Revenge of the supply chain: why hardware still matters in the age of AI

The promise of AI has already resulted in massive amounts of investment and spending, leading to huge valuations for firms seen as well-positioned to capitalise on the boom. The root of this innovation is in software development and the continuous refinement of algorithms. But what does the focus on software mean for physical hardware and associated supply chains? Using the example of the credit-card-sized Raspberry Pi computer, this blog argues that supply chains are very much here to stay — and as important as ever.

Some time in the mid-2000s, in the Computer Laboratory at the University of Cambridge, a crisis was unfolding. A slow-burning, gradual crisis, but a crisis nonetheless. Eben Upton had resolved to do something about it. The number of qualified applicants who had applied to study Computer Science had fallen by about 50% in the first five years of the millennium. This slow, year-on-year atrophy of the cohort threatened to bleed the innovation of the sector, and therefore society, at a time of profound change. So Upton and his Raspberry Pi Foundation colleagues settled on a plan: “to build something that would fit into children’s lives … something that encourages them to get hands on, that doesn’t just become a black box, with which they can get right down to the metal without meeting any artificial barriers between them and the hardware.”^[1]

The result was a compact, self-contained computer: the Raspberry Pi. It allows its owner to interact directly with hardware through exposed GPIO (General Purpose Input/Output) pins, run a full Linux operating system on an energy-efficient processor, and handle HD graphics on a single credit-card-sized board without requiring external components or interface cards. The Raspberry Pi went on to be remarkably successful, selling millions of units worldwide in its first decade and growing from an ambitious, educational project into a global commercial enterprise. Beginning as a strictly non-profit endeavour in 2008, in 2012 it created a commercial subsidiary; in 2024, this subsidiary successfully listed on the London Stock Exchange as Raspberry Pi Holdings. By March 2025, the company had sold over 68 million single-board computers.^[2]

Eben Upton had always been a self-proclaimed “software guy”, and had envisioned the various types of free software that people would write using the devices. The Raspberry Pi has since found a wide array of uses: high-altitude balloon tracking and imaging; microgravity space experiments; AI wildlife monitoring and anti-poaching devices; and autonomous marine ocean drone navigation, among many others. Over time, the profile of the company’s average consumer has evolved considerably. In 2024, 70% of the company’s sales were to industrial customers, primarily for embedded applications, with the remaining 30% to the enthusiast and education sectors. The company has even branched out into directly supplying its industrial consumers with semiconductor units, to such a degree that in 2025 semiconductor unit volumes exceeded its traditional board sales^[3].

Fast-forward to 2026, where the story of the humble Raspberry Pi experienced its latest twist. Throughout this year, demand for the devices surged, driven by industrial and commercial customers increasingly integrating Raspberry Pi platforms and proprietary microcontrollers into AI, robotics and automated manufacturing systems. This momentum was further amplified by rapid international expansion across key original equipment manufacturer (OEM) markets, particularly in the United States and China. Specific, accessible AI uses such as integration with AI agents like OpenClaw, at a time of Mac Mini shortages, also served to fan the flames of demand. The result was that by June 2026, Raspberry Pi Holdings had seen the value of its stock more than triple since the start of the year.

Background: the AI boom

Periods of dramatic societal change — political, social, technological — are only ever fully known in retrospect. They are a kind of mirage whose contours can only be properly discerned at a distance. The current AI boom is no exception. The full extent of its impacts, the second- and third-order consequences, the degree to which it might all be a ‘bubble’: these dynamics will likely not be clear for many years.

Nevertheless, certain facts are beyond doubt. Huge capital expenditure has taken place under the banner of AI, driving unprecedented venture-capital inflows and colossal valuations for AI firms, with even infant start-ups regularly valued in the billions of dollars. Some such cutting-edge start-ups are so new that they are probably younger than many common house spiders (life expectancy 1-2 years), and certainly younger than many queen bees (3-4 years). (Statistics on the insect community’s annual economic output are not currently available.) These trends, of immense spending and commensurate expectations, may indeed be more bug than feature — short-lived and ultimately reversed — but what is certain is that they are driving substantial growth at least in the short-term.

Also certain is the fact that this current economic momentum has been catalysed by software innovation rather than hardware. It is algorithmic software breakthroughs that create a demand for compute power and are in turn forcing tech companies into a physical-hardware arms race, driving hundreds of billions in real-world infrastructure spending. In the Dotcom bubble of the late 1990s, physical hardware like fibre infrastructure was built out ahead of demand; by contrast, physical AI infrastructure (data centres, chips, compute) remains far behind demand. The Dotcom boom entailed a catastrophic oversupply, where up to 97% of global bandwidth sat unused, crashing network prices before software and consumer broadband were mature enough to utilise the excess capacity.^[4] Conversely, in the AI boom, industrial capacity is chasing technological innovation that mainly manifests digitally and is already well integrated into consumer devices and software, driven by a not-inconsiderable amount of pre-existing consumer demand. All signs point to this remaining the case, at least until such a time that the likes of Jony Ive — who has partnered with OpenAI to create a ‘‘hardware ecosystem’’ for the AI giant^[5] — delivers compelling physical AI devices that can shift the locus of innovation back in the direction of physical hardware.

These are truly uncharted waters, with no clear reference points. Indeed, so far out to sea are we that bouts of queasiness for the entire economic arrangement are percolating amongst the shipmates (market traders) and quartermasters (fund managers). Investors are growing increasingly nervous about the sustainability of big spending by AI groups, with reports of large outlays being punished by traders, including most recently a broad and deep sell-off in late July 2026.^[6] But beneath the vicissitudes of market forces and the dizzying pace of innovation, one factor remains constant: hardware is paramount, and its supply chains are inescapable.

The hardware landscape

The dramatic stock market rise of Raspberry Pi is a direct downstream consequence of increased AI enthusiasm. The company also raised prices several times over the last year, citing rising component costs and the global shortage of memory chips, driven in part by demand from AI data centres. But this position in the AI ecosystem cuts both ways: Raspberry Pi’s stock price also suffered in the July sell-offs. Manufacturers of RAM (random access memory), once one of the cheapest electronic components, have seen levels of demand soar with the revelation that the supply of high-end, high-bandwidth memory for AI did not meet forecasted needs. Cloud-service providers finalising their memory requirements for 2026 and 2027 made clear that these physical components would be a key bottleneck for hyperscalers like Amazon and Google, and component manufacturers have duly reaped the rewards.

One result is that consumer-electronics prices are rising precipitously, particularly games consoles, laptops and smartphones. Micron, previously one of the biggest sellers of RAM, announced in December it would stop selling its ‘‘Crucial’’ brand to focus on AI demand.^[7] While AI hyperscalers undertake huge spending in the hope of even greater future returns, component manufacturers are enjoying the benefits and achieving record profits today.

A similar but perhaps even more significant bottleneck comes in the form of large electrical power transformers (LPTs). Grid operators, utilities and project developers are being asked to deliver capacity for a wave of new load from data centres, EV charging networks and electrified buildings, while at the same time integrating a record pipeline of new generation from utility-scale solar, wind and battery storage. Meanwhile, there is an equally pressing need to replace decades-old equipment that is approaching, or has already surpassed, the end of its designed service life. All three jobs land on the same piece of infrastructure: the transformer.^[8] Wait times for LPTs have surged from roughly 24 months, before 2020, to as long as 4 to 5 years today. More compact distribution transformers are experiencing the same squeeze, albeit on a smaller scale and with generally shorter wait times. Compounding this are two additional factors from further down the chain: a global scarcity of GOES (grain-oriented electrical steel) and pressure on the specialised curing cycles required for transformer oil.

Underappreciated physical — in this case, metallurgical — constraints such as these represent a hidden layer determining, to one extent or another, the deployment of transformative technologies. The chain can be followed even further. For instance, AI training clusters require uninterrupted uptime, meaning that gigawatt-scale data centres must be equipped with extensive (usually diesel) backup-generator systems. This is the noisy, dirty underbelly of AI models. Hence, niche companies specialising in areas such as custom-engineered acoustic baffles or exhaust silencers for these generators are securing multi-year supply contracts. Similarly, heavy industrial air-filtration firms that traditionally supplied mining trucks are now retrofitting data-centre generator enclosures with heavy-duty, multi-stage units.^[9] Even gas turbines, once near-obsolete, are reported to have an up-to-seven-year backlog in the US. Siemens Energy, which cut thousands of jobs as gas-turbine demand plummeted in the 2010s, has said Q1 2026 was its strongest ever for orders.^[10]

The macroeconomic, political and geostrategic implications of these processes are many and varied. But evidence of the vital geopolitical importance of physical hardware is ubiquitous. Whether in the growing tendency towards ‘‘friendshoring’’ (concentrating supply-chain networks around political and economic allies), the increasing imposition of export controls on high-tech goods like semiconductor-manufacturing equipment, or the inability of Donald Trump to wrest control of the Strait of Hormuz and end the war in Iran, the same point is ultimately made: supply chains are here to stay. Technological change may be reshaping economies and reinventing the dynamics of infrastructure in entirely novel ways — but it is also, ironically, reinforcing traditional, entrenched paradigms of hardware and supply-chain dependency.

Sources

^[1] Byfield, Bruce. 2026. “Meet Raspberry Pi’s Eben Upton » Linux Magazine.” Linux Magazine. 2026. https://www.linux-magazine.com/Online/Features/Meet-Raspberry-Pi-s-Eben-Upton.

^[2] “Reports - Investor Relations - Raspberry Pi.” 2024. Raspberrypi.com. https://investors.raspberrypi.com/reports.

^[3]Sants, Arthur. 2025. “Raspberry Pi Increases Semiconductor Sales to OEM Customers.” Investorschronicle.co.uk. Investors” Chronicle. September 24, 2025. https://www.investorschronicle.co.uk/content/c68e2d6b-b1e0-48c9-8cc6-20626f7cef72.

^[4] “Journal on Telecommunications and High Technology Law (JTHTL) - Archive - Volume 4.” 2026. Jthtl.Org. 2026. http://www.jthtl.org/articles.php?volume=4. P. 62.

^[5] Constantino, Tor. 2025. “AI Experts React To Merger Of OpenAI And Jony Ive To Create AI Devices.” Forbes, May 22, 2025. https://www.forbes.com/sites/torconstantino/2025/05/22/ai-experts-react-to-merger-of-openai-and-jony-ive-to-create-ai-devices/.

^[6] “Chip Stocks Tumble as AI Sell-Off Deepens.” 2026. FinancialTimes. Financial Times. July 28, 2026. https://www.ft.com/content/f8c03b5b-e194-4236-82c3-389b6f5dd7ae?syn-25a6b1a6=1.

^[7] “Micron Announces Exit from Crucial Consumer Business | Micron Technology.” 2025. Micron Technology. https://investors.micron.com/news-releases/news-release-details/micron-announces-exit-crucial-consumer-business.

^[8] POWER. 2026. “Beating the Transformer Bottleneck: Remanufacturing, Build-to-Stock, and Smart Procurement.” POWER Magazine. June 2026. https://www.powermag.com/partner-content/beating-the-transformer-bottleneck-remanufacturing-build-to-stock-and-smart-procurement/.

^[9] Murray, Clara. 2026. “The Unlikely Corporate Winners of AI.” @FinancialTimes. Financial Times. https://www.ft.com/content/5ede9d4d-3989-49b5-a282-4722c8d8fc59?syn-25a6b1a6=1.

^[10] “Siemens Energy - Earnings Release Q1 FY 2026.” 2026. Siemens-Energy. 2026. https://www.siemens-energy.com/global/en/home/press-releases/earnings-release-q1-fy-2026.html.