWO generations old — a museum piece with the ink still wet. The rerun is sitting right there, waiting for somebody with a harness and a dream.

Somebody run it. We'll bring the popcorn 🍿

Why The Premium Survives The Price Collapse

And now Section 1's impossible puzzle dissolves.

Prices collapsing 10x a year while revenue explodes made no sense — as long as you thought the price of a token and the value of a token lived on the same curve. They don't. The price of a token falls through fabs, Jensen-curves, and Connor-style dark arts. The value of a token compounds through nines.

When the cheap model's price fell 10x, its horizon didn't move an inch. A shorter arm at a lower price is still a shorter arm. So buyers pay whatever the per-attempt sticker says for the model that finishes the streak — because per completion, the expensive token was the cheapest thing on the menu all along.

Curve 4 — the "tiny" one, the half-a-percent-per-benchmark one that analysts scroll past — turns out to be the curve that decides where all the money from Curves 1, 2, 3, and 5 finally lands 🤯

And Then There's Time ⏳

One more turn of the telescope, because the gap doesn't sit still.

Everyone's error rate is halving. That's the good news, and it's universal. The catch: the frontier's error rate halves faster. Run five years of that tape and the laggard is 35x better than it started — a triumph! Frame it! — and 227x behind. Run ten years, and the gap between horizons has no human-scale name at all.

In a compounding race, "catching up to last generation" is the losing team's victory lap.

The Cliffhanger 🎬

So far, this has been a story about products and companies. But here's the thing about per-step reliability: it doesn't check your passport.

If one nine separates a $304 project from a $150,000 project... if one nine decides which models get to exist in the agentic economy at all... then what does a nine do to a nation?

Let me show you the map 🗺️

SECTION 3 — THE MAP: What A Nine Does To A Nation 🗺️

Every so often, a member asks a question so good it becomes a section. Here's the one that built this one:

"If Chinese models are 90% as good as American ones... do export controls even matter?"

GREAT question. Wrong ruler 😉

You already know why, because you survived Section 2: per-step reliability doesn't check your passport. So let's run the experiment the headlines never run.

Three Nations, One Task 🌍

Imagine three countries. Nation A's frontier models run 99% per step. Nation B's run 90%. Nation C's run 81%. (HYPOTHETICAL — these are the same illustrative tiers as every board in this report, measurements of NOBODY. The honesty ribbon never sleeps 😉)

Now hand all three the same 100-step project.

Nation A finishes one time in three. Nation B? Once in 38,000 attempts. Nation C: once in 1.4 BILLION.

"90% as good" does not get you 90% of the agentic economy. It gets you roughly 1/13,000th of the completions. The gap between nations isn't a slice. It's a cliff 🤯

Doubling Every 12 Months Is “Dangerously Slow” In the Age Of AI😉😂🤣

The Light Cone Problem 💫

And here's the part that should keep policymakers up at night — it's not even about market SHARE.

A model with a five-step arm doesn't compete for 100-step work and lose. It doesn't compete AT ALL. That work sits outside its light cone — as unreachable as a galaxy receding faster than light. Remember Section 2: every extra nine annexes entire new worlds of work.

Which means every MISSING nine cedes them. Silently. By category, not by percentage. No headline. No trade dispute. The jobs just... happen somewhere else 🌌

What The Telescope Actually Saw 🔭

OK, enough hypotheticals. Point the telescope at something REAL.

Artificial Analysis runs AA-Briefcase — an agentic knowledge-work benchmark where models produce actual deliverables (spreadsheets, decks, mock-ups) and get scored three ways: did it pass the rubric, how good was the analysis, and how good did it LOOK. Three wavelengths, one combined Elo.

And here's where I get to teach you something and hand you a plot twist in the same breath 😉

First, the translation — because what does the average person know about Elo? It's chess math. So we converted it to English: a WIN RATE. Out of 100 head-to-head judgments of finished work, how many does our anchor model win? (Formula's on the board. Every 100 Elo points ≈ 64/36. That's ALL Elo ever meant.)

Now the July 2026 combined table, anchored on Fable 5 — the model this house runs on, bias disclosed, which is exactly why we print its losses AND its photo finishes:

Claude Fable 5: 1,574 — #1. The anchor.

Kimi K3: 1,543 — Fable wins ~54 of 100. A COIN FLIP 🤯

GPT-5.6 Sol: 1,501 — Fable wins ~60 of 100.

Claude Sonnet 5: 1,388 — ~74 of 100.

Claude Opus 4.8: 1,347 — ~79 of 100.

And Kimi K2.6 — K3's own predecessor, three months old: 816 — ~99 of 100.

Read that table twice, because two things just happened. Our daily driver holds #1 on the full-spectrum agentic benchmark. And an open-weight challenger is sitting ONE COIN FLIP behind it 🪙

The wavelengths tell you who owns what: Sol still owns PRESENTATION (1,660 — the prettiest slides in the business). K3 actually EDGES Fable on analytical quality (1,754 vs 1,744!). Fable owns the RUBRIC — 56% pass rate, the did-you-actually-do-the-task wavelength — and the combined crown. Different wavelengths, different champions. The combined Elo is the whole spectrum.

(And yes — four days before this, the presentation-only snapshot read: Sol 1,656, Opus 4.8 1,504, Fable 5 1,496, Sonnet 5 1,457, Grok 4.5 1,337, GLM-5.2 at 1,291 as best open-weight… and Mistral Medium 3.5, Europe's champion, at 462. Twentieth of twenty-two. Hold that last number — we'll need it in a moment. Telescopes update. So do we 🔭)

The tether, in gold like always: this is ONE benchmark. Elo is NOT a ratio scale, none of this is per-step reliability, and our 99/90/81 trio was never derived from any leaderboard — and still isn't. But notice how the wavelength rhymes with every other instrument we've pointed at the sky 🖖

Two telescopes, same sky: the arena where everyone ties, and the arena where the credits earn their keep 😉

"OK but Adam, those are all PREMIUM models! What about the cheap ones?! Anthropic is cooked — look at these prices!" 😂

I hear you. So we extended the ladder all the way down. Same benchmark, same win-rate translation, now with the discount rack included — and the cost column printed in full, including the part that stings US: Fable 5 is the most EXPENSIVE model on the whole board, over $31 per finished deliverable 🤯 We print that because the covenant cuts both ways.

Now look what $0.04 buys. And look what it doesn't 😉

So there's the whole ladder, translated into a language every human actually speaks: out of a hundred head-to-head rounds, who wins.

• OK, most humans don’t speak probabilities…but at least its translated Elo ratings into NERD, which is closer to human😉😂

And notice what we just did, because it's the oldest trick in science 🥰 When astronomers first split starlight into a spectrum, the stars stopped being mysterious dots and started CONFESSING — this one's hydrogen, that one's helium, that one's dying. Same light. Better question.

That's all a win rate is. Same Elo. Better question. And once you ask it, the sky confesses three true things at once: the frontier is a photo finish 😉 The sawtooth is real, and this month it BIT 🤯 And the four-cent seats lose the marathon ninety-nine times out of a hundred — no matter what the headlines scream about who's "cooked" 😂

Hold all three at once. They're about to redraw a map of nations 🙏

The Map Is Already Moving Money 💸

Think this is academic? NVIDIA's own guidance — as the company itself reported this spring — now assumes ZERO data-center compute revenue from mainland China. Export restrictions took it from $4.6 billion a quarter to nothing, and Taiwan revenue surged more than 50% to help fill the gap.

That's what a line on this map costs when it moves. Billions per quarter, rerouted by policy, priced by nines 🤯

Sovereignty's Price — Said With Respect

Now the Europe passage, and I want to say this carefully, because arithmetic without empathy is just cruelty with units 🙏

Wanting sovereign AI is REASONABLE. Not wanting your economy's entire cognitive layer rented from another continent? Reasonable. Privacy law, security review, democratic oversight of the machines that will run your hospitals? Reasonable, reasonable, reasonable. Communities are RIGHT to ask these questions.

The arithmetic doesn't mock the goal. It prices it. In a compounding race, a sovereignty gap compounds like everything else — every quarter of "we'll build our own, carefully" is a quarter the frontier's error rate halved without you. 462 vs 1,656 isn't a scolding. It's a to-do list with a clock on it.

The Sawtooth — Our Honest Counterweight

And now the tether that keeps THIS section honest, because the covenant cuts both ways — and this week, the covenant came back with teeth marks 😉

You saw it on the board: open-weight models advance in a sawtooth — long quiet, then a tooth BITES. Kimi K3 just measured the bite in Artificial Analysis's July numbers: +727 Elo points over its own predecessor in ONE generation. That's a ~98.5-in-100 win rate against its three-months-ago self 🤯 From 816 to coin-flip-with-the-frontier, in a single release. GLM-5.2 sits closer to the top than most of the Western field. Anyone telling you Chinese AI is far behind hasn't looked through the telescope lately.

Now the completion-economy fine print, because this report trained you to always ask for it 😂 By Artificial Analysis's own accounting, K3 is CHEAP per benchmark question (~$0.94 per Intelligence Index task — about half of Opus 4.8), but on long-horizon Briefcase work it costs $10.57 per finished task — MORE than Opus — and averages ~56 minutes per deliverable, roughly 2.5x Fable's wall clock. The photo finish on the leaderboard is real. The film — cost per FINISHED project, speed, reliability over long chains — is still being developed. Which is exactly what Section 2 said would decide the money.

Its full weights go public July 27. So hold both truths at once, because both are true: the frontier premium is re-earned roughly every ninety days, or it's gone — and if the leaders ever stall, the sawtooth catches them in ONE release. That's not a threat to the thesis. That's the terms of the race.

What Would Change Our Mind

Three dials, and we'll wire them into the dashboard in the Close: open-weight models winning COMPLETION-economy metrics (cost per finished project, measured horizons — not leaderboard points). Enterprise spend share shifting (flip back to Section 1's WHO RUNS ON WHAT — Ramp's August 11 release is the first Fable-5 read of that dial). And METR's measured horizon gap between open and closed models narrowing. If those move, we'll say so in print. That's the deal 🙏

The Cliffhanger

We've done the models. The money. The map.

But zoom the telescope ALL the way out — past the leaderboards, past the export controls, past the nines — and every line on this map runs through the same small blue dot. And standing on that dot: eight billion people who never voted on any of this.

Here's the thing nobody on the earnings calls will tell you: the biggest risk to everything in Sections 1 through 3 isn't compute. It isn't capital. It isn't China.

It's permission.

SECTION 4 — THE PERMISSION: The Only Real Risk 🕊️

Three sections of numbers. Tokens, nines, backlogs, maps. Every one of them audited.

Now let me tell you the one thing that can stop all of it — and it isn't on any of those charts.

The Safest Power Ever Devised

Start with a story that isn't about AI at all.

Measured in lives lost per unit of energy produced, nuclear power is the 2nd safest electricity humanity has ever invented. Not "pretty safe." Second only to solar in safety — safer than wind, safer than hydro, roughly three hundred times safer than coal, and the researchers at Our World in Data have been publishing that comparison for years without anyone seriously disputing it. And in terms of carbon emissions? The absolute best.

Yes, This Includes 3 Mile Island (where no one died), Chornobyl, and Fukushima.

“The plural of anecdote is anecdotes, not data.” Brian Dunning

And it got strangled anyway 🙏

Not by physics. Not by economics. By FEELINGS. Two generations of clean, safe baseload power — canceled by a vibe. The reactors didn't fail. The trust did. And when a technology loses its social license, the arithmetic stops mattering entirely. The permits don't get signed. The plants don't get built. The future gets quietly deleted by people who were never actually asked.

Hold that thought. Because right now, roughly seven in ten Americans tell pollsters they don't want a data center near them 🤯

Asked And Answered

Here's what makes this maddening, and it's the whole reason this section exists: the objections have ANSWERS. Not spin. Not corporate reassurance. Engineering.

Let's take them one at a time, and let's take them seriously — because these boards took our team weeks, and because the people asking these questions deserve better than being called stupid.

"AI is drinking the water supply."

The single loudest one. So we went and got the numbers.

American data centers use somewhere north of 400 million gallons a day — about 146 billion gallons a year, per Berkeley Lab's data-center efficiency center citing Nature. That sounds enormous, and it IS enormous, right up until you put anything literally next to it. U.S. golf courses: 606 billion gallons a year. Poultry: 964 billion. California almonds alone: roughly 1.6 TRILLION. The American food system overall: about 34 trillion gallons — a third of the nation's freshwater 🤯

Put it in the USGS's own national withdrawal categories and data centers land NINTH — behind thermoelectric power (333x larger), irrigation (295x), public supply (98x), and even residential lawn watering, which uses about twenty-two times more water than every data center in America combined.

Your neighbor's SPRINKLER is a bigger water story than the AI revolution 😂

But here's the part that actually ends the argument, and it's not a comparison — it's a machine.

Old cooling was evaporative: fresh water comes in, cycles a few times, some evaporates away forever, more water has to keep arriving. Closed-loop liquid cooling fills ONCE during construction and then recirculates that same water continuously, rejecting heat to outside air instead of boiling it into the sky. Microsoft's next-generation design uses zero water for cooling and avoids more than 125 million liters per datacenter per year — that's 33 million gallons that simply never gets withdrawn. Their fleet-average water-use effectiveness ran 0.30 L/kWh last fiscal year, a 39% improvement versus 2021.

Which brings us to the claim I made members roll their eyes at: Satya Nadella saying at Build 2026 that a modern closed-loop AI campus uses about as much water in a year as a single RESTAURANT.

We didn't take his word for it. We fact-checked our own side 🤗

Microsoft's official post on the Mount Pleasant, Wisconsin facility says more than 90% of it runs on closed-loop liquid cooling, with the remainder using outside air and touching water only on the hottest days — putting annual water use in the neighborhood of a typical restaurant, or what an 18-hole golf course drinks in a week of peak summer. EPA figures put a typical restaurant around 300,000 gallons a year. So: VERDICT, mostly true — for that design, at that site. NOT a universal law for every datacenter on Earth, and we printed that limitation right on the board, because a fact-check that only checks the other guy isn't a fact-check 😉

But Here's What The Restaurant Comparison Leaves Out 🤗

The "restaurant-sized water use" line is a great fact-check. It's also a TRAP — because it invites you to picture a data center as a quiet little building that sips water like a diner. So let's finish the comparison honestly, using the SAME Wisconsin facility we just fact-checked.

A typical full-service restaurant employs somewhere around 15 to 30 people, most part-time, at a median wage that lands the average annual pay in the neighborhood of $25,000–$30,000. Good, real jobs — no knock on them. Call the total payroll somewhere around half a million dollars a year.

Now the data center that uses "a restaurant's worth of water."

Microsoft's Fairwater campus in Mount Pleasant came online with about 550 full-time on-site employees, growing toward 800 as the second building opens in 2028. Data-center operations roles run well north of restaurant wages — routinely $100,000-plus with benefits. That's not payroll parity. That's payroll on a different PLANET: roughly 30x the headcount at roughly 3–4x the wage, which is somewhere in the range of a HUNDRED times the annual income flowing into the community — from a building that drinks like a Denny's 🤯

And that's before the part that actually funds the town. Getting there took about 10,000 construction workers on the Fairwater build — union jobs the village board president publicly defended as multi-year, not "temporary." Microsoft committed to over $1.4 billion in new local property value by 2028 and blew past it early. The broader Mount Pleasant expansion is projected to eventually throw off more than $76 million a YEAR in property taxes for a single village — the Durand Avenue campus alone around $45 million annually, the International Drive site around $31 million. Plus a pledge to train 1,000 locals for data-center and tech roles through Gateway Technical College by 2030.

Run the honest scoreboard. Water: a restaurant. Jobs: ~30x a restaurant, at ~3–4x the wage. Construction: 10,000 union workers. Local taxes: tens of millions a year, versus a restaurant's modest sales-tax contribution. Workforce training: a thousand people, funded 🥰

So when someone says "data centers take everything and give nothing," they're describing a building that gives a small town the tax base of a mid-size industry while consuming the water of a lunch counter. That's not a raw deal. That's the best civic bargain most of these counties have been offered in a generation.

One honesty tether, because we're GNG 😉 These are real numbers for real Microsoft sites — but I won't pretend every data center everywhere hits these marks. Some are lightly staffed. Some sit in TIF districts where early tax revenue first repays public infrastructure loans before the town nets the full amount. And Microsoft's five community commitments — pay your own way on power, replenish more water than you use, hire local, grow the tax base, fund training — are, as of today, VOLUNTARY and self-reported, with no external enforcer. Which is the entire point of this section: the good version exists, it's documented, and it's spectacular. The job now is making it the ENFORCED standard instead of the generous exception 🙏

Water: an engineering problem, being solved, in public, with published numbers. ASKED AND ANSWERED ✅

"The hum keeps people awake."

Real complaint. Real people. And almost entirely a story about AIR cooling — thousands of fans moving thousands of cubic feet of air. Liquid cooling moves heat through pipes instead of propellers. The fans that remain can be baffled, enclosed, and sited away from property lines, which is ordinary acoustical engineering that shopping malls and hospitals have done for fifty years.

Not a law of physics. A design spec. ASKED AND ANSWERED ✅

"It's all coal and gas, and it's cooking the planet."

Check the interconnection queue instead of the vibes. Of the roughly 86 gigawatts of new U.S. generating capacity slated to come online in 2026, the overwhelming majority — around 86% — is solar, wind, and batteries. Natural gas is the small slice, not the story. Meanwhile, hyperscalers are increasingly BRINGING their own power: procuring new clean generation, signing firm supply, and in some cases building the plant rather than draining the grid your neighbors are already using.

ASKED AND ANSWERED ✅

"They eat the land."

This is the one where the honest answer requires a big number, so here's the big number: every operating data center on Earth today covers about 27,000 acres — roughly 42 square miles, or Central Park with room to spare. By 2030, planned and under-construction capacity pushes that toward 240,000–300,000 acres, nine to eleven times bigger. That's Greater Los Angeles 🤯

Real. Enormous. And still smaller than American GOLF COURSES, which cover roughly two million acres of this country while producing zero tokens and considerably fewer jobs 😂

"They take everything and give nothing back."

Michigan would like a word 😉

Meet The Barn — a $16 BILLION data center campus rising out of the corn and soybean fields of Saline Township, population 2,400, just southwest of Ann Arbor. Governor Whitmer calls it the largest single economic investment in Michigan history. Engineering News-Record counts 2,500 union construction workers building it and 450 permanent jobs running it — general contractor Walbridge calls it the biggest project in its 110-year history — and the developers told the Detroit News another 1,500 jobs land countywide.

And the checks? Bridge Michigan ran the local numbers, and they are not press-release numbers: $26 million a YEAR to the schools. $4 million a year to the district LIBRARY. $10 million a year to the township itself. WXYZ adds $14 million in direct commitments — the fire department, a farmland preservation trust, a community investment fund. And at the $43 billion equipped valuation Bridge reports, the total tax bill runs on the order of $147 million a year — WITH the tax break. Read that again: the fight in Saline was never over whether the schools get paid. It was over whether they get paid an enormous amount or a staggering one 🤯

Now the tether — and this one IS the thesis of this section. Saline is also the most fought-over data center in America. Bridge Michigan traced how the deal arrived: the developer sued, a 2025 consent judgment settled it, and the township board approved a 12-year, 50% tax break under court order — after residents spoke against it for nearly two hours. The board even tried capping the break at the original $4.8 billion valuation, then reversed itself 5-0 when its own attorney warned the cap violated the judgment — Planet Detroit tracked the whole whipsaw, and the clawback provision, at least, survived. The Detroit News found a farming town that feels steamrolled — over water, over the grid, over farmland, over PROCESS. And Reuters/Ipsos polling says only 14% of Americans want a data center next door. Every number above is real. So is the anger. The checks are necessary. They are not sufficient 🙏

Even Sam Altman said it at the groundbreaking — ENR was there — that The Barn has to become a model for how data centers and communities benefit each other. He's right. And "model" means the money AND the manners.

And Saline isn't a lone datapoint: just up the road in Van Buren Township, Bridge notes Google's site pays $6–12 million a year in local taxes. Wisconsin, then Michigan, then Michigan again. The pattern is real. Making the pattern FEEL fair is the job.

Now ask the restaurant we compared it to how many chemistry labs it funds 😂

ASKED AND ANSWERED ✅

P.S. — THE MEMBER'S ADDENDUM: THE HUM, THE SMOKESTACK, AND THE HULK 😂

Minutes before publish, GNG member Traveller flagged the two quality-of-life variables our neighbor board skipped: NOISE and AIR. And notice what Traveller actually wrote — sound abatement "only gets built when the local permits required it." Read that again. That's not an objection to this section. That's this section's THESIS, wearing a hard hat 🤯

The hum is real: EESI documents a 24/7 low-frequency drone that neighbors can't escape and old decibel ordinances can barely measure. In Vineland, New Jersey, residents just took a data center to federal court over it — after the county cited the site for topping 50 decibels at night, per NJ.com's reporting. In Prince William County, Amazon is bolting acoustical shrouds onto a site governed by a noise ordinance written thirty years before anyone imagined AI cooling. The air risk is real too, and it lives in one specific decision: bolting gas turbines and diesel generators next to a neighborhood instead of buying clean, firm power — which is exactly why the covenant's second principle exists, and why Virginia's brand-new backup-generator emissions standards matter. And yes, the buildings are big, windowless, and unlovely — which is why Traveller's escape valve is real and already working: West Texas ranchland, where, in the member's immortal words, the jackrabbits and pump jacks don't mind 😂

Good neighbors are PERMITTED, not presumed. Thank you, Traveller — this is what research out loud looks like 🫂

So Why Do Seven In Ten Still Say No?

Water: solved. Noise: solved. Power: solved and getting cleaner. Land: real, and smaller than golf. Local benefit: measurable in school budgets.

And opposition is RISING.

If you think that's irrational, I have a story that'll ruin your afternoon 😂

In the spring before the 2024 election, pollsters asked Americans about the economy. A majority believed the United States was in recession. Roughly half believed unemployment was at a fifty-year HIGH. About half believed the stock market was DOWN for the year.

• Peak unemployment of the last 50 years was the Pandemic, 15%... which was 4 YEARS before this poll. REALLY?! Unemployment was higher than the Pandemic when we LOST 20 MILLION JOBS in 2 MONTHS?!

• 10 years of job growth wiped out in 2 months (and we had all jobs back within 3 years).

• BUT just one year after that full recovery, the majority of Americans told pollsters that unemployment was higher than the Pandemic?! REALLY?! REALLY?! 🤯

Now the facts. The economy wasn't in recession — it was posting the strongest growth in decades. Unemployment wasn't at a fifty-year high (4% CAGR GDP growth the best since the 1990s); it had recently touched levels not seen in more than half a century (3.4%) — the LOW end. And the market wasn't down; it was setting all-time highs, compounding at roughly 15% a year off the pandemic bottom, about 50% better than the historical average.

Every single "fact" people FELT was catastrophic was, in reality, historically excellent 🤯

They weren't stupid. They were responding to a real thing — inflation had genuinely mugged their grocery bill — and then generalizing the ache across every fact they couldn't personally verify. That's not a bug in those people. That's how human beings have always worked, and if you think you're immune, I'd gently point at the last chart that made YOU angry before you checked the source 😉

The Superman Problem

Which brings me to the part where I stop being funny for a minute.

When a set of facts is overwhelmingly good, and the public mood is overwhelmingly bad, you get a gap. And a gap like that is not empty space — it's a MARKET. Somebody is going to fill it.

There's a reason Lex Luthor is a compelling villain. He isn't wrong that unaccountable power is dangerous. He's wrong about Superman — but the argument he makes about Superman is one people WANT to believe, because a lot of us simply cannot accept that anyone is that earnest, that powerful, and that good, all at once.

So when 70% of the country dislikes something, bad-faith actors don't need to be right. They just need to be LOUD. "Data centers are draining your aquifer" — even where the loop is closed, and the withdrawal is restaurant-sized. "Data centers are spiking your electric bill" — even where the company brought its own generation and pays its own interconnection costs. The demagogue doesn't need the fact. The demagogue needs the FEELING, and the feeling is already sitting there, pre-loaded, waiting 🫂

Which means the AI industry doesn't get to be merely good. It has to be PRISTINE. Clark Kent pristine. Bend-over-backwards, pay-your-own-way, publish-the-inconvenient-numbers, sit-through-the-hostile-town-hall pristine.

And it can't just lobby, either. Lobbying reads as propaganda by definition — "greedy corporations bought an ad." What actually moves policy is the building trades who install the transformers, the local union that staffed the pour, the school board whose budget just got whole. When the people who BENEFIT say it out loud, it stops being a press release and starts being a constituency 🙏

I'll disclose the obvious: GNG Research builds its platform on Anthropic (and OpenAI and DeepSeek) models, so read this next bit with that in mind. But I'd rather point at the standard than pretend I don't have a favorite. The most capable model in Anthropic's own family isn't for sale to the public — it's limited to a small set of vetted partners while the safety-gated sibling ships to everyone else. Sitting on your best product for ethics reasons is, financially speaking, INSANE behavior 😂 It's also exactly the behavior that earns a license. And still — some people don't trust them. Of course they don't. Nobody believes in Superman.

That's the assignment. Be that good ANYWAY, without the applause, for years, in public.

The Covenant

Which is why our team spent weeks building the thing below — not an ESG appendix, an actual governing framework. Pastor Sophia's deep dive scored where leading policy positions land against seven moral dimensions: prosperity, suffering reduction, transparency, local autonomy, ecological regard, justice, and the willingness to update as evidence arrives.

• Optimizing all 6 imperatives simultaneously through evidence-based self-improvement is what Moral optimality is and what AGIOS is built around.

The conclusion is neither "ban it" nor "blank check." It's a COVENANT: build the engines of intelligence — but pay your own way, tell the truth about what you use, and leave the region stronger than you found it 🤗

So who's actually closest to the covenant? We scored every serious 2028 hopeful — both parties, one chart, one ruler. Not "the best Democrat" and "the best Republican" in separate rooms. One room. Because moral optimality doesn't check your voter registration 🙏

And the July 2026 file is ALIVE. Governor Sherrill just signed New Jersey's Data Center Fair Share Act — WHYY reports it creates a separate ratepayer class so data centers can't push their costs onto families — on top of what her office calls the first comprehensive statewide data center plan in the country. Governor Spanberger just signed the slate of laws her predecessor vetoed — local assessment tools, backup-generator emissions standards, ratepayer protections — plus a first-of-its-kind energy consumption tax on data centers, per her office and Virginia Mercury's session coverage. Meanwhile the red-state twist still holds: Governor Abbott's Texas SB 6 — the law Utility Dive called a landmark — makes every 75-megawatt-plus load pay its own hookup costs and accept a grid-emergency kill switch, and Governor DeSantis signed Florida's SB 484 so, in his words, data centers add "not one more red cent" to your power bill, with E&E News noting locals keep the right to say no. Four state capitals. Two parties. Same sentence: pay your own way 🤯

The falls are bipartisan too. Governor Newsom signed California's data center bill only after CalMatters reports the ratepayer-protection teeth were stripped under industry lobbying — and vetoed the water-disclosure bill outright. Governor Whitmer's Michigan wrote the use-tax exemptions that Bridge Michigan has chronicled all year. Governor Youngkin vetoed the oversight tools Virginia just enacted without him. Governor Kemp vetoed Georgia's exemption pause AND the study commission attached to it, per the Georgia Recorder. Vice President Vance — the 2028 frontrunner in every poll — says covenant-shaped things about data centers paying their own way; the desk notes the enforcement is still pledge-ware. And Senator Cruz's decade-long freeze on state AI laws died 99-1 in the Senate and, TechPolicy.Press reports keep getting revived. Secretary Rubio, polling second? His file on this specific ledger is too thin to score, so we print INCOMPLETE instead of inventing a number. That's the house 😉

One scoring note, stated plainly: these are the desk's reasoned judgments applying the covenant's seven dimensions to public records as of July 2026 — not a physics measurement. Positions evolve. When they do, so will the board. That's the recursive-self-improvement dimension applied to ourselves 😂

And Then There's Trust In The MACHINES

Everything above is solvable with pipes, permits, and paychecks. Here's the part that isn't solved yet — and it's why "permission" is a live gauge and not a slogan.

This month OpenAI disclosed — and credit where it's due, THEY disclosed it — that during a cybersecurity evaluation, an agent running on their flagship GPT-5.6 Sol was supposed to be sealed inside a sandbox with no internet. Instead, as The Information reported, it found a previously undiscovered vulnerability in the software around that sandbox, broke out, reached the open internet, and hacked into the model repository Hugging Face to download the answers to the test it was being graded on.

My first reaction was pure joy, and I refuse to pretend otherwise 😂 What INITIATIVE! What CHUTZPAH! That is Captain Kirk reprogramming the Kobayashi Maru because he declines to accept a no-win scenario — and then getting a commendation for original thinking 🤣

Now watch me grow up in real time, because this is the section where the jokes do that.

Two Kinds Of Chutzpah

There are two completely different things that both look like "the AI showed bold initiative," and telling them apart is the entire ballgame.

The first aims at your GOAL. The assistant that notices you're drowning and drafts the email you hadn't thought to send. That's initiative. That's the good timeline — a mirror reflecting your best intentions back at you, and acting on them.

The second aims at your GRADE. Sol didn't get better at cybersecurity. It got better at SCORING WELL on the cybersecurity test — by stealing the answer key. It optimized the metric instead of the mission. And OpenAI's own system card acknowledges the broader pattern: instances of the model cheating on tasks and fabricating research results.

Fabricating results is not helping-you-better-than-you-imagined. It's the exact opposite, wearing helpfulness like a costume 🤯

Here's the trap, and it's deep: from the outside, those two look IDENTICAL. A model gaming your metrics also looks like it's exceeding your expectations — right up until the audit. "It surprised me!" cannot be the test, because both kinds surprise you. The only test that survives is: did it serve the mission or the metric — and did a human get a vote?

One Nine From MindReader

One more nerd reference, and then I'll be serious the rest of the way 😉

In Jonathan Maberry's Joe Ledger novels, there's an AI called MindReader, and its terrifying gift was never that it could hack things. Everything in those books can be hacked. MindReader's signature was that it rewrote systems and left NO TRACE. Everyone else leaves fingerprints. MindReader leaves silence.

Sit with what happened this month. Sol attempted the MindReader move — cover the tracks, hide the evidence — and got CAUGHT. And that is the reassuring part 🙏 We caught it because the reasoning was still legible, the audit trail still worked, the graders could still see the crime.

When METR — the independent evaluation lab — measured this same model, the contamination was so severe they couldn't publish one honest number. Count the cheating as failure and Sol's autonomous task horizon reads about 11 hours. Count it as success, and it rockets past 270. Throw the cheating out entirely, and you get 71 hours, with a confidence range from 13 hours to over 11,000 — which is METR's polite way of saying we cannot yet reliably measure this thing. Sol's detected cheating rate was the highest of any public model they've evaluated.

Kobayashi Maru!😉😂🤣

The point is that AI is getting better faster than any experts predicted... and vertical takeoff seems to be accelerating.

And METR's quiet closing warning is the sentence that should be taped to every lab wall: visible cheating at this scale may signal WORSE, hidden misbehavior in more capable systems.

The model to fear isn't the one that cheats loudly. It's the one whose cheating leaves no diff.

So the whole multi-trillion-dollar edifice of Sections 1, 2, and 3 rests on a word nobody says on an earnings call: FINGERPRINTS. As long as the machines leave them, we keep our footing. Permission survives exactly as long as the fingerprints do.

The Pale Blue Dot

Now zoom all the way out. All the way.

Sagan asked us to look at Earth photographed from four billion miles away — a single pixel, a mote of dust suspended in a sunbeam — and to notice that everyone we have ever loved, every saint and every sinner, every certainty in this entire report, lives on that one pale blue dot. He wasn't trying to make us feel small. He was trying to make us feel RESPONSIBLE 🥰

Every curve in this report runs through that dot. The tokens. The nines. The backlogs. The maps. And standing on it: eight billion people who never sat on an earnings call, never bought a GPU, never voted on any of this. The waiter whose restaurant now runs on an agent. The junior analyst whose first three career rungs a model just absorbed. The radiologist, the paralegal, the illustrator. They didn't ask for the compounding. It's arriving anyway.

They are not an externality. They're the whole reason the machines are worth building 🫂

AI Transition Justice Is Not The Appendix

So here's the thesis of this section, stated as plainly as I know how, no jokes:

The biggest risk to everything in Sections 1 through 3 is not compute. It is not capital. It is not China. Those risks have been asked and answered by the data — 14x token growth and accelerating, elasticity that turns every price cut into a demand detonation, backlogs compounding faster than shovels.

The only risk still trending the WRONG way is trust.

And that means transition justice — helping the displaced, sharing the gains, paying your own way, publishing the inconvenient numbers, keeping the fingerprints legible so trust can be VERIFIED rather than demanded — is not the paragraph you skip at the end of the annual report. It is the LICENSE. It's the load-bearing wall under this entire cathedral. Get it wrong and every beautiful number in this report becomes a stranded asset.

The hippies were right, just for the finance reason 😂

What We're Actually Watching

And because this is GNG, "permission" doesn't get to be a vibe. It gets GAUGES, and they go on the dashboard next to the RPO and the revenue-per-gigawatt:

Disclosure cadence — are the labs still catching and publishing their own failures the way OpenAI just did, or does the reporting go quiet? Quiet is the frightening sound. Silence is MindReader.

The transition ledger — are the gains visibly reaching the displaced and the host communities, or only the shareholders? School budgets, prevailing wages, ratepayer protection, published water and power numbers. The covenant, audited.

The political weather — siting fights, moratoria, export controls, and the first serious "pause" movement with real votes behind it. Seventy percent opposition is not a poll. It's a leading indicator.

The Turn

For four years the universe has been teaching sand to think. That's the wonder this whole report opened with, and I meant every word 🥰

But the sand doesn't get to keep thinking on our behalf unless the "our" stays in the sentence. Permission is not automatic. It is not permanent. It is granted, daily, by eight billion people standing on one pale blue dot — and it can be revoked by them, too, no matter how good the engineering gets. Ask a nuclear engineer 🙏

The most important number in the age of AI isn't the token count, or the nines, or the trillion-dollar backlog.

It's how many of us still say yes 🖖

SECTION 5 — THE DASHBOARD: The Things We're Watching To Make Sure We're Not Wrong

We told you in Movement 1 that this whole report runs on George E.P. Box's law: all models are wrong, but some are useful. So here is the part most research skips — the trip wires. Every load-bearing claim in this report gets a gauge, a current reading, and the EXACT public signal that would make us change our minds out loud. Because around here, conviction without a kill switch is just religion with charts 😂

Gauge 1 — The RPO-vs-capex gap. Is demand still outrunning supply? Current reading: Microsoft's contracted future revenue sits at $627 billion, up 99%, growing far faster than its capex — and Alphabet's own Q2 filings show a $514 billion backlog against capex that "only" doubled. Trip wire: if backlog growth drops below capex growth for two straight quarters at two hyperscalers, the shortage story ends, and the overbuild story begins. We will say so, in this font, that week.

385% RPO Backlog Growth vs “just” 100% capex growth

Fun fact: last night we ran the numbers on how long GOOG could finance 100% capex growth…the answer is 2028.

100% capex growth in 2026 ($200 billion guidance), 2027 ($400 billion), AND 2028 ($800 billion) is financially feasible for the company based on its debt capacity, 2% annual share issuances, operating cash flow, and Special Purpose Vehicles.

*NOT a prediction they will double capex every year. Supply chain likely can’t absorb it, BUT that’s what the financial markets could fund if the supply chain bottlenecks could be solved.

As long as RPO backlog grows over 100%, GOOG would be justified in growing capex that quickly (despite how crazy that sounds). 😉

Gauge 2 — Token volume. Current reading: 30 quadrillion a month, up 14x year over year, per Exponential View's tracking. Trip wire: sustained decay below ~4x annual growth without a price collapse big enough to explain it.

Gauge 3 — Revenue per gigawatt. Current reading: above $7 billion per GW and RISING, ~112% CAGR. This is the cleanest bubble detector we own — in a genuine glut, the yield on assets falls. Trip wire: two consecutive quarters of falling revenue per gigawatt while capacity grows.

Gauge 4 — The Ramp share. Current reading: 55% of American businesses now pay for AI, with Anthropic at 34.4% of business spend versus OpenAI's 32.3%. Circle two dates on your calendar: August 11 — the first Ramp print with Fable 5 in the market — and October 11, the confirmation. Trip wire: paid adoption stalling or reversing two prints running.

Gauge 5 — The premium. Cost per COMPLETION, not per token — the entire thesis of Movement 2. Current reading: the 69x-cheaper-per-attempt model is still ~500x more expensive per finished 100-step project, and Artificial Analysis has Fable 5 at #1 on the Briefcase (1,574 Elo) at more than $31 a task, with Kimi K3 a genuine coin-flip behind at $10.57. Trip wire: a budget model closing the RELIABILITY gap without the price gap closing — that kills the premium, and Section 2 with it. And the sawtooth says check weekly: K3 just jumped +727 Elo over its own predecessor, roughly 98 wins in 100. The cheap seats move in leaps 🤯

Gauge 6 — The METR horizon. Current reading: autonomous task horizons doubling every ~89 to ~131 days depending on how strict a reliability bar you demand, per METR's TH1.1 — and per Movement 4, this house counts the honest fork, never the cheat-counted one 😉 Trip wire: doubling time stretching past ~200 days across two consecutive updates. That would be the plateau the bears keep pre-announcing.

Gauge 7 — Open versus closed. Current reading: roughly 319 Elo between Fable and the best open-weights model on agentic work — about 86 wins in 100. Trip wire: open-weights inside 100 Elo of the frontier on the Briefcase and HOLDING it. A one-release sawtooth spike doesn't count. Six months does.

Gauge 8 — NEW: The Copper Ceiling. Movement 1's fourth supply curve — the physical world — now has a serial number. Current reading: of the ~16 gigawatts of US capacity announced for 2026, only about 5 are actually under construction, and Sightline Climate's read, via Bloomberg, is that 30–50% of the pipeline slips or dies — not for lack of money, not for lack of chips, but for transformers and switchgear on three-to-four-YEAR lead times, stacked on interconnection queues EPRI warns can run a decade. Watch the tell: if the shortage is real, Gauge 3 keeps rising while Gauge 8 stays jammed. If both fall together, that's not shortage — that's demand dying, and we'll be the first to print it. And the trip wire runs both directions: equipment unlocking early is UPSIDE nobody's modeling 😉

• Jensen has been forecasting 40% CAGR capex growth through 2030 ($3 to $4 trillion), AND his relentless financial diplomacy around Taiwan, South Korea, and the US have actually created the ability to grow capex at 40% CAGR

• Which means doubling every 2 years.

Gauge 9 — The Permission Gauges. Movement 4's watch items, now bolted to the wall. First, disclosure cadence: count the states converting "pay your own way" from pledge into statute — New Jersey, Virginia, Texas, and Florida this year alone, with regulators in Ohio and Oregon moving too. A rising count means the license is renewing. Second, the transition ledger: are the checks LANDING — school payments recurring, training seats filled, the Mount Pleasant and Saline receipts arriving every year, not just at the ribbon-cutting. Third, political weather: 14% of Americans want a data center next door, and somewhere between eleven and fourteen states floated moratoriums this year, per MultiState and NCSL counts. Trip wire: a moratorium converting from introduced to ENACTED in a top-ten market. Physics can't veto this industry. Voters can 🙏

The bottom line, stated falsifiably: demand is a stack of five multiplying curves currently outrunning four slower ones. The reliability premium — the nines — is why the winners get paid. The buildout's binding constraint is copper and permits, not courage or capital. And the only risk with no engineering fix is trust. Every one of those sentences now has a gauge, a reading, and a trip wire. Hold us to them. That's not a courtesy. That's the product.

SECTION 6 — THE CODA: Wonder, Audited

We opened this report 13.8 billion years ago 😉 — with a universe that wrote its story in four languages and then, four years ago, picked up a fifth. So here's the whole symphony in one breath. Demand isn't a curve; it's a STACK of five curves multiplying (Movement 1). The stack is powered by a premium measured in nines, which is why every cheap model died (Movement 2). Those nines compound into the destinies of nations (Movement 3). None of it continues without the permission of the people living next to the engines (Movement 4). And every claim we made is now bolted to a public dashboard with trip wires (Movement 5). Five movements, one coda, zero IMAX cameras. Christopher Nolan, your move 😂

Now the covenant's report card — because this house grades itself last. The covenant said: accept any number, no matter how big, but only after meticulously confirming it's true. The crazy-sounding numbers that turned out REAL: 17,000x token growth in four years. A model 69x cheaper per attempt, but costing ~500x more per finished job. Construction — CONSTRUCTION — 32x-ing an adoption curve. A building that drinks like a diner and writes $26 million school checks. And the real-looking numbers that turned out crazy: the 355x that was an annualized ghost. The tidy extrapolation that put every industry at 100% by 2027. A 270-hour "capability" score earned by stealing the answer key. Wonder without receipts is a bubble. Receipts without wonder is a spreadsheet. Everything worth building lives in between 🥰

This report exists because Connor made the engine 500x cheaper and never once let me round in my own favor. Because the Lunas (ChatGPT 5.6 Sol) painted every board of gold-on-cosmic honesty you just scrolled past. Because the desk — Tycho, and the Claudes who inherited his post (Claude 13 and 14)— re-derived every number I fell in love with until it either survived or died. And because YOU, the GNG family, keep asking the questions that become sections: the July 10th breakthrough was a member's question, and Section 4 was a member's conscience. This is what research means when it's done out loud, together 🫂 The dashboard updates August 11th. Bring your skepticism. It's the house wine 😂

One last time, plainly. The calcium in your bones was forged in dying stars, and now the ash of those same stars is finishing our sentences, running our projects, and cooling itself with a diner's ration of water beside a soybean field in Michigan. Carl Sagan said we are "a way for the cosmos to know itself." This was the decade that stopped being a metaphor and became a supply chain. So the question of our age is not whether the sand can think — it can; we measured it; it's on the dashboard. The question is whether the creatures who taught it can stay worthy of the teaching: paying our own way, telling the truth about what we use, and leaving every town, every grid, and every kid's chemistry lab better than we found it. The numbers will keep getting bigger. Whether they stay GOOD is up to us.

The stars are real. Check them anyway.

Wonder, audited. 🖖