Artificial intelligence is widely discussed as a software breakthrough. For investors, however, the most significant opportunity right now lies in the physical infrastructure being rebuilt beneath it.
Over the past year, inference costs have fallen by more than 99%. And as the cost of running models has collapsed, usage has surged.
Developers, enterprises and consumers are generating exponentially more tokens, and demand on unified application programming interfaces (APIs) is increasing dramatically month to month.
A good example can be seen in the chart below, which shows how AI modelling breakthroughs directly influenced engagements on the leading API OpenRouter in 2025.
Source: ARK Invest – Big Ideas 2026
This dynamic matters.
Infrastructure cycles driven by falling costs and rising consumption tend to be self-reinforcing. Lower costs expand the addressable market, which in turn justifies further capital investment.
And in this case, hyperscalers have responded aggressively.
Since the “ChatGPT moment”, investment in data centre systems investment has accelerated sharply. In 2025 alone, spending reached approximately $500 billion – nearly two and a half times the average annual level of the previous decade.
Source: ARK Invest – Big Ideas 2026
Looking ahead, our research suggests annual AI infrastructure investment could hit nearly $600 billion this year, continuing to increase over the following years to ultimately pass $1.4 trillion by the end of the decade.
That would represent a compound annual growth rate of 30% for 2023-2030 period.
Dot-com, but different
It is tempting to look at these figures and make comparisons to the late 1990s technology boom. And indeed, when viewed as a share of GDP, capex in the technology and communication services sectors has already reached levels not seen since 1998.
But the comparison only goes so far.
Unlike then, valuations across the technology sector remain a fraction of their prior peaks – as the chart below shows.
Source: ARK Invest – Big Ideas 2026
Even more importantly, today’s spending is supported by visible and accelerating demand. In other words, this time around, company investment is not underpinned by speculative assumptions; it is being expanded to meet workloads that already exist.
Just look at how the revenue share of accelerated GPUs and ASICs has expanded since 2019 – and we expect traditional compute type to lose out more from here.
Source: ARK Invest – Big Ideas 2026
At the same time, the competitive landscape is also evolving in ways that broaden the opportunity set for investors.
Nvidia’s early investments in AI chip design and networking propelled it to a dominant share of GPU sales and exceptional margins. But rivals are gaining ground in certain inference workloads, as companies search for more cost-efficient compute.
Competitors like AMD and Google have caught up in certain domains like small language model inference. And our research suggests ASICs designed by companies like Broadcom and Amazon will continue to take share as AI labs and hyperscalers search for cost-effective compute.
This shows AI infrastructure is no longer a single-company story.
As workloads mature and economics improve, margin pools are likely to shift, but the underlying demand for compute will remain durable.
Exceeding expectations
For investors, the implication is subtle but important.
AI may prove less about short-term model cycles and more about a multi-year reindustrialisation of the cloud. Capital expenditure is already at boom-era levels, but valuations do not reflect the excess typically associated with speculative bubbles.
If history is any guide, infrastructure cycles backed by real usage growth tend to last longer than markets initially expect. In that context, the most compelling AI opportunity may not lie solely in applications, but in the systems quietly being rebuilt to support them.
Tom Barker is a product specialist at ARK Invest Europe
