Comparisons between the artificial intelligence boom and the dotcom bubble have become increasingly common as technology companies commit unprecedented amounts of capital to chips, data centres and computing infrastructure.
The comparison is not without merit. Both periods have been characterised by high expectations, rapidly increasing investment and uncertainty about whether future revenue will justify current spending.
However, the most useful lesson from the dotcom era is that a transformative technology can create lasting economic value while simultaneously destroying capital for businesses that invest on the wrong assumptions.
During the telecommunications boom of the 1990s, operators invested heavily in fibre-optic infrastructure in anticipation of extraordinary growth in internet traffic. Much of the capacity remained unused for years, and several of the companies that financed it ultimately collapsed.
Yet the infrastructure itself did not disappear. The so-called dark fibre later supported broadband services, cloud computing, video streaming and the digital economy that followed.
This pattern has led some analysts to argue that speculative bubbles can accelerate the construction of infrastructure that the market would otherwise build too slowly.
According to Francois van der Merwe, CEO and founder of Otinga, business leaders should nevertheless separate their belief in the long-term value of AI from the financial assumptions underpinning individual investments.
“The technology is not fake, but the valuations and the underlying value are two different things. Boards should treat them that way,” he said. A company can believe that AI will reshape its industry without assuming that every provider, hardware purchase or long-term contract will deliver an acceptable return.
This distinction is especially important because the physical assets supporting the AI boom differ significantly from the fibre infrastructure installed during the dotcom era. Fibre-optic cables can remain economically useful for decades. Graphics processing units, or GPUs, are replaced by newer generations much more quickly. Improvements in performance, energy consumption and model efficiency may reduce the competitiveness of hardware long before the end of a conventional depreciation schedule.
“The fibre from the 1990s had a useful life measured in decades. The GPU does not behave that way,” Van der Merwe said.
Companies acquiring their own AI infrastructure must therefore be cautious about treating it as a long-life asset. A business case may appear attractive when the cost is spread over five or six years, but the economics change materially when the realistic useful life is shorter. “Whenever an AI implementation needs hardware on-site, depreciate it over three to four years, not the longer schedules some are booking,” he said.
This does not mean every organisation should avoid owning infrastructure. Certain businesses face strict data-sovereignty, latency, security or regulatory requirements that may justify local deployment.
For many organisations, however, consuming computing power through a cloud or managed service could be more prudent than purchasing hardware directly. Renting may appear more expensive when compared over a narrow one-year period. It nevertheless gives a company the ability to change providers, adopt more efficient models, adjust capacity or exit an initiative without being left with stranded equipment.
The value of this flexibility increases when technology and pricing models are changing rapidly.
“Unless you have hard data-sovereignty requirements, buy the outcome, not the iron,” Van der Merwe said.
South African businesses must consider this issue within the realities of the local operating environment. Currency volatility can increase the cost of imported technology, while power, cooling, maintenance, security and scarce specialist skills can add significantly to the total cost of operating AI infrastructure locally.
A hardware investment is therefore not limited to its purchase price. It may create long-term commitments relating to facilities, energy, technical support and future upgrades.
Cloud consumption introduces its own risks, including unpredictable usage costs, vendor dependence and data-governance concerns. The decision should therefore not be reduced to a simple choice between cloud and on-site infrastructure.
Boards should instead assess which approach gives the organisation the best combination of business value, flexibility, control and manageable risk.
Van der Merwe believes AI investments should also be assessed against a shorter and more demanding return threshold than many conventional technology projects.
“Hold AI investments to a clear bar: a two-to-five-times return over 24 months. That buffer accounts for the complexities of deployment,” he said. A 24-month horizon forces executives to be clear about how value will be generated. It also reduces their dependence on long-term assumptions about model capabilities, vendor pricing, demand and hardware performance.
This discipline does not require businesses to abandon strategic thinking. An organisation can build a long-term AI capability through a sequence of shorter investments, each of which delivers measurable value and creates the foundation for the next stage.
Boards should also consider how exposed a proposed investment would be during a market correction. If vendor funding declined, compute prices changed or a new generation of technology emerged, would the organisation still be able to justify the asset or contract?
The companies that survived the dotcom collapse were not necessarily those that avoided the internet. Amazon and Google built businesses around real usage, improving economics and disciplined execution. Many failures were companies whose valuations, debt levels or operating models depended on expectations that never materialised. The same principle is likely to apply to AI. “You can believe completely in AI’s durable value while refusing to fund it on fragile hardware assumptions,” Van der Merwe said.
The AI boom may ultimately leave behind infrastructure, skills and capabilities that support decades of innovation. However, that does not make every investment made during the boom a sound one.
For business leaders, the objective should not be to determine whether the entire market is a bubble. It should be to ensure that their organisations can continue benefiting from AI after valuations correct, vendors consolidate and today’s hardware has been replaced.
Francois van der Merwe is the CEO and founder of Otinga.io and convenes the AI Strategy-to-Results Executive Bootcamp at Henley Business School Africa.
