Why & How Every Company Will Build Its Own AI: My TechCrunch Disrupt Recap

Speaking at TechCrunch Disrupt last week was one of the best rooms I’ve been in all year. I was really looking forward to having Matt Burns moderate the talk.

More than 10,000 people came through Moscone West for 200+ sessions and 300+ startups. TechCrunch framed the week as "a big focus on building AI in the real world, not just hype and theory."

On Wednesday, I took the “Industry - Real World AI” Stage for a session called “Why and How Every Company Will Build Its Own AI.” The audience was mostly founders and builders, and they didn’t come for theory. They came with real-world questions and looked for pragmatic solutions.

This is my recap, built to skim: why every company will build its own AI, what to build, how to do it, the questions from the room, and what I took away.

Why: The three reasons every company will build its own AI

In May, the CTO of a top 10 global bank told us: "We're using frontier AI models so that we don't fall behind." By August it became: "We are moving to open models." That's the whole shift. The safe bet in May became the way to fall behind.

From 2023 until about June 2026, most enterprises ran on GPT, Claude or Gemini APIs, including the Gemini my team was building at Google. Now they're moving to specialized intelligence they build and own. Three reasons.

1. Differentiation: Rented AI won't be your moat; only owned specialized AI can be

If a company wants to win in this AI era, it has to build its own AI. Prompting the same models from Anthropic, OpenAI or Google as everyone else isn't a strategy for differentiation or building IP. It’s the opposite. It leaks your IP to frontier labs and keeps you from building your own.

As AI becomes a bigger part of what you build, prompting the same model as your competitors shrinks your differentiation. Build on your own expertise and data and you differentiate right away. Run it as a flywheel: the more you use it in production, the more the model improves, and the gap compounds.

Satya Nadella made this case in June in a post called "A frontier without an ecosystem is not stable." My TL;DR: build, own and compound your intelligence. Don't rent it. Don’t outsource it.

Revolut built its own AI lab and reports much better results on fraud and credit risk. Its head of AI, Pavel Nesterov, put it well: you can't rely on third-party blueprints. Coinbase achieved similar results about a week ago as well. Harvey wrapped frontier models for a long time, then said a few months ago it's building its own, for more intelligence at lower cost. Kirkland & Ellis, a law firm, said AI will be key to its differentiation, so it can't use the same AI as its competitors.

2. Economics: It makes no sense to pay for a thousand skills you don't use

A model built to be good at a thousand things has to be big, slow, and expensive. You need one thing done exceptionally well. A specialized model can be 5 to 100 times smaller, faster, and cheaper. That means higher margins, and a competitor on that curve can outprice you.

3. Control: For your critical AI, you can’t be at a vendor's mercy

Access can be cut for geopolitical reasons. Terms change. Models get deprecated and you're forced to migrate. Some labs now compete with their own customers. And if you don't control the model, you don't control where your data goes or how long it's kept.

What: Build and own the AI that matters, rent the rest

You don't need to replace every GPT, Claude and Gemini call with your own model. Be intentional. The shift is about moving business logic out of brittle code and prompts and into a model trained for your domain. The harness around it gets much thinner.

To choose where, score each use case on two axes: volume, roughly what you spend on it, and how much it differentiates your business.

Low-differentiation work like drafting emails can stay on frontier models. High volume and high differentiation, like a bank's risk analysis or a personal banking agent, is what you build first. Highly differentiating but low volume is still worth building. And if a common task like document processing has no off-the-shelf option with the right quality or economics, build that too.

How: Use AI to build AI

The bottleneck is orchestration, not GPUs or point solutions

Building specialized AI means defining the problem, building the harness and tools, writing evaluations, curating data, training, deploying, and improving. From small startups to the largest enterprises, we heard the same thing. The blocker isn't GPUs, open-weight models, or point solutions. It's the effort and expertise to run the whole process end to end.

Automate the AI engineer's loop

Every company needs an AI engineering platform that both AI experts and domain experts can use. We built ours around an agent. Give it a prompt you already use. It builds the tools and evaluations, finds where the model falls short, curates data to fix it, trains, and repeats. Then it deploys, monitors quality with the same evaluations, and keeps improving. Work that took weeks or months can often take 15 to 30 minutes, plus a couple of hours of training, for the moderate-size models most enterprises train.

Own the recipes, not just the weights and data

Whatever you use, insist on full transparency and sovereignty. You should own not just the weights and the data but also the recipe for every step that can run on an open-source stack, so you can walk away with all of it. This is too important to compromise with anything less.

Questions from the room

Matt Burns opened with a few questions. Then the audience took the mic.

Why not just Claude plus a context layer?

Matt had recently met a group of AI companies in New York. Every one used open-weight models for client-facing tasks and Claude with a custom context layer internally. My answer: it depends on the task. We use Claude and GPT internally too, mostly for coding. But context inserts narrow knowledge for the specific question. It doesn't give the model a deep understanding of how your business makes decisions. Also, context engineering may be acceptable for prototypes, but it's often not for production. Teams keep inflating context to fix issue after issue until it's unreliable and unmaintainable. That's when they start training.

Coinbase beat Opus 4.5 for $100. Why does anyone need you?

I wasn't surprised by Coinbase's results. A model asked to do a thousand things, without your data, can't beat a specialized one. What we hear isn't about the cost of training on a GPU that could be just $100 for the size of models most enterprises train. The true “cost” is the sheer amount of human AI engineering effort and expertise required. That could be in the order of tens or hundreds of thousands of dollars. We turn that cost down by 10x-100x by automating AI engineering with AI itself.

Why did OpenAI shut down its build-your-own-model offering?

It mostly let you train a model and left the rest of the engineering to you. The decision also came before the inflection point we started seeing in June. And one model that learns from everyone is simply better for a lab than many forks.

Who buys, and who's the right fit?

Entry points range from engineers who find us to executives. The fit is any company whose differentiation is its AI, not its ML infrastructure. If you are one of the top 5 frontier AI labs and ML infra is your edge, build it yourself.

Which modalities, and is synthetic data included?

Open source covers text and images. The enterprise platform is text-to-text for now. Data synthesis is built in, because synthesis, training, and evaluation form one closed loop. Automate two of the three, and it still takes more than a month.

Where does the moat live: data, evals or harness?

All of them. Evals are critical: they define what good means and steer every round of improvement. With zero data, you can still distill a model, ship it, and let production data improve it.

What about knowledge in people's heads?

That's why the tools have to work for domain experts, not just AI PhDs. Our head of marketing builds one-off apps with coding agents and specialized AI with Oumi.

When should a business-first company with a small team start?

The sooner the better. The earlier your model compounds in production, the harder you are to catch. Waiting any longer is the bigger risk, not starting now.

Key takeaways

Your specialist intelligence is going to be the moat. You can’t rent it or outsource it. Build it.

The debate has moved from whether to how. Hardly anyone is debating anymore whether owning your AI makes sense. They are asking what AI to prioritize building and how to get there.

Money isn't the barrier. Skills are. Coinbase's $100 shows training on a GPU can be cheap. The bottlenecks are putting in the effort and having the expertise to do so. That’s why Oumi automates AI engineering with AI itself.

Agents were the loudest topic all week. Agents are having a significant impact and most discussions on AI involve agents. On November 4 at our Fall Launch, we’ll show how Oumi automates the building of specialized AI agents. It’s 35 minutes, online, at 10 AM PT: Register to join us at oumi.ai/events/fall26.

Thank you to TechCrunch for having us, to Matt for the exceptional moderation, and to everyone who packed the room.

Can’t wait for the next one!

Manos