See the whole. Cut the cost.

Onthos Infrastructure

Your firm’s information is often fragmented, leading to trapped knowledge, partial answers and repeated cost.

Onthos builds a living ontology from your firm’s information, turning it into connected, evidence-linked intelligence. Users, applications and AI agents can ask questions across the whole corpus with significantly higher accuracy than RAG, at a fraction of the LLM compute cost of reading documents directly.

Book a demoRead the white paper
What is Onthos

From raw information to evidence-linked intelligence

Onthos builds a living ontology from your firm’s sources of information: the companies, products, technologies, business models and themes it describes, the relationships between them, and the metrics that measure them. It distils that information into typed facts, each linked to its original source. As new information arrives, the ontology evolves, building a connected body of knowledge that people, applications and AI agents can query and reuse.

Example

From scattered information to typed facts

Every fact is typed against the ontology and linked to the passage that supports it. Hover or tap a fact to bring its source document to the front.

Three documents · your corpus holds thousands
Q2 FY26 earnings call transcriptTranscript

Thank you, and good morning. Revenue for the quarter was up on a sequential basis, with the data-centre segment now the largest contributor for the first time in the company's history.

We are positioning our co-packaged optical interconnect as the lower-power alternative to pluggable transceivers for AI clusters, and we now have qualification programmes running with two of the five largest hyperscalers. Those programmes are progressing to plan, although we would caution that they are not yet revenue.

Management acknowledged that qualification cycles for new silicon photonics remain long, typically twelve to eighteen months, and that pricing pressure from incumbent transceiver suppliers has intensified over the quarter.

On capital expenditure, we expect to bring the second fabrication line in Penang online in the first half of next year, which

Optical interconnect for AI data centresIndustry note

Optical interconnect has moved from a networking line item to a gating factor in AI cluster design, as copper reaches its reach and power limits at 800G and above.

We expect the co-packaged optics market to grow at more than 38% CAGR through 2029, driven by hyperscaler capital expenditure and power constraints inside the rack. Pluggable modules will persist in the access layer, but the scale-up fabric is where the growth sits.

The principal risk to adoption is thermal: placing lasers beside switch silicon exposes them to heat that shortens laser life, and most vendors have yet to prove field reliability at scale.

Among the listed names with credible roadmaps we note Northgate Photonics, Lumen Fabric and two incumbents that

Northgate Photonics and Meridian CloudPress release

LONDON and SEATTLE. Northgate Photonics today announced a multi-year supply agreement with Meridian Cloud for co-packaged optical interconnect across Meridian's next-generation AI regions.

Under the agreement, Northgate will supply optical engines for Meridian's AI data centres in Northern Virginia, Dublin and Singapore from the second half of 2027, with volumes scaling through 2029.

The companies said the deployment is expected to reduce interconnect power per bit by roughly a third compared with pluggable modules, and that Meridian has taken an option to extend the agreement to a fourth region.

Northgate's chief executive said the agreement validated the company's decision to prioritise reliability testing over

Living ontologyGrows with new filesCompanyCompanyProduct & serviceTechnologyTheme · TIICPlace

Industry and theme classification follows the Theia Insights Industry Classification (TIIC) taxonomy.

Distilled summary

Northgate Photonics is positioning its co-packaged optical interconnect as a lower-power alternative for AI clusters and has a multi-year supply agreement with Meridian Cloud starting in the second half of 2027. It targets a market forecast to grow at more than 38% a year through 2029, where laser reliability remains the principal risk to adoption.

Typed factsHover or tap a fact to trace its evidence
Subject
Northgate co-packaged optical interconnect
Claim
Positioned as the lower-power alternative to pluggable transceivers for AI clusters
Status
Qualification with two of the five largest hyperscalers
CompanyNorthgate PhotonicsProduct & serviceCo-packaged optical interconnectTechnologySilicon photonicsMicro themeOptical interconnect
Metric
Co-packaged optics market growth
Value
More than 38% CAGR through 2029
Driver
Hyperscaler capital expenditure and rack power constraints
Major themeAI infrastructureMicro themeOptical interconnectProduct & serviceCo-packaged optics
Exposure
Laser reliability under switch-silicon heat
Status
Field reliability at scale unproven for most vendors
TechnologySilicon photonics lasersMicro themeOptical interconnect
Supplier
Northgate Photonics
Customer
Meridian Cloud
Scope
Multi-year supply for three AI regions from the second half of 2027
CompanyNorthgate PhotonicsCompanyMeridian CloudPlaceNorthern Virginia · Dublin · SingaporeMicro themeAI data-centre build-out

Why Onthos

The fragmented information contains intelligence. Unlocking it takes an NLP pipeline built around an ontology.

Investment teams ask questions that span multiple sources of information. Answering them means processing raw files, extracting facts and serving them at scale. An ontology is what turns extracted facts from fragments into knowledge: every fact gets a stable identity and type, and evidence accumulates across the corpus.

Onthos is that pipeline.

Direct LLMRAGOnthos
Query pathQuery → read files → answerQuery → retrieve passages → interpret → answerFiles → build ontology and distil facts; query → retrieve facts → answer
Knowledge unitRaw textChunksLinked, typed facts
ContextLimited to text that fits within the model’s context windowLimited to the top N passages retrievedTyped facts from the whole corpus
EvidenceSources from text if explicitly instructedSources from retrieved passagesSources from across the corpus
CostHigh per query: reprocesses full document textRepeated per query: retrieves and interprets passagesUpfront understanding; low cost per query

Direct LLM: files supplied directly to a large language model. RAG: retrieval-augmented generation, which retrieves the top relevant passages before generating an answer.

Cost

Cut the compute cost of AI at scale

Direct LLM and RAG pipelines process document text again for every question, so compute spend grows with usage. Onthos reads each document once at ingestion and answers from typed facts. See our white paper, ‘The case for ontology’, for more details.

~0.04%of the input-token cost of reading everything
~2×the answer accuracy of RAG on thematic screening questions

Fewer tokens per question means less compute and energy for every query your teams, applications and AI agents run. As your document base grows, each question still draws only on the facts it needs, so your AI workflows scale sustainably.

Client journey

Use it in 3 steps

You do not need to build or run a new NLP pipeline. Keep your files where they are, connect them to Onthos and use the outputs in the tools you already run.

Connect your file sources

Connect your existing data stores. Your information stays in place. Onthos accesses only the sources you approve.

Configure the permissions

Choose which sources to process, the processing cadence, who can access the results, region and deployment model. Onthos handles parsing, extraction, ontology creation, classification, provenance and freshness within those boundaries.

Use the endpoints

Onthos stores the structured facts in a queryable knowledge base and serves them through query APIs and SDKs for users, applications and AI agents.

FAQ

Common questions

How can we reduce LLM costs for AI search over our documents?

Stop paying to reread them. Onthos reads each document once at ingestion, distils it into typed, evidence-linked facts and answers questions from those facts. In our white paper, that took an average of 2,720 input tokens per question, against 5–6 million to have an LLM read every filing.

How does Onthos compare with RAG?

Per question, the cost is similar: both send a small context to the model. The difference is accuracy. RAG only sees the top passages it retrieves, so it cannot tell a true negative from a retrieval miss. Onthos classifies every company in the corpus upfront, and on thematic screening questions it scored roughly twice RAG’s accuracy.

How does Onthos reduce the compute and energy used by AI workflows?

Compute and energy scale with the tokens a model processes. Because Onthos answers from a small set of relevant facts rather than the full text, each question processes a fraction of the tokens, and so uses a fraction of the compute.

Does Onthos replace our LLM or AI agents?

No. Onthos sits beneath the models and agents you already use, serving them typed facts through query APIs and SDKs. You keep your choice of model and the tools you already run.

Onthos Infrastructure

Connect your sources. Keep your stack.

Turn your firm’s fragmented information into connected intelligence for end users and AI workflows.

Talk to the teamRead the white paper