AI Value is Time Value

Last night Colette Kress, Nvidia’s CFO, noted that capital expenditure (capex) by the top five hyperscalers is now expected to reach nearly $800 billion in 2026 and $1.3 trillion in 2027, against a cloud industry backlog above $2 trillion.  Those numbers are extraordinary and they are numbers which make headlines.  But underneath the headline numbers are two quieter developments which potentially matter more.  Firstly, buyers are broadening:

• Eli Lilly, Roche and Bristol Myers Squibb are building dedicated ‘AI factories’ for drug discovery;

• Mercedes-Benz is putting Nvidia's self-driving stack into a production car;

• Samsung cited a 20x speed-up in chip lithography;

• Jane Street and Hudson River Trading are running quantitative trading on the platform; and

• Sovereigns are building national computing capabilities.

Secondly, and less noticed, the amount of useful AI work a single graphics processing unit (GPU) can perform is changing rapidly as software (best evidenced by DeepSeek’s model improvements) and systems integration (as evidenced by Nvidia’s new Vera Rubin range versus the older Blackwell GPU) gets better at squeezing more out of the raw hardware.

The Value of Time

Nvidia revealed that Bristol-Myers Squibb is building an AI factory in a "fast follow to the Roche and Lilly build-outs" and revealed $8 billion in trailing-twelve-month automotive revenue and a combined $7 billion across financial services, manufacturing and healthcare.  In short, AI infrastructure spending is escaping Silicon Valley and into capital-intensive, physical world industries.

Thinking about why capex commitments are being made across a range of disparate industries, the answer is more consistent than it first appears.  Lilly, Roche and Bristol-Myers Squibb are buying compressed drug-discovery and development cycles.   Samsung is buying compressed lithography and chip-design cycles, while Jane Street and Hudson River Trading are buying compressed reaction and computation time.  Automotive manufacturers are buying compressed simulation and autonomous-driving iteration cycles. 

None of these companies is buying computing power for its own sake: they are buying time.  They are developing the ability to do in months what used to take years.  This points directly at another interesting question: what happens to the value of the asset these companies are renting or building, once the software running on it gets much better at compressing that same time?

A Strange Feature of Computing Capacity

Normally, when demand for a productive asset explodes the response is simple: build more of it.  This is exactly what is happening with the growth of data centres.  However, computing power is an unusual asset, because its effective supply can expand without anyone building another silicon wafer: software and better systems integration can also manufacture capacity.

In the case of software, DeepSeek’s Multi-Head Latent Attention (MLA) technique has reduced memory per word produced by AI to ~70KB, versus 190–330KB for comparable models using the standard approach (a 3x to 5x reduction in memory requirements) and is an example that meaningful computing capacity can be manufactured in software.  In the case of systems integration, Nvidia’s Vera Rubin addresses a range of bottlenecks including doubling the speed at which tokens are routed between subnetworks on chips and increasing memory per processor to prevent the processor resting idle while it waits for data.  This is expected to deliver a 10x reduction in cost per token processed.  Given this, the true effective AI capacity can be expressed as:

Effective AI capacity = raw silicon compute × software efficiency x system integration efficiency

Almost all investor attention goes into forecasting the first term (GPU shipments, data-centre construction, power availability).  The second and third terms may matter just as much but are far harder to forecast because they depend on the next clever piece of engineering rather than a visible construction timeline.  A trillion dollars of data centre does not necessarily buy a trillion dollars' worth of computing scarcity, if next year's software gets substantially more useful work out of the same installed base.

An Unusual Capital Cycle Problem

This creates an asymmetry where GPUs and data centres are financed today against assumptions about future utilisation, pricing, useful life and computational intensity per task.  Software improvements can invalidate those assumptions after the capital has already been committed.  Crucially, the same technological improvement does not impact the ultimate user.  For example, Eli Lilly does not particularly care whether discovering a molecule takes half as many GPUs, discovering the molecule is unambiguously good news.  Given this, it is worth mapping where economic risk plausibly sits along the value chain:

• Consumers of tokens and applications should benefit: software businesses whose input costs fall as the cost of computing behind them cheapens may be able to either pass on savings to take market share or keep them as margin.

• Physical-world buyers are arguably even better placed: they are paying for an outcome (a faster discovery, a better chip, a safer driving model) so more efficient computing is pure benefit (although this may also benefit competitors!).

• Asset-light renters of infrastructure might benefit: businesses which buy wholesale computing capacity and resell it with a value-added element plausibly benefit from falling wholesale costs without carrying the hardware depreciation risk.

• Hyperscalers may or may not be beneficiaries:  They are enormous owners of exactly the assets at risk from an efficiency shock, but simultaneously the largest sellers of proprietary tokens and models (Azure OpenAI, Google Gemini, AWS Bedrock), capturing model-layer economics a pure infrastructure view would miss.

• Pure infrastructure owners are at risk: the independents whose economics rest on long-dated take-or-pay contracts look like the cleanest exposure to utilisation and depreciation risk, with the least offsetting business to fall back on.

A Jevons Paradox

Despite the issues outlined above, none of this is a straightforward bear case for infrastructure because there is a chance that cheaper computing cost accelerates total usage by enough to grow the total value of AI consumption.  This is a Jevons Paradox, a term coined in 1865 when William Jevons noted that Watt’s more efficient steam engine, despite using less coal per unit of work versus its predecessors, was increasing coal consumption as cheaper power created uses for steam which were previously uneconomic.  Within AI, the paradox is repeating given trends in token pricing and volume thus far:  Alphabet disclosed processing 3.2 quadrillion tokens a month in May 2026, up from 480 trillion a year earlier and 9.7 trillion two years earlier.

That said, the paradox doesn’t consider the pace of demand growth (only that overall demand grows).  In the late 1990s, Telecoms companies laid ~85 million miles of internet cable across the US betting that internet bandwidth demand would grow into capacity.  In 2001, only about 5% was carrying traffic, and by 2008 only ~35% of the capacity was filled by which time many of the builders were bankrupt. 

Conclusion

The entire debate therefore reduces to a single question: does cheaper AI create demand faster than efficiency creates capacity? If it does, today’s extraordinary data centre build-out may prove entirely rational, even as the cost of computing collapses. But if efficiency wins, the paradox is more uncomfortable: AI can transform the economy, demand can explode, but the owners of the infrastructure can still earn poor returns.

CDAJ

Risk Warning & Disclaimer

This article is provided for general information purposes only and is not intended to constitute investment advice, investment research, or a personal recommendation. It does not take into account the investment objectives, financial situation, or particular needs of any individual or entity.

The views expressed are those of Ptarmigan Capital as at the date of publication and are subject to change without notice. References to investment concepts, asset classes, or portfolio construction approaches are included for illustrative purposes only and should not be construed as a recommendation or a solicitation to buy or sell any security or financial instrument.

Drafting tools, including artificial intelligence and automated systems, may be used as part of our content production process, however all materials are subject to human review and approval.

The value of investments and any income derived from them may fall as well as rise, and investors may receive back less than they originally invested. Past performance is not a reliable indicator of future results. Where investments involve overseas assets, returns may be affected by movements in exchange rates.

Ptarmigan Capital Limited is an employee-owned investment management firm providing discretionary investment management services to private clients, trusts, charities, and family offices. This article is intended to explain the firm’s perspective on bespoke investment management and the factors some clients consider when choosing such an approach. It does not relate to any specific investment product or strategy.

Ptarmigan Capital Limited is authorised and regulated by the Financial Conduct Authority (FRN 940407). Registered in England and Wales (Company No. 12715470). Registered office: 17 Cavendish Square, London W1G 0PH.