The AI Marathon: Why Lasting Value May Matter More Than Leading the Race
AI capability is still advancing. For investors, the harder question is which businesses can convert that progress into trust, adoption and durable returns.
By Jura Capital
The race to build the world’s most powerful artificial intelligence has rewarded speed, scale and ambition. But leadership is proving temporary, costs continue to rise and safety concerns are becoming harder to separate from commercial strategy.
Medicine and smartphones offer useful parallels. Both moved from periods of extraordinary invention into more measured phases, where evidence, regulation, distribution and recurring revenue became increasingly important. AI may now be approaching a similar transition.
For investors, the next winners may not simply be the companies that produce the smartest model first. They may be those capable of turning rapidly changing technology into trusted products, durable customer relationships and sustainable returns.
Progress Has Been the Product
For much of the past three years, the AI industry has operated according to a simple assumption: the most capable model will attract the most users, capital and influence.
That assumption encouraged an extraordinary development race.
OpenAI established an early consumer lead with ChatGPT. Google, Anthropic, Meta, xAI and a growing group of Chinese developers responded with increasingly capable alternatives. New releases have regularly arrived with claims of better reasoning, stronger coding, lower costs or improved performance on carefully selected benchmarks.
The result has been genuine progress. Stanford’s 2026 AI Index documents rapid gains on Humanity’s Last Exam, a benchmark specifically designed to remain difficult for advanced AI systems. Research from METR also indicates that the length and complexity of tasks frontier systems can complete has been increasing rapidly.
AI is not running out of capability.
But the commercial meaning of being “best” is becoming less obvious.
By March 2026, Stanford found that models from Anthropic, xAI, Google and OpenAI were competing within a narrow band on Arena-style human-preference rankings. Alibaba and DeepSeek were not far behind. In practical terms, several companies are now competing within the same top tier.
Leadership still matters. It simply appears to expire more quickly.
The Fastest-Moving Crown in Technology
A chart of the leading AI models since 2023 would not show one company steadily extending an insurmountable advantage. It would show repeated changes in leadership as new models overtook previous releases.
Even that chart would need careful explanation.
LMArena ranks models using people’s preferences in blind comparisons, making it a useful indicator of how users experience different systems. But no single leaderboard can definitively identify the world’s “best” model. Performance varies across coding, reasoning, factual accuracy, creative writing, speed and cost.
Researchers have also warned that public leaderboards can be influenced by testing behaviour and optimisation. One study identified 27 private model variants tested by Meta before the release of Llama 4, illustrating how strongly developers may optimise around prominent evaluation systems.
That does not make the competition meaningless. It reveals how narrow the lead can be.
A model can top one benchmark and trail on another. It can be the most capable option in March and look comparatively ordinary by June. The commercial danger is clear: enormous sums can be committed to winning a position that may last only until the next release.
LMArena snapshots
The AI race has more than one finish line
The highest-ranked model from each developer, with rank one at the top. Leadership rotates by task and over time.
Overall preference
OpenAIJun 25
OpenAIDec 25
GoogleMar 26
AnthropicSep 26
Anthropic
Coding
OpenAIJun 25
OpenAIDec 25
AnthropicMar 26
AnthropicSep 26
Anthropic
Mathematics
OpenAIJun 25
OpenAIDec 25
AnthropicMar 26
GoogleSep 26
Anthropic
Source: LMArena historical leaderboard dataset, text style-control subset, snapshots frozen 13 September 2026. Arena moved from Elo to Bradley-Terry in January 2024, made style control the default in May 2025 and introduced frequency re-weighting in July 2025. Rankings measure human preference within each category, not universal capability.
What Medicine Teaches Us About Useful Friction
Medicine offers a valuable comparison, although not because pharmaceutical innovation has reached a natural peak.
New drugs move through staged research, clinical trials, regulatory review and continued monitoring after approval. The process is deliberately slower than the underlying science because discovery alone is not enough. A treatment must also demonstrate safety, effectiveness, consistent manufacturing and an acceptable balance between benefit and risk.
The US Food and Drug Administration describes a multi-stage development process spanning discovery, pre-clinical research, clinical trials and regulatory review. Monitoring continues after launch because clinical trials cannot reveal every possible effect in advance.
This friction can delay valuable treatments. It can also prevent unproven or unsafe ones from reaching millions of people.
AI is different. A new model can be deployed globally almost immediately, updated without a physical recall and embedded inside products used for education, healthcare, finance and public services. The distance between invention and adoption is dramatically shorter.
Yet the underlying principle increasingly applies: greater capability creates a greater need for evidence.
Leading AI companies have already acknowledged this. At the Seoul AI Summit, 16 organisations made voluntary frontier-safety commitments, including commitments to assess risks and define circumstances in which a model would not be developed or deployed.
The difficulty is that voluntary restraint comes with a commercial cost.
If one company delays a release for additional testing while a competitor launches, the cautious company risks losing customers, attention and investor confidence. Every participant may recognise the value of slowing down, while each remains individually incentivised to keep moving.
This is why AI safety is not simply an ethical question. It is a coordination problem.
Capability and convergence
Better models. Shorter leads.
SWE-bench resolved
4.4% → 71.7%
2023 to 2024
Gap between first and second
4.9% → 0.7%
2023 to 2024
The gap from first to tenth also narrowed from 11.9% to 5.4% by early 2025.
Source: Stanford AI Index 2025, Technical Performance. Progress is rapid while a wider group clusters near the frontier.
The Smartphone Lesson: Innovation Does Not End, but Value Moves
Smartphones offer a second and more commercial parallel.
The early smartphone market was defined by obvious leaps: touchscreens, application stores, better cameras, faster mobile internet and entirely new ways of accessing services. Consumers replaced devices frequently because each generation offered a visible improvement.
Eventually, the product matured.
Innovation continued, but annual changes became less transformational for many users. Replacement cycles lengthened and manufacturers increasingly competed through ecosystems, services, brand loyalty, financing, premium positioning and control of distribution.
The handset stopped being the entire product. It became the gateway to a much larger commercial relationship.
AI could follow a similar pattern.
If leading models become increasingly close in capability, customers may care more about how well a system integrates with their existing work, what happens to their data, whether its answers can be trusted and how reliably it performs at scale.
The model may remain essential while becoming less visible.
That would move value towards the businesses controlling distribution, cloud infrastructure, specialist data, workflow integration and customer relationships. It could also favour smaller models that solve a particular problem reliably and economically over frontier systems designed to perform almost everything.
There is already evidence of this shift. The cost of using a model capable of matching GPT-3.5 on a prominent benchmark fell from $20 per million tokens in November 2022 to just $0.07 by October 2024, according to Stanford’s 2025 AI Index. That represents a reduction of more than 99% in less than two years.
Rapidly falling prices are excellent for adoption. They are more complicated for businesses whose advantage depends primarily on selling access to intelligence.
Where value moves
Breakthrough technology becomes lasting infrastructure
As a market matures, value tends to move beyond the invention itself and into the systems people rely on every day.
01
Breakthrough
A new capability captures attention.
02
Infrastructure
Tools make it dependable and accessible.
03
Distribution
Products reach customers at scale.
04
Trusted use
Habit, confidence and recurring value emerge.
The smartphone market followed this path. AI may do the same, with durable value accumulating around reliable products and customer relationships.
The Economics Are Beginning to Demand Proof
The financial stakes make this transition increasingly important.
Microsoft’s quarterly capital expenditure reached approximately $41 billion, while Meta projected annual 2026 capital expenditure of between $130 billion and $145 billion. Microsoft has also acknowledged that much of its spending is directed towards relatively short-lived assets, including processors that will eventually need replacing. Axios
This is not conventional software investment.
Traditional software can often be reproduced for another customer at minimal additional cost. Frontier AI requires data centres, electricity, advanced chips, cooling systems and continual investment in new generations of hardware. The leading companies are not only funding innovation. They are financing an industrial build-out.
That creates two competing pressures.
Moving quickly can secure developers, enterprise contracts, consumer habits and control of emerging standards. But moving too quickly can produce waste, regulatory intervention, safety failures or infrastructure that becomes outdated before it earns an acceptable return.
The winner cannot simply arrive first. It must earn enough from its position to pay for staying there.
The durability test
Capability opens the door. The business model keeps it open.
The strongest AI businesses may be those that connect technical progress to a repeatable economic engine.
Useful product
Solves a real problem
Trusted adoption
Earns confidence and habit
Durable economics
Supports recurring returns
For investors, the question shifts from who leads today to who can convert innovation into a trusted, enduring business.
Why Agreement Is Rational and Difficult
A slower, more deliberate development cycle could benefit the industry.
It would give companies more time to evaluate systems, develop standards, improve reliability and build products customers are prepared to pay for. It could reduce the risk that a high-profile failure destroys confidence across the entire sector.
It could also protect long-term returns by limiting a cycle in which every capability gain immediately triggers another round of spending.
But an agreement is only valuable if competitors follow it.
AI is being developed by companies and countries that do not share identical commercial priorities, political systems or attitudes towards risk. A unilateral pause may improve safety while weakening the position of the organisation that observes it.
This creates an uncomfortable reality. The industry may collectively benefit from restraint while its individual participants continue to benefit from speed.
Effective coordination will therefore require more than general promises. It is likely to need shared evaluation standards, greater transparency, independent testing and sufficiently broad international participation to prevent responsible developers from carrying all the competitive cost.
Regulation that is too weak may fail to change behaviour. Regulation that is too rigid may protect established companies by making it prohibitively expensive for smaller challengers to comply.
The objective should not be to stop progress. It should be to make the consequences of releasing a system part of the cost of developing it.
What The Market Actually Repriced
The clearest evidence that this debate is now a financial one arrived in the market itself.
Within days of prominent AI leaders arguing for a slower pace of capability development, close to a trillion dollars of market value moved. The reaction was not uniform, and the pattern is the most instructive part of it.
The companies that fell hardest were the ones that sell the buildout: chip designers, semiconductor equipment manufacturers, memory producers and the investment vehicles built around them. The companies that held firm, or rose, were the ones that run software on top of that infrastructure and sell it to existing customers.
Reported weekly share price moves
The sellers of the buildout fell. The users of it did not.
Approximate moves during the week that calls for a slower development pace dominated the news agenda. Figures are indicative of direction rather than precise close-to-close returns.
Companies that sell the infrastructure
Korean and Japanese memory manufacturers also fell sharply.
Companies that run software on top of it
The split matters more than the size of the moves. No published capital expenditure plan changed during the week.
What is striking is what did not happen. No capital expenditure programme was reduced. No chip order was reported as cancelled. Every one of the executives calling for restraint continued to fund their existing plans.
Investors did not reprice demand. They repriced the assumed rate of growth in that demand, and that alone was enough to remove hundreds of billions from the most capital-intensive part of the chain.
That is the signature of a market valued on acceleration rather than earnings. When a business is priced for demand that compounds indefinitely, a change in expected pace does the damage that a change in results would normally do.
Whether the development pace should slow is a serious question, and it deserves to be argued on its own terms rather than through share prices. But the week was a useful drill for portfolios. Many investors discovered they held a direct exposure to the AI construction cycle, rather than to AI adoption, and had not consciously chosen it.
The distinction is worth making before it is made for you. Exposure to the buildout is a bet on the pace of spending. Exposure to adoption is a bet on the usefulness of the output. They will not behave the same way in a slower cycle.
The Jura View
The AI investment case is moving into a more demanding phase.
For the past three years, investors could reasonably reward companies for technical progress, user growth and the scale of their ambition. The next phase will require more evidence of conversion: from capability to adoption, adoption to revenue and revenue to durable returns.
We believe temporary model leadership will become a weaker source of competitive advantage as top-tier performance converges and costs fall.
The more defensible positions may sit elsewhere:
- Distribution that makes an AI product difficult to replace
- Proprietary data that improves a specific commercial outcome
- Infrastructure supported by contracted demand
- Integration within important customer workflows
- Trust, security and regulatory approval in sensitive industries
- Business models that improve as the underlying models become cheaper
This does not mean the model developers will lose. Some may combine frontier research with infrastructure, distribution and customer relationships so effectively that they capture value across the entire system.
But investors should distinguish between leading the race and owning the course.
The company with the smartest model today may attract the headlines. The company that customers still depend upon in ten years may create the more valuable business.
Conclusion: The Marathon Begins After the Breakthrough
AI remains one of the most important technological developments of this generation. Its capabilities are still improving rapidly, and dismissing that progress would be a mistake.
It would be equally mistaken to assume that every improvement creates a durable commercial advantage.
Medicine shows why consequential innovation eventually requires stronger evidence and monitoring. Smartphones show how value can move from the original breakthrough towards ecosystems, services and distribution. AI is likely to develop its own version of both transitions.
The defining question is changing.
It is no longer only: who can build the most intelligent system first?
It is also: who can deploy it responsibly, integrate it deeply, earn trust and generate returns for long enough to finance the next generation?
The sprint created the AI industry.
The marathon will decide who captures its value.
Referenced research and data sources
- Stanford AI Index 2026: Technical Performance
- METR: Task-Completion Time Horizons of Frontier AI Models
- LMArena Leaderboard
- NeurIPS 2025: The Leaderboard Illusion
- US FDA: Drug Development and Review Definitions
- US FDA: The Drug Development Process
- UK Government: Frontier AI Safety Commitments
- Stanford AI Index 2025
- Axios: Meta and Microsoft report rising AI expenditure
This article is for information only and does not constitute investment advice. Private-market investments are high risk and illiquid. Capital is at risk.