Wednesday, 5 August 2026
The next phase of artificial intelligence will depend as much on power stations, semiconductor factories and data centres as it does on software, according to technology leaders speaking at AI4 2026.
During a discussion on the future of computing infrastructure, former Intel CEO Pat Gelsinger, now a general partner at venture capital firm Playground Global, joined Sachin Katti, Head of Compute at OpenAI, to examine the physical systems required to support increasingly capable AI models.
The discussion reflected one of the central themes emerging from the conference: artificial intelligence may appear to users as software, but its continued expansion depends on an extensive industrial network of chips, electricity, cooling systems, communications infrastructure and highly specialised manufacturing.
AI4 lists Gelsinger as a general partner at Playground Global and Katti as OpenAI’s Head of Compute. (Ai4 2026)
Chips become the fuel of the AI economy
Gelsinger described semiconductors as the underlying fuel of an economy increasingly driven by AI-generated tokens.
Every model training exercise, business application and user request ultimately runs on physical processors. As the use of AI grows, so does demand for computing capacity throughout the infrastructure chain.
However, supplying that capacity is considerably more complicated than simply producing additional chips.
Advanced semiconductor factories are among the most expensive and technically complex industrial facilities ever developed. They take years to plan and build, while leading-edge manufacturing depends on a relatively small group of companies capable of producing advanced processors, memory systems and chipmaking equipment.
This concentration has allowed semiconductor manufacturers and equipment suppliers to capture a significant share of the economic value created by the current AI investment cycle.
Gelsinger argued that companies operating at the silicon level have so far been among the clearest financial beneficiaries of the AI expansion.
Compute is not yet a standard commodity
Although computing power is often compared with oil, the speakers cautioned that AI compute has not yet become a fully interchangeable commodity.
Different chips offer different performance, energy use, memory capacity and suitability for particular workloads. Infrastructure designed for training a large model may not be equally effective for operating that model and answering millions of user requests.
Katti explained that demand remains strong because greater computing capacity continues to support advances in model performance.
His comments reflected the principle commonly associated with AI researcher Rich Sutton’s The Bitter Lesson: over time, approaches that make effective use of increasing computational power have frequently outperformed systems based mainly on handcrafted human knowledge.
For OpenAI, this makes access to computing infrastructure a strategic requirement rather than an ordinary purchasing decision.
OpenAI has described compute as the essential input that allows it to train more capable models, support growing usage, improve reliability and reduce the long-term cost of providing AI services. (OpenAI)
Existing GPUs remain too inefficient
Despite the rapid development of AI hardware, Gelsinger said the present generation of graphical processing units remains highly inefficient in its use of electricity and memory.
GPUs became the dominant hardware for AI because they were the best available option for processing many calculations simultaneously. However, they were not originally created specifically for the enormous training and inference demands now being placed upon them.
The next generation of AI chips will therefore need to deliver significantly more useful work from each unit of power.
Memory is one of the most important constraints. AI processors frequently spend time waiting for information to move between memory and the computing elements of the chip.
Bringing more memory closer to processors and increasing memory bandwidth could substantially improve performance and lower the cost of inference—the process through which a trained AI model produces answers, images or other outputs.
Gelsinger suggested that new processor architectures could eventually transform inference economics by very large margins.
Energy capacity may limit AI expansion
The discussion repeatedly returned to electricity as the ultimate constraint on AI growth.
A company can order processors and construct data centres, but the equipment cannot operate without sufficient and reliable power. This means that available energy capacity may determine how quickly countries can expand their AI industries.
“In a digital AI economy, economic capacity equals energy capacity,” Gelsinger told the audience.
He argued that the United States and other Western economies need to increase electricity generation while improving transmission systems so power can reach new data-centre locations more quickly.
The challenge is not simply the total amount of electricity produced. Developers must secure grid connections, transmission capacity, land, planning permission and local community support before a major AI facility can begin operating.
OpenAI has similarly said that large-scale AI infrastructure depends on coordination between utilities, energy companies, semiconductor manufacturers, cloud providers, construction businesses, investors and public authorities. (OpenAI)
Data-centre development needs to become faster
Katti said the entire infrastructure stack must become easier and quicker to build.
At present, companies can face long delays when moving from chip design to production and from identifying a data-centre site to bringing the facility online. Power availability, construction capacity and equipment supply can each become a bottleneck.
One possible response is to develop more standardised and repeatable data-centre designs. Greater use of modular buildings, pre-engineered systems and on-site energy generation could shorten construction programmes and reduce dependence on lengthy grid expansion.
However, the industry also faces a timing problem.
AI models and algorithms can evolve more quickly than semiconductor development. A specialised processor designed for today’s dominant AI architecture may be less suitable by the time it reaches commercial production several years later.
This creates a risk that companies could invest heavily in hardware optimised for workloads that have already changed.
Infrastructure investment measured against world GDP
Gelsinger predicted that expenditure on AI infrastructure could eventually be measured as a proportion of global gross domestic product.
Trillions of dollars are already being directed towards semiconductor capacity, data centres, electricity generation and network infrastructure. Yet the present economics remain difficult because AI systems require enormous investment before operators can recover their costs through commercial services.
The answer, he argued, cannot be investment alone.
The industry must achieve dramatic improvements in processor efficiency, energy consumption, memory performance, construction speed and operating costs. Incremental gains will be insufficient if AI demand continues growing at its current rate.
The panel identified four interconnected priorities: expanding infrastructure, improving computing efficiency, increasing energy availability and creating more sustainable economics.
Silicon companies currently capture the greatest value
The speakers also considered which parts of the AI supply chain are currently receiving the greatest financial benefit.
While model developers and enterprise software providers expect to build valuable businesses, semiconductor designers, advanced manufacturers and chip-equipment suppliers have so far occupied particularly strong positions.
Companies such as TSMC and ASML have built expertise and supplier networks over several decades, creating capabilities that cannot be reproduced quickly simply by providing additional capital.
This presents a strategic problem for governments seeking to establish or rebuild domestic semiconductor industries. Funding new factories is important, but successful production also depends on technical knowledge, specialist workers, equipment suppliers, materials providers and long-term customer demand.
The development of a competitive semiconductor ecosystem is therefore likely to take considerably longer than the construction of an individual manufacturing plant.
AI becomes an industrial and property challenge
The panel’s conclusions extend beyond the technology sector.
As computing requirements expand, AI is becoming a major factor in energy policy, industrial development and commercial property markets. Data-centre operators will require increasingly large sites with access to electricity, fibre networks, water or alternative cooling systems and skilled labour.
Locations capable of offering these conditions could attract substantial investment. Regions without sufficient power capacity or efficient planning systems may struggle to participate fully in the AI economy.
The debate is therefore shifting from whether AI demand will grow to whether physical infrastructure can be delivered quickly and economically enough to support it.
The future of artificial intelligence will not be decided by algorithms alone. It will also depend on whether the world can produce enough chips, generate enough electricity and construct enough specialised property to keep the machines running.
© 2026 cij.world
