Author : Anulekha Nandi

Published on Apr 19, 2024

Tech giants wield vast computational resources, fostering complexity and vendor reliance, underscoring the need for local capability development for smaller businesses to access computing power

AI economies of (hyper)scale

AI systems require vast amounts of data and computational power for model development and training leading to the emergence of a newer category of vertically integrated service, platform, and infrastructure providers known as hyperscalers. This provides flexible hardware resources like networking, storage, and servers, the environment in which to develop, manage, and host applications, and provides remotely accessed software based on enterprise user needs. Together they make the conditions enabling AI development and production and represent critical infrastructures and resources that are indispensable for further innovation of AI applications. These offerings are the result of the exponential scale required by computational and data-intensive processes of AI development. 

Hyperscale denotes a technical term i.e. the ability of the IT architecture to scale in response to demands placed on the system as well as an epithet for some of the largest cloud service providers and their consolidation and market concentration across the AI development pipeline. Hyperscalers like Amazon Web Services, Google Cloud Platform, and Microsoft Azure store vast amounts of data and computational power to develop and train advance AI models. These three companies alone command 67 percent of the worldwide market

With industrial, sectoral, and organisational transformations led by generative AI, the demand for hyperscale cloud computing has increased. This increased demand coupled with data sovereignty policies across jurisdictions means that major players are investing in hyperscale facilities spanning the offering of infrastructure, platform, and software as a service backed by physical infrastructural investments in data centres. However, this mutually-reinforcing feedback loop is predicated on vast economies of scale enjoyed by vertically integrated technology companies which raises questions on resource accessibility for downstream innovation. 

This increased demand coupled with data sovereignty policies across jurisdictions means that major players are investing in hyperscale facilities spanning the offering of infrastructure, platform, and software as a service backed by physical infrastructural investments in data centres.

Dynamic capabilities and resource consolidation 

Hyperscalers initially scaled to meet their own need but now compete on making computational capacity commercially available on demand. Their scale of product and service capabilities has come to be determined by the way they have managed to vertically integrate critical data resources from their existing business models into product and service offerings. These firms have rich data resources which incentivises them to leverage and invest in AI with large databases reducing the computational cost of training models and increasing predictive accuracy as datasets grow. This means organisations with bigger datasets have lower costs and better returns in AI production. The use of AI reinforces the use of the cloud industry with growth in AI leading to higher upstream demand for their cloud services. Many cloud providers also own AI platforms, allowing them to control a large portion of the industry.

The use of AI reinforces the use of the cloud industry with growth in AI leading to higher upstream demand for their cloud services. Many cloud providers also own AI platforms, allowing them to control a large portion of the industry.

The data advantages mean that these companies enjoy significant economies of scale in growth areas of machine learning like natural language processing. Training AI models can be a cost-intensive endeavour running into millions of dollars which places many companies and start-ups at a disadvantage. However, hyperscalers have almost zero marginal cost of deploying AI because of the positive feedback loops from data, scale, process transformation, and augmented capabilities. Their cloud service offering reinforces their dominant position due to positive feedback loop as the use of their platform lead to more data generation which feeds into their capabilities. Hyperscalers often cement their dominant positions through strategic partnerships, e.g. Microsoft Azure has strategic partnerships for foundational models both with OpenAI for ChatGPT and Meta for Llama 2. AWS’ foundational model Bedrock is supported by models from AI21, Stability AI, and Cohere in combination with Amazon’s Titan models. However, many of these models from generative AI unicorns like Cohere are Google Cloud customers demonstrating complex multi-layered linkages cross-cutting the ecosystem and concentrating around market leaders. While some argue that such market concentration is not an issue as these companies can act as infrastructure providers for downstream application developers, they also involve challenges of increasing complexity, vendor lock-in, and unpredictable pricing. These create entry barriers for downstream innovation to flourish, moreover, data issues in large foundational models creep into applications developed on top of them. The globally interlinked scale of hyperscaler operation and extensive market penetration across geographies undermine the development of national capabilities leading some countries to enact policies to protect and foster domestic competition in this space.

These create entry barriers for downstream innovation to flourish, moreover, data issues in large foundational models creep into applications developed on top of them. 

Democratising innovation

The dominant position of US-based hyperscalers has become strategic policy issues in the EU with France wanting to strengthen support for indigenous hyperscalers. The trade body Cloud Infrastructure Service Providers in Europe (CISPE) filed a complaint with the EU antitrust body against Microsoft’s licensing practices which they argue favour its own Azure platform limiting options for users to run their workload. This comes amidst complaints from Germany’s NextCloud, France’s OVHCloud, Italy’s Aruba, and an association of Danish cloud service providers. However, apart from 26 small European cloud service providers CISPE also includes AWS. While Microsoft is trying to resolve these complaints through bilateral talks, it comes in tandem with concerns of some EU member states about foreign hyperscalers winning sensitive contracts in Europe, particularly with regard to data access provisions e.g. US Cloud Act 2018 which requires firms with connections to the US to hand over data stored offshore which contravenes MLAT (Mutual Legal Assistance Treaty) agreements for bilateral cooperation and assistance. However, despite concentration at the top, the AI community is also exploring alternative and additional options provided by AI unicorns that can work alongside them through co-specialisation. With the Chinese government’s support for AI clusters paying off digital dividends, the US government’s non-interventional stance has led many academics to call for government involvement to reduce the reliance on hyperscalers.

According to some trends, estimates predict that the next hyperscaler might emerge out of India, with the latter generating terabytes of monthly data on account of 5G penetration, OTT consumption, social media usage, and the rise of generative AI. Data service companies like Yotta and E2E networks are looking to expand their data centre footprint aggressively in the country. This is complemented by companies like Reliance, Tata, and Adani entering the data centre space through strategic partnerships which might lead the next global hyperscaler to emerge out of India in the next 7-10 years. However, while these companies might have the capacity for deep financial investments which will help them provide computational power and server space, the three top contenders still have a competitive advantage when it comes to large granular datasets which forms the lifeblood of AI systems and models. This would require public investment, policy, and regulatory support to create digital public infrastructures for AI to make computational resources accessible to start-up communities to reduce the barriers and costs of innovation. 


Anulekha Nandi is a Fellow at the Observer Research Foundation

The views expressed above belong to the author(s). ORF research and analyses now available on Telegram! Click here to access our curated content — blogs, longforms and interviews.

Author

Anulekha Nandi

Anulekha Nandi

Dr. Anulekha Nandi is a Fellow in Technology, Economy, and Society at ORF. Her primary area of research includes digital innovation management and governance focusing ...

Read More +

Related Search Terms