TDWG2026
Home / Talks / SAL B
talk · Tuesday 22 September · 15:15–⁠15:30 · SAL B

Responsible and GreenAI: Does our community need Gigantic Data Centers to deliver?

Patricia Mergen · Responsible AI, Open Digital Curation, and Round-Tripping for Biodiversity Data

Recording time 6:48:31–7:03:55Open on Vimeo ↗

Summaries and key moments are machine-produced. Check the recording for the speaker’s exact words.

The short versionBiodiversity researchers mostly need AI-ready data rather than giant data centres, and AI use should be weighed against its environmental cost.

Overview

What this was about

Patricia Mergen reflected on AI compute needs and trustworthy AI based on her work in the EU AI Factories project. Contrasting SETI@home with today's investment race in data centres and GPUs (US, EU, China), she noted EU sovereignty levels are utopian and trustworthy-AI requirements can block research, while compute dependency disadvantages poorer countries. Although digitising billions of specimens could need millions of GPU hours, researchers she interviewed mostly train on laptops; data cleaning and AI-readiness is the real bottleneck, so she argued for biodiversity thematic data labs and a TDWG role. She closed on environmental costs, including PFAS in data-centre cooling, urging people to think twice before using AI.

Why it matters. Challenges assumptions behind large AI infrastructure investment and points to data standardisation as the community's main AI need.

Key ideas

In the room

  • SETI@home-style volunteer computing would be unlikely today due to data scale, cyber-security fears and desire for autonomy.
  • Massive investment in data centres is a competition between US, EU and China, with few analyses of actual community needs.
  • EU sovereignty certificates (levels 0-4) require fully EU components at the highest level, which is currently unrealistic.
  • Trustworthy AI requirements sometimes block research; UNESCO discussions highlight compute dependency of countries without their own infrastructure.
  • GPU targets: Amazon claims 2.5 million GPUs in 2025; EU hopes for 1.2 million by 2030; China 1.6 million by 2029.
  • ChatGPT estimated analysing 3 billion specimen images might need up to 500 PB and 1-50 million GPU hours, but researchers (e.g. TETTRIs cascade grants) mostly run models on laptops.
  • The main time cost is finding, cleaning and making data AI-ready; AI Factories plan thematic data labs, and she is pushing for a biodiversity lab where TDWG standards could play a role.
Jump in

Notable moments

In their words

Transcript

Automatically generated captions can contain mistakes, especially in names and technical terms. Times are relative to the room recording.

Read the transcript ↓
Loading transcript…
Keep exploring

Related talks.

See the map ↗
discussion · 25 Sept · 11:45–⁠12:00 · SAL B

Discussion: designing institutional biodiversity knowledge and data science centres

Interoperable institutional knowledge spaces depend less on technology than on people, shared storage and semantics, incentives and governance.

Unidentified session chair, Lars Vogt
talk · 25 Sept · 11:30–⁠11:45 · SAL B

What is the Role of "AI" in the Centers for Biodiversity Informatics of the Future?

Biodiversity informatics centres are about people. AI raises the importance of standards, storage, openness and distributed risk, and centres should fuel researchers' curiosity rather than be driven by the technology.

Matthew Yoder
discussion · 24 Sept · 15:15–⁠15:30 · SAL C

Panel discussion: From Mobilizing Data to AI-Ready Knowledge

AI-readiness is purpose-dependent, and the hardest remaining problems are social: trust, attribution, governance and data sovereignty, where ontologies and knowledge graphs offer some control and explainability.

Hilmar Lapp, Robert Guralnick
Programme abstract

Responsible and GreenAI: Does our community need Gigantic Data Centers to deliver?

Patricia Mergen

Both the public and private sectors invest billions in AI. Key players notably in the US, China and the EU try to outcompete each other by releasing the fastest and most powerful Large Language Models (LLMs) and building the largest data centers to cope with the increasing demand for computational power. AI Factories are end-to-end systems that continuously turn data into AI models and AI-powered services at scale. For example, in February 2025, the EU launched the InvestAI initiative to mobilise €200 billion in broader AI investment, including a dedicated €20 billion European fund for AI Gigafactories. In the US, figures around 500 billion US$ invested in AI infrastructures are cited, with similar figures being estimated from China. Looking at the global scale, with the digitisation efforts and open science policies, billions of data points and media resources around biodiversity are made available and potential candidates for AI applications. One may argue that our discipline is actually in need of such large resources to deliver. When sizing down to specific taxon levels or smaller geographic areas, training datasets of several 100 000 to less than 1 Mio of records managed to deliver accurate AI models for automated species recognition based on images or sounds. When consulted, researchers often answer that they can run the algorithms on their local servers, even local computers, without being in need of heavy computing power from large data centers. Could Small Language Models (SMLs) dedicated to a specific discipline or question, including fact-checking, taking into account ethical and legal aspects, do the job and become a model for a greener and more rational use of AI?

Submitted to the conference programme before the talk, so the talk itself may differ.