Intel and DigitalBridge form Articul8: how to test enterprise AI security claims

Intel and DigitalBridge announced the formation of Articul8 AI on January 3, 2024, an independent company built with intellectual property and technology developed at Intel. Its proposition was a full-stack enterprise generative AI platform that could keep customer data, training and inference inside the enterprise security perimeter and deploy in the cloud, on premises or across both.

The release establishes who participated, where the technology came from and which commercial options were offered. It does not publish pricing, detailed architecture, comparative metrics, contractual terms or an independent security assessment. That absence does not invalidate the platform; it defines what a buyer still needs to verify. The transferable skill is to turn “secure” and “private” into a data diagram and a test plan.

Which company was formed and what Intel retained

Intel's official release described Articul8 as a separate company. Arun Subramaniyan, previously vice president and general manager in Intel's Data Center and AI Group, became chief executive. Intel retained an equity stake and commercial alignment with the company, but corporate independence and strategic collaboration do not mean full ownership.

DigitalBridge Ventures was the lead investor. Intel and a syndicate including Fin Capital, Mindset Ventures, Communitas Capital, GiantLeap Capital, GS Futures and Zain Group also took stakes. DigitalBridge's investor communication repeats those terms. No valuation, ownership percentages or amount raised were disclosed, so assigning a number or a specific financing motive would go beyond the source.

The platform had been incubated inside Intel and was deployed early at Boston Consulting Group. The release said it later reached customers in financial services, aerospace, semiconductors and telecommunications. It did not identify those customers, user counts, project duration or quantified outcomes. “Deployed” establishes existence, not general effectiveness.

On premises, cloud and hybrid describe location

Articul8 offered to keep customer data, training and inference inside the enterprise perimeter. It also supported cloud, on-premises and hybrid options. These properties matter for organizations with restricted information, but “on premises” answers only where part of the system runs. It does not by itself reveal who manages credentials, which telemetry leaves, where updates are downloaded or which provider can connect for support.

A useful diagram traces at least six flows: source documents, user prompts, embeddings, tuned weights or adapters, generated answers and technical logs. For each, record location, encryption, authorized identities, retention and backup destination. If one component calls an outside service, the real perimeter includes that service even when the main interface is inside the customer's data center.

Intel's archived Spanish release uses precise language: customer data, training and inference remain within the perimeter. A buyer should ask “training” to be split into pretraining, tuning on internal data and subsequent learning from usage. These are different processes and may follow different rules.

Optimized for Intel did not mean locked to Intel

The platform launched optimized for Intel Xeon Scalable processors and Intel Gaudi accelerators. At the same time, the announcement said it would support a range of hybrid infrastructure alternatives and called its technology hardware agnostic. Both can be true: initial optimization need not prevent a multiplatform ambition.

What was missing was a public compatibility and performance matrix. To test independence, request the models, versions, drivers and features supported on each architecture; repeat the same workload; measure throughput, latency, energy and total cost; and document what is lost when hardware changes. “Agnostic” should mean that a substitution path exists, not that every path performs equally.

Portability also covers application assets. A customer should know whether prompts, indexes, evaluations, logs and tuning artifacts can be exported in useful formats. A solution may run on several kinds of servers and still create dependency through its management layer.

A full stack does not guarantee an outcome

Intel described Articul8 as a turnkey platform for building, deploying and managing generative applications. That scope may prevent every team from separately integrating models, document retrieval, controls, observability and infrastructure. Yet a platform does not automatically make generated output correct.

Testing should begin with a bounded use. For answers over internal manuals, build questions with known answers and source documents, add questions that lack an answer and measure whether the system cites the right passage, abstains when evidence is absent and preserves permissions. A quality average is insufficient if an employee can retrieve a document they are not allowed to see.

Security, accuracy and utility must be measured separately. Security asks who can access data and what can leave. Accuracy compares claims with sources. Utility asks whether output reduces work without transferring errors. A system can protect data perfectly and hallucinate; it can also answer accurately while logging sensitive input in the wrong place.

The release is issuer evidence

The descriptions “fast,” “secure,” “cost-efficient” and “production-ready” came from Intel, DigitalBridge and a BCG quotation included in the same announcement. They were statements by parties launching, funding or working with the technology. The document contained no reproducible benchmark, audit report or study by a third party without that relationship.

Responsible diligence labels each piece of evidence by origin. The release confirms the company's formation, leadership, investors and proposition. A contract should confirm data and support obligations. An audit should test controls. A buyer's evaluation should measure the use case. No layer replaces the others.

The date should also be retained. On January 3, the company had just become an independent entity even though its technology had been incubating for some time. Prior team experience is not the same as the new company's operating history. Incidents, availability, support and financial capacity require accumulated evidence.

Twelve questions before a pilot

First: what data reaches the model, retriever and logs? Is it used to improve shared products? Who can decrypt it? What leaves the customer network? Second: which models and hardware work in each deployment mode? How are they updated? Can a customer pin a version and roll back?

Third: how are document permissions enforced during retrieval? Does an answer cite its source? What happens when evidence is missing? How are data, evaluations and configuration exported when a contract ends? Each answer should point to a clause, setting or test rather than another adjective.

Then run a pilot with synthetic or low-risk data, identities holding different permissions and adversarial cases. Observe network calls, inspect logs, attempt to access prohibited material and measure recovery from failure. Only then introduce more sensitive information.

The transferable skill is to draw the perimeter a platform promises and find every arrow crossing it. Articul8 offered valuable control and deployment choices, but the release was the beginning of verification. “On premises,” “hybrid” and “secure” become useful information when each can be demonstrated through a data path, an obligation and a test.