The AI stack: a framework for a holistic national AI strategy
Written by Timothy Liljebrunn and Armin Catovic
In 1998 the Swedish Government appointed an IT Infrastructure Commission to investigate the need for information and communication infrastructure. By 2005, Sweden ranked no.1 in Europe for full online availability [1]. Sweden’s strategy to build one of the best broadband systems in the world involved a multi-pronged approach. Firstly, it involved a significant investment in the communication infrastructure, including the creation of an alternative open backbone network connecting all municipalities. Secondly, it included economic support to municipalities, households, as well as businesses, when connecting to broadband networks. This multi-layered strategy enabled entrepreneurs, businesses, and individuals to flourish, essentially positioning Sweden as a world-class digital leader with success stories like Spotify, Skype, Klarna and Kry, to name a few.
Just as the broadband internet had an immense impact on how we conduct our work and lives, the current wave of artificial intelligence (AI) technologies, particularly spearheaded by the multi-modal large language models (LLMs), is unequivocally believed to be the bringer of the next dawn of societal and economic change. To capture this new wave, AI Sweden recently proposed an AI Strategy [2] that presents an admirable effort to accelerate the country toward significant AI adoption. Yet, the strategy is solely adoption-centric — it focuses on adopting AI technologies within the existing ecosystem, which is dominated by US companies.
The primary issues with an adoption-only AI strategy are twofold: first, it fails to address all the necessary areas for ensuring long-term resilience and control, which are essential for capturing the full benefits of the technology. Second, it supports a “bandwagon hypothesis” that assumes society and the economy will automatically benefit from technological advancements simply through widespread adoption. However, as Daron Acemoglu and Simon Johnson so eloquently put it in their book Power and Progress:
“There is nothing in the past 1,000 years of history to suggest the presence of an automatic mechanism that ensures gains for ordinary working people when technology improves. On the contrary, there are plenty of episodes in which technological breakthroughs were associated with no improvement — or even significant deterioration — in working conditions and living standards for most people” [3].
As a nation, such a critical situation must be handled wisely. The scope of a national AI strategy for Sweden must therefore be multilayered and address not only the adoption of the technology but also everything that is happening “down the AI stack”. Sweden must emulate its multi-pronged broadband internet strategy of the late 90s, in order to ensure that our society dictates its own fate and instills its own values when building AI systems, but also just as importantly, that its people and workforce are “augmented” with AI, rather than replaced by AI that is controlled by a handful few.
The AI stack — five layers that a national strategy needs to address
Below, we present a version of the “AI Stack” that forms a basis for further analysis and to guide policy and decision-makers. It is important to note a few things. Firstly, nearly all of the top-3 companies (by market capitalization, or valuation), within each layer, are non-European. Secondly, it is beneficial to think of the first two layers equivalent to the “broadband infrastructure” when comparing the current AI situation with the national broadband strategy of the late 90s. When constructing our AI stack, we’ve been inspired by the Open Systems Interconnection (OSI) layered approach to computer networks.
Layer 1 — Hardware: Design and production of CPUs / GPUs, which entails everything from raw resources and rare earths, to fabrication. American companies NVIDIA and Intel have complete market dominance across GPUs and CPUs respectively. Interestingly enough, Qualcomm-based AI chips (another US company) consisting of ARM based processors, are currently favoured by Microsoft in its deployment of their new AI PCs [4]; ARM is originally a UK company, but currently held by Japanese SoftBank Group. On the chipset production front, the market is heavily dependent on TSMC (Taiwan) and ASML (Netherlands). More recently, specialized AI chips from companies such as Groq that optimize memory accesses and implement language processing units (LPUs) have risen in popularity. Sweden is no stranger to specialized chipset production, with companies such as Ericsson, SAAB, Axis Communications and Qamcom all having highly-competitive custom made silicon.
Layer 2 — Cloud Infrastructure: Data centers consisting of computing infrastructure as well as associated data storage and platform-as-a-service (PaaS) solutions. This layer also entails energy supply, cooling, interconnectivity, security and redundancy. This layer is dominated by American companies such as Amazon (AWS), Microsoft (Azure), and Google (GCP). However, most recently, Swedish based Evroc began spearheading an AI-first PaaS approach to building cloud infrastructure in Sweden and Europe. Thanks to its significant investments in renewables (nuclear, hydro, solar and wind), ample real estate, and mild climate with low occurrence of natural disasters and extreme weather, Sweden is well positioned to host a large number of highly performant data centers.
Layer 3 — Software & Libraries: AI software frameworks and libraries such as PyTorch, TensorFlow and JAX, as well as linear algebra libraries, and optimizers — foundational software used for building and executing LLMs. Nearly all of today’s LLMs are built using PyTorch. While PyTorch is technically open-source software, it is largely driven by Meta (US). In recent years, HuggingFace (France) has also risen in standing, partly due to its well executed transformers library, and partly due to its strategic placement as an AI hub. LLMs have more recently paved the way to new types of platforms and libraries, such as LlangChain and LlamaIndex (all US based). Together with more bespoke platforms and libraries, such as Pruna AI (model pruning/optimization), Outlines (structured output enforcement), we are in a new era of LLMOps or GenAIOps.
Layer 4 — Foundational Models: Multi-modal LLMs and associated data (for pre-training as well as labeled/alignment data), as well as various sub-technologies such as context/attention management and sampling strategies. American companies including OpenAI (GPT-series of models), Anthropic (Claude) and Google DeepMind (Gemini) are very well positioned, even though the French company Mistral serves as a strong competitor. There are also “dilution” players, such as Meta, with its release of LlaMA based open-weight foundation models.
Layer 5 — Applications & Services: Applications leveraging the underlying foundation models, such as ChatGPT, Perplexity AI, Glean, Sana AI, and Freepik, to name a few. This layer also includes players where the application is intimately tied to the underlying foundation model, almost being one and the same; for example, Midjourney and Luma AI. Applications and services need not be just pure software and apps — but can also reside within the more physical / tangible domain, such as robot humanoids currently being built by Tesla (US), Figure (US) and 1X (Norway). In general, Sweden (and EU) have the strongest representation within the application and services layer.
Catching the wave is both about offense and defense
This is not just about avoiding being left behind. Just as it was in the early internet wave, AI today also presents an immense opportunity — a $15.7 trillion opportunity [5]. As one of the world’s leading providers of clean energy [6], Sweden is well-positioned to contribute to the sustainable development of the Cloud Infrastructure layer. With the second highest number of researchers per capita globally [7], we should be able to create value across Layers 1–4. Moreover, with several tech role models, there is energy and inspiration for startups and companies to innovate in Applications & Services layer.
A thorough analysis of all of the layers is needed to ensure we build defensibility and secure access to what we need to capture this opportunity. Areas such as talent, regulation, capital, research, and leadership must be examined for each layer. Microsoft just announced a $3.2 billion investment in Swedish infrastructure [8], which is a great starting point, but one must understand their motivations and incentives, and how this translates to the local economics as well as what it means in terms of Sweden’s AI defensibility in the long term. Now is the time to accelerate our efforts and truly take charge of this new technological wave, securing our digital and environmental leadership for the future.
To close it off, we can once again quote Acemoglu and Johnson:
“New techniques and machines are not gifts descending unimpeded from the skies. They can focus on automation and surveillance to reduce labor costs. Or they can create new tasks and empower workers. More broadly, they can generate shared prosperity or relentless inequality, depending on how they are used and where new innovative effort is directed” [3].
About the Authors
Armin Catovic is a Secretary of the Board at Stockholm AI, and a Vice President and Senior AI Engineer at the EQT Group.
Timothy Liljebrunn is an entrepreneur and previously a management consultant, and an active member of the Stockholm AI community.
References
[1] Norwegian and Swedish broadband initiatives (1999–2005), Kjell Hansteen, HØYKOM report №505, Oslo, September 2005, https://www.forskningsradet.no/siteassets/publikasjoner/1127199375708.pdf
[2] An AI Strategy for Sweden, https://strategy.ai.se/
[3] Power and Progress — Our Thousand-Year Struggle Over Technology and Prosperity, Daron Acemoglu & Simon Johnson, Revised paperback edition 2024
[4] Microsoft announces new PCs with AI chips from Qualcomm, CNBC, Published May 20, 2024, https://www.cnbc.com/2024/05/20/microsoft-qualcomm-ai-pcs-snapdragon-arm-processors.html
[5] Sizing the prize — PwC’s Global Artificial Intelligence Study: Exploiting the AI Revolution. Published 2017. https://www.pwc.com/gx/en/issues/data-and-analytics/publications/artificial-intelligence-study.html
[6] Enerdata — Share of renewables in electricity production. Published 2022 https://yearbook.enerdata.net/renewables/renewable-in-electricity-production-share.html
[7] Tee world bank — Researchers in R&D (per million people). Published 2021 https://data.worldbank.org/indicator/SP.POP.SCIE.RD.P6?end=2022&most_recent_value_desc=true&start=1996&view=chart
[8] Microsoft tillkännager en investering på 33,7 miljarder kronor under två år i moln- och AI-infrastruktur och ett AI-kompetenslyft för en kvarts miljon människor i Sverige, Microsoft, Published 3 June 2024, https://news.microsoft.com/sv-se/2024/06/03/microsoft-tillkannager-en-investering-pa-337-miljarder-kronor-under-tva-ar-i-moln-och-ai-infrastruktur-och-ett-ai-kompetenslyft-for-en-kvarts-miljon-manniskor-i-sverige/
