According to market intelligence provider IDC’s “Data Age 2025” White Paper, published in 2018, the public cloud – cloud services such as AI tools offered across multiple customers by a single provider – will account for 26% of the world’s digitally-stored information by 2025, double the share in 2016.
Around 30% of the global datasphere is expected to be generated in realtime. This would potentially be absorbed for deep learning training or inference.
These may seem like abstract figures, but the infrastructure challenge is enormous. Bottlenecks in data centres, such as packed interconnects and obstructions that risk system failures, are proving one of the biggest roadblocks to AI’s adoption.
In a signal the developing world will soon demand more cloud data capacity, Teraco, Africa’s biggest operator of independent data centres, has begun work on a 38-megawatt facility. Teraco has seen its NAPAfrica subsidiary and the continent’s main internet exchange point connecting hundreds of African telecoms and content providers reach 1.5 terabit a second in peak traffic, according to the Financial Times’s feature last month.
Nigeria, whose economy has historically plagued by recurrent power outages, has announced plans for its own AI research hub, the Centre for Artificial Intelligence and Robotics.
Isa Pantami, the minister of communication and the digital economy, has pledged to engage young Nigerians into AI with the aim of fuelling innovation that might give its economy a lift.
Should the centre launch – and caution should be exercised as infrastructure projects in the country do have a history of going awry, then it might add more competition for Africa’s developmental dollar, pitting the US and its public clouds against China’s, which owns around 20% of African debt and has added the region to its Belt and Road infrastructure initiative.
The Economist article used an analogy for today’s “technopolitics”, comparing software platforms to strategic territorial assets – a country containing mountains in the path of long-distance transit routes, for instance or onshore constrictions for maritime traffic. Each is a potential flashpoint for international relations.
The rise of tech giants, such as Google, Microsoft, Apple, Amazon and Facebook in the US and Alibaba, Tencent, SoftBank and Baidu in Asia, provides scope for software platforms to challenge or support governments and lawmakers. In January 2020, Google and Facebook threatened to discontinue Australian media services if legislation was passed to force them to negotiate licences with content providers.
China was once more accepting of western IT hegemonies than it is now. There was little alternative in the early days of the internet but the country’s drive for tech sufficiency is driven by five-year plans and US embargos.
The European Union, meanwhile, has amended its laws to prevent software hosted by US public clouds from engulfing its citizens’ rights.
An AI cloud bottleneck can only exacerbate the division of internet norms along geographical lines especially as AI does not have a monopoly on global cloud storage.
Public cloud applications, such as AI model development, are anticipated to account for more than 40% of overall cloud storage by 2025, according to IDC estimates cited by The Economist in November. This was up by about 10 percentage points from 2020.
Growing pressures on public clouds helps explain the investment boom for edge-focused AI chips. These edge computing trends target intermediary data centres closer to the end-user in a bid to reduce latency in application.
Grand View Research has pegged the size of this market at $1.8bn in 2019 and expects 21.3% annual growth from 2020 to 2027. As our main case study observed, the segment may now absorb early-stage venture dollars from cloud-targeted AI training products.
The landscape in the AI edge space begins to look less like the world’s interconnected geography, and more like an arrangement of remote islands – to extend the Economist’s metaphor. Some end-devices will require distinct specifications to implement AI directly, and there will potentially be rewards from protocols that harmonise edge or end-based inference better with the cloud. That helps explain Nvidia also taking a tilt at this market, with a stripped-down variant of its A100 – the EGX A100 – with just a single A100 processor tailored for edge settings.