In March 2016, a computer made a move on a Go board in Seoul that experienced players initially struggled to comprehend. Move 37, played by DeepMind's AlphaGo against the South Korean champion Lee Sedol, defied centuries of accumulated knowledge about the game. What appeared to be an error turned out to be an extraordinary calculation.
AlphaGo won the match. The achievement was celebrated as a breakthrough in machine intelligence, although its economic significance was not immediately apparent. Few businesses had any practical reason to reconsider their operations because a computer had defeated a Go champion.
A decade later, the implications are considerably harder to ignore.
Artificial intelligence has moved from demonstrating what machines might accomplish to changing what businesses are willing to pay for computing, knowledge and specialised work. It has helped turn semiconductor manufacturers into some of the world's most valuable enterprises, encouraged extraordinary investment in digital infrastructure and brought capabilities once confined to research laboratories within reach of ordinary companies.
But there is a distinction between creating a technological revolution and creating an economic one. The former can be demonstrated by increasingly capable machines. The latter must eventually appear in productivity, corporate earnings and the wider standard of living.
Understanding the past ten years requires examining both.
From digital platforms to computational power
In 2016, the technology industry's most successful businesses largely derived their economic strength from digital platforms. Search engines monetised attention, social networks accumulated users, electronic commerce expanded distribution, and cloud computing allowed businesses to rent infrastructure instead of building it themselves.
Artificial intelligence was already part of this system, powering recommendations, fraud detection, advertising and forecasting. But it was generally a supporting capability within established commercial models.
The publication of the Transformer architecture in 2017 helped change the technical possibilities. The public arrival of ChatGPT in November 2022 then made those advances commercially intelligible to a much wider audience. Sophisticated language capabilities could be accessed without specialist training, opening new markets for software and accelerating demand for the computing systems behind it.
The financial consequences were extraordinary.
Nvidia reported annual revenue of $5.01 billion in its 2016 financial year. By fiscal 2026, revenue had reached $215.9 billion. Not all of that expansion can be attributed to AI, but demand for accelerated computing and data-centre infrastructure became the principal force behind its recent growth.
Amazon Web Services followed a different but complementary trajectory. Its revenue increased from approximately $12.2 billion in 2016 to $128.7 billion in 2025. The expansion of cloud computing preceded the generative AI boom, but it also provided much of the infrastructure through which businesses could deploy increasingly sophisticated AI systems.
Together, these developments reveal how the technology market has changed. Consumer reach and software distribution remain valuable, but computational capacity has acquired a strategic importance of its own. Semiconductors, advanced networking, data centres and access to electricity increasingly determine who can develop and operate the most demanding systems.
Capital has followed that shift. According to Stanford University's 2026 AI Index, global corporate investment activity related to artificial intelligence reached approximately $581.7 billion in 2025, including private funding, mergers and acquisitions, and minority stakes.
The figure is striking, though it requires careful interpretation. It is not a measure of new productive capacity, nor can it be treated as an equivalent contribution to economic output. It records the scale of financial commitments made in anticipation of future returns.
That distinction matters because the technology industry's largest investment cycle in years is being financed partly on the expectation that other industries will become substantially more productive.
The infrastructure is being built. The economic returns remain uneven.
What happens when intelligence becomes cheaper?
For all the attention paid to the enormous sums invested in AI infrastructure, a less visible development may prove equally consequential: the rapidly declining cost of using the technology.
Stanford's 2025 AI Index found that the price of querying language models delivering performance comparable to GPT-3.5 on a standard benchmark fell from approximately $20 per million tokens in November 2022 to just seven cents by October 2024.
That is a reduction of more than 280 times in roughly two years.
The economic importance lies not simply in making existing software cheaper. Lower costs allow businesses to perform activities that were previously uneconomic, experiment with new services and extend sophisticated capabilities to customers who could not previously afford them.
This is particularly relevant to smaller enterprises. For decades, the advantages of scale included the ability to employ specialised analysts, engineers and technical teams. AI cannot eliminate the need for such expertise, but it can reduce the cost of accessing some of the capabilities those professionals provide.
A smaller company may be able to develop software, analyse customer behaviour or automate complex administrative work without expanding its workforce at the same rate as its operations. What once required substantial fixed expenditure can increasingly be purchased as a service.
The result is a potentially important change in the economics of business formation and expansion.
There is also an apparent contradiction: AI is becoming cheaper to use even as total expenditure on the infrastructure supporting it continues to rise. The explanation is that lower unit costs encourage wider adoption, while developers pursue increasingly demanding applications.
Yet falling prices introduce a challenge for investors. As computational intelligence becomes more widely available, its scarcity value may decline in certain markets. Businesses that rely exclusively on selling access to capabilities offered by numerous competitors could face pressure on their margins.
The more durable opportunities may emerge where AI is integrated into activities that are difficult to replicate: proprietary industrial processes, specialist knowledge, established customer relationships or complex operational systems.
In other words, the economic value of intelligence will not necessarily remain concentrated where the technology itself is created.
The productivity question
The most important test of artificial intelligence is whether it can increase output without requiring a proportionate increase in labour and capital.
There is already evidence that it can.
A study by Erik Brynjolfsson, Danielle Li and Lindsey Raymond, published in The Quarterly Journal of Economics in 2025, examined the deployment of a generative AI assistant among 5,172 customer-support employees. It found an average productivity increase of 15 per cent, measured by customer issues resolved per hour.
The distribution of those gains was particularly revealing. Less experienced workers benefited substantially more than their highly experienced colleagues, suggesting that AI can make elements of accumulated professional knowledge accessible to employees who have not yet acquired comparable expertise.
This has consequences beyond customer service. One of the longstanding constraints on economic productivity is the time required to develop specialised skills. Technologies that help workers perform certain complex tasks earlier in their careers could alter the relationship between experience, labour costs and productive capacity.
But the results of one workplace study cannot be extended mechanically across an entire economy.
A company might use AI to prepare reports faster without improving its underlying performance. Another might redesign its operations, reduce processing times, improve customer retention or enter markets that were previously beyond its capacity.
Both have adopted the technology. Only the second has clearly begun to change its economic potential.
The distinction becomes even more important outside conventional office work.
In scientific research, DeepMind's AlphaFold demonstrated how machine learning could accelerate the prediction of protein structures, opening new possibilities for biological investigation. In manufacturing and logistics, AI can improve forecasting, identify equipment failures and help manage complex production networks.
These applications are economically significant not because they imitate human conversation, but because they address problems whose costs extend through entire industries.
Still, there is a distance between demonstrating a technical capability and earning a commercial return from it. Scientific advances must pass through lengthy development processes. Industrial applications require integration with existing systems. Employees need training, organisations need to change, and management must establish where the technology genuinely improves performance.
For investors, these practical constraints deserve as much attention as the pace of technical progress.
The relevant question is no longer whether AI can perform impressive tasks. It is whether those capabilities can generate additional revenue, stronger margins or more efficient use of capital at a cost that justifies their deployment.
That is where technological progress begins to become measurable economic growth.
The limits of an extraordinary expansion
There is another reason to distinguish technological promise from economic return: artificial intelligence is far from an inexpensive industry to sustain.
Its software may be increasingly accessible, but its physical requirements are formidable.
The International Energy Agency estimated in April 2026 that global data-centre electricity consumption had reached approximately 485 terawatt-hours in 2025 and could rise to around 950 terawatt-hours by 2030.
Behind those figures lies a growing competition for energy, advanced semiconductors, engineering expertise, land and financial capital.
These constraints are already influencing corporate investment decisions and industrial policy. They also complicate the assumption that demand for AI will necessarily translate into attractive returns for every company supplying it.
Expanding computing capacity requires substantial upfront expenditure. The profitability of that capacity depends on utilisation, pricing, operating costs and the useful economic life of equipment that may face rapid technological obsolescence.
Strong demand is encouraging. It is not, by itself, proof of a sound investment.
The same caution applies at the national level.
Countries are increasingly treating AI infrastructure as a strategic asset, with implications for competitiveness, energy policy and technological sovereignty. For economies with substantial energy resources or financial capital, the opportunity to participate in this expansion is considerable.
Yet owning infrastructure and developing a productive AI economy are different achievements.
A country can attract investment in data centres without necessarily developing competitive domestic technology businesses. It can acquire sophisticated systems without improving the efficiency of its industries. It can announce ambitious programmes while leaving the institutional weaknesses that obstruct economic development largely untouched.
The countries that benefit most will be those capable of converting technological access into domestic capabilities, stronger enterprises and higher-value economic activity.
A related question concerns employment. The International Monetary Fund has estimated that nearly 40 per cent of global employment is exposed to AI, with the proportion considerably higher in advanced economies. Some occupations may become more productive; others will face displacement or substantial changes in their requirements.
This is not a secondary consideration. The distribution of productivity gains will influence consumption, income inequality and the long-term political sustainability of the transition.
An economy in which AI raises corporate profitability without creating sufficient opportunities for workers could experience technological success alongside considerable social and economic strain.
The quality of growth matters as much as its pace.
The decade ahead
The past ten years have established something that was far from certain in 2016: advanced machine intelligence can be developed at scale, distributed commercially and applied to an expanding range of economic activities.
What remains less certain is how quickly those capabilities will translate into sustained productivity growth across the wider economy.
Financial markets have already assigned enormous value to the companies building the infrastructure. Businesses have begun discovering practical efficiencies. Governments are committing capital and political attention to securing their place in the emerging technological order.
The next phase will be judged by a more demanding measure.
It will depend on whether manufacturers produce more efficiently, whether smaller companies can compete more effectively, whether scientific advances lead to commercially useful discoveries, and whether investment in computing delivers returns that justify its considerable cost.
For investors, the distinction between companies selling the tools and companies using them profitably will become increasingly important. For governments, the challenge will be to ensure that technological capacity produces economic development rather than becoming an expensive symbol of modernisation.
In Seoul, the significance of AlphaGo's unexpected move became clear before the game was over. Economic transformations reveal their consequences far more slowly.
A decade later, we have learnt how to produce increasingly capable machine intelligence. We are only beginning to discover how to translate it into enduring economic value.
The greatest achievement of the next ten years may therefore have little to do with building a machine that thinks more impressively.
It may be building an economy that makes better use of what machines can already do.
