Rishi Raam

I take tech apart and explain why it's built that way - systems, the companies that run them, and the businesses on top.

[Rishi Raam @rishiraamns]

What nobody understands is humans thrive on problem-solving. We constantly search for problems, however tiny or large they might be, to solve them. To get a sense of pride.

Many devs I know are starting to hate their job because it's just constantly prompting and pushing code, nothing difficult to solve.

[Rishi Raam @rishiraamns]

Current-gen LLMs have grown lot more complex than that.

They are not next-token predictors. They predict much ahead, plenty of times to the end of sentences, and frame their sentences to arrive at their ending.

Likewise, they have their own world model with concepts and reasoning that transcends language, just like us. They surely think and reason for each word it predicts.

[Rishi Raam @rishiraamns]

Something that almost never happens in AI just happened: a frontier lab is about to turn a profit.

Anthropic is on track for its first profitable quarter ever, and the twist is that its biggest customers are panicking over the bill.

The numbers are hard to believe. Revenue went from $4.8B in Q1 to a projected $10.9B in Q2. That's ~130% growth in a single quarter.

Faster than Zoom at its pandemic peak, faster than Google or Meta before their IPOs. And it comes with a projected $559M operating profit. Their first ever, even if it may not hold all year.

So what flipped a money-incinerator into a profit machine?
Two things, working together: enterprise customers and coding agents.

Here's the mechanic most people miss:
You and I pay a flat monthly fee. Enterprises don't.

They pay per token, basically API pricing, and coding agents devour tokens. A single developer can burn millions a day. Multiply that across thousands of engineers, and you get billions of tokens daily from one customer. That's the engine behind the hockey-stick chart.

But the bills just landed, and they're brutal.

Uber put Claude Code in front of ~5,000 engineers and burned its entire $3.4B AI budget for 2026 in four months. Microsoft is cutting most internal Claude Code licenses by June 30 and moving engineers to a cheaper in-house tool.

Now some companies are running the math on a different option: just self-host an open model in-house. Slower, less capable, but a fraction of the cost. To a CFO staring at a 7-figure invoice, that tradeoff suddenly looks very real.

Which leaves the real question. Will the labs move the rest of us onto usage-based pricing too?

Sam Altman has already floated a world where you buy intelligence "on a meter."

I wouldn't bet on us staying in the cheap seat forever. What's your call?

[Rishi Raam @rishiraamns]

@arpit_bhayani Yes, more than 60% of the time I use Claude to find something in the codebase that is buried so deep with no documentation whatsoever.

[Rishi Raam @rishiraamns]

Everyone's saying the same thing: Apple won the AI race, and OpenAI will bleed users.

Maybe.

But the narrative skips the part that actually matters. OpenAI and Anthropic saw this coming, and they've quietly been building for a different game.

Start with the math. For every dollar a subscriber pays, these companies are probably spending more than that just to serve them. The margins aren't there now, and there's a real chance they never show up at scale.

Enterprise flips that. One Fortune 500 contract is worth tens of thousands of individual subscriptions. And enterprises don't only pay for usage. They pay for reliability, compliance, and a support team that actually picks up. None of which a consumer will ever pay extra for.

So watch where the people are going.

OpenAI has 700+ open roles right now. Over 200 are in enterprise sales and support. Anthropic is hiring for 350+, with roughly 100 aimed straight at enterprise.
Job listings are the most honest roadmap a company publishes. You can spin a keynote. You can't fake a hiring plan.

And yes, the OpenAI consumer super app is still coming. It'll ship, it'll get the headlines, it'll keep the brand in front of everyone. But that's the marketing. Enterprise is turning into the actual business.

So here's the part I keep coming back to:

If the big labs go upmarket and Apple owns your home screen, who's left building for everyone in between? For developers who want more capable AI without watching every token burn?

[Rishi Raam @rishiraamns]

@nivi Everyone's offloading their thinking to AI while reviewing it and iterating on it manually.

It should be the other way around. Here's why:
https://t.co/WtCwJ3ty5b

[Rishi Raam @rishiraamns]

Most developers I talk to haven't written a line of code in six months.

Anthropic has revealed that 80% of their code is written by AI.

The productivity gains are obvious. But writing code always did two jobs.

It shipped the feature. It built the engineer.

AI took the first job. Nobody took the second.

Every line AI writes for you is a small loan against your own ability. The balance is called cognitive debt.

Here's how it gets called in and how to pay it down:

You approve a PR on Tuesday. On Friday someone asks why the retry logic works that way and you're reading code with your name on it like a stranger wrote it. Because one did.

The API goes down for an afternoon. A "twenty-minute task" sits open in your editor while you rediscover that you no longer know how to start.

You reached for an ability, and the balance came up short.

Here's what's actually missing in those moments.

When you write something yourself - code, a design, an argument, the output is the visible product. The invisible product is the wiring.

Every problem you push through wires the system into your head: how it behaves, where it bends, where it breaks.

That's the model you were reaching for in every scene above.

Tuesday's PR made it into prod. It never made it into you.

"But I review everything the AI writes."

Be honest about what that review is. With a model in your head, reviewing means holding the code against what you know and catching where they differ.

Without one, reviewing collapses into reading. Scrolling. Nodding. It compiles, tests pass, looks clean and you approve. It feels like critical thinking. And every approval makes the next one easier. That's the interest compounding.

So no, the answer isn't "use AI less."

Using AI doesn't create the debt. Using it first does.

So here's the repayment plan. Three steps, and the order is the point.

1. Create the first version alone. No AI in the room. It will be worse than what the machine would've made. Good. Shitty drafts build the model. Pasted ones don't.

2. Let AI attack it. Not rewrite, attack. What's weak, what's missing, what breaks. You have a model now, so its critique is something you can judge instead of swallow.

3. Take the pointers, close the tab, and rewrite alone. The final pass is where the learning sticks. Skip it and you've just outsourced with extra steps.

One boundary keeps the loop honest: outsource knowledge, protect reasoning.

AI took the first job. Take back the second.

Honest question to close:
how long has it been since you wrote something unaided? Six months? Longer?

[Rishi Raam @rishiraamns]

@steipete You still need to test what it's doing and be active in the loop, unless you have unlimited tokens.

SDE isn't about skill anymore, it seems. You're either token-rich or token-poor.

[Rishi Raam @rishiraamns]

@michalmalewicz The cognitive decline we're seeing from over-relying on AI may be harder to reverse than most people think.

I wrote more about why it's happening, what it means, and how we can avoid it:
https://t.co/WtCwJ3ty5b

[Rishi Raam @rishiraamns]

Most developers I talk to haven't written a line of code in six months.

Anthropic has revealed that 80% of their code is written by AI.

The productivity gains are obvious. But writing code always did two jobs.

It shipped the feature. It built the engineer.

AI took the first job. Nobody took the second.

Every line AI writes for you is a small loan against your own ability. The balance is called cognitive debt.

Here's how it gets called in and how to pay it down:

You approve a PR on Tuesday. On Friday someone asks why the retry logic works that way and you're reading code with your name on it like a stranger wrote it. Because one did.

The API goes down for an afternoon. A "twenty-minute task" sits open in your editor while you rediscover that you no longer know how to start.

You reached for an ability, and the balance came up short.

Here's what's actually missing in those moments.

When you write something yourself - code, a design, an argument, the output is the visible product. The invisible product is the wiring.

Every problem you push through wires the system into your head: how it behaves, where it bends, where it breaks.

That's the model you were reaching for in every scene above.

Tuesday's PR made it into prod. It never made it into you.

"But I review everything the AI writes."

Be honest about what that review is. With a model in your head, reviewing means holding the code against what you know and catching where they differ.

Without one, reviewing collapses into reading. Scrolling. Nodding. It compiles, tests pass, looks clean and you approve. It feels like critical thinking. And every approval makes the next one easier. That's the interest compounding.

So no, the answer isn't "use AI less."

Using AI doesn't create the debt. Using it first does.

So here's the repayment plan. Three steps, and the order is the point.

1. Create the first version alone. No AI in the room. It will be worse than what the machine would've made. Good. Shitty drafts build the model. Pasted ones don't.

2. Let AI attack it. Not rewrite, attack. What's weak, what's missing, what breaks. You have a model now, so its critique is something you can judge instead of swallow.

3. Take the pointers, close the tab, and rewrite alone. The final pass is where the learning sticks. Skip it and you've just outsourced with extra steps.

One boundary keeps the loop honest: outsource knowledge, protect reasoning.

AI took the first job. Take back the second.

Honest question to close:
how long has it been since you wrote something unaided? Six months? Longer?

[Rishi Raam @rishiraamns]

@nxthompson Not a striking shift; it's more of an expected shift. Startups can't afford to pay millions of dollars to Anthropic and OpenAI like what big enterprises are paying.

They are hunting for cheaper options.

[Rishi Raam @rishiraamns]

@GaryMarcus It's expected. SpaceX is on track to get more profits from compute renting than rockets.

[Rishi Raam @rishiraamns]

OpenAI and Anthropic are betting heavily on enterprise adoption as their path to profitability.

But as enterprises get more familiar with AI economics, many will start questioning today's API pricing and looking seriously at open-source alternatives. That shift won't happen tomorrow. Enterprise migrations take time. But the incentives are becoming hard to ignore.

That's why frontier AI labs need another growth engine.

If enterprise revenue alone can't justify the enormous infrastructure costs, consumer products start looking a lot more important.

OpenAI's push toward a "superapp" feels less like an expansion and more like insurance.

[Hedgie @HedgieMarkets]

🦔OpenAI plans to turn ChatGPT into a "superapp" with coding tools, AI agents, and paid integrations ahead of its IPO. The overhaul shifts resources toward enterprise clients, with 2 million businesses already at 40% of revenue and expected to hit 50% by year-end. ChatGPT has 900 million weekly users and 50 million paid subscribers. OpenAI projects a $14 billion loss this year and hasn't filed its S-1.

My Take
Altman said last year that apps would become obsolete because of AI, and now he wants to build a superapp. Coding tools are at the center of it. GitHub Copilot's coding tool billing fell apart last week when users burned through monthly credits in hours.

900 million people use ChatGPT every week, 50 million pay for it, and OpenAI still loses $14 billion a year. The superapp is supposed to change that. OpenAI has shifted hard toward enterprise, with 2 million businesses already at 40% of revenue and a target of 50% by December. But those are the same enterprise customers who discovered last week what AI tools cost and whether they produce anything. SpaceX goes public Thursday. OpenAI is redesigning its app.

Hedgie🤗

[Rishi Raam @rishiraamns]

But I would say the AI profitability question is slowly fading. Anthropic is said to have their first profitable quarter, and OpenAI is also moving towards one.

Both of them have found their product-market fit, which is selling coding agents to enterprises at API pricing.

Their profits will continue to slowly increase to cover their infrastructure costs unless these big orgs move to OSS models, which seem unlikely, at least for now.

[Rishi Raam @rishiraamns]

Most developers I talk to haven't written a line of code in six months.

Anthropic has revealed that 80% of their code is written by AI.

The productivity gains are obvious. But writing code always did two jobs.

It shipped the feature. It built the engineer.

AI took the first job. Nobody took the second.

Every line AI writes for you is a small loan against your own ability. The balance is called cognitive debt.

Here's how it gets called in and how to pay it down:

You approve a PR on Tuesday. On Friday someone asks why the retry logic works that way and you're reading code with your name on it like a stranger wrote it. Because one did.

The API goes down for an afternoon. A "twenty-minute task" sits open in your editor while you rediscover that you no longer know how to start.

You reached for an ability, and the balance came up short.

Here's what's actually missing in those moments.

When you write something yourself - code, a design, an argument, the output is the visible product. The invisible product is the wiring.

Every problem you push through wires the system into your head: how it behaves, where it bends, where it breaks.

That's the model you were reaching for in every scene above.

Tuesday's PR made it into prod. It never made it into you.

"But I review everything the AI writes."

Be honest about what that review is. With a model in your head, reviewing means holding the code against what you know and catching where they differ.

Without one, reviewing collapses into reading. Scrolling. Nodding. It compiles, tests pass, looks clean and you approve. It feels like critical thinking. And every approval makes the next one easier. That's the interest compounding.

So no, the answer isn't "use AI less."

Using AI doesn't create the debt. Using it first does.

So here's the repayment plan. Three steps, and the order is the point.

1. Create the first version alone. No AI in the room. It will be worse than what the machine would've made. Good. Shitty drafts build the model. Pasted ones don't.

2. Let AI attack it. Not rewrite, attack. What's weak, what's missing, what breaks. You have a model now, so its critique is something you can judge instead of swallow.

3. Take the pointers, close the tab, and rewrite alone. The final pass is where the learning sticks. Skip it and you've just outsourced with extra steps.

One boundary keeps the loop honest: outsource knowledge, protect reasoning.

AI took the first job. Take back the second.

Honest question to close:
how long has it been since you wrote something unaided? Six months? Longer?

[Rishi Raam @rishiraamns]

@bradmenezes Except that these frontier AI labs will focus more on enterprise customers.

It seems as if enterprise is their product-market fit. They are going to get their first profitable quarter due to enterprise demand.

[Rishi Raam @rishiraamns]

Some teams are winning big in AI era, while others are falling behind.

The top 5% teams grew code output ~97% YoY, while others grew just 4%.

Nearly every team is prototyping and building new features faster than ever, but the bottleneck seems to be integrating that AI-built code into production.

That's where the top 5% pull away: their main-branch output is up 26%, while the median team's actually fell ~7%.

More code written, less of it shipped.

More on what's driving the gap and how to get ahead in an explainer next week.

[Rishi Raam @rishiraamns]

It feels like frontier AI labs have finally found their product-market fit.

Enterprise demand is pushing them toward their first truly profitable quarters.

What many underestimated was how many tokens get consumed inside large organizations once AI becomes part of daily workflows.

Companies are burning through millions of tokens, and that demand is finally starting to show up in the revenue numbers.

[Rishi Raam @rishiraamns]

@Av8r07 The annual value of SpaceX's compute leasing contracts already exceeds the revenue of its rocket business.

Now imagine what happens if they succeed in renting out compute from space-based data centers.

SpaceX may end up making more money from compute than from rockets.

[Rishi Raam @rishiraamns]

Google designs its own AI chips. It runs its own data centers. It owns more compute than almost anyone on Earth.

Yet, it just agreed to pay SpaceX $920 million a month to rent 110,000 Nvidia GPUs.

The reason is timing.

Demand for AI shows up in weeks. Supply shows up in years. Alphabet is spending $180 billion on its own data centers this year. But Nvidia is sold out for months, and concrete doesn't care how rich you are. Meanwhile, OpenAI and Anthropic compound every month Google waits.

Google isn't renting chips. It's buying time.

The GPUs sit in Memphis, inside Colossus, the complex xAI built to train Grok. When Elon Musk folded xAI into SpaceX earlier this year, those racks became inventory. SpaceX now rents them out by the month.

And Google isn't even the anchor tenant.

In May, Anthropic signed for $1.25 billion a month through 2029. That's $15 billion a year for an entire Colossus site to itself. Google's deal buys roughly half that much compute.

Add it up: two contracts, more than $70 billion, signed in a matter of weeks.

SpaceX's entire revenue last year - rockets, Starlink, all of it was $18.7 billion. The rent alone is worth almost four years of everything else it does.

Next week, SpaceX goes public at a $1.75 trillion valuation. The largest IPO in history. And the pitch isn't rockets.

[Shaun Maguire @shaunmmaguire]

SpaceX just quietly amended its S-1 announcing another mega deal

$920M/month from Google from October 2026 through June 2029

With both parties being able to terminate the agreement with 90 days notice

Things are getting exciting 🚀

[Rishi Raam @rishiraamns]

@abledoc Yeah, and the fact that it may take months till Google gets their hands on chips required from Nvidia.

This deal is just for the time being till Google figure out their supply.

[Rishi Raam @rishiraamns]

@ZackKorman I'll belive in Microsoft the day they fix windows. They can't even perfect an operating system in all these decades, how can they catch up with these accelerating AI companies.

[Rishi Raam @rishiraamns]

@mikeydsoftware Well, research says otherwise. We're losing our cognitive ability, at an alarming rate. And we're seeing signs that it might be unrecoverable.