Neolab

A neolab, also written neo-lab or NeoLab, is an informal term for an AI research company organized around a technical research program.^[1] Its meaning varies: some investors restrict it to research-first model developers, while other participants include application-oriented companies.^[2][5]

The term describes an organization and its research strategy, rather than a particular model architecture. Its boundaries are disputed even among participants. A company can conduct substantial AI research while declining the label.^[4]

Meaning and boundaries

The following are examples of how named participants use the term, not competing technical standards.

Leonis explicitly excludes Harvey from its research-first definition because the company's model training serves its legal software business. This is a classification criterion, not a finding that Harvey does no research.^[2]

Agrawal likewise describes substantial work on search models while rejecting the neolab label for Parallel. Research activity alone therefore does not settle the classification.^[4]

Research strategies

Companies discussed as neolabs pursue different objectives. The examples below describe stated research programs; they do not establish that the companies have achieved their goals or share one development method.

Thinking Machines presents deployment as a source of feedback for research and commits to sharing technical work through papers, code and blog posts. Its stated approach differs from Safe Superintelligence's emphasis on concentrating on a single long-term objective without ordinary product-cycle pressures. These are the companies' descriptions of their operating approaches, not guarantees about publication practices or future results.^[6][7]

The research relationship can also extend beyond the lab's own product. In a presentation hosted by Sequoia, Harvey describes collaborations with multiple specialized providers on post-training and related research. It names partners working on model training, enterprise knowledge and inference techniques, explaining that different teams pursue different research bets. This illustrates one use of "neolab" for a research partner within a larger application-development process.^[10]

Compute and research costs

The economics of a research company depend partly on what it is trying to train and test. Ben Cottier and colleagues' study of frontier models, revised in February 2025, estimates costs through hardware, energy and cloud-rental approaches and also considers research staff and experiments. Accelerator hardware and staff are major components in the frontier-model cases it examines; server equipment, interconnects and electricity add other expenses.^[11]

The study distinguishes the cost assigned to a final training run from wider model-development spending. Buying a cluster and allocating a fraction of its useful life to one run are different accounting choices. Renting compute gives another estimate, influenced by the rental rate. A dollar figure for a completed training run consequently should not be read as the full budget needed to establish and operate a research company. The paper studies frontier-model development, not a representative sample of neolabs, so its findings do not set a minimum budget for every company using that label.^[11]

NEA's investment thesis emphasizes resources beyond cluster size. It argues for specialized training data, task environments and reliable checks on model outputs, including reinforcement learning systems with domain-specific feedback. That is an investor's account of potential competitive advantages. It does not establish that every successful laboratory must use the same training procedure or that greater spending alone produces better research.^[1]

Financing and evidence

Venture capital can finance a research team before sales provide a way to evaluate its business. In his analysis of neolab investment, Tu separates the price paid at entry, the value achieved at an eventual exit, and dilution from later financing or employee equity. His scenarios illustrate how capital requirements and changes in ownership affect returns. They are conditional calculations, not observed returns for the category or predictions for a particular laboratory.^[3]

Leonis argues that research-company valuations can move ahead of comparable public evidence. Hiring, compute commitments and private evaluations offer different signals from customer retention or revenue. Its analysis warns that a financing schedule demanding rapid milestones can also influence research priorities. These are claims about investment and governance incentives, not evidence that a specific lab has misrepresented its progress.^[2]

Company announcements provide a narrower kind of evidence. Noetive's September 16, 2026 announcement reports a $41 million seed round and describes relationships with design partners. Such an announcement documents the company's stated financing and development plans; it does not independently validate every claimed capability or customer outcome. A financing event, a deployed product and a reproducible research result answer different questions about a company's development.^[9]

Relation to wider AI research

Academic work on research resources predates much of the neolab discussion. In a 2020 preprint, Nur Ahmed and Muntasir Wahed analyzed 171,394 papers from 57 computer-science conferences. They reported increasing participation by large firms and highly ranked universities after the rise of deep learning, and examined unequal access to computing resources as an explanation for that divergence. The authors called this a "compute divide." Their study concerns institutional participation in AI research; it does not measure the later neolab startup cohort.^[12]

Ahmed and Neil Thompson's 2023 policy analysis describes the overlap between basic and applied AI research and the concentration of data, computing resources and researchers in industry. It argues for supporting academic participation and independent examination of consequential AI systems. This provides context for debates about who can afford research and who can assess its results. It does not show that creating a new private lab necessarily improves access, reduces concentration or serves the public interest.^[13]

The practical meaning of "neolab" therefore depends on the particular source and company. A useful description identifies the research objective, the relationship between research and products, and the evidence available for the work, rather than treating the label itself as a measure of technical quality.^[4][7]

References

1. ^1 ^2 ^3Thomas Joshi, Madison Faulkner, Lila Tretikov and Andrew Schoen. "The Neolab Wild West". NEA, June 17, 2026.
2. ^1 ^2 ^3 ^4Liang Wu, Jenny Xiao and Jay Zhao. "We Finally Got Our Flying Cars, They're Called Neolabs". Leonis Capital, September 2, 2026.
3. ^1 ^2 ^3 ^4Kevin Tu. "Presuming Success: The Math Behind Neolabs". Signal Processing, February 2, 2026. Investor analysis with illustrative assumptions.
4. ^1 ^2 ^3 ^4Sequoia Capital. "Parallel's Parag Agrawal: Building a New Web for AI Agents". Training Data, episode 98. Interview transcript, accessed September 29, 2026.
5. ^1 ^2Sequoia Capital. "Box's Aaron Levie On Reinventing Yourself in the AI Age and Enterprise Diffusion". Training Data. Interview transcript, accessed September 29, 2026.
6. ^1 ^2Safe Superintelligence Inc. Company statement. Accessed September 29, 2026.
7. ^1 ^2 ^3Thinking Machines Lab. Company research and product principles. Accessed September 29, 2026.
8. ^Sequoia Capital. "Memory and Continual Learning: Engram's Dan Biderman and Jessy Lin". Training Data, episode 90. Interview transcript, accessed September 29, 2026.
9. ^1 ^2Noetive. "Noetive Emerges From Stealth With $41 Million Seed to Build the Intelligence of Record for the Physical Economy". September 16, 2026.
10. ^Sequoia Capital. "Building Frontier AI at the Application Layer". Presentation transcript, accessed September 29, 2026.
11. ^1 ^2Ben Cottier, Robi Rahman, Loredana Fattorini, Nestor Maslej, Tamay Besiroglu and David Owen. "The Rising Costs of Training Frontier AI Models". arXiv:2405.21015, version 2, February 7, 2025; first submitted May 31, 2024.
12. ^Nur Ahmed and Muntasir Wahed. "The De-democratization of AI: Deep Learning and the Compute Divide in Artificial Intelligence Research". arXiv:2010.15581, October 22, 2020. Preprint.
13. ^Nur Ahmed and Neil C. Thompson. "What Should Be Done About the Growing Influence of Industry in AI Research?". Brookings, December 5, 2023.