The Digital Markets Act

1. The basics

Regulation (EU) 2022/1925. Adopted September 2022, applicable from 2 May 2023, with the first gatekeepers designated in September 2023 and their compliance deadline in March 2024.

The six firms designated first were Alphabet, Amazon, Apple, ByteDance, Meta and Microsoft. Booking was added later.

2. Gatekeepers, and how you become one

The DMA applies to gatekeepers providing core platform services. Designation runs on quantitative presumptions in Article 3, which a firm can rebut.

Size. Annual EU turnover of at least 7.5 billion euro in each of the last three financial years, or an average market capitalisation of at least 75 billion euro in the last financial year — and the service is provided in at least three Member States.
Gateway position. At least 45 million monthly active end users in the EU and at least 10,000 yearly active business users in the EU, in the last financial year.
Entrenchment. The user thresholds met in each of the last three financial years.

Core platform services are a closed list: online intermediation services, search engines, social networking, video-sharing platforms, number-independent interpersonal communication services such as messaging, operating systems, web browsers, virtual assistants, cloud computing, and online advertising services.

3. The obligations

Three articles, and the distinction between the first two matters.

Article	Character	Content
Article 5	Self-executing. Applies as written	No combining personal data across services without consent, and no cross-use of it. Business users may offer different prices and conditions elsewhere, which is the anti-steering rule. Business users may communicate with and conclude contracts with customers off-platform. No requiring app developers to use the gatekeeper's identification, payment or browser engine. No tying one core service to another. Transparency on advertising prices and remuneration
Article 6	Susceptible of further specification through dialogue with the Commission	No self-preferencing in ranking. Allow uninstalling pre-installed apps and changing defaults. Allow third-party app stores and sideloading. Interoperability with hardware and software features. No using a business user's non-public data to compete against it. Data portability, and real-time data access for business users. Access on fair, reasonable and non-discriminatory terms to search ranking, query, click and view data for rival search engines. Fair access conditions for app stores, search and social networks
Article 7	Specific	Interoperability of number-independent interpersonal communication services, that is, messaging

4. The compliance machinery, and four public corpora

Article	Mechanism
Article 8	General duty to comply, and to demonstrate compliance
Article 11	Annual compliance report, with a non-confidential version published
Article 13	Anti-circumvention. No behavioural or structural workaround of the obligations
Article 14	Notify every intended acquisition in the digital sector or involving data collection, regardless of whether it meets any merger threshold
Article 15	An independently audited description of the gatekeeper's consumer profiling techniques, with a non-confidential overview published
Article 18	Systematic non-compliance can trigger a market investigation and behavioural or structural remedies, including divestiture

Investigative powers mirror the competition toolkit — requests for information, interviews, inspections, interim measures and commitments — and the Commission can open market investigations to designate new gatekeepers, to find systematic non-compliance, or to bring new services into scope.

5. Enforcement, and the institutional difference

Fines reach 10 percent of total worldwide turnover, and 20 percent for a repeated infringement. Periodic penalty payments run up to 5 percent of average daily turnover.

Private enforcement before national courts remains available alongside.

6. The relationship to Articles 101 and 102

The DMA is formally independent of competition law. It is not an application of Article 102 and does not require proof of dominance, market definition, or effects. Its stated aims are contestability and fairness rather than consumer welfare.

7. How this connects to ATLANTIS

The DMA is not ATLANTIS's subject matter. Its value to you is as the clearest existing proof of the project's premise.

Verifying whether a gatekeeper self-preferences in its own search results is not a doctrinal question. It is a measurement question: rank outcomes have to be observed at scale, across queries, over time, and compared against a counterfactual. The same is true of interoperability, of data portability actually working, and of whether an app store's access terms are fair in practice. No team of lawyers can read their way to those answers.

8. What the panel brings to this chapter

Panel member	Why the DMA is their territory
Georgiana Mirza	Digital ecosystems, competition law and market regulation, and common EU data spaces at the intersection of fundamental rights, competition and innovation. Gatekeeper ecosystems and data access are the centre of her field. Of the four, she is the most likely to ask you a DMA question
Thibault Schrepel	Measured the openness of AI foundation models with Pott across GPT-4, Llama 3, Gemini, Mistral and MidJourney, producing policy conclusions from a constructed metric. Co-authored with Alex Pentland at MIT on foundation model competition and ecosystem dynamics. Has written on the DMA's overlap with competition law, and on open source as a competitive structure. His complexity science framing — markets as systems that evolve rather than settle — is most at home in platform ecosystems
Catalina Goanta	Platform governance and the measurement of platform behaviour at scale. Her large-scale empirical work on disclosure practices across platforms is methodologically the same problem as verifying a gatekeeper obligation

9. What is unexplored, and five projects you could run

Project 1. Code the Article 11 compliance reports

The gap: every gatekeeper publishes an annual non-confidential compliance report describing how it satisfies each obligation. Across seven gatekeepers and several years that is now a substantial recurring corpus, and nobody has systematically coded it. The method: structured extraction per obligation per gatekeeper per year — what is claimed, what is measurable, what is asserted without evidence — then track how the claims change across years and across firms facing the same obligation. Why it works: the corpus is public, text-based, and repetitive enough for reliable extraction with manual validation. Your NLP and entity extraction work maps straight onto it. Produces: an empirical account of what compliance claims actually look like, and the first measure of whether the Article 6 regulatory dialogue converges or drifts.

Project 2. Build the Article 14 acquisition dataset

The gap: Article 14 requires gatekeepers to notify every intended acquisition in the digital sector regardless of merger thresholds. That produces something that has never existed — a near-complete census of gatekeeper acquisitions, including the sub-threshold deals that are exactly the killer acquisitions merger control was criticised for missing. Nobody has assembled it as a dataset. The method: compile the notifications, link to company registries and funding data, then characterise what gets bought — stage, sector, proximity to the acquirer's core service. Why it works: the data is reported by obligation rather than gathered by inference, so the selection problem that plagues acquisition studies largely disappears. Produces: a dataset other researchers will use, which is a strong early-career asset, plus the first direct evidence on whether sub-threshold acquisition patterns look defensive.

Project 3. Extend the openness metric down the stack

The gap: Schrepel and Pott measured the openness of foundation models at the level of licences and model access. The stack has more layers than that — compute, training data provenance, weights, fine-tuning rights, API terms, and the downstream right to fork. Nobody has measured openness across all of them, or linked an openness score to observable contestability outcomes. The method: extend the existing framework, apply it to a current model set, then test whether openness predicts anything — entry by downstream developers, derivative model counts, switching. Why it works: it is a direct continuation of his own published methodology, which is a smart thing to propose because it shows you read the work rather than the vacancy. Your fine-tuning and open-weight experience with TinyLlama, Gemma and Llama 3 is the practitioner knowledge the scoring needs.

Project 4. Gatekeeper ecosystems as graphs

The gap: the DMA designates services, but power sits in the ecosystem — how services interlock, where data flows between them, which acquisitions extended which position. Structural power is a graph property and is rarely measured as one. The method: build the graph with services and corporate entities as nodes and ownership, data flows, default placements and integrations as edges, then measure centrality, clustering and the structural effect of removing a single edge, which is what an interoperability or unbundling remedy actually is. Why it works: it operationalises the complexity science framing in a form a lawyer can read, and remedy design is exactly a question about which edge to cut. Produces: a method for evaluating structural remedies before imposing them.

10. Your CV, mapped onto this chapter

What you have	Where it lands	Why it fits
NLP and entity extraction — TF-IDF, GloVe, WordNet, POS tagging, regex pipelines, and the hybrid legal summarisation and extraction paper	Projects 1 and 5, coding the compliance reports and the profiling descriptions	Both corpora are repetitive legal prose with a stable structure, which is the easiest and most reliable extraction target there is
Graph methods	Project 4, ecosystem graphs and remedy simulation	Structural power is a network property; an unbundling remedy is an edge deletion
Open-weight and fine-tuning work — TinyLlama, Hugging Face, Gemma, Llama 3, NVIDIA NIM	Project 3, extending the openness metric	Scoring openness credibly requires knowing what model access actually restricts in practice, not only what a licence says
Time series and forecasting — ARIMA, LSTM, Bayesian structural time series	Projects 1 and 2, tracking claims and acquisition patterns over time	Both corpora are longitudinal, and the interesting questions are about change rather than levels
Full-stack data engineering — FastAPI, SQLAlchemy, Alembic, Docker, Postgres-shaped schemas	Building datasets other people use, which is what Projects 1 and 2 produce	A dataset nobody can query is not a contribution. You have shipped systems at 28,000 monthly users
Legal training in company law and competition fundamentals	Reading a compliance report critically rather than literally	The research value is in the gap between what is claimed and what is verifiable, and spotting that needs both halves

11. If you remember eight things

Regulation (EU) 2022/1925, applicable May 2023, compliance from March 2024.
It is ex ante: standing obligations on named firms, rather than case-by-case proof after the fact.
Gatekeepers providing core platform services, designated on Article 3 thresholds — 7.5 billion euro turnover or 75 billion market cap, plus 45 million end users and 10,000 business users.
Article 5 obligations bind as drafted; Article 6 obligations are specified through dialogue with the Commission; Article 7 is messaging interoperability.
Fines to 10 percent of worldwide turnover, 20 percent for repeat infringement.
Enforcement is centralised at the Commission. National authorities do not apply it, so the entire monitoring burden sits in one place.
Articles 11, 14 and 15 generate three public, recurring, structured corpora that nobody has coded.
The DMA is your proof that the law has already outrun manual enforcement, which is why a legal regime for computational enforcement is urgent rather than speculative.