An applied AI lab is a research company that starts from work a business already does, trains models for that work, and measures them on real traffic. It earns from its first deployment. Frontier labs build the most capable general models, and neolabs raise large sums for a research bet before they have a product.
This essay sets the three side by side, with facts sourced and dated as of October 2026. Ertas AI publishes this site and is an applied AI lab in Brunswick, Melbourne, so it serves as the worked example in the last part.
What is an AI research lab?
An AI research lab is an organisation, or a team inside one, whose main work is developing and testing new AI methods and models. Labs sit in universities, in government science agencies and in companies, and they publish, deploy or sell what they find. Our map of AI research labs in Australia shows all three settings in one country.
Among companies, three kinds of lab now carry most of the attention. They differ in what they build, how they are funded and when they start selling.
| Kind of lab | What it builds | When it sells | Examples |
|---|---|---|---|
| Frontier lab | The most capable general-purpose models | From launch, through apps and APIs | OpenAI, Anthropic, Google DeepMind |
| Neolab | A research bet on a future breakthrough | Often after the research lands | Safe Superintelligence, Thinking Machines Lab |
| Applied lab | Models trained for specific, repeated work | From its first deployment | Ertas AI |
The three overlap at the edges. A frontier lab runs applied teams, a neolab can ship a product early, and an applied lab publishes research. The table describes where each one starts.
What do frontier AI labs do?
A frontier AI lab builds the most capable general-purpose models of its time and sells access to them, usually through a chat app and an API. The Frontier Model Forum, an industry body, defines a frontier model as "a general-purpose model that outperforms, based on a range of conventional performance benchmarks or high-risk capability assessments, all other models that have been widely deployed for at least 12 months" (Frontier Model Forum, read 29 October 2026).
Frontier labs train on general data at very large scale, so one model serves millions of users across almost every task. They also carry most of the money in the field. New Enterprise Associates (NEA) put the aggregate private valuation of AI research labs at roughly $1.68 trillion, with 93.6% of it in frontier generalists, and named OpenAI, Anthropic, xAI and Safe Superintelligence as absorbing almost all of it (NEA, 17 June 2026, citing PitchBook valuations as of 30 April 2026).
What are the big 3 AI labs?
No official list exists. The phrase usually means OpenAI, Anthropic and Google DeepMind, the developers of ChatGPT, Claude and Gemini. The closest thing to a formal roster is the Frontier Model Forum, whose members as of 29 October 2026 are Amazon, Anthropic, Google, Meta, Microsoft and OpenAI (membership page).
What is an AI neolab?
A neolab is a research-first AI startup, usually founded by researchers who left an established lab, that raises large sums to pursue a long-term breakthrough before it has a product. Business Insider described neolabs on 22 October 2026 as "the lavishly funded, research-first AI startups founded by top talent who defected from the likes of OpenAI or Google DeepMind" (Business Insider).
The numbers behind the word, as reported in October 2026:
- 29 neolabs had raised a combined $23 billion, by a count from Radical Ventures, and fewer than 35% had released a commercial product (Dealroom, 18 October 2026, summarising The Economist).
- Just over half of those 29 had released any code, often research demos only (same source).
- Safe Superintelligence was valued at $32 billion and has said it will ship no product until it reaches "superintelligence" (same source).
The Economist ran the story on 15 October 2026 under the headline "AI neolabs are raising huge sums. Most have little to show for it".
Investors are split on the model. In the same Business Insider piece, Menlo Ventures partner Matt Murphy named neolabs when asked which sector he thought was most overhyped, while Radical Ventures' Aaron Rosenberg said, "We're confident that the best neolabs will attract ongoing support." NEA's June essay makes the bullish case: it expects the next generation of venture outcomes to come from "research-first companies that own a domain-biased corpus, run an integrated RL and verification loop perhaps inside a high-stakes domain", and notes that neolabs raise $50 million to $500 million before they have revenue.
What does "applied AI" mean?
Applied AI means using AI methods on a defined problem in a real setting, and judging the result by whether it works there. The research questions come from the work: what a model needs to know to do this task, what size of model holds the quality bar, and what it costs per task to run.
Universities use the word for the same instinct. Deakin University's Applied Artificial Intelligence Initiative and Macquarie University's Applied AI Research Centre both build their research around industry problems (AI research labs in Australia, checked 28 October 2026).
What is an applied AI lab?
An applied AI lab is a company that does AI research in order to deploy it: it picks problems from work businesses already do, trains models for that work, and measures them in production against the alternative the business would otherwise use. Three traits set it apart from the other two kinds of lab:
- The problem comes first. The research agenda follows the tasks businesses repeat at volume, so each result has a buyer waiting for it.
- Deployments are the test. A method counts once it holds on a real business's traffic, measured against that business's own quality bar.
- It earns from the first deployment. Each build brings in revenue, and each model that keeps running brings in more, so the research is funded by the work it improves.
Venture investors have started describing a neighbouring idea. Aspiring for Intelligence argued on 25 September 2026 that enterprises are becoming producers of their own models, which it called the "neo neo lab", where "the 'lab' is less about pretraining and more about owning the process that turns what an organization knows into how a model behaves" (Aspiring for Intelligence). An applied lab gives a business that process without the business having to build a lab of its own.
How does an applied AI lab work? Ertas AI as the example
Ertas AI runs on a simple loop: the lab's research picks the problems, design-partner deployments prove the method on real traffic, and each deployment leaves a reusable piece in the product.
Research picks the problems. The lab's research runs under three themes: right-sized specialists, models that learn from traffic, and agentic unit economics. The third theme points at where the money goes. In OpenRouter's public data for the 30 days to 12 October 2026, classification, extraction and transformation took 20.8% of classified model spend. Those are repeated tasks with checkable outputs, which makes them the first candidates for a trained model.
Research gets published with its limits. In the lab's QMSum study (arXiv 2609.25028, 18 August 2026), a model trained on Ertas scored 35.41 ROUGE-1 against 32.43 for the best of five hosted proprietary models on query-focused meeting summarization. The hosted models ran zero-shot under one concise prompt, their longer outputs contribute to the gap on this reference-overlap metric, and no human or factuality evaluation was run. The paper says so next to the numbers.
Deployments prove the method on real traffic. A design partner brings a task its AI agents repeat. The lab trains a specialist model for that task and serves it through the adaptive model endpoint, an OpenAI-compatible API. The endpoint improves through continual learning: new versions train on the agents' traffic in the background, are scored against the customer's evaluations, and go live only with their sign-off. If a new version falls short, the current model stays in service. The endpoint is currently invitation only, and we are developing a more streamlined version for wider access.
Each deployment leaves a piece in the product. An evaluation harness, a data-preparation step or a training recipe built for one deployment becomes part of the endpoint and the Ertas platform, so the next team starts further along. That is the path from lab to product: the deployments fund the research, and the research becomes the product.
For the business on the other side, the case for a trained model rests on two cost dimensions. A trained specialist runs at a lower price per token than the frontier model it replaces, and it needs fewer tokens per task, because the instructions and examples repeated in every prompt are trained into the model. Our cost-per-task calculator shows both effects on your own numbers, and the note on AI startup gross margins works through them with real prompt sizes.
How is an applied AI lab different from a neolab?
The two differ on where the problem comes from and when revenue starts. A neolab chooses a research bet, raises enough to pursue it for years, and builds a product once the research lands. An applied lab takes its problems from businesses' existing work and is paid as each model goes into service, so its research and its revenue grow from the same deployments.
Both are research-first in their own way, and both can be independent of big tech and universities. The difference shows in the questions each one asks first: a neolab asks what new capability it can create, and an applied lab asks which repeated task it can do better, faster and cheaper with a model trained for it.
Is applied AI worth it?
Applied AI is worth it where a task repeats at volume, its output can be checked, and its cost shows up on a monthly token bill. Those conditions give a trained model a quality bar to meet and a cost to beat. A task that changes every week, or runs a few thousand times a month, is better served by prompting a frontier model while it settles; our post on whether fine-tuning is dead sets out where that line falls.
A business can run a frontier model and trained specialists side by side: the generalist for open-ended work, each specialist for the job it was trained on. Our explainer on foundation models and task-specific models covers how that split works.
Frequently Asked Questions
What is a neolab in AI?
A neolab is a research-first AI startup, usually founded by researchers who left an established lab, that raises large sums for a long-term breakthrough before it has a product. As of October 2026, Dealroom reported 29 such labs had raised $23 billion combined, with fewer than 35% having released a commercial product.
What does an applied AI lab do?
An applied AI lab does AI research in order to deploy it. It picks problems from work businesses already do, trains models for that work, measures them on real traffic, and earns from its first deployment.
What is an example of an AI lab?
OpenAI, Anthropic and Google DeepMind are frontier labs; Safe Superintelligence and Thinking Machines Lab are neolabs; Ertas AI in Melbourne is an applied AI lab. University institutes and government science agencies run AI research labs too, such as those on our map of Australian AI labs.
How many frontier AI labs are there?
There is no fixed count, because "frontier" moves as models improve. The Frontier Model Forum's membership is one reference point: six firms as of 29 October 2026 (Amazon, Anthropic, Google, Meta, Microsoft and OpenAI).
What are the top 10 AI research labs in the world?
Rankings depend on the measure. By private valuation, NEA found that frontier generalists held 93.6% of roughly $1.68 trillion as of 30 April 2026, led by OpenAI, Anthropic, xAI and Safe Superintelligence. By published research, university and corporate labs rank differently, so pick the measure that matches your question first.
What is the difference between an applied AI lab and a frontier lab?
A frontier lab builds one general model for millions of users and sells access to it. An applied lab trains specialist models for specific work a business repeats and serves them for that business.
The Short Version
Frontier labs build the most capable general models. Neolabs raise large sums for a research bet and sell later, if the bet lands. An applied AI lab starts from work businesses already do, trains models for it, proves them on real traffic and earns from the first deployment. Ertas AI is an applied AI lab in Brunswick, Melbourne; its research page lists what it has published, and the about page explains how the lab works.
Sources, read 29 October 2026: Dealroom (18 October 2026); The Economist (15 October 2026, headline); Business Insider (22 October 2026); NEA, "The AI Neolab Wild West" (17 June 2026); Aspiring for Intelligence, "The Rise Of The Neo Neo Lab" (25 September 2026); Frontier Model Forum about and membership. Our research: the QMSum study and the OpenRouter structured-task report.