AI Policy Primer (September 2024)
Issue #14: Compute, energy, and competitive measures
In this edition of the AI Policy Primer, we have pieces on investment in compute, work assessing AI’s energy footprint, and moves by the EU to boost the competitiveness of its AI capabilities. As usual, leave a comment or let us know if you have any feedback at aipolicyperspectives@google.com.
What we’re reading
Compute investments add up
What happened: Trials in the TSMC Arizona plant reportedly put the fab’s productivity on par with some of the firm’s locations in Taiwan. The plant, which will begin commercial operations in the first half of 2025, was the subject of intense scrutiny just one month ago. According to the report, in trial production, the yield rate - how many usable chips a company can produce during a single manufacturing process - is similar to comparable facilities in the southern Taiwanese city of Tainan.
What’s interesting: The news comes at a busy time for chip companies and data centre developers. Intel, the American semiconductor giant facing stiff competition, announced greater autonomy for its foundry business. In a press release, the firm said that Intel Foundry would be established as an independent subsidiary to provide “future flexibility to evaluate independent sources of funding.” Earlier this month, Intel was also awarded up to $3 billion from the CHIPS and Science Act, which seeks to bring chipmaking to the U.S. via the “Secure Enclave” program in partnership with the Department of Defense.
Technology firms in the US also continue to invest in compute capacity to train large models, while in the UK, new developer DC01UK submitted plans for a £3.8bn data centre in Hertfordshire - although questions were raised about the structure and experience of the group.
For their part, policymakers are considering how to best assess and manage this growing demand for data centres and chips and what it means for energy supplies. In the US, the Department of Energy recently convened a Working Group on Powering AI and Data Center Infrastructure, which published a set of recommendations ranging from advancing different types of efficiencies for LLM training and inference to exploring new types of generation, storage and grid technologies to power data centres. It also noted that any efforts to predict energy demand were “fraught with uncertainties” (the subject of our next update below).
Looking ahead: Investment in compute capacity and energy infrastructure will continue to increase dramatically. In the next two years, we may see several developers implement plans to train models requiring over 1 gigawatt of power - equivalent to approximately 300 utility-scale wind turbines.
What we’re reading
Power consumption under the spotlight
What happened: Tim Fist at the Institute for Progress assessed a recent estimate from a Washington Post article which claimed that GPT-4 consumes 0.14kWh of energy to produce a 100-word email. Fist suggests that the estimate, which was produced in collaboration with researchers at the University of California, Riverside, was off by a factor of at least 350x.
What’s interesting: If correct, Fist’s analysis would put the Washington Post article in line with other past claims about dramatic energy usage, or greenhouse gas emissions, from digital technology that were subsequently questioned or debunked. For example, George Kamiya, then at the IEA, explained why claims by an NGO, The Shift Project, during COVID-19 that watching 30 minutes of Netflix generated 1.6kg of CO2 were off by 90x. Jonathan Koomey and Eric Masanet have also cautioned about regular missteps in this area, such as conflating increased data, or internet use, with increased energy use, which can ignore important questions such as whether demand affects peak capacity.
Experts have created tools to estimate the energy use and emissions from AI models more reliably, though mainly focussed on training rather than inference. These estimations can be complex and require information that is not always available, such the efficiency of the data centre, the source of energy, the type of chips used, and the training protocols.
Beyond these methodological challenges, any estimates for the energy used to train or deploy an AI model – and the emissions generated – will only ever offer a partial answer to the broader question about how an AI model will impact the environment. A lifecycle approach also requires considering other types of emissions, such as the embodied emissions from building a data centre or device, and any indirect effects on emissions from applications that AI enables. These estimates, in turn, would need to be considered against counterfactual scenarios that do not use AI, as few actions in the modern economy produce no emissions.
One of the main reasons for disagreements in this space is the upfront energy costs from training AI models are growing and tend to appear quickly, with relative certainty, while the downstream benefits - from a more efficient Internet to potentially helping to enable new kinds of materials for solar panels or batteries - are potentially much more consequential, but also less immediate and certain.
Looking ahead: As models increase in size – and despite the emergence of more efficient training and inference procedures – require more energy, we expect interested third parties to continue to develop new methodologies to estimate the energy required to train and run AI models. Wide-ranging debates about what this may mean for emissions is likely to follow.
Policymakers taking action
EU eyes competitive measures
What happened: The European Union is heading into a new mandate and things are changing fast. The new regulatory agenda will be key to follow. At the outset of the recent Draghi report, authored by European eminence grise Mario Draghi, the paper suggests that Europe needs to invest massively to compete with the rest of the world - and not just regulate - but really grow a tech sector of its own. Draghi did not pull punches: “Technological change is accelerating rapidly. Europe largely missed out on the digital revolution led by the internet and the productivity gains it brought: in fact, the productivity gap between the EU and the US is largely explained by the tech sector. The EU is weak in the emerging technologies that will drive future growth. Only four of the world’s top 50 tech companies are European.”
What’s interesting: Europe is at a crossroads. The old hypothesis was that regulation would provide Europe with a seat at the global table. Will the new Commission still agree that this is the case? Efforts such as the AI factories programme, which will allow AI developers to build on the EuroHPC network of supercomputers, attempt to put the EU on the path to compete with the US. But will it be enough? The AI factories are meant to help startups and public sector efforts - but without access to capital markets and growth mechanisms, will startups stay in Europe? And will the code of practice for AI now being drafted help the European efforts to focus on productivity growth?
Looking ahead: This is the question the European Union needs to answer: whether it will double down on regulation as a competitive advantage, or if it will pivot to policy interventions seeking to bolster innovation. For that reason, we may see the European Commission double down on opening the European market for AI, and implementing the AI-act in a way that allows for a transatlantic market to emerge over time.