Sofina: the companies behind AI that gets to work

A payment has just failed on a merchant's website. Its team can examine several dashboards, investigate the cause and decide on a correction. It can also ask an artificial intelligence assistant to investigate and prepare that correction. Once the assistant is allowed to act, the questions change: where does it find the data? What may it access? Which system will carry out its decision?

Three direct Sofina holdings help explain this shift. Supabase provides application infrastructure. Cyera secures data and access. Primer coordinates merchants' payments. Developments announced this autumn make their comparison timely: Supabase wants databases suited to agents; Cyera is opening a sales channel to Anthropic customers; Primer already describes actions executed after human approval.

For Sofina shareholders, the interest is economic. AI performing a task creates demand for software and services. The questions are which services businesses will pay for, at what margin, and with what consequence for the holding company.

Start with an order to understand three businesses

Imagine our merchant. A customer asks about an order. The assistant must retrieve the record, read the relevant information and retain a record of its reply. Later, the payments team asks why transactions were declined, then approves a change in the payment route.

These operations involve different businesses. Distinguishing them reveals what each company actually sells.

This is an educational example constructed by Opulion. It does not describe a joint offering, an integration between the three companies or a customer using all three products. An agent here means software able to use tools to perform a task within the permissions it receives.

What Sofina actually owns

All three companies appear in Sofina Direct, the holding company's direct investment portfolio. This relationship is distinct from exposure obtained through funds to which Sofina commits capital. Investing directly alongside other shareholders also does not, by itself, establish a technical fund-manager-led co-investment arrangement.

The profiles were accessed on 20 October 2026; they do not specify the reference date for these percentages. Sofina also states that it is represented in Primer's decision-making bodies and led its 2026 funding round alongside other investors.

These percentages do not give each investment's weight in Sofina's net asset value. They cannot be added: each company has its own capital and value. The profiles do not establish amounts invested, individual carrying values or detailed economic rights. Rapid Supabase growth could therefore have a modest effect on the holding company; a larger ownership percentage in Primer does not prove that its euro value is greater.

Supabase: multiply applications, then win their paying business

Supabase sells developers the services that make an application work: databases, user authentication and file storage. Its appeal to a business lies in the technical work avoided and service continuity as an application grows.

On 2 October, Supabase announced the acquisition of Turso, which specialises in lightweight databases. Its stated aim is to support numerous small projects created by agents and their progression to larger applications. Supabase says it already launches more than a million databases a week. This is company-reported activity, neither a count of paying customers nor revenue.

The economic issue requires no coding knowledge. A prototype may last a few hours; an application used daily needs to retain data, remain available and handle more requests. Commercial value develops when the customer chooses to keep paying for these services.

Pricing makes this transition visible: a free plan, followed by Pro starting at $25 a month, with resources and services that can incur additional charges. This entry price, checked on 20 October, is not average revenue per customer.

Our interpretation: the issue is retaining applications that become useful. Automated software creation can widen customer acquisition, while also multiplying abandoned projects and service costs. Conversion to paid plans, spending by retained customers and margins after hosting costs will matter. Popularity opens the door; it does not pay the bill.

Cyera: useful AI needs limited permissions

An agent's risk extends beyond an incorrect answer. Excessive access may allow it to inspect records unrelated to its task or modify information it was only meant to read. For our merchant, checking an order's status should not grant access to the entire customer database.

Cyera helps businesses identify sensitive data and manage access risks. On 3 September, it announced completion of its acquisition of Oasis Security, valued at $1 billion. Oasis contributes management of software and machine identities: the expansion combines knowledge of data with knowledge of permissions. The announced amount concerns Cyera's transaction, not a billion-dollar Sofina investment.

A second development concerns sales. In an announcement dated 16 October on its website, Cyera says its platform is available in Claude Marketplace. Anthropic customers can allocate part of their committed spending to Cyera.

The commercial rationale is understandable: security software may become easier to buy through an existing budget channel. This is not guaranteed revenue. The announcement quantifies no additional orders. Cyera describes Agent Guardian as providing agent visibility and intercepting actions that breach rules; its product presentation is not an independent assessment of effectiveness.

The question for Sofina has two parts: will this expanded offering convince customers, and can Cyera serve them at a cost that leaves a durable margin? Acquisition integration and competition from other security vendors matter alongside marketplace availability.

Primer: the benefit must appear in payments collected

A merchant may use several payment providers depending on countries, cards or payment methods. Primer connects them through a platform for monitoring transactions and configuring their processing. This is payment orchestration: coordinating existing routes and rules rather than making teams manage everything separately.

In May, Primer announced a $100 million funding round led by Sofina. This is the entire round, not the holding company's contribution alone. Primer said it wanted to develop its Companion assistant from supporting decisions towards executing actions within merchant-defined limits.

In its orchestration guide accessed in October, Primer says teams can approve recommendations and allow Companion to deploy changes, for example to payment-routing rules. Its dedicated Companion page emphasises analysis, recommendations and monitoring, while stating that the user decides. These presentations do not establish the execution capabilities available to each customer. Executing an approved action and having general autonomy remain different levels. The May ambition and features described today should be considered separately.

Why pay for this service? Because a merchant can measure its benefit: more orders actually paid for, lower fees or time saved. That benefit must then be compared with the platform's cost and additional risks. Payment volume processed is never, by itself, Primer's revenue.

A recovered order does not become pure profit

Take an entirely fictional example, unrelated to Primer's or any customer's performance. Over a period, a merchant receives 100,000 orders with one initial payment attempt per order. Its share of orders successfully paid for improves by one percentage point following a processing improvement. Assume the additional orders are distinct, average €80 and are fulfilled.

This €20,000 is a hypothetical economic benefit for the merchant before the specified costs. The portion that would pay Primer would depend on the contract; Primer's margin would in turn depend on its own costs. The sources examined do not provide the information needed to quantify either step.

The route to Sofina's shareholders

These companies share a role in customers' everyday operations. An application retains its data; a security policy governs access; a payment rule affects transactions. Our hypothesis is that such operational proximity can help a supplier become durably useful. It still has to translate into paying activity and profitability.

A disposal or distribution can subsequently bring cash to Sofina. An accounting revaluation of a holding alone does not.

A widely used business may remain expensive to develop. It may become profitable after funding rounds have reduced its original shareholders' ownership. Sofina may also invest more capital to maintain its position. Finally, rights attached to private-company securities may vary between investors: multiplying a disclosed percentage by the latest funding-round valuation would create misleading precision.

This distinction builds on our analyses of Sofina's four cybersecurity businesses and turning private investments into cash. Here, the new focus is the economics of services agents need and the evidence investors should look for.

These indicators form our monitoring framework; the documents examined do not disclose all of them. Individual weights in Sofina's net asset value also remain unknown from these materials. We can identify a value-creation mechanism without deriving a numerical contribution to the holding company.

Shareholders can redirect their curiosity accordingly. Behind each new AI demonstration, they can ask what the business using it will need to buy to make it work over time. At Sofina, Supabase, Cyera and Primer provide three specific settings to examine: retaining information, governing its use and completing a transaction. The next decisive step will be evidence of lasting revenue and a margin that survives.