The most underrated AI startups of 2026

The AI startup conversation in 2026 is dominated by a handful of names. OpenAI, Anthropic, and xAI between them account for the majority of venture capital, media coverage, and public attention. That is understandable given their scale, but it creates a large blind spot. Some of the most consequential work in AI is happening at companies that most people outside the developer community have not heard of.

This is a list of AI startups that are underrated relative to their impact. Not underfunded or unknown within their niche, but genuinely under-discussed given what they have built and where they sit in the stack.

Exabase

Exabase is a data layer for AI agents covering memory, search, file storage, extraction, and isolated workspaces. It grew out of Fabric, a consumer knowledge management product, which means the infrastructure was battle-tested against real users before it was offered as a platform.

The memory engine, M-1, holds state-of-the-art results on both LongMemEval and BEAM at every evaluated scale, including 10M tokens, the equivalent of roughly a year of daily agent conversations. Both results were achieved with Gemini 3 Flash while every competitor used Gemini 3 Pro, a model four to six times more expensive. M-1 also consumed approximately 20% fewer tokens per query than the next best system on BEAM.

Exabase is underrated because the agent memory category is new enough that most developers have not yet evaluated it seriously. The benchmark results suggest they should. The broader platform, combining memory with Deep Search, Extract, Resources, and Bases, covers ground that most teams are currently assembling from five or six separate tools.

exabase.io

Braintrust

Braintrust is an AI evaluation and observability platform that treats evals as the centre of the development workflow rather than an afterthought. It raised an $80M Series B in 2026 at a reported $800M valuation, which signals category interest, but it remains far less discussed than the agent frameworks and model providers it supports.

The core loop is straightforward: production traces become evaluation test cases, regressions are caught before they reach users, and prompt or model changes are tested against real inputs rather than synthetic benchmarks. Notion, Stripe, Dropbox, and Replit are among its customers.

Braintrust is underrated because evaluation is not as visible as inference or orchestration. It does not generate demos or product announcements. It prevents failures that would otherwise be invisible until a user reports them. That is a harder thing to market but a more important thing to have.

braintrust.dev

Factory

Factory builds AI agents for software engineering workflows. It ranked number one on the Terminal-Bench benchmark by outperforming major competitors in multi-agent collaboration, debugging, and infrastructure tasks. Its customers include Bayer, EY, MongoDB, and NVIDIA.

Factory is underrated because the coding agent space is dominated by Cursor in terms of attention. Factory is solving a different problem: not helping an individual developer write code faster, but automating engineering workflows at the team and organisational level. That is a less photogenic problem and a larger market.

The company has raised around $70M from Lux Capital, Sequoia, NEA, J.P. Morgan, and NVIDIA. Fast Company named it one of the most innovative AI companies of 2026.

Groq

Groq built custom silicon designed specifically for LLM inference. The Language Processing Unit delivers extremely fast inference at competitive cost, and the NVIDIA licensing deal in 2026, reportedly around $20B, validated the architecture at the highest possible level.

Groq is underrated not because people have not heard of it, but because the conversation about inference has been dominated by GPU cloud providers like CoreWeave and Fireworks, while Groq represents a fundamentally different approach: purpose-built hardware rather than optimised software on general-purpose chips.

Whether Groq scales as an independent inference provider or becomes the technology underneath other platforms, the LPU has proven that custom inference silicon is a viable category. The company is raising $650M in 2026 to fund its transition to an inference cloud business.

Zep

Zep approaches agent memory through a temporal knowledge graph called Graphiti. Every fact gets timestamped, so the system tracks not just what was true but when it was true and whether it still is. This makes it particularly strong for workloads where context evolves over time: support agents tracking customer issues, project assistants following status changes, and any agent where the answer to the same question changes depending on when you ask it.

Zep is underrated because it sits in the shadow of Mem0 in the agent memory category. Its approach is architecturally distinctive and the temporal graph is a genuine differentiator for the use cases it targets. The underlying Graphiti engine is open source.

Harvey

Harvey is a legal AI company that has built deep vertical AI for law firms and in-house legal teams. Its $3B+ valuation and rapid adoption across AmLaw firms make it one of the most successful vertical AI companies in any domain.

Harvey is underrated in the broader AI conversation because vertical AI does not generate the same attention as horizontal platforms. Building AI that understands legal workflows, citation requirements, jurisdictional nuance, and the specific failure modes that matter in legal practice (a hallucinated case citation is not just wrong, it is sanctionable) requires depth that general-purpose models cannot replicate out of the box.

Harvey demonstrates that the most valuable AI applications may not be the most general ones.

Glean

Glean is an enterprise search platform that connects to a company's internal tools and makes everything searchable through a single interface. It sits across Google Workspace, Slack, Confluence, Jira, Salesforce, and dozens of other data sources, with permission-aware retrieval so users only see results they are authorised to access.

Glean is underrated because enterprise search is perceived as a solved or boring problem. It is neither. The permission model alone is a significant engineering challenge, and the breadth of connectors required to cover a real enterprise's data landscape is substantial. Glean is reportedly valued at over $4B and growing rapidly.

FAQs

How were these companies selected?

We looked for companies that are doing consequential work relative to the attention they receive. The list is subjective and reflects our view of where the most interesting work is happening in AI infrastructure and applications in 2026.

Is Exabase on the list because this is the Exabase blog?

Yes, and the benchmark results are the reason we think the inclusion is defensible. M-1 leads both major memory benchmarks at every evaluated scale using a cheaper model than every competitor. The methodology is published and reproducible. See the research papers for LongMemEval and BEAM.

Why is OpenAI not on this list?

OpenAI is not underrated. It is the most discussed AI company in the world. This list is specifically about companies whose impact exceeds their visibility.

Why is Anthropic not on this list?

Same reason. Anthropic is valued at roughly $965B in 2026 and is extensively covered. It is not underrated by any reasonable definition.

Are these companies competitors with each other?

Some are. Exabase and Zep both operate in the agent memory space with different architectures. Groq competes with inference platforms like Fireworks and Together. Others, like Braintrust and Factory, solve different problems and are complementary rather than competitive.

What does "underrated" mean here?

It means the gap between how much attention a company gets and how much impact it is having is larger than it should be. Some of these companies are well-known within their specific developer community but barely discussed outside it.

Should I evaluate all of these?

Only if they are relevant to what you are building. A team building a legal AI product should look at Harvey. A team that needs agent memory should evaluate Exabase and Zep. A team shipping LLM features without an eval pipeline should look at Braintrust. The list is not a recommendation to use all of them.

Where can I learn more about the agent memory space specifically?

See the best agent memory platforms 2026 comparison, which covers the full landscape with benchmark results, feature matrices, and pricing.

Will this list be updated?

We may update it as the landscape evolves. The AI infrastructure space moves quickly and companies that are underrated today may not be in six months.

What about hardware companies like Cerebras?

Cerebras completed the largest tech IPO of 2026 at roughly $66B. It is no longer underrated. Groq is included because despite the NVIDIA licensing deal, the inference cloud business is still emerging and under-discussed relative to its potential.