From Compute to Cost: How AI's Next Phase Could Reshape Investment Winners

Artificial intelligence has been one of the defining investment themes of the past three years, with the first phase of the boom driven by a simple premise: more powerful models required more computing power. That fuelled unprecedented investment in GPUs, high-bandwidth memory, networking equipment and data centres, rewarding the companies supplying the infrastructure behind AI.

Today, that narrative is beginning to evolve.

As AI models become more efficient and the cost of training and deploying them falls, the competitive focus is shifting beyond raw performance towards commercial scalability and affordability. The question is no longer just who can build the most powerful model, but who can deliver AI at the lowest cost.

For investors, this could mark the beginning of a new phase. If efficiency becomes as important as computing power, the next generation of AI winners may look very different from the first.

The Compute Race That Built the AI Boom

The first phase of the AI cycle was defined by a simple principle: more capable language models required exponentially greater computing resources. Each successive generation demanded larger training clusters, greater processing power, higher-bandwidth memory, faster networking infrastructure and increasing amounts of electricity.

The scale of this investment has been unprecedented. The four largest U.S. hyperscalers, Microsoft, Alphabet, Amazon and Meta, are collectively guiding to approximately USD 725 billion in capital expenditure in 2026, up 77% from a record USD 410 billion in 2025. Individual guidance includes roughly USD 200 billion for Amazon, USD 190 billion for Microsoft, USD 175–185 billion for Alphabet, and USD 115–135 billion for Meta. The majority of this spending is expected to be directed towards expanding AI compute capacity, data centres and supporting power infrastructure.

This wave of investment established AI infrastructure as one of the defining investment themes of the decade. Investors rewarded the "picks and shovels" of AI, namely the companies supplying the hardware and infrastructure underpinning model development and deployment. Semiconductor manufacturers, memory producers, networking providers and data centre operators all benefited from the race to build capacity ahead of rapidly growing demand. Nvidia's data centre revenue reached a record USD 75.2 billion in a single quarter, representing year-on-year growth of 92%, highlighting the extent to which value creation has been concentrated at the infrastructure layer of the AI ecosystem.

Yet while infrastructure providers captured much of the early economic value, a key question remained: could the companies purchasing this computing capacity generate sufficient commercial returns to justify the scale of their investment?

The Cost Race Begins

Several developments over the past year have begun to reshape the AI investment narrative.

The performance gap between frontier models has narrowed considerably. Independent benchmarking suggests the latest models from OpenAI, Anthropic, Google and other leading developers now perform within a relatively tight range across many common tasks, making it increasingly difficult for users to distinguish between them. As performance becomes more comparable, competition is shifting towards a new metric: cost.

The economics of AI deployment have changed just as dramatically. GPT-4 launched in March 2023 with pricing of USD 30 per million input tokens and USD 60 per million output tokens. By mid-2026, comparable capability is available from efficient open-weight and mid-tier models for less than USD 0.50 per million tokens on a blended basis, representing a decline of well over 95% in approximately three years.

Competitive pressure has been further intensified by the rapid advancement of open-source AI models, particularly those developed by Chinese companies. DeepSeek's late-2024 release demonstrated that high-performing models could be delivered at a fraction of incumbent pricing, prompting OpenAI, Google and Anthropic to compete not only on model capability but also on inference costs. As a result, improving efficiency has become just as important as increasing model scale.

While training frontier models remains highly compute-intensive, the fastest improvements are occurring in inference, the process of running trained models to generate responses. Because inference accounts for the vast majority of enterprise AI usage, reducing its cost has become a major competitive advantage.

Two technological advances are driving this transition. The first is model distillation, where the capabilities of a large, computationally intensive model are transferred into a smaller model that performs specific tasks more efficiently. The second is the growing adoption of smaller specialised models, designed to excel in targeted applications such as customer support, software development and document analysis, where broad general-purpose reasoning is often unnecessary. Together, these innovations are reducing the computational resources required to deliver practical AI applications.

The cost differential between open-source and proprietary models is now becoming significant enough to influence enterprise purchasing decisions. Meta's Llama 4 Maverick is available through managed API providers for approximately USD 0.20 to USD 0.85 per million tokens, compared with roughly USD 2 to USD 15 per million for leading proprietary models from OpenAI, Anthropic and Google. Organisations that route routine workloads through open-source models while reserving frontier models for more complex reasoning have reported cost savings of between 60% and 80% without a measurable reduction in output quality. At that level, lower inference costs become more than a technical advantage; they become a meaningful driver of technology budgets.

The competitive frontier is therefore evolving. Rather than competing solely on benchmark performance, AI developers are increasingly being judged on cost per token, inference efficiency, latency and their ability to deliver commercially scalable AI solutions.

The Investment Winners May Start to Broaden

As AI moves from experimentation to commercial deployment, investors are placing greater emphasis on measurable returns. Deloitte's State of AI in the Enterprise 2026 found that while 74% of organisations expect AI to drive revenue growth, only 20% can currently demonstrate a measurable impact. The message is not that AI has failed, but that businesses are becoming more disciplined about where and how they invest. That shift may broaden the opportunity set beyond the companies that powered the first phase of the AI boom.

The first phase of the AI boom largely favoured infrastructure providers such as Nvidia, AMD, Micron and Broadcom. These companies benefited directly from strong demand for graphics processors, high-bandwidth memory and networking silicon used to train and deploy increasingly sophisticated AI models. Their importance within the AI ecosystem remains unchanged as deployment costs fall. However, the marginal buyer is becoming more cost-conscious, with greater emphasis placed on performance per dollar and efficiency per watt rather than simply maximising computing power.

Potential beneficiaries of the next phase include enterprise software companies such as Microsoft, Salesforce, ServiceNow and Adobe. As inference costs decline, embedding AI capabilities into existing software platforms becomes commercially viable across a much broader range of applications. Lower model costs improve the economics of AI-enabled features, allowing software providers to expand monetisation opportunities while leveraging their established customer bases, existing workflows and pricing power, without the capital intensity of developing frontier AI models.

The acceleration in enterprise AI adoption is already becoming evident in industry data. Menlo Ventures' 2025 State of Generative AI in the Enterprise report, based on a survey of approximately 500 U.S. enterprise decision-makers, found that enterprise spending on generative AI more than tripled from USD 11.5 billion in 2024 to USD 37 billion in 2025. Gartner forecasts global AI spending will reach approximately USD 2.59 trillion in 2026, up from USD 1.76 trillion in 2025. Within that total, spending on AI agent software is expected to increase from USD 86 billion to USD 206 billion, making it one of the fastest-growing segments of the AI market and highlighting the expanding role of enterprise software in the next stage of AI adoption.

Does Cheaper AI Mean Less Demand for Chips?

This is one of the most important questions facing AI investors today, and the answer is far from straightforward. The intuitive view is that if AI models become significantly more efficient, demand for GPUs, memory and supporting infrastructure should moderate over time. Fewer chips per query and lower computing requirements could eventually reduce the pace of AI infrastructure investment.

The counterargument draws on Jevons Paradox, which suggests that improvements in efficiency often increase, rather than reduce, overall consumption. As AI becomes cheaper to train and deploy, businesses may embed it across customer service, software development, healthcare, finance and manufacturing. Lower inference costs could also unlock new applications, including AI agents, autonomous workflows and always-on digital assistants, that were previously uneconomic.

Recent industry data supports this possibility. At NVIDIA's GTC conferences, CEO Jensen Huang argued that reasoning models and agentic AI could require substantially more computing than previously expected, initially suggesting workloads around 100 times greater than earlier assumptions before later increasing that estimate to approximately 1,000 times. Deloitte's State of AI in the Enterprise 2026 also found that 23% of organisations are already deploying agentic AI at scale, with adoption expected to reach 74% within the next two years.

Ultimately, the question is not whether AI demand will continue to grow, but where the economic value created by that growth is likely to accrue. If AI adoption accelerates faster than computing requirements decline, the long-term outlook for semiconductor and infrastructure providers could remain stronger than many expect.

What Investors Can Watch

Rather than reacting to every AI headline, investors should focus on several indicators that are likely to shape the next phase of the AI cycle:

• Hyperscaler capital expenditure. Spending plans from Microsoft, Amazon, Alphabet and Meta remain the clearest leading indicator for AI infrastructure demand.
• Inference costs. Continued declines in cost per token signal ongoing efficiency gains and broader commercial adoption.
• Enterprise adoption. AI usage metrics disclosed by software companies during earnings season can provide valuable insight into real-world deployment.
• Semiconductor demand. Trends in GPU orders, high-bandwidth memory and advanced packaging remain important indicators of infrastructure investment.
• AI monetisation. The market is increasingly focused not on how much companies spend on AI, but on whether those investments translate into revenue growth and improved margins.

The Next Winners May Look Different

The compute race is far from over, but it is no longer the only driver shaping the AI investment landscape. As the technology matures, the focus is expanding beyond model capability towards cost efficiency, commercial scalability and sustainable returns. This progression is consistent with the evolution of previous technology cycles, where long-term value creation ultimately shifted from enabling infrastructure to widespread commercial adoption.

For investors, the key takeaway is that AI remains a compelling long-term structural growth theme. Rather than diminishing the investment opportunity, improving cost efficiency has the potential to broaden it. As AI becomes more affordable to deploy, the next generation of winners may not be the companies building the most powerful models, but those best positioned to commercialise AI at scale and generate enduring economic value from its adoption.