Skip to content
Back to all essays
George's Takes

Build the Model, or Build the Business?

A five-person lab trained a 65-billion-parameter model from scratch. An AI video company reached a $1 billion annualized revenue pace by owning the workflow. The difference matters to anyone building a career or product in AI.

By George PuEssay16 min read

Published

An unfinished precision mechanism and a finished field instrument sit on separate platforms joined by a narrow wooden bridge.
AI-generated object-led editorial still life about connecting technical invention to a useful, deliverable product; not a documentary photograph.

This month, a 5-person lab released a 65-billion-parameter model it trained from scratch.

A couple of months earlier, an AI video company said it had passed $1 billion in annualized revenue, 18 months after launch - mostly by running other people's models.

Opposite bets. Both working.

Meanwhile, some of the most famous names in AI are finding out how brutal the model race has become. Mira Murati's lab shipped a 975-billion-parameter model and said up front it wasn't the best. Ilya Sutskever's lab is valued at $32 billion and still hasn't shipped anything at all. NVIDIA is organizing other labs around frontier open models designed to run on its hardware.

It's ugly out there, even for the best of the best.

If you run a small AI lab, you have to decide which lesson applies to you. Do you build the model, or build the business?

I'm in the middle of that decision at Vinci - with 2 of our own pretraining runs going as I write this. Before I share where I've landed, here's what the labs making these bets actually did, and what it cost them.

The headline number is lying to you

Every conversation about training a model starts with one question. How much did it cost?

It sounds simple. It isn't, because people answer it with completely different numbers.

DeepSeek-V3 became famous for a figure of about $5.6 million. But read the report. That number is the cost of the official training run, priced at an assumed $2 per GPU-hour, and it explicitly excludes the earlier research and experiments that got them there.

Now compare Ricursive. When it released Ripple-65B Preview this month, it put the cost at $32 million - and that's everything. The experiments, the data work, the debugging, the synthetic data, the recipe development, and the final pretraining and post-training.

Put those two numbers side by side and you'd think one lab was nearly 6 times more efficient. You'd be comparing a receipt for the last step with a bill for the whole journey.

For a small lab, the real cost is the whole journey. The runs that failed. The data you cleaned and then threw out. The evaluation work. The engineering to make the pipeline survive a crash. Deployment.

And the biggest cost of all never shows up on an invoice: everything your team didn't build while they were busy training.

A cheaper training run can still be the most expensive decision you make, if it delays the thing your customers actually needed.

The number I care about isn't the cost of a model. It's the total cost of delivering an outcome a customer accepts. Token prices and benchmark scores feed into that. Neither one replaces it.

The 5-person lab that did it anyway

Ripple is the strongest case for training your own model. So let's take it seriously.

A team that grew from 1 person to 5 built a dense 65-billion-parameter model from scratch, trained on 12 trillion tokens, and released it with open weights. That's a real achievement. It kills the lazy version of my argument - that small labs simply can't do foundation work anymore.

They can.

But look at why Ricursive did it. Ripple isn't a general chatbot looking for a reason to exist. It's the general-purpose version of a model built to run their chip-design platform. They had a specific product, and they designed the model around it.

And small headcount didn't mean small money. 5 people still spent $32 million.

So the lesson I take from Ripple isn't 'every ambitious lab needs its own foundation'. It's the opposite. The labs that should train from scratch are the ones that can finish this sentence: we need our own foundation because [specific reason], and nothing available does it well enough.

Ricursive could finish that sentence. That's why it worked.

Other ways to win

The interesting labs right now aren't all trying to be the best model at everything. They're picking a layer and winning there.

Liquid picked a constraint. Its LFM3-1.7B is built for local agents, tool calling and structured outputs - small enough to run on a laptop or a phone. Liquid's own model card says it's not recommended for agentic coding or complex multistep reasoning, because of its small size. That's not a weakness. That's a company that knows exactly what it's for.

Prime Intellect picked post-training. INTELLECT-3 starts from an existing open base, GLM-4.5-Air, and improves it with fine-tuning and reinforcement learning. Their contribution is the training infrastructure, the evaluation environments and the recipe. And before anyone thinks post-training is the cheap option - they used 512 H200 GPUs.

Hugging Face picked openness. SmolLM3 is a 3B model with the whole recipe written down - data choices, training stages, all of it. That's worth something beyond any one product, because the next lab gets to start where they left off.

None of these is the one right strategy. That's the point. Architecture, special data, training methods, deployment constraints, access to users - each one can be an edge.

What you can't do is skip the choice. 'We'll build a general foundation and figure out the edge later' is how a small lab burns its runway arriving in fourth place.

The company that skipped the race

Then there's Higgsfield. To me, it's the most interesting AI company of the year, and it barely competes in the model race at all.

Higgsfield makes AI video and image tools for creators, brands and studios. It does build its own models - its Soul model for images, among others. But its enterprise plan bundles more than 35 image, video and audio models, most of them from other companies: Seedance, Kling, Nano Banana Pro, GPT Image.

What Higgsfield owns isn't the foundation. It's the workflow, the creative tools, and the customer.

And it worked. Higgsfield says it crossed $1 billion in annualized revenue 18 months after launch, with more than 32 million users, nearly 1 million paying subscribers, and 390 of the Fortune 500 using it. In August it raised $400 million at a $5.4 billion valuation.

It went its own way. It didn't try to out-train OpenAI or Google. It built custom models where they made the product better, used everyone else's where they didn't, and spent its energy on the thing that actually pays: getting the product into people's hands.

The risk is real - when you depend on other people's models, they can change the price or compete with you. But Higgsfield made a bet that the customer relationship is worth more than the foundation. So far, the market agrees.

Even the best of the best

If you think the model race is only hard for small labs, look at what happened this year to the people with the most talent, the most money and the most famous names in the field.

Thinking Machines. Mira Murati, OpenAI's former CTO, founded Thinking Machines in February 2025 and raised a $2 billion seed round at a $12 billion valuation - one of the largest seed rounds in Silicon Valley history. Its founding team included many of the people behind some of OpenAI's best post-training work.

Then came the churn. Before the company shipped a single model, there was a wave of departures, including 2 co-founders who left for OpenAI in January.

In July, it finally released Inkling: a mixture-of-experts model with 975 billion total parameters, about 41 billion active, trained on 45 trillion tokens, with open weights. It's a serious model.

And Thinking Machines said so itself. Its launch post called Inkling 'not the strongest overall model available today, open or closed'. On agentic coding, its own launch table put Inkling at 63.8 on Terminal-Bench 2.1.

Here's the part that should keep every lab founder up at night. Less than 4 months later, Ripple - 5 people, $32 million - reported 80.5 on that same benchmark. Inkling's own number was 63.8. Both are company-reported and benchmark harnesses matter, so take the gap with a grain of salt. But a lab valued at $12 billion, with one of the most talented teams ever assembled, got out-scored on that test by a chip-design startup.

NVIDIA. NVIDIA launched a Nemotron Coalition with other labs to build open frontier models together, and it keeps releasing open Nemotron models, weights and research.

Its flagship Nemotron 3 Ultra model card reports 56.4 on Terminal-Bench 2.1. That is a serious open model, and it still trails the company-reported Inkling and Ripple results on this one test.

But NVIDIA doesn't need Nemotron to win. Every strong open model that runs best on NVIDIA hardware sells more NVIDIA chips. The models are marketing for the real product. A small lab has no chip business to pay for its models.

Safe Superintelligence. Ilya Sutskever co-founded SSI in June 2024, promising one product: a safe superintelligence, with nothing shipped before it. It raised $1 billion that September, then $2 billion more at a reported $32 billion valuation. A co-founder, Daniel Gross, left for Meta's superintelligence lab in 2025.

More than 2 years in, SSI has not publicly released a model, product or paper. In July, NVIDIA signed on as a strategic partner, and Sutskever said: 'We have research that is worthy of scaling up.'

To be fair, nothing public doesn't mean nothing's cooking. Sutskever has been right about this field before, more than once. But the rest of the market isn't waiting for him.

And even OpenAI shuts things down. In April, OpenAI discontinued Sora's web and app experiences. It later discontinued the API too.

Sit with that for a second. The most famous AI company in the world built one of the best video models in the world - and walked away from it as a product. The AI video company that's winning right now is Higgsfield, the one that never needed to win the model race.

Why OpenAI can afford to lose money and you can't

Now look at the 2 companies that are winning.

OpenAI says its products now reach more than 1 billion weekly active users and 2.5 million businesses. In February it reported more than 50 million consumer subscribers. Those point-in-time groups are not perfectly comparable, but the rough ratio is still about 1 consumer subscriber for every 20 weekly users. The free audience is enormous, and every free question still consumes compute.

Anthropic took the other road: an enterprise-heavy business that also has a consumer product. Its annualized revenue pace topped $65 billion by the end of July, according to Reuters' source. Its later IPO filing says customers generating at least $100,000 in trailing revenue accounted for 74% of revenue in the first 9 months of 2026. The filing also shows $1 billion of operating income in the second quarter, followed by a $567 million operating loss in the third. Distribution is powerful. It is not magic.

Different strategies. Same advantage: distribution.

This is the part people get wrong about the $20 a month plan. It isn't a pricing mistake. It's a land grab. When you already have the users, the habit and the brand, losing money on heavy users is just the cost of staying the default. Every person who opens ChatGPT or Claude first thing in the morning is a person your new model has to pull away from something they already use, already trust and already pay for.

And that's the real reason the model race is so ugly for everyone else. A better model alone doesn't win anymore. A slightly better model with no distribution loses to a good-enough model that's already on a billion phones.

The winners aren't just the labs with the best model. They're the ones who own the front door.

Where I've landed

I've spent years telling people to own or be owned.

Own your customers. Own your distribution. Own your compute. Don't build your whole business on someone else's platform and act surprised when they change the price.

So here's the irony. Even as I write that pretraining is expensive and hard to justify, we're doing it.

Right now, Vinci has 2 pretraining runs going from scratch: a 4.85-billion-parameter model and a 1-billion-parameter model. Hundreds of millions to billions of tokens have gone through them so far, and more are coming.

By frontier standards, that's tiny. Ripple trained on 12 trillion tokens. That's the point. We're not racing anyone. We're learning.

And the learning is real. Earlier this month we reviewed our first from-scratch run. The loss kept falling - on text the model hadn't seen, it kept getting better at predicting what came next. For a small lab, that's a real moment. It means the thing works.

The rest of the review was less fun. The final evaluations weren't done. The checks on which data and settings produced which checkpoint were incomplete. We hadn't finished verifying that the test text didn't overlap with the training text. And one attempt to continue a run failed before any useful training happened.

None of that means the model stopped learning. Those are engineering and measurement gaps, not a ceiling. And every one of them taught us something about how the whole system works, end to end.

So I think of our pretraining as a learning journey, not a product launch. We probably won't publish these checkpoints. We'll keep improving them, find the use case where our own foundation actually beats what's available, and then release our work to the public.

And we're adopting too. Where someone else's model already does the job, we use it. Owning the ability to build doesn't mean refusing to use what's already built.

But here's the honest math.

Without hundreds of millions of dollars, I can't see how a small lab gets its money back from training a general-purpose LLM from scratch. The free and cheap alternatives keep getting better while you train. You can spend a fortune and arrive exactly where everyone else already is.

And even if I had hundreds of millions - or billions - I'd still put most of it into go-to-market, not into another LLM. Product. Distribution. Customers. The part Higgsfield got right, and the part OpenAI and Anthropic are spending billions to own.

That's the part of 'own or be owned' I had slightly wrong in my own head. Owning something doesn't mean rebuilding it yourself. It means you're never trapped. You understand the whole system, you could build it if you had to - and you put your money where owning it actually changes the outcome.

For most AI companies right now, the thing worth owning is the customer.

The best case for training more

I don't want to argue against a strawman, so here's the strongest case for doing more pretraining, not less.

Serving costs can justify it. Research that counts the cost of running a model, not just training it, finds that if you expect enough demand, it can pay to train a smaller model for longer. You spend more up front and save every time someone uses it. At real scale, that adds up.

Some things you can't get any other way. Unique data. Full control over what went into the model and where it came from. A new architecture. A deployment environment the available models serve badly. Any of those can justify starting from zero.

And independence is worth something on its own. A country or a company that can build its own foundation has options the others don't. That value doesn't show up as a better benchmark score. It shows up the day someone else changes their terms.

I believe all of that.

But each reason needs its own evidence.

If the argument is independence, name the dependency you're removing. If the argument is cheaper serving, show a realistic estimate of demand. If the argument is a better foundation, compare it against a strong alternative, under the same conditions, on the same tasks.

'Sovereignty' is a reason. It isn't a blank cheque.

How I'll decide at Vinci

So here's how I'm going to make the call. Not as a rule for everyone - as the rule for us.

1. Start with the user and the task. A general assistant, a coding agent and a model running on someone's laptop need completely different things. 'Better model' means nothing until you say better at what, for whom.

2. Find the best thing already available, and measure the real gap. Not just answer quality. Reliability, speed and cost too. If an available model already does 90% of what the user needs, the question is how to close the last 10%, not how to rebuild the first 90.

3. Try the cheaper levers first. Better data. Better post-training. Better evaluation. On an existing foundation. For near-term performance, that's where our next unit of effort goes before anyone proposes a bigger from-scratch run.

4. Make the next experiment test a reason our own foundation could matter. Maybe our data gives a repeatable improvement that adapting someone else's model can't match. Maybe a new architecture changes what it costs to run. Maybe a customer needs provenance an existing base can't provide. Small runs can test those ideas cheaply - knowing a win at small scale still has to hold up at full scale.

5. Keep 2 budgets, and never mix them. One for learning. One for shipping. Our current runs are learning runs, and I can say that plainly without pretending they're a commercial breakthrough. Learning is a legitimate goal. It just needs a different definition of success than a product does.

6. Decide in advance what earns more compute. For me, that's a measured advantage, a credible explanation for why it exists, and a plan to check that it survives outside the original experiment. Not 'we've already spent too much to stop'. That's the most expensive sentence in research.

Not closing the door

I'm not telling anyone to give up on foundation research. Someone has to do it, and some of the most important work in this field came from small teams that refused to accept the default. Ripple just proved a small team still can.

And we're doing it at Vinci, right now.

But the bar has moved. If an available model already gives users most of what they need, a small company has to explain why rebuilding that foundation beats improving the remaining gap - or just getting the product into people's hands.

So before we scale any of it up, I'm going to ask one question. I'd suggest every small lab ask it too.

What will owning this foundation let us do that we can't do well enough today?

When we can answer it, we'll scale up and share what we built. Until then, we'll keep learning on small runs, use the best models available - and build the business.

Share this:

New essays and letters, by email

Subscribe if you would like to hear from me when I have something worth sending. Everything here remains free to read.

Optional. Unsubscribe anytime.