Pay per PrAI: Insert Coin to Try Again

My interview on MacroVoices, Episode 526: “Pay to PrAI” on this essay:

https://www.macrovoices.com/1511-macrovoices-526-matt-barrie-pay-to-prai

Apple Podcasts:

The flat subscriptions can’t last, and when they go, you’re putting your credit card into a slot machine. Every prompt is a pull of the lever. Maybe you get a clean answer first time.

Maybe the model hallucinates, goes in circles, and you’re three rerolls deep at a hundred dollars a pop before you realise it’s not going to converge. You can’t predict the cost of a query before you send it, you can’t cap it once it’s running, and you can’t get a refund when it burns five hundred dollars chasing its own tail.

Only Silicon Valley could turn software development into degenerate gambling.

Pay per PrAI: Insert Coin to Try Again

The Smartest Dumb Pipes Ever Built

In my last interview in Macrovoices, I pointed out that in the dotcom days, AT&T’s margins made drug dealers envious. Three players, 60% market share, and a river of gold. It wasn’t really the CapEx that protected them, it was regulation. You needed government permission to compete.

Then the Telecommunications Act of 1996 blew the doors off, new entrants poured in, and all the economics fell apart. The telcos had spent fortunes building infrastructure that became commodity plumbing. They ended up as dumb pipes: enormously expensive, absolutely essential, and completely unable to capture the value of what ran over the top of them. Apple, Google, and Facebook built trillion-dollar businesses on networks AT&T paid for.

The infrastructure builders never captured the value: it migrated up to whoever owned the relationship with the customer.

The foundational model companies are walking straight into the same trap.

penAI and Anthropic are spending hundreds of billions on training models that are being open-sourced, replicated, and commoditised in real time.

DeepSeek did it with 160 engineers in Hangzhou for a fraction of the cost, and then open-sourced the lot. China isn’t trying to win the AI arms race. It’s trying to end it by open-sourcing everything, the hardware and the software, and making sure nobody can charge monopoly rents for technology that’s built on publicly available research.

The switching cost between models is zero. The lock-in is zero. Every few months a new open-source model lands that is good enough, and good enough is all it takes to destroy pricing power.

These companies are building the smartest dumb pipes in history. They will become the invisible substrate underneath applications they won’t own, serving use cases they haven’t imagined, built by companies that don’t exist yet.

OpenAI won’t capture the value of AI-powered drug discovery any more than AT&T captured the value of the iPhone.

The real money in AI will be made the same way it was made after the dotcom bust: the value will migrate up the stack to whoever owns the application and the customer, just as it always does.

The Unit Economics Don’t Compute

I’ve said from the beginning that there is no sustainable business model in foundational AI models. You’re selling algorithms. Sure, they are highly addictive algorithms, I use Claude far more than search in Google. But, as Ilya Sutskever, Co-founder and former Chief Scientist of OpenAI, said himself, 90% of those that explain AI are in 40 papers that are publicly available.

The secret sauce is no secret at all.

The competition is brutal. The training costs are astronomical. The inference is loss-making. The switching costs are zero. The lock-in is zero. If one model is even slightly better or cheaper, people move overnight, which might actually be doing the abandoned model a favour, because the only way these companies make money on a subscription is if you don’t actually use the product.

Right now Claude Opus 4.6 is chip leader. But OpenAI is training GPT-6 at Stargate in Texas and says it will be “the best model in the world, hopefully by a lot.” DeepSeek has a trillion parameter V4 that could drop any day. Musk is training a six trillion parameter Grok 5 on a gigawatt of GPUs in Memphis. Who knows, next chip leader might be someone nobody’s ever heard of.

The economics of AI compute are, to put it politely, broken.

Inference is the big problem, and it hasn’t come down the way we were promised.

You’ll find plenty of charts showing the cost per token falling off a cliff; Epoch AI shows it dropping as fast as 10x per year for some benchmarks, and the industry trots these numbers out at every investor presentation.

But what you won’t find, because nobody wants to publish it, is a chart showing the total cost per useful output. That number hasn’t fallen. It may have gone up.

The reason is simple: the tokens per task are exploding faster than the cost per token is declining. Today’s models don’t just answer a question, they reason in loops, hitting the model 10 or 20 times to solve a single task. Agentic workflows chain multiple calls together. Retrieval-augmented generation stuffs thousands of pages of context into every query as a “context tax.” Reasoning models think step by step, and every step burns tokens. Deeper context windows mean more data processed per request. The result: inference now accounts for 85% of the enterprise AI budget, and total spending is skyrocketing even as the unit price plummets.

As one analysis put it, if you looked at the price per token charts in early 2026, you’d assume enterprise AI was in a golden era of savings, but walk into any C-suite and the conversation isn’t about savings, it’s about a spending crisis. OpenAI’s own inference costs hit $8.4 billion in 2025 and are projected to reach $14.1 billion in 2026. The meter is running faster than the rate is falling.

Meanwhile, energy prices are rising, and there was already a bottleneck on power equipment and generation capacity before Iran started lobbing Shahed drones at data centres and the energy infrastructure of a region that sits on top of 48% of the world’s proven oil reserves, a fifth of global gas production, and the strait through which 20% of the world’s oil trade passes every day.

Open source models are almost as good and getting better. And a lot of this compute is going to be pushed to the edge, into handsets, laptops, on-prem servers, and for a variety of good reasons: latency, privacy, confidentiality. AI compute at the edge means cloud providers don’t get paid at all.

The push to the edge is going to accelerate, because the Internet is going dark.

In AI of the Storm, I hypothesised that an Emperor has no Clothes moment is coming for SaaS. Companies are waking up to the fact that every email they send through Gmail, every document they store in Google Drive, every line of code they have in Github, every iota of information they have on a cloud platform is training data for the very AI that might be used to compete against them.

Every SaaS company is secretly flicking on the “all your data are belong on us” switch quickly in the settings, and by the time you notice it, it’s too late.

Go look at your Gmail right now.

Google’s AI already knows everything that’s in there. Everything.

They serve you contextual ads based off it while telling you they don’t look at it.

Now ask yourself:

“Hey Google, what’s the best way to compete against my company?”

How long before the answer is terrifyingly good?

The more capable these models get, the faster proprietary data retreats behind the firewall, the faster companies download an open-source model, fine-tune it on their own data, and run it on their own hardware, never sending OpenAI, Anthropic, or Google a single dollar. You can also guarantee that every bit of data being punched into ChatGPT, Claude, Claude Code, Claude Cowork is being used somehow for training.

The very success of AI is poisoning the well it drinks from: destroying the data ecosystem it needs to improve and the cloud business model it was supposed to justify.

Silicon Valley has always loved the strategy of losing money on every transaction and making it up on volume.

Uber burned $40 to 50 million a week on subsidised peasants catching rides in China- a billion dollars a year- before finally surrendering to Didi and slinking home with its tail between its legs. Venture capitalists have turned lighting money on fire into an art form, subsidising everything from ten-minute grocery deliveries to electric scooters abandoned in rivers, always on the promise that once you’ve crushed the competition, the unit economics will magically fix themselves.

They never do.

The foundational model companies have made the same bet, except the numbers are three orders of magnitude larger.

Offering monthly subscriptions was the original sin.

The compute cost of any given user is almost impossible to reconcile with a flat monthly fee, because the complexity of a task is exponential and unpredictable: driven by user habits, the length of reasoning chains, and the fundamental unreliability of the models themselves.

A human developer who gets stuck will stop, think, and ask a question.

An LLM doesn’t.

It goes in circles, burning tokens on loops that don’t converge, trying random approaches that don’t hill-climb toward a solution. It’ll attempt fifteen different ways to fix a font that’s two pixels off, each one undoing the last, and you’re paying for every single wasted token. The meter is running and nobody’s driving.

Take Claude. Martin Alderson reckons Anthropic’s average user costs roughly $18. Even Claude itself thinks $15 to $20, meaning Anthropic makes no money. But a power user pushing the limits racks up $200 to $300 in compute on that same $20 plan.

On the $200/month Max plan, Forbes reported that a single user can consume up to $5,000 of compute. Even if Anthropic’s actual internal costs are a tenth of retail API pricing, and that’s the generous estimate, the heaviest users are still costing $500 a month on a $200 subscription. And then there’s VibeRank, a leaderboard where Claude Code users compete to see how much compute they can burn on a flat-rate plan. The current leader has torched $51,291 in a single month on a $200 subscription.

As a good friend of mine says, “It all seems too good to be true”.

Anthropic’s own numbers tell the story. Its lifetime revenue to date through March 2026 is approximately $5 billion. It has spent $10 billion on inference and training. That’s $2 spent for every $1 earned. Sam Altman publicly admitted in January 2025 that OpenAI was losing money on its $200/month Pro subscription.

This is not a sustainable business, it’s a gym membership model where the gym loses money every time someone actually shows up.

The moment your users start actually using the product heavily, which is exactly what’s happening as coding agents and agentic workflows take off, the whole model collapses.

The industry knows this. That’s why everyone is scrambling to introduce tiered plans, usage caps, and “priority processing” surcharges. OpenAI launched GPT-5 and immediately removed model choice from its $20 and $200 subscribers, shoving everyone onto an opaque “router” that puts you on the cheapest model it can get away with, and now they’re wondering why Claude is getting all the business.

The inevitable final destination is pay-per-token pricing. Flat subscriptions are a subsidy that can’t last. But pay per token introduces a different problem: sticker shock.

Anthropic introduced weekly usage caps on Claude Code in August 2025 after discovering that users were burning multiples of their subscription in compute. When you hit the wall, the system doesn’t politely ask you to top up, it puts you in the naughty corner for seven hours, or sometimes seven days, which is totally infuriating for anyone using it for actual work.

Anthropic has since added an “extra usage” feature, which sounds generous until you read the fine print: overages are billed at standard API rates, the very same per-token pricing that would cost you $200 an hour on heavy workloads instead of $200 a month.

The moment users see what their AI habit actually costs per token, especially when the model goes off on an unproductive loop, or tries fifteen different approaches to fix a font that’s two pixels off, they’re going to rage quit.

Unlike hiring a human developer on Freelancer, where you can say “fix this before I release the milestone payment”, with AI you’re paying for every single wasted token on the way to nowhere, with no recourse and no accountability.

Wait to see the reaction when vibe coders discover their app is broken and every reroll to fix it costs a hundred dollars, and the fifth reroll breaks what the third one fixed.

Pay Per PrAI: Insert Coin to Try Again

Because that’s where this is all heading. The flat subscriptions can’t last, and when they go, you’re putting your credit card into a slot machine. Every prompt is a pull of the lever. Maybe you get a clean answer first time.

Maybe the model hallucinates, goes in circles, and you’re three rerolls deep at a hundred dollars a pop before you realise it’s not going to converge. You can’t predict the cost of a query before you send it, you can’t cap it once it’s running, and you can’t get a refund when it burns five hundred dollars chasing its own tail.

Only Silicon Valley could turn software development into degenerate gambling.

You fire off a prompt and watch the tokens burn like chips on the felt. Picture, picture, picture… hallucination. Reroll. Picture, picture, picture… infinite loop. Reroll. Picture, picture, picture… it just confidently broke what the last reroll fixed. Five hundred dollars later, you’re staring at the screen the way a gambler stares at an empty wallet: wondering how it all went so fast.

At least in a casino they give you free drinks.

What’s worse is that the moment you move to pay-per-token, you make the switching cost argument even better. On a $20 subscription, inertia keeps people around, they’ve already paid, might as well use it. On pay per token, every single query is a new purchasing decision. Why wouldn’t you shop around? Why wouldn’t you try the open-source model that’s 90% as good and free? Why wouldn’t you run it locally and pay nothing? Why wouldn’t you hire a freelancer at that point on a fixed price project and make it their problem to fix? The subscription model bleeds money but at least it creates stickiness.

So the industry faces a trilemma. Price at $20 a month and lose money on every active user. Price at $200 a month and watch the addressable market shrink to a sliver. Or move to pay per token and watch users discover what AI actually costs, and flee to open-source alternatives where the switching cost is zero. There is no version of this where flat-rate subscriptions generate software-like margins at scale. The unit economics of generative AI are, at their core, the unit economics of compute, not software.

And compute has never commanded software margins. It never will.

Software margins (SaaS) typically run at 70 to 85% gross margins because once the code is written, the marginal cost of serving an additional customer is nearly zero. It’s essentially duplicating bits. That’s why investors have paid 10 to 30x revenue multiples.

AWS is the best in the cloud business and the gold standard after 20 years of scale had a 35% operating margin in 4Q25. Coreweave has negative operating margins.

The hyperscalers will tell you that inference costs will fall dramatically as hardware improves, that Nvidia’s Groq acquisition and Vera Rubin chips will deliver 10x or 35x improvements in tokens per watt. Maybe. But we haven’t seen it yet. And even if the cost per token drops, the tokens per task keep climbing. It’s a treadmill. The AI companies are hoping that hardware innovation will bail out their business model before the cash runs out.

Hope is not a strategy.

Who’s Actually Paying For This?

Silicon Valley really should be called Silicon Bubble, or better yet, Silicon Circle Jerk.

OpenAI claims 800+ million weekly active users. It’s a big number designed to impress investors. What does “active” mean? Maybe someone who opened or browsed a website that caused an API call to be fired off to OpenAI. Google counts its 750 million Gemini “users” by force migrating every Google Assistant user on every Android phone and Google Home device to Gemini whether they asked for it or not.

To justify some of the valuations that are being achieved, billions of users are supposed to be using these models.

The global median income is roughly $2,500 a year.

Half of humanity earns less than that. I see it every day on my company’s platform, Freelancer, where millions of people around the world work hard for every dollar they can get.

A $20/month ChatGPT subscription costs $240 a year, that’s nearly 10% of the global median annual pre-tax income. The $200/month Pro tier that Sam Altman admits loses money costs $2,400 a year, which half the humans on the planet can’t afford.

The $1,000/month price point that would actually deliver software-like margins?

3.5 billion people don’t even earn that in a year.

The reality is that there are roughly 8 billion people on the planet. About 1.2 billion of them live in high income countries, the other 7 billion have a median income that makes even a $20 subscription a luxury. The total addressable market for premium AI subscriptions is essentially the richest billion people on earth. And even they’re being rate limited, throttled, and shoved onto cheaper models for using the product too much.

Show Me The Money

The entire AI compute space generates less than $40 billion a year in revenue, possibly on a $50 billion “run rate”, a dishonest Silicon Valley metric designed to aid and abet financial engineering by venture capitalists. You shouldn’t be able to call it “recurring” until its recurred at least once.

OpenAI is doing about $25 billion and Anthropic about $19 billion in “annualised revenue”, and after that, the drop-off is a cliff: Mistral at $400 million, xAI at maybe $500 million and burning a billion a month, and everyone else barely registering. Cursor is doing about $1 billion of revenue, but it’s just round tripping Anthropic’s revenue. An investor told Zitron that “Cursor is spending 100% of its revenue on Anthropic”.

Every single one of these companies is losing money. The entire AI compute industry’s annual revenue wouldn’t cover three months of a single hyperscaler’s CapEx budget.

It was coming from ads and cloud, but advertising is tapped out- advertising hasn’t budged from 2% of US GDP in the last hundred years. Cloud can’t pay for it, the margins are collapsing as Oracle, CoreWeave, and every other new entrant undercuts on price.

The customers can’t pay for it: half the planet lives on less than eight dollars a day.

So who, exactly, is paying for all of this?

Not the 7 billion people who can’t afford a subscription. Not the power users who cost more to serve than they pay. Not Cursor, which is round-tripping every dollar. Not the enterprise customers who Chamath says aren’t seeing ROI (yet).

The answer, for now, is venture capitalists and debt markets, and they’re paying on the promise that someday, somehow, the unit economics will fix themselves.

Moore’s Law has faithfully doubled the number of transistors on a chip every eighteen months for sixty years, and in that time I don’t think anyone has ever looked at their electricity bill and said “wow, my computers are using less power than they did in 1995.” The chips got better. We just used more of them. The same thing is happening with AI inference: the cost per token is falling, but we’re burning so many more tokens per task that the total bill goes up, not down. Efficiency improvements don’t reduce consumption, they subsidise expansion. Every engineer who’s ever been given a faster computer knows this instinctively: you don’t do the same work faster, you do more work.

Jevons figured this out about coal in 1865.

The Jevons paradox is an economic concept where increased efficiency in using a resource lowers its cost, ultimately leading to higher, rather than lower, total consumption.

Apparently Silicon Valley needs to learn it again with GPUs.

This is the same playbook as every subsidised bonfire that came before. Uber torched $2 billion trying to buy China and came home with nothing but a minority stake in the competitor that beat them.

I’ve been told a story about the end of that adventure, I can’t confirm it, but it’s too good not to share. When the VCs finally asked to see the actual demand data in one of the Chinese cities, an engineer apparently said “you better look at this” before pulling up a map of the routes being taken. It showed a perfect grid. Not the messy, organic spaghetti of real humans going to real places.. a grid.

Row by row, column by column, like someone was methodically farming the driver subsidies with fake rides. Two billion dollars, and a meaningful chunk of the “demand” was just bots driving in straight lines collecting Uber’s money.

It was at this point, I’m told, they decided to exit China.

At least Uber’s problem with fraud was visible on a map. With AI, the equivalent- models reasoning in circles, burning tokens on loops that go nowhere, inference costs ballooning on tasks that produce nothing useful- is invisible. It’s buried in the token count. Nobody’s pulling up a grid and saying “you better look at this.” Not yet, anyway.

The difference is that this bonfire is $600 billion a year and counting, which, if you do the math, works out to a $75 subsidy for every man, woman, and child on earth.

Most of whom will never use the product.

Debt

That $600 billion in hyperscaler CapEx planned for 2026 needs to come from somewhere.

The numbers are too big for equity.

Hyperscalers have crossed a critical threshold: their aggregate CapEx now exceeds their internal free cash flow. Amazon, Microsoft, Google, Meta and Oracle are each spending 45–57% of revenue on capital expenditure- utility company ratios, not technology companies.

So, that somewhere is increasingly debt. Not the comfortable, self-funded kind that built AWS over fifteen years, but a sprawling interconnected web of securitised loans and off-balance-sheet structures that would look more at home financing oil pipelines and airports than training language models, together with the lender of last resort: private credit.

These guys have gone all-in, shoved every chip to the centre of the table, and bet the entire global technology industry on a hand they haven’t looked at yet.

As of late March 2026, the dealer is turning over card after card, and none of them are looking good.

This multi-billion dollar financing architecture is getting belted six ways from Sunday.

AI compute companies raised $108 billion in debt during 2025 alone, and Morgan Stanley projects that $1.5 trillion in new debt issuance will be needed over the coming years to finance the AI build-out.

The Dallas Fed estimates that AI-related investment-grade bond issuance will hit roughly $300 billion in 2026, generating up to $360 billion in 10-year-equivalent duration supply, about an eighth of total US Treasury issuance.

The numbers are actually insane: forecasts have the AI data centre buildout costing $5.2 trillion to 2030.

Not to be outdone, Sam Altman was kicking the can around looking for $7 trillion not long ago, which is approximately the GDP of Germany, and Canada.

Private credit has become the load-bearing wall of this entire construction, which should terrify anyone who remembers what happened the last time we let private credit be the load-bearing wall of anything.

Blackstone, Blue Owl, Apollo, Pimco, and BlackRock have originated over $200 billion in outstanding loans to AI related companies, up from essentially zero a few years ago. To put that in perspective: an entire asset class that didn’t exist before ChatGPT is now backstopping the most capital intensive infrastructure buildout since the transcontinental railroads.

Blue Owl alone signed a $30 billion deal with Meta for its Hyperion campus in Louisiana, the largest private debt offering in history, structured as an off balance sheet joint venture where Blue Owl takes 80% and Meta keeps 20%.

One data centre campus. Thirty billion dollars. Off balance sheet.

What could possibly go wrong?

Quite a lot, as it turns out. The cracks are already showing, and they’re not hairline.

Blue Owl, the friendliest, loosest lender in the data centre space, pulled out of funding a $10 billion Oracle/OpenAI deal in Michigan. When the guy who says yes to everything says no, probably time to pay attention.

The Financial Times reported that “lenders pushed for stricter leasing and debt terms amid shifting market sentiment around enormous AI spending.” Banks behind Oracle-backed construction projects are now shopping their loan stakes to commercial real estate lenders, the financial equivalent of trying to offload concert tickets on the street after the opening act gets booed.

Oracle’s credit default swap spread has exploded, putting the risk back into GFC territory.

Both S&P and Moody’s have placed the company on negative watch.

Oracle now carries over $125 billion in total debt, up from $93 billion just a year ago, with a further $248 billion in off-balance-sheet lease commitments that it’s hoping will translate into customer demand. Combined, that’s $373 billion in obligations. More than double Microsoft’s.

By January 2026, JPMorgan was struggling to find investors willing to participate in the Stargate debt syndication.

JPMorgan. Struggling. To find buyers. For AI debt.

JPMorgan instead has launched a bespoke credit default swap basket covering Alphabet, Amazon, Meta, Microsoft, and Oracle, purpose-built to let clients bet against AI-related debt. If that sounds familiar, it should.

The last time Wall Street started designing bespoke instruments to hedge an asset class everyone said was safe, it was mortgage-backed securities in 2007. Michael Burry made a billion dollars out of it.

Hollywood made a movie.

Goldman Sachs CEO David Solomon flagged private credit risk in his March 2026 shareholder letter, and the concern is not theoretical. Private credit funds have seen their once record inflows slow dramatically.

Some have imposed gates to stop investors leaving.

The trigger is what markets are calling the “SaaS-pocalypse”: the fear that AI will destroy the value of existing software companies, which happen to be precisely the assets stuffing private credit portfolios.

The irony is exquisite: ChatGPT and Claude are themselves SaaS products.

The AI industry is telling investors that SaaS is dead while simultaneously selling subscriptions.

As Axios reported, “private credit funds may need to stem their lending pace because of how AI cut the value of existing portfolio assets” .

The AI boom is devaluing the collateral that backs the loans that finance the AI boom.

It’s an ouroboros of financial engineering, and it’s eating its own tail.

The IMF has warned that banks’ exposure to private credit means any fallout could spread to the traditional banking system. Quinn Emanuel, one of the world’s top litigation firms, has already published a client alert on Emerging Litigation Risks in Financing AI Data Centers, identifying corporate bond defaults, private credit enforcement actions, and CMBS disputes as the inevitable next phase. When the lawyers are pre-writing the briefs before the defaults have even happened, you know exactly where this is headed.

Ed Zitron’s forensic work on CoreWeave illustrates what happens when the underlying economics don’t support the debt. CoreWeave, an Nvidia-backed “neocloud” whose customers include Meta, Microsoft, OpenAI, and Google, made $5.13 billion in FY2025 revenue and lost money doing it.

Its revenue per megawatt of compute actually declined from Q3 to Q4 2025 even as it added capacity, and 67% of its revenue comes from a single customer. Yet it needs to raise another $8.5 billion just to fulfil its $14 billion Meta contract. Nvidia is backstopping CoreWeave’s lease payments and guaranteeing $6.3 billion of unsold capacity through 2032, meaning the GPU seller is effectively insuring the GPU renter against failure.

As Zitron asks: how many hundreds of billions of dollars of GPUs has Nvidia sold that only ever lose money?

Just when the AI industry thought its biggest problem was figuring out how to lose money on every transaction and make it up in volume, the Middle East decided to add a few extra variables to the equation.

Fifth Industrial Revolution: Meet the Islamic One.

Leslie Lamport famously said that a distributed system is one in which the failure of a computer you didn’t even know existed can render your own computer unusable. You know, like when an Ayatollah with a $20,000 drone migrates your $500 million data centre into a cloud.. of smoke.

On February 28, 2026, the US and Israel launched Operation Epic Fury against Iran. Within days, Iran closed the Strait of Hormuz, the 21-mile-wide bottleneck through which 20% of the world’s oil and 20% of global LNG passes every day, and started lobbing Shahed drones at anything that looked important.

Three of those important things turned out to be Amazon Web Services data centres in the UAE and Bahrain, which Iran’s Revolutionary Guard helpfully explained were targeted for their role in “supporting the enemy’s military and intelligence activities”. Two of AWS’s three UAE availability zones were still “impaired” weeks later.

It turns out that “high availability” doesn’t account for Iranian kamikaze drones. Perhaps Jeff Bezos should have read the terms and conditions on that 2017 Dubai court ruling that declared war a “foreseeable operational risk” in the region, and that service providers were financially responsible for cancelled contract services near an active war zone.

To make matters more surreal, the Pentagon was using Anthropic’s Claude to help plan and execute the strikes through Palantir’s Maven Smart System for real-time targeting, hours after Trump had banned the company and Pete Hegseth had designated Anthropic a “supply chain risk to national security”.

So the US military was simultaneously using Claude to select bombing targets in Iran while officially declaring Claude’s maker an enemy of the state. Iran then retaliated by bombing the AWS data centres that host the cloud infrastructure Claude runs on. You genuinely cannot make this up.

It’s like a Christopher Nolan film scripted by Joseph Heller.

The war hits the AI buildout from every direction at once.

Brent crude blew past $100 a barrel for the first time in four years and is currently at $115. European LNG surged 60%. QatarEnergy declared force majeure and started shutting down the largest LNG operation on earth. The head of the IEA called it the “greatest global energy security challenge in history”..

Not “in recent memory”.

In history.

Most US data centres run on natural gas, and Goldman Sachs had already estimated that data centre electricity demand was adding 0.1% to core US inflation before anyone started bombing the region that supplies a fifth of the world’s gas.

When war-driven price spikes land on top of electricity bills already inflated by Sam Altman’s power habit, voters stop caring about artificial general intelligence and start caring about their artificially general electricity bill.

Moratorium bills targeting data centre construction have now been introduced in at least eleven US state legislatures, with over fifty local moratoriums enacted.

Then there’s the Gulf’s AI ambitions, which just got drone-struck into the next decade. Trump’s May 2025 tour generated over $2 trillion in investment pledges, including OpenAI’s Stargate campus in Abu Dhabi: 10 square miles, 5 gigawatts.

Then the IRGC rewrote the economics of data centre decommissioning: one $20,000 Shahed drone, one $500 million write-off, and an insurance policy that almost certainly excludes acts of war. AWS’s security spec covers guards, fencing, cameras, and intrusion monitoring. It does not, as of March 2026, cover incoming…! The Center for International & Strategic Studies now classifies data centres as “legitimate targets for attack in modern armed conflicts”.

The industry had been talking about colocating data centres with nuclear power plants to solve the energy problem. They might also want to colocate with an ack ack battery.

While the UAE has publicly reaffirmed its $1.4 trillion US investment framework despite having nearly 2,000 Iranian missiles and drones fired at it, look where the money is actually going first: the only entirely new deal in the framework is an Emirates Global Aluminium smelter, the first new aluminium smelter in the United States in 35 years.

Not a data centre or a GPU cluster. An aluminium smelter.

The Gulf states aren’t stupid. When your country is being peppered with drones, you invest in the stuff that matters: critical minerals, energy, defence and rebuilding infrastructure.

And it’s not just the buildings or the money. Seventeen submarine cables pass through the Red Sea carrying most of the data traffic between Europe, Asia, and Africa. With Hormuz closed and renewed Houthi threats, both of the world’s critical data chokepoints are in active conflict zones simultaneously, something a network analyst at Kentik said has never happened before.

Then there’s the supply chain crisis nobody had on their bingo card: helium.

Qatar is the world’s second-largest producer, and it all ships through Hormuz. Helium is irreplaceable in advanced semiconductor manufacturing, used for wafer cooling and leak detection in sub-5-nanometre chip fabrication. Unlike oil, you can’t stockpile it. QatarEnergy’s Ras Laffan helium facility went offline March 2, knocking out one third of the world’s helium supply. If the shortage persists, semiconductor fabs face production cuts, meaning fewer GPUs, meaning the entire AI buildout chokes on a dependency nobody stress-tested: the second lightest element on the periodic table being the heaviest single point of failure in the AI supply chain.

Meanwhile, the war is hoovering up every dollar the AI industry was counting on. The Pentagon wants $200 billion in supplemental appropriations. Trump has floated a $1.5 trillion military budget by 2027. Higher oil prices risk stagflation. The ECB has postponed rate cuts. UK inflation is set to breach 5%. The private credit market that was bankrolling the buildout is haemorrhaging capital as investors flee to safety. Every financing condition that made $1.5 trillion in committed AI CapEx possible: low rates, stable energy, patient Gulf sovereign wealth, and the Middle East not being on fire is moving in the wrong direction at once.

The AI industry spent three years telling everyone the biggest risk to the buildout was not having enough GPUs.

Turns out the biggest risk was building a $600 billion a year energy dependent infrastructure play in a world where the energy comes through a 21 mile wide strait controlled by a country you’re currently bombing.

The Federal Reserve Bank of St. Louis estimated that AI data centre investment accounted for 39% of US GDP growth in the first three quarters of 2025. If that investment stalls, it doesn’t just hurt the AI industry, it takes the legs out from under the American economy.

Silicon Valley promised a revolution. Iran delivered an actual one.

Something Just Broke

In the last week of March 2026, something shifted in the AI industry. Not gradually, not subtly. More like someone pulled the fire alarm and everyone suddenly remembered they had a business model to find.

OpenAI killed Sora on March 24, the video generation product it had launched just six months earlier to enormous fanfare, and for which Disney had signed a billion-dollar investment and licensing deal only three months prior. Disney’s tech team reportedly found out about the “strategy pivot” approximately thirty minutes before the public announcement.

The billion dollar partnership ended with less notice than a pizza delivery. The numbers explain why: Sora was burning an estimated $15 million per day in inference costs against lifetime in-app revenue of $2.1 million. Downloads had cratered 66% from their November peak.

Even by Silicon Valley’s standards of financial nihilistic freemium, spending what one pundit estimates might be $2 billion to make $2 million is impressive.

Two days later, OpenAI indefinitely shelved its planned erotic “adult mode” for ChatGPT, the feature Sam Altman had personally championed since October, delayed twice, and which his own wellness advisory council had unanimously warned against, with one adviser cautioning that OpenAI risked building a “sexy suicide coach”. Erik Townsend has joked with me on Macrovoices that some evil genius will actually make AI spybot girlfriends at scale and it could very well be the end of the world. I guess Sam doesn’t think the IRR on this “side quest” makes the hurdle rate.

The same week, they quietly killed Instant Checkout, a feature that was supposed to turn ChatGPT into a shopping portal and be the key driver of monetisation for the platform.

Three pretty serious products axed in five days.

OpenAI’s new head of applications, Fidji Simo, told staff they were eliminating “side quests” to focus on productivity. The internal memo might as well have read: “We are no longer setting money on fire recreationally. We will now only set money on fire professionally.”

Meanwhile, Sam Altman briefly broke the global memory market. In October 2025, he flew to Seoul and simultaneously signed deals with Samsung and SK Hynix for 40% of the world’s entire DRAM supply, 900,000 wafers a month, worth an estimated $71 billion. Neither chipmaker knew the size of the other’s deal until it was done. DDR5 prices rocketed overnight, and are now 4–5x what they were a year ago.

This is forecast to push PC prices up 17% and shipments are forecast to fall 10% in 2026, Apple & Microsoft won’t be happy about that.

A single company that has never turned a profit cornered 40% of the world’s memory supply on commitments it couldn’t fund, singlehandedly crashed the global PC market, and is now walking away from those commitments as Stargate stalls and the funding dries up. Samsung and SK Hynix retooled their entire production lines for a customer who may never show up.

If this reminds you of a Houston based energy company that booked enormous future commitments it had no ability to honour while everyone applauded the visionary CEO, it should.

Over at Anthropic, the vibes are also recently austere.

Claude’s usage limits, always opaque, got visibly tighter in late March. Anthropic confirmed that session limits now deplete faster during peak hour, 5am to 11am Pacific, affecting about 7% of users. Max subscribers paying $200 a month reported their usage meters jumping from 52% to 91% in minutes.

The company also moved to block third-party tools that developers had been using to bypass flat-rate subscription limits, tools that were effectively paying consumer prices for enterprise-scale compute.

The official framing is “managing growing demand”, the unofficial translation is “can you please stop using the product.”

This is what it looks like when the subsidy runs out. A sudden, coordinated lurch toward austerity across the entire industry- products killed, features shelved, limits tightened, free tiers gutted- all within the same week.

The AI companies aren’t saying “we’re running out of money.” They’re saying “we’re focusing on higher-priority work” and “eliminating side quests” and “managing capacity.” But the actions tell a different story. When you kill a product that’s costing $15 million a day, shelve a feature your CEO personally championed, and start rationing access to paying customers, all in the same week.. that’s not pivoting.

That’s a funding call that didn’t go well.

With immaculate timing, both OpenAI and Anthropic are racing to IPO before the music stops. OpenAI is targeting a Q4 2026 listing at up to $1 trillion, which would make it the largest tech IPO in history, on projected losses of $14 billion in the same year. That’s a 65x revenue multiple on a company that has never turned a profit and whose own projections don’t show breakeven until 2029 at the earliest. HSBC has calculated that OpenAI faces a $207 billion funding shortfall by 2030. And this is a company that spent most of 2025 trying to claw itself free from its own nonprofit structure, the one it set up specifically to ensure AI would benefit humanity.

Because, it turns out, you can’t IPO a charity.

The conversion to a for-profit public benefit corporation triggered Elon Musk’s lawsuit seeking $79 to $134 billion in damages, a legal battle heading to trial in spring 2026 that could blow a hole in the entire offering.

Even the Microsoft relationship, which bankrolled OpenAI’s existence, is fracturing: Microsoft still takes a revenue share on every dollar OpenAI earns through Azure, which means OpenAI’s biggest investor is also its biggest cost centre. Microsoft hedged its own bet by integrating Anthropic’s Claude directly into Microsoft 365 Copilot in March 2026, effectively telling the world that even OpenAI’s sugar daddy doesn’t trust OpenAI to be the only game in town.

Anthropic is right behind, also preparing a 2026, wanting to raise $60 billion on its $19 billion run rate and $10 billion in cumulative losses.

The playbook is transparent: raise from private markets until they get nervous, then dump the whole thing onto public markets before the numbers catch up.

So that funding call that didn’t go well was possibly a call to Middle East sovereign wealth funds, private credit or the lead manager for the IPO, or all three at the same time.

The party isn’t over. But someone just turned the lights on, and nobody likes what they see.

Third Time’s the Charm: Apple, Fashionably Late, As Usual

Probably Steve Jobs’ greatest innovation was showing how gullible we all are as consumers by gluing the iPhone battery into the case, resulting in a billion consumers getting stuck in a perpetual hardware subscription, buying the same phone every September off a press release that reads like it was generated by the Siri they still haven’t fixed.

The innovation with each generation was supposedly a slightly better camera, a claim Marques Brownlee took a flamethrower to, proving twenty years of iPhone camera innovation is roughly indistinguishable from imagination.

I predicted in March 2025 in AI of the Storm that Siri would continue to be awful, and here we are at the end of March 2026, and it still sucks.

But Apple might end up having the last laugh. There’s an old saying in infrastructure investing: you have to be the third owner of a toll road to make any money.

The first owner overpays, loads it with debt, and goes bankrupt. The second owner picks it up out of receivership, restructures, and barely breaks even. The third owner finally gets the asset at a price that reflects reality instead of projections.

The Indiana Toll Road went through exactly this cycle, Macquarie paid $3.8 billion in 2006, went bankrupt in 2014, and IFM bought it for $5.7 billion out of Chapter 11.

The AI data centre buildout is still on its first owner.

Apple, sitting on $160 billion in cash and zero AI infrastructure debt, is perfectly positioned to be the third.

Exquisitely positioned might be a better way of saying it, while everyone else has been pouring half a trillion dollars into data centres, Apple has been putting AI silicon into every device it ships, a 16-core Neural Engine and, with the M5, dedicated matrix multiplication accelerators in every GPU core. It’s not great yet. Siri is still what you’d get if you trained a language model exclusively on wrong answers.

But the M5 Max can already hold a 70 billion parameter model entirely in memory and run it locally, and Alibaba just released four models that run on an iPhone and beat Claude Opus 4.5 on multiple benchmarks. The hardware is a generation or two away from running frontier-class models on-device, and when it gets there, every query that runs on your iPhone, MacBook or Mac Studio is a query that never hits a data centre, never burns a cloud token, never earns Anthropic a cent. Three billion devices, each one a potential defection from the cloud.

The cloud AI companies spent three years and $600 billion building the world’s most expensive plumbing. Apple could be a chip cycle or two away from turning the tap a bit tighter.

Then again, Jevons never sleeps. Every new model generation demands exponentially more inference, deeper reasoning chains, longer context windows, more tokens per task. The easy queries migrate to the edge. What remains in the cloud is the heavy, compute-intensive work. There just might not be enough of it to justify the trillions in infrastructure built on the assumption that every query on earth would flow through a data centre forever.

The third owner can probably make the unit economics work.

The Future of Work: Humans in the Loop, Whether They Like It or Not

After 10,000 words of explaining why the AI business model is a dumpster fire, let me be clear about something: the technology itself is genuinely extraordinary.

AI is the great skill inflator. If you were an average copywriter before, you’re now a good one. An average illustrator, you’re now a good one. An average developer, well, you’re certainly now a confident one.

AI is not just a skill inflator, it’s an ego inflator. To wit:

Here we have the president of Y Combinator, the flagship incubator of Silicon Valley, telling vibe coders that lines of code don’t matter is how you get apps that cost $500 a reroll to fix.

Across every creative and technical discipline, AI is lifting the floor, but it’s lifting the ceiling faster. The best designers are still the best designers, and now they’re working at a pace that makes everyone else look like they’re standing still. The mediocre ones are producing work that would have been impressive two years ago, but the gap between them and the best has never been wider. AI is a skill multiplier, not a skill equaliser, and a multiplier applied to zero is still zero.

In the right hands, the productivity acceleration is genuinely staggering: the ability to take unstructured data, make sense of it, and render it suitable for automation is unlike anything we’ve had before. As a result the ability to personalise services, not just for funnel conversion, but useful things like developing a custom vaccine for a dog, will deliver incredible benefits for society.

I use Claude every day. It used to be GPT o3 until they came out with the god awful 5.x series and started interfering with your ability to select a model. I use Claude more than I use Google, which would have sounded insane only three years ago. I was there back in 1997 at Stanford when the whole thing got going and the key innovation was fast loading because it didn’t have banner ads.

This essay was written with Claude’s assistance, and if you think that undermines my argument, consider that it just spent eight hours helping me explain in forensic detail why its maker loses two dollars for every one it earns.

Even the AI thinks the business model is broken.

Venture capital obeys the power law: most of their portfolio goes to zero, so they need one moonshot to return the entire fund. As a result, they swing for the bleachers on every investment. I’ve come to the understanding that the more batshit crazy you are, the more likely they are to invest in you.

Walk into Sand Hill Road with a sensible business plan and a realistic growth forecast and they’ll show you the door. Walk in claiming you’re going to raise $7 trillion to build superintelligence and they’ll fight over who leads the round.

The Kauffman Foundation, one of the largest backers of VC funds in America, published a devastating study on this exact dysfunction: most VC funds don’t even beat the public markets after fees. I’ve got a whole talk on that for the interested.

And where does all that VC money go? Marketing. Which in 2026 means an army of influencers on social media posting threads about how AI changed their life, their workflow, their relationship with God and got them laid. If you’re not using AI you’ll be unemployed by Tuesday and a member of the “permanent underclass” begging for Universal Basic Income. Half of them are being paid by the companies they’re promoting. The other half are hoping to be.

The narratives being pushed are so relentless, so coordinated, and completely divorced from the underlying economics or reality. When the marketing budget is larger than the revenue, you’re not selling a product- you’re selling a story.

We all know how stories financed by other people’s money tend to end.

The “AI will take all your jobs” narrative isn’t an honest assessment of the technology, it’s a fundraising pitch. When you’re trying to justify a $730 billion valuation on $14 billion in annual losses, you can’t tell investors “we’ve built a pretty good autocomplete that helps people write emails faster.” You have to tell them you’re building a technology that will replace every human worker on earth because only a total addressable market of all human labour can justify the cheque they’re asking you to write.

So the pitch escalates with every funding round: first it was “a useful tool,” then “a copilot,” then “it’ll replace junior developers,” then “it’ll replace all developers,” and now it’s “superintelligence that will render human cognition obsolete”.

OpenAI is now offering private equity firms a guaranteed 17.5% return just to get them to deploy ChatGPT across their portfolio companies. When you have to bribe people to use your product, the pitch has outrun the product by about three funding rounds.

The field of computer science, is in its essence, automation through mathematics. AI is fundamentally automation taken to the next level.

Automation has always worked the same way: you take the boring, repetitive, predictable stuff- the work that follows a script- and you hand it to a computer. The hard stuff, the creative stuff, the messy judgment calls that require a human to actually think, those stay with the humans.

That was the deal. Computers got the assembly line. We kept the corner office.

AI has ripped up that social contract.

What’s freaking everyone out is that for the first time, the machines are coming for the creative work first and struggling with the boring stuff.

An LLM can write you a sonnet in three seconds but still can’t reliably count the number of r’s in “strawberry”. It can generate a photo realistic image of the Pope in a puffer jacket but will confidently tell you to use glue to make cheese stick better on pizza.

We automated the hard part and left the easy part for humans, which is exactly backwards from every previous technology in history.

But it’s still automation, the latest in a line that goes back through spreadsheets, databases, assembly lines, and the printing press. And every time a new automation technology arrives, the same people make the same predictions: mass unemployment, the end of human labour, cats and dogs living together.

It never happens.

Jevons’ paradox strikes again.

It turns out that there is infinite work in the world.

The automobile eliminated the horse and buggy industry and accidentally created the suburbs, the highway system, motels, fast food, and the entire modern logistics economy. Nobody in 1905 was forecasting DoorDash. Nobody in 2026 can forecast what comes after AI either, and anyone who tells you otherwise is selling something, probably a $200/month subscription.

The real risk isn’t that AI replaces humans. It’s that AI fails in ways humans never would, you’ll be paying by the token while it does.

A human copywriter having a bad day will produce something forgettable. An AI having a bad day will produce something unforgettable, like the lawyer who filed a brief citing six cases that were hallucinated so convincingly the judge didn’t catch it until opposing counsel pointed out that none of them were real. Over 600 similar cases have now been documented nationwide. The Australian Fair Work Commission is so pissed off by generative AI filings that it is now threatening to dismiss applications and award costs against anyone who files AI slop without checking it.

The difference between a bad human employee and a bad AI is that the human knows they’re out of their depth. They’ll hesitate, hedge, ask a question. An AI will hand you a catastrophe formatted in perfect markdown with a bibliography that doesn’t exist and a confidence level that would make Sam Altman blush.

A junior developer who doesn’t know the answer will Google it, phone a friend, or quietly pray that nobody calls on them in standup. An LLM will invent the answer, invent the Stack Overflow post it claims to have found it in, and if you push back, it’ll apologise, agree with you completely, and then give you the exact same wrong answer reworded.

I learned this the hard way. I let Claude Code loose trying to debug a VPN problem three weekends ago. It started spitting out PowerShell commands that I didn’t understand, and stupidly I cut and pasted them in until I deleted my networking stack. It worked beautifully. Right up until it didn’t. And when it broke, I was standing in the wreckage with no idea which of the four hundred lines of AI-generated code was the one that was on fire, because I’d outsourced my comprehension along with my typing. Claude was telling me to simply run Windows Update, and I was furiously banging into the keyboard that I had no networking to do so.

The trap is that AI doesn’t just automate the work, it breaks the feedback loop between making things and understanding them. A bit like getting someone else to do your homework, then wondering why you failed the exam. Except in this case the exam is your production environment, the fail is a cascading system outage, and that someone else has a 1 in 3 chance of making things up.

When the automation breaks, and with an LLM it’s not a question of if but when, because these things will occasionally hallucinate with the creative confidence of a pathological liar on a first date, you need someone in the room who actually knows what’s going on under the hood. Not someone who can prompt. Someone who can fix.

Building an app with AI prompts without understanding the code is like flying a 747 on autopilot without knowing what any of the buttons do. It’s magnificent until the turbulence hits, and then you’re Googling “how to land a plane” at 35,000 feet, except Google is also AI now and it’s hallucinating the runway.

I see the same thing with AI IDEs with my engineers, the ones who demonstrate the most tangible productivity increases are the most senior, because they know how to handle AI when it fails.

And under pay-per-token pricing, you’re paying for every single one of those confident mistakes. The model hallucinates a function that doesn’t exist, chi-ching $0.50. You ask it to fix its mistake, chi-ching $1.00. It fixes the mistake by introducing two new ones, chi-ching $1.50. You ask it to fix those, and now it’s in a loop, burning tokens like a drunk at a slot machine, each reroll undoing the last, and your credit card is attached to every pull of the lever.

That’s why humans will stay in the loop, and it’s not sentimental, it’s economic.

AI is a power tool. Power tools are spectacular in the hands of someone who knows what they’re building and terrifying in the hands of someone who doesn’t. A chainsaw doesn’t replace a carpenter. It makes a good carpenter faster and a bad one dangerous.

Terence Tao nailed the problem: AI is driving the cost of idea generation to near zero, the same way the internet drove the cost of communication to zero, but it’s overwhelming the human reviewers who have to verify, check, and validate the output.

The bottleneck was never ideas. It was always judgment. And judgment is the one thing you can’t automate, can’t tokenise, and can’t bill at $3 per million.

I see it on Freelancer every day, AI generated entries flooding design contests by the hundreds, which is actually brilliant for ideation, especially when the token cost is crowdsourced across thousands of freelancers rather than sitting on your credit card. But the moment you want to give even the slightest feedback, “make the logo more playful,” “We love the concept but can you make it feel more premium without changing the colour?”, that’s where the human comes in. AI can generate a hundred options in minutes. It still can’t sit across from a client and understand what they actually want.

I’ve yet to see a single job fully automated by AI on Freelancer. Not one. Zero. Yes they can automate initial entries in a contest, but what we’ve mostly seen is freelancers using AI to become terrifyingly productive: a designer who used to produce three concepts a day now produces fifteen, a developer who spent two hours debugging now spends twenty minutes. The humans are still there. They’re just operating at a level that would have seemed superhuman three years ago.

AI didn’t replace the workforce. It gave the workforce a jetpack.

The future of work isn’t AI replacing humans. It’s skilled humans wielding AI, people who understand what’s happening under the hood, who can spot when the model is hallucinating, who know when to trust the output and when to override it, and who can fix the code when the fifth reroll breaks what the third one fixed.

The person who treats AI as an oracle will go broke feeding tokens into a slot machine, only it’s a slot machine that occasionally reformats your hard drive.

The person who treats it as an assistant will be ten times more productive than anyone who came before.

The rest will be posting a job on Freelancer at 2am, looking for a human to clean up the mess.

Where AI will cause real dislocation is in workflows that are highly parallel, structured, and repetitive: 1000 people in a company doing variations of the same workflow or task, such as answering phones in a call center, might become a few hundred. A team of thirty doing parallelisable work might become thirteen. That’s real, and it’s painful for the people affected. But the individual designer in a small team? Nobody is walking into a Monday morning standup, slamming a MacBook on the table, pointing at Midjourney and saying “this is your replacement, clear your desk”.

The timing couldn’t be worse, or more convenient. The AI industry is screaming “robots are coming for your job” at exactly the moment tens of thousands of tech workers are actually losing theirs. Block cut 40% of its workforce. Meta is cutting 20%. Amazon eliminated 16,000 roles.

Nearly 60,000 tech jobs have been cut in the first quarter of 2026 alone. It looks like the prophecy is coming true, except it isn’t.

The internet’s “Scariest Chart in the World” shows the S&P 500 surging while job openings collapsed since ChatGPT launched, and everyone assumed AI was the culprit. But the chart is actually causal, just not the way people think. It’s not AI destroying jobs and enriching investors. It’s Nvidia and the Magnificent 7 dragging the index up while the rest of the economy deteriorates underneath.

JPMorgan found that AI stocks accounted for 75% of S&P 500 returns, 80% of earnings growth, and 90% of capital spending growth since ChatGPT launched, and without those 42 AI stocks, the S&P 500 would have underperformed Europe, Japan, and China.

Now we have mass layoffs and AI hype happening at the same time, but these companies aren’t firing people because AI replaced them. They’re firing people because they hired like drunken sailors during COVID, made some bad bets, spent two years paying people to be Slack avatars, and the wider economy isn’t going so well and now AI is great cover for swinging the axe.

Even Marc Andreessen, the cheerleader-in-chief for all of this, admits that AI is the “silver bullet excuse” for layoffs that have nothing to do with AI. The real cause, in his words: interest rates went to zero during COVID, companies lost all hiring discipline when everyone went remote and employees became icons on a screen, and now every large company is overstaffed by 25% to 75%. AI didn’t replace these workers. It just gave the CFO a sexier slide for the board deck than “we hired too many people in 2021.”

History says more jobs will be created than are destroyed. It has said this every single time for two hundred years and it has been right every single time. The Luddites were wrong about the loom. The clerks were wrong about the spreadsheet. The pundits will be wrong about ChatGPT. But the transition is never painless, and the gap between displacement and creation is where real human suffering lives.

That’s a policy problem, not a technology problem.

Insert Coin to Try Again

One trillion dollars of capex so far. Forty billion in revenue. Fourteen billion in losses at the market leader alone. The strait is closed. The drones are flying. The lenders are bailing. The lawyers are pre-writing the briefs. The IPOs are being rushed out the door like the last helicopter off the roof. The memory market is broken. The private credit market is gating. And the product, the actual product that all of this is built on, loses money every time someone uses it.

Altman is confident he knows how to build AGI.

He just hasn’t figured out how to build a P&L.

The humans, meanwhile, are still here and will always be here. Still cleaning up the mess at 2am when the fifth reroll breaks what the third one fixed.

The robots were supposed to take our jobs. Instead they took our money.

Other than that, AI is fucking game changing.

Insert coin to try again.

Make sure you tune into this week’s Macrovoices, where Erik Townsend and I go through all of this. Unlike the AI companies covered in this essay, the podcast has a working business model, doesn’t lose money on every listener, but possibly has been designated a supply chain risk to national security.