Executive Summary

The global artificial intelligence landscape is currently defined by a profound paradox: while large-scale models demonstrate increasingly human-like capabilities in dominant global languages, they simultaneously exacerbate a digital divide for low-resource languages such as Scottish Gaelic (Gàidhlig). This white paper provides a comprehensive strategic framework for AI industry leaders, linguistic researchers, and public sector decision-makers to bridge this gap. The analysis centers on two primary technical and operational innovations: the Adaptive Transfer Scaling Law (ATLAS) and the community-first “Data Flywheel” acquisition model propose by Unityflow AI.

The fundamental challenge addressed is the “Curse of Multilinguality”, a phenomenon where the performance of individual languages in a massively multilingual model degrades due to parameter competition and model capacity constraints. For Scottish Gaelic, this technical hurdle is compounded by extreme data sparsity across text, speech, and parallel bilingual corpora. However, recent breakthroughs in scaling law research and sovereign AI initiatives suggest that these barriers are no longer insurmountable.

The proposed solution integrates the ATLAS framework, which provides a mathematical foundation for optimizing model size and data mixtures by leveraging cross-lingual transfer from synergistic high-resource languages. When coupled with an operational “Data Flywheel” which utilizes community engagement portals like ThunderScribe, automated web scraping, and optical character recognition (OCR), this approach creates a self-reinforcing cycle of data growth and model improvement.

Strategic value is further evidenced by a 2025 case study on Scottish Gaelic Automatic Speech Recognition (ASR), which demonstrated that hybrid HMM-DNN systems, when enhanced by self-supervised learning (SSL) on unlabeled audio, can achieve a 32% relative reduction in Word Error Rate (WER) over leading end-to-end models like OpenAI’s Whisper. With the Scottish Languages Act 2025 granting Gaelic official status and the “Tog” initiative targeting Gaelic as an untapped economic asset, the impetus for technological intervention has reached a critical threshold. This report outlines the roadmap to achieve a sovereign Gaelic AI capability that preserves cultural heritage while unlocking an estimated £82 million to £149 million in annual economic value.

Industry Background and Current Challenges

The evolution of Natural Language Processing (NLP) over the last decade has been characterized by the “scaling hypothesis” the observation that model performance improves predictably with increases in compute, parameters, and data volume. However, the practical application of these scaling laws has historically been an English-centric endeavor. The industry now faces a transitional period where the focus must shift from “scale at any cost” to “optimized scaling for the long tail” of the world’s 7,000+ languages.

The Dominance of English-Centric Scaling

Since the seminal work on Chinchilla scaling laws, researchers have relied on the assumption that data is an abundant, high-quality resource. For English, this held true through the exploitation of massive web crawls like Common Crawl. However, for languages like Scottish Gaelic, the “digitized” footprint is significantly smaller than the actual linguistic output of the community. This discrepancy creates a “low-resource” trap: because there is insufficient digital data to train large-scale models, these languages are excluded from the performance gains of modern AI, leading to further digital marginalization.

The current industry standard for multilingual support involves massive models trained on hundreds of languages simultaneously. While this provides broad coverage, it introduces the “Curse of Multilinguality”. As more languages are added to a fixed-capacity model, the parameters must “specialize” in multiple grammars, scripts, and vocabularies, often leading to a plateau or decline in performance for any single language compared to a dedicated monolingual model of the same size.

The Sovereignty Movement in AI

In response to the limitations of general-purpose models, a “Sovereign AI” movement has emerged. Nations and linguistic communities are increasingly recognizing that relying on third-party, English-first infrastructure poses risks to cultural heritage, data privacy, and economic competitiveness. Initiatives like the UK-LLM project, which utilizes the Isambard-AI supercomputer to build reasoning models for Welsh and other UK minority languages, represent a strategic shift toward local, high-quality AI development.

Scottish Gaelic is positioned at the center of this movement in Scotland. The passage of the Scottish Languages Act 2025 has provided the political and legal mandate to ensure that Gaelic thrives in the digital age. However, the technical path forward requires more than just political will; it demands a rigorous, data-driven approach to scaling that accounts for Gaelic’s specific linguistic complexities and resource limitations.

Problem Analysis: The Curse of Multilinguality and Data Sparsity

To understand why simple fine-tuning of existing models often fails for Scottish Gaelic, it is necessary to analyze the mathematical and linguistic roots of the problem. Data sparsity for Gaelic is not merely a lack of words; it is a structural deficiency in the “AI-readiness” of available linguistic materials.

Mathematical Constraints of Multilingual Models

Linguistic Complexity and Tokenization

Gaelic presents specific morphological challenges that exacerbate data sparsity. Consonant mutations (lenition and eclipsis) change the beginning of words based on their grammatical environment. For example, the root word cat (cat) becomes a’ chat (the cat). Standard subword tokenizers (like Byte Pair Encoding or WordPiece), which are often trained on English-dominant datasets, fragment these Gaelic mutations into meaningless sub-units. This fragmentation increases the “effective sequence length” and requires more data for the model to learn that cat and chat refer to the same semantic concept.

The Multi-Modal Data Gap

The data sparsity problem for Gaelic is divided into three distinct silos:

1. Textual Sparsity: Modern, clean, deduplicated text is rare. Web-scraped data from sources like MADLAD-400 or OSCAR often contains high levels of noise, non-Gaelic text, or machine-translated gibberish. Achieving the “one million word” threshold of clean text is seen as a baseline for useful fine-tuning, but even this is several orders of magnitude smaller than the corpora used for major languages.

2. Speech Sparsity: While hours of Gaelic broadcast audio exist in archives like the School of Scottish Studies, very little is paired with time-aligned, accurate transcripts. Automatic Speech Recognition (ASR) requires diverse acoustic data including various accents, background noise levels, and spontaneous conversation which is currently non-existent in model-ready formats.

3. Parallel Data Sparsity: Machine translation and instruction-following capabilities depend on parallel Gaelic-English sentence pairs. Currently, these are limited to specific genres like legal text or educational materials, making models struggle with conversational or technical domains.

Proposed Solution: The ATLAS and Data Flywheel Integration

The integration of the Adaptive Transfer Scaling Law (ATLAS) with a community driven “Data Flywheel” provides a robust, scientifically grounded methodology for overcoming Gaelic’s resource limitations. This approach moves beyond simple data collection and enters the realm of “optimized scaling”.

1. The ADAPTIVE TRANSFER SCALING LAW (ATLAS) Framework

ATLAS is an adaptive variant of scaling laws that is specifically designed for multilingual and data-constrained settings. Its primary innovation is the concept of “Effective Data Exposure”, which weights different data sources based on their empirical transfer benefits to the target language.

For Scottish Gaelic, the ATLAS strategy involves:

• Repetition-Aware Modeling: Unlike Chinchilla, which assumes data is only seen once, ATLAS models the diminishing returns of data repetition. This is vital for Gaelic, where a small, high-quality corpus must be used over multiple training epochs.

• Synergy Mapping: Using the Cross-Lingual Transfer Matrix to identify “helper” languages. Research shows that sharing a script (Latin) and a language family (Indo-European) are the strongest predictors of positive transfer. Gaelic benefits significantly from co-training with English, French, and Spanish, as the model can share surface-level representations and sub word vocabularies.

• Capacity Optimization: ATLAS provides the formula for the “Compute-Optimal” scaling of parameters when adding languages. To double the language coverage while maintaining Gaelic performance, the model size should increase by approximately 1.18x and total training data by 1.66x.

2. The community-first “Data Flywheel”

The “Data Flywheel” is an operational model that ensures data acquisition is not a one-off event but a self-reinforcing cycle.

• Acquisition (The “Capture” Phase):

• Community Portals: Projects like the Community Enterprise digital portal and “ThunderScribe” enable speakers to contribute authentic voice and text data.

• Mobile-First Channels: Using ubiquitous tools like WhatsApp or dedicated mobile apps allows for the collection of spontaneous, daily speech that broadcast archives lack.

• Automated Harvesting: Building custom web-scraper connectors for high-yield sources like news sites (BBC Alba) and social media.

• Processing (The “Refinement” Phase):

• Gaelic Linguistic Toolkit: Integrating lemmatization and morphological analysis to normalize consonant mutations before they enter the training set.

• OCR for Archives: Converting scanned documents into clean text using an OCR pipeline optimized for Gaelic diacritics and historical fonts.

• Validation (The “Quality Gate” Phase):

• Implementing strict quality benchmarks: 95% language identification precision, <2% duplicate rate, and ≥85% inter-rater agreement on manual transcripts.

• Training and Redeployment:

• Improved models are used to build better data-collection tools (e.g., more accurate auto-transcription prompts), which in turn lowers the friction for future community contributors.

3. Modeling Strategy Under Sparse Labels

The modeling strategy is explicitly low-resource aware, diverging from the “standard” industry practice of simple fine-tuning.

• Hybrid HMM-DNN Approach: For speech recognition, a hybrid system (using Kaldi) is recommended. This allows for the integration of a strong Language Model (LM) trained on text and a Pronunciation Lexicon, which provides “guardrails” for the neural acoustic model.

• Self-Supervised Learning (SSL): Utilizing unlabeled audio through SSL pre-training (e.g., XLS-R 300M). This adapts the model to Gaelic acoustics and code-switching patterns before any labeled data is introduced.

• Semi-Supervised Learning Loop: Running a baseline recognizer on unlabeled audio, filtering the outputs by confidence to create “pseudo-labels,” and then retraining. This effectively increases the training set without the cost of human transcription.

Case Study: Scottish Gaelic ASR (2025)

The effectiveness of this integrated framework was validated in a landmark 2025 case study conducted at the University of Edinburgh. The research challenged the prevailing industry assumption that fine-tuning massive end-to-end (E2E) models is always superior for minoritized languages.

Experimental Setup and Methodology

The researchers compared traditional E2E models (OpenAI’s Whisper-Turbo) against an optimized hybrid HMM-DNN system. The hybrid system leveraged:

1. SSL Features: Replacing standard acoustic features with XLS-R 300M representations.

2. Unlabeled Data: Continued pre-training on 103 hours of Gaelic teaching videos and narratives.

3. Language Modeling: An RNN-LM rescored n-gram model trained on all available Gaelic text.

Performance Results

The hybrid system achieved a 32% relative WER reduction over the best fine-tuned Whisper model.⁶ This result is statistically significant and provides a powerful technical justification for the “Sovereign AI” approach. It demonstrates that linguistically informed models, when paired with self-supervised adaptation, can outperform models that rely solely on massive scale and generic English-centric pre-training.

Strategic Value and Economic Impact

Investing in Scottish Gaelic language technology is not only a cultural preservation effort but a strategic economic catalyst for the Highlands, Islands, and the wider Scottish economy.

Economic Quantification

Research by Highlands and Islands Enterprise (HIE) and other bodies has quantified the “Gaelic Economy” as a significant contributor to national prosperity.

• Direct Economic Value: Estimated at £82 million to £149 million annually.

• Regional Impact: In Glasgow alone, Gaelic-related activities contribute approximately £21.6 million per year and support over 700 full-time equivalent (FTE) jobs.

• Productivity Gains: Improved language technology is forecast to increase international sales for Highlands businesses by £62.8 million and combined annual turnover by £236.2 million.

The “Tog” Initiative and Official Status

The official recognition of Gaelic on St. Andrew’s Day 2025 has fundamentally changed the risk-reward profile for AI investment. The “Tog” initiative (meaning “to raise”) explicitly targets job creation and local prosperity through the language. Under the Scottish Languages Act 2025, public bodies are empowered to designate “Areas of Linguistic Significance” where targeted policy and technological support can be deployed to ensure the language remains a primary medium of communication and commerce.

Comparison with the Welsh Experience

Scottish Gaelic can draw strategic insights from the “Cymraeg 2050” strategy in Wales.

• Trio Writing Methodology: The Welsh Government utilizes a “trio writing” process involving a content designer, a translator, and a user researcher to design bilingual digital services from the ground up. This avoids the “translation lag” and ensures that digital tools sound natural and authentic rather than like machine-translated English.

• Infrastructure Investment: Wales has successfully partnered with Microsoft to integrate Welsh into tools like Teams and Copilot, demonstrating that proactive data-sharing can motivate big tech companies to support minoritized languages.

  
  Competitive Positioning & Defensibility
  

  

  Moat Components

• Community data exclusivity

• Dialect-stratified corpus

• Sovereign data trust governance

• Public-sector embedded infrastructure

• Continuous data flywheel advantage

Even if large providers release Gaelic support, depth, dialect coverage and governance remain differentiators.

Governance: The Gaelic Data Trust

A permanent Gaelic Data Trust ensures:

• Clear contributor consent frameworks

• IP licensing transparency

• Public-private data usage rights

• Dialect equity oversight

• Ethical review board with native speakers
  
  This structure aligns with the mandate of the Scottish Languages Act 2025 and prevents extractive AI practices.
  

  Cost & Compute Strategy

• Fine-tune multilingual checkpoints under ATLAS crossover thresholds

• Target <10B parameter sovereign models

• Leverage UK-based compute clusters

• Prioritise energy-efficient “Green AI” training

Estimated model size: 2B–7B parameters
Target hosting: regional cloud or public sector clusters

This reduces total cost of ownership while preserving performance.

Implementation Roadmap: Weeks 1 to 15

The transition from data sparsity to technological parity requires a disciplined, time-bound execution plan. The following roadmap is based on the roadmap UnityflowAI submitted at CivTech 11.2 exploration stage.

Phase 1: Governance and Ingestion (Weeks 1-6)

• Ethics and Consent: Establishing a “Gaelic Data Trust” to manage consent capture, privacy checks, and license tracking. This ensures all data collected through community portals is usable in both research and commercial contexts.

• ThunderScribe Setup: Launching the ingestion backend for WhatsApp, mobile apps, and IVR (Interactive Voice Response) telephony.

• Labeling Schema: Defining the metadata taxonomy, including dialect tags (e.g., Uist, Harris, Argyll), speaker age groups, and recording environment markers.

Phase 2: First Data Sprint (Weeks 7-9)

• Community Seeding: Engaging 20-30 active contributors to record prompted phrases and conversational segments.

• ASR Baseline: Training the first recognizer using existing DASG text and CLTW audio to establish a Word Error Rate (WER) baseline.

• QA Routine: Implementing the first round of inter-rater agreement checks to ensure transcription quality meets the 85% threshold.

Phase 3: Corpus Harvest and OCR (Weeks 10-12)

• Web Scraper Connectors: Running automated scrapers on BBC Alba news, Gaelic Wikipedia, and available forum data.

• OCR Prototype: Testing character error rates (CER) on scanned archival texts and implementing a “lexicon pass” to correct common OCR errors.

• Parallel Pair Mining: Using bilingual sentence embeddings (e.g., LASER or MUSE) to mine parallel Gaelic-English text from comparable government and educational websites.

Phase 4: Final Sprint and MVP (Weeks 13-15)

• Second Annotation Sprint: Scaling up to reach the target of 20 validated audio hours.

• Model Refinement: Implementing the semi-supervised learning loop using the newly collected audio to refine the ASR system.

• Publication: Releasing the first “Release Candidate” dataset bundle, including clean text, validated audio, and baseline model performance metrics.

Risks, Limitations, and Considerations

A realistic strategy must account for the technical and social risks inherent in low-resource AI development.

1. The Crossover Point for Pre-training

The ATLAS framework identifies a specific “crossover point” in compute efficiency. For 2B parameter models, it is more efficient to fine-tune a general multilingual checkpoint (like Unimax) if the training budget is below 144B to 283B tokens. Given that Gaelic’s clean text corpus is currently far below this threshold, the strategic priority must remain on continued pre-training and fine-tuning of existing base models rather than training from scratch.

2. The “Digital Extinction” of Dialects

There is a significant risk that AI models will over-index on “Standard Broadcast Gaelic,” leading to poor performance for speakers of more conservative or isolated dialects. If a healthcare assistant in South Uist fails to understand a local speaker because it was trained solely on Inverness-based news text, the technology fails its inclusion mandate. Stratified sampling and dialect-aware metadata are the only technical mitigations for this risk.

3. Participation and Maintenance Risk

The “Data Flywheel” depends on continuous community participation. If the portal becomes dormant after the initial accelerator phase, the model will struggle to keep pace with evolving language use (e.g., new technical terms or slang). Sustainability requires the establishment of a permanent “Gaelic Language Data Hub” or Trust, possibly supported by the £35.7 million allocated to the Scottish Languages Act.

4. Technical Bias and Ethics

Generative AI models are prone to “hallucinations” and algorithmic bias. For a minoritized language, these biases can be particularly damaging as they may distort cultural narratives or introduce linguistic patterns that are unnatural to native speakers. Ethical oversight, involving Gaelic speakers in the “human-in-the-loop” evaluation phase, is a non-negotiable requirement for public sector deployment.

Future Outlook and Trends (2026-2030)

The next five years will see a rapid acceleration in the capabilities and efficiency of multilingual AI.

Multimodal Fusion and Generative Dubbing

The trend toward “Multimodal Fusion” like integrating text, images, and audio into a single model will allow for revolutionary applications in cultural preservation. Generative dubbing technologies, similar to those recently piloted by YouTube, could allow English-language educational videos or health guides to be automatically and naturally dubbed into Gaelic using the speaker’s original voice characteristics, dramatically increasing accessibility for Gaelic learners.

Green AI and Model Optimization

The push for “Green AI” and efficiency over pure scale will benefit low-resource languages. As training costs for massive models become prohibitive, the industry is focusing on “DeepSeek” style optimizations by building smaller, highly capable models (under 10B parameters) that provide sovereign reasoning abilities at a fraction of the cost of GPT-sized systems. This matches Gaelic’s resource reality, as these smaller models are easier to host locally on public sector infrastructure.

Instruction-Tuning for Minoritized Languages

Following the model of “BanglaLlama” or the UK-LLM Welsh model, the creation of “Gaelic-Instruct” datasets will be the next frontier. By translating high-quality English instruction corpora and post-editing them with Gaelic speakers, we can create AI assistants that don’t just “translate” but actually “reason” and “follow instructions” in Gaelic.

Conclusion and Strategic Recommendations

Scottish Gaelic stands at a technical and political crossroads. The “Curse of Multilinguality” and extreme data sparsity represent significant barriers, but the integration of Adaptive Scaling Laws with a community-driven Data Flywheel offers a scientifically proven path to digital sovereignty.

For AI Industry Leaders

• Move Beyond Fine-Tuning: Recognize that for minoritized languages, simple supervised fine-tuning is sub-optimal. Prioritize hybrid systems and self-supervised adaptation (SSL) to unlock the latent potential of unlabeled audio and sparse text.

• Utilize ATLAS Parameters: Use the transfer matrix and iso-loss frontiers provided by ATLAS to optimize the compute-efficiency of multilingual training mixtures. Incorporate helper languages based on empirical synergy scores rather than just geographic proximity.

For Linguistic Researchers

• Develop Gaelic-Instruct Corpora: Focus on the creation of high-quality, human-validated instruction-tuning data. This is the primary lever for lifting Gaelic task performance in Large Language Models.

• Standardize Evaluation Benchmarks: Build robust, multi-genre evaluation sets (e.g., FLORES-style) that truly reflect real-world Gaelic usage, including conversational and code-switched speech.

For Public Sector Decision-Makers

• Establish a National Data Hub: Ensure the longevity of the CivTech “Data Flywheel” by establishing a permanent, well-funded repository for Gaelic language data.

• Prioritize Sovereign AI Infrastructure: Invest in domestic supercomputing and local hosting (e.g., Azure, AWS, Google Cloud or regional clusters) to maintain control over Gaelic linguistic assets and ensure compliance with the Data and AI Ethics Framework.

• Implement “Areas of Linguistic Significance”: Use the powers granted by the Scottish Languages Act 2025 to deploy Gaelic AI pilots in health, education, and local government, measuring impact not just in WER reductions but in social and economic revitalization.

The successful implementation of this framework will do more than just build a better speech recognizer; it will ensure that Scottish Gaelic remains a vital, modern medium of human expression in the 21st century. By achieving these data and modeling milestones, Scotland can transform a cultural treasure into a digital powerhouse, proving that in the age of AI, no language needs to be left behind.

Works cited

1. How AI is leaving non-English speakers behind - Stanford Report, accessed on February 5, 2026, https://news.stanford.edu/stories/2025/05/digital-divide-ai-llms-exclusion-non-english-speakers-research

2. ATLAS: Practical scaling laws for multilingual models - Google Research, accessed on February 5, 2026, https://research.google/blog/atlas-practical-scaling-laws-for-multilingual-models/

3. A Practitioner’s Guide to Building ASR Models for Low-Resource Languages: A Case Study on Scottish Gaelic - ResearchGate, accessed on February 5, 2026, https://www.researchgate.net/publication/392466418_A_Practitioner’s_Guide_to_Building_ASR_Models_for_Low-Resource_Languages_A_Case_Study_on_Scottish_Gaelic

4. Gaelic and Scots gain official status | Scottish Legal News, accessed on February 5, 2026, https://www.scottishlegal.com/articles/gaelic-and-scots-gain-official-status

5. SDS Gaelic Language Plan 2025-28 - Skills Development Scotland, accessed on February 5, 2026, https://www.skillsdevelopmentscotland.co.uk/media/qw4lb5dc/sds-gaelic-language-plan-2025-28-english-version-for-consultation.pdf

6. A Practitioner’s Guide to Building ASR Models for Low-Resource Languages: A Case Study on Scottish Gaelic - arXiv, accessed on February 5, 2026, https://arxiv.org/html/2506.04915v1

7. How AI’s Failure on Linguistic Diversity is Deepening Global Inequality - RESET.ORG, accessed on February 5, 2026, https://en.reset.org/ais-echo-chamber-why-linguistic-diversity-matters-for-a-just-digital-future/

8. Reaching Across the Isles: UK-LLM Brings AI to UK Languages With NVIDIA Nemotron | Bangor University, accessed on February 5, 2026, https://www.bangor.ac.uk/news/2025-09-15-reaching-across-the-isles-uk-llm-brings-ai-to-uk-languages-with-nvidia-nemotron

9. AI breakthrough brings Welsh into the digital age - Nation.Cymru, accessed on February 5, 2026, https://nation.cymru/news/ai-breakthrough-brings-welsh-into-the-digital-age/

10. Actions to improve Gaelic and Scots language skills: FOI release - The Scottish Government, accessed on February 5, 2026, https://www.gov.scot/publications/foi-202500486935/

11. 14 Arts, Culture and Language - Scottish Islands Data Overview (2025) - gov.scot, accessed on February 5, 2026, https://www.gov.scot/publications/scottish-islands-data-overview-2025/pages/16/

12. CommonLID: Re-evaluating State-of-the-Art Language Identification Performance on Web Data - arXiv, accessed on February 5, 2026, https://arxiv.org/html/2601.18026v1

13. Welsh language technology and AI | GOV.WALES, accessed on February 5, 2026, https://www.gov.wales/welsh-language-technology-and-ai

14. Gaelic language plan. 2023-2028 - Scottish Enterprise, accessed on February 5, 2026, https://www.scottish-enterprise.com/media/ustldjim/scottish-enterprise-gaelic-language-plan-2023-2028-eng.pdf

15. Advanced Technology and Innovation - Digital strategy for Scotland: sustainable digital public services - delivery plan 2025-2028 - gov.scot, accessed on February 5, 2026, https://www.gov.scot/publications/digital-strategy-scotland-sustainable-digital-public-services-delivery-plan-2025-2028/pages/6/

16. There needs to be a reason to use Gaelic beyond box-ticking - The National, accessed on February 5, 2026, https://www.thenational.scot/politics/25696161.needs-reason-use-gaelic-beyond-box-ticking/

17. ARGYLL AND BUTE COUNCIL COMMUNITY SERVICES COMMITTEE EDUCATION & CUSTOMER SUPPORT SERVICES 28 AUGUST 2025 SCOTTISH LANGUAG, accessed on February 5, 2026, https://www.argyll-bute.gov.uk/moderngov/documents/s222675/Scottish%20Languages%20Act%202025%20-%20CSC%20Report%2014082025%20Pre-Agenda%20Briefing%20of%20the%20Community%20Services%20C.pdf

18. Artificial Intelligence and the Welsh language: The Welsh Language Commissioner’s Regulatory Policy Statement, accessed on February 5, 2026, https://www.welshlanguagecommissioner.wales/media/hscpv33c/20250722-datganiad-polisi-rheoleiddiol-deallusrwydd-artiffisial-ar-gymraeg-saesneg.pdf

19. Artificial Intelligence and Culture - UNESCO, accessed on February 5, 2026, https://www.unesco.org/sites/default/files/medias/fichiers/2025/09/CULTAI_Report%20of%20the%20Independent%20Expert%20Group%20on%20Artificial%20Intelligence%20and%20Culture%20%28final%20online%20version%29%201.pdf

20. EMEA AI Model Builders | NVIDIA Developer, accessed on February 5, 2026, https://developer.nvidia.com/ai-models/emea