The layer with no factory: what Europe can actually document about Lovable

I. Why the application layer is the part of the stack Europe forgot to argue about

European sovereignty policy is organised around things that are expensive and visible: lithography in Veldhoven, 300 mm wafers in Crolles, cells in Dunkirk, detectors in Moirans. Every one of those anchors has an address, a permit file and a payroll. That physicality is why the industrial instruments work at all — a subsidy can be attached to a site.

The application layer has none of those properties. It is where the value of the whole stack is finally captured, and it is the layer with the lowest capital intensity, the shortest time to revenue and the weakest attachment to any jurisdiction. A company can be incorporated in Delaware, operated from Stockholm, hosted by an American hyperscaler, serve customers in ninety countries and be described by every European trade publication as a European champion — all of which can be true at once, and none of which is a sovereignty claim.

Lovable is the sharpest available specimen of that geometry, which is why this dossier treats it as evidence rather than as an example. The mechanism worth naming is not "Europe produced a unicorn". It is this: at the application layer, growth outruns institutions. Revenue can quadruple in seven months; a company register updates once a year; a national ecosystem report is written annually; an EU legislative cycle takes three years. By the time any of those instruments can describe the company, the company has changed shape twice.

II. Two founding dates, two jurisdictions, one narrative

The registry record is where the European-champion framing first comes apart, and it does so on documents rather than on opinion.

The Swedish operating company is identifiable: Lovable Labs Sweden AB, organisation number 559506-1739, registered address Regeringsgatan 25, 111 53 Stockholm, industry code SNI 62100. Sweden's official business register (Näringslivsregistret, Bolagsverket) shows this entity with a 2024 registration. Meanwhile the group is consistently described as founded in 2023 by Anton Osika and Fabian Hedin, and the company's own corporate profile names the group entity as Lovable Labs Incorporated — a United States corporate form, not a Swedish one.

We do not reconcile these. We print them as a genuine, sourced discrepancy: a 2023 founding of a United States parent and a 2024 registration of the Swedish subsidiary are not a data error, they are a corporate structure. And we state the limit of our own work honestly — the register view we could open gives a registration year, not a certified incorporation date, and Bolagsverket exposes no stable citable URL for the underlying filing. So no exact founding day appears anywhere in this dossier.

One European publication has made the structural point explicitly, arguing that Sweden's most celebrated AI company "is actually a US company — and that's Europe's real problem". A reader may find that framing sharp. The underlying fact it rests on — a United States parent above the Swedish operating company — is documented, and it is the fact that matters for anyone drawing sovereignty conclusions from a growth chart.

III. The capital trail, including the round nobody announced

The funding record is unusually well documented, which makes the one undocumented figure stand out.

• Pre-seed, reported 7 October 2024: US$7.5m (reported as approximately €6.83m at the time), raised under the company's earlier identity as GPT Engineer. Worldwide figures, in United States dollars.
• Series A, 17 July 2025: US$200m led by Accel, with 20VC, byFounders, Creandum, Hummingbird, Visionaries Club and angels participating, at a reported US$1.8bn valuation. This is stated in the company's own announcement and corroborated the same day by independent technology press.

Two things about that round deserve to be printed rather than smoothed. First, the pre-close reporting shows a live negotiation rather than a fixed number: a US$1.5bn valuation was reported on 12 June 2025, then "over US$150m at a near US$2bn valuation" on 2 July 2025, before the announced outcome of US$200m at US$1.8bn on 17 July. We print the sequence as three dated reports of a moving process, not as three competing estimates of one fact.

Second, there is a currency contradiction we will not average away. One European publication reports the same round at €1.5bn while the company, the leading United States technology press and the business press report US$1.8bn. At mid-2025 rates €1.5bn is roughly US$1.65bn, so this is not a conversion artefact — it is a reporting discrepancy between named publishers, and it stays visible here.

Then the gap. In March 2026 the business press described Lovable as "valued at US$6.6bn". We found no announced priced round — no company post, no filing, no investor announcement — supporting that mark. We therefore record it as a press-reported valuation with no disclosed round behind it in our evidence, and we refuse to call it a Series B. If a secondary transaction or an unannounced round produced it, that transaction is not in the openable record.

IV. Revenue: the fastest curve in European software, and what it is not

On 23 July 2025 the company stated on its own blog that it had "officially passed US$100m in ARR — in just 8 months since our first US$1m", alongside "more than 10 million projects" built and "100,000 per day". Independent technology press reported the milestone the same day, one outlet adding 2.3 million users, another framing it as the fastest software company ever to reach the mark. All of these figures are worldwide and in United States dollars.

On 11 March 2026 the company confirmed to technology press that it had crossed US$400m in annual recurring revenue in February 2026 — and, in the same reporting, declined to say whether it was still projecting US$1bn for the year, a target its chief executive had stated publicly in August 2025. That refusal is a data point in its own right, and it is why no US$1bn figure appears in this dossier except as a withdrawn-from-comment projection.

Three disciplines apply to every number above, and they are not decorative.

Every figure is a self-reported run-rate. Annual recurring revenue is a company-defined operating metric, not a filed account. No audited statement for either entity was located in this pass. A US$400m ARR claim and US$400m of recognised revenue in a filed annual report are different objects, and only the first exists in our evidence.

The same article can contradict itself. One March 2026 business-press piece headlined the milestone as revenue "doubling in a few months" while its body reported a 33% jump in a month. Both statements are from the same publisher on the same day. We print the internal inconsistency instead of picking the more impressive half.

Third-party trackers disagree with the company. A secondary blog states US$206m ARR at eleven months; an individual's social post claimed US$75m at seven months with forty staff. Neither is a company statement or a dated press report, both conflict with the company's own timeline, and neither is used for any argument here. They are recorded to show what a reader will encounter when they search this company, and why aggregator figures are not evidence.

V. The headcount question, and why the productivity arithmetic fails

The most widely repeated claim about Lovable is a ratio: enormous revenue divided by a tiny team. That ratio cannot be computed responsibly from the current record, and this is where the dossier is most useful.

Five published values exist:

• 146 employees — company-confirmed to technology press, March 2026.
• 539 people — a professional-network platform's own headcount display, an undated live snapshot derived from member profiles.
• between about 40 and 120 — several undated third-party aggregators, tertiary encyclopaedia infoboxes and individual social posts. None of them is a company statement or a dated report by a named publisher, so none of them is cited in this dossier's source list; they are noted only because a reader searching this company will meet them first.

The 146 and the 539 are the serious problem: they describe the same month, one from the company and one from a platform counting self-identified employees, and they differ by a factor of nearly four. No reconciling source was found. Consequently this dossier prints no revenue-per-employee figure for Lovable, and readers should treat any publication that does — however striking the number — as having quietly chosen a numerator and a denominator from incompatible sources.

That refusal is the finding. The single most cited fact about the AI application layer, the claim that it collapses the headcount required to build software, is currently unmeasurable at its most famous instance.

VI. What the platform runs on, and what its own legal documents promise

The infrastructure record is clearer than the corporate one.

The backend layer is Supabase: the vendor publishes a Lovable customer study quoting a Lovable product manager, its own documentation describes the integration, and its announcement of the platform's managed backend states that "every project created in Lovable Cloud is powered by Supabase behind the scenes". On 3 June 2026, a Google Cloud press release announced an expanded collaboration under which Lovable would use Gemini models and Google's AI-optimised infrastructure, dated Stockholm.

So the shape of the dependency is legible: a Swedish-operated, United States-parented application layer, built on an American backend platform and an American hyperscaler, with a frontier-model supply that is not European either. Nothing in that sentence is a criticism of the company. It is a description of what "European software success" now means at this layer, and it should be read next to the semiconductor and battery anchors, where the physical asset really is in Europe.

On data, we hold to the documents. Lovable publishes a signed Data Processing Agreement, last updated 17 November 2025, which explicitly states that it "does not establish a joint controllership arrangement under Article 26 of the GDPR" and that each party remains solely responsible for its own compliance. We found no EU data-residency guarantee in the text we could open — no commitment that European customer data is stored and processed exclusively inside the European Union. As with any vendor, availability in a European cloud region and a contractual residency commitment are different objects; only the first is documented here, and a regulated European buyer should ask for the second in writing.

Two further absences are stated rather than filled. We found no source substantiating a Cloudflare relationship, and no named European enterprise customer beyond the Supabase case study — no bank, retailer or public body. For a company at this revenue scale, the absence of named enterprise references in the openable record is itself worth recording.

VII. What this actually does to engineering work

The anchor this dossier discharges is not about a company's valuation; it is about what the tooling layer does to the labour market that European industry recruits from. Here the evidence supports less than the discourse claims, and precision matters more than enthusiasm.

What the record does support: 10 million projects and about 100,000 created per day as of July 2025, and 2.3 million users that month rising to a chief executive's statement of nearly 8 million users by 11 November 2025. Read carefully, those are adoption figures, not outcome figures. A project created is not an application shipped, maintained, secured or operated.

What follows for a career, stated only as far as the evidence reaches: when the first working version of an interface can be produced by describing it, the scarce skill moves away from producing that first version. It moves toward the work the platform does not do — specifying the behaviour precisely enough to be judged, reviewing generated code against a security and data-protection standard, owning the database schema and its access rules, and taking accountability when a system that nobody typed line by line fails in production. A junior engineer whose value proposition was "I can build the screen" now competes with a text box. A junior engineer whose value proposition is "I can tell whether this is correct, safe and maintainable" does not.

We refuse the stronger version of that claim. No source in this pass measures a change in European hiring volumes, junior-developer demand, salaries or time-to-delivery attributable to this class of tooling. The mechanism is visible in the adoption numbers and in the product's own description of itself; the labour-market effect is not yet measured, and we will not print a number for it.

VIII. Refused in print

This dossier does not state, imply or reconstruct: an exact founding date for either entity; a certified Bolagsverket incorporation record; any Series B or any priced round behind the reported US$6.6bn valuation; any audited revenue, profit or cash figure; any single headcount; any revenue-per-employee ratio; any EU data-residency guarantee; any Cloudflare relationship; any named European enterprise customer beyond the published Supabase study; any EU AI Act classification or obligation for this company, on which no source was found; any claim that this company is named inside the European Parliament's December 2025 study on European software and cyber dependencies, which we could not verify by full text; and any measured effect on European engineering employment.

Several pages relevant to this subject returned HTTP 403 to our automated fetcher — a Swedish register aggregator, a company database and a business magazine. None of them is used to establish any fact in this dossier, and none appears in the source list below.

IX. Disclosure of interest

CareerOn's editorial platform is itself built with Lovable. We consider that a reason to publish this dossier under a stricter standard than usual rather than a reason to avoid the subject: every figure above comes from a named, dated, openable source, no vendor material is used to establish an outcome, the growth narrative is not endorsed, and the two findings least flattering to the platform's own supplier — the United States parent above the Swedish company and the absence of a published EU data-residency commitment — are printed in the body rather than in a footnote. If a reader wishes to discount this dossier because of that relationship, they now have the fact needed to do so.

X. What a European decision-maker should take from this

Three conclusions survive the evidence.

First, the application layer will keep producing "European champions" whose sovereignty content is thin. Not because founders are disloyal, but because the capital, the incorporation advice and the infrastructure defaults all point one way, and nothing in current European policy addresses a layer with no factory to subsidise.

Second, growth metrics at this layer are self-reported by construction. ARR is not an audited account, valuation marks appear without rounds, and headcount can differ fourfold between the company and a platform counting its own employees. Any institution using such numbers in a strategy document is building on operating metrics, and should say so.

Third, the skills consequence is real but not yet measured. The demonstrable shift is upward, from producing a first version to judging, securing and owning systems. That is a training and standards question for European employers today — and it deserves better evidence than a growth chart before it becomes a hiring policy.