Upstream Firms Have Seen Exceptional Gains, But These Will Not Last Forever
Today, most of the earnings in the AI value chain are still captured by upstream players. These are the companies that supply the hardware and semiconductors needed to build AI infrastructure, including GPU players such as Nvidia and AMD, CPU players such as Intel, as well as memory and foundry companies.
Figure: Upstream AI Companies Have Seen Exceptional Gains Year-to-Date
Source: Bloomberg, 2 Jul 2026
These upstream suppliers have enjoyed exceptional revenue growth, strong margins, and sharp share price gains. The reason is simple: hyperscalers have been spending heavily on building data centers to meet surging AI demand. Each data center requires large amounts of GPUs, CPUs, memory chips, and networking equipment. In the AI sector value chain, hyperscalers are the midstream players buying the hardware and turning it into usable computing capacity.
Table: AI Industry Value Chain
Source: Tiger Brokers
However, upstream growth cannot continue indefinitely on its own. The “pick-and-shovel” trade only works so long as the midstream still has sufficient cash to fund continued AI capex. Today, that cash position is becoming tighter. Data center spending has risen to such elevated levels, leaving many hyperscalers with operating cash flow only just covering their capex and free cash flow approaching negative territory. Hyperscalers have also repeatedly seen their share prices penalized by investors for high capex guidance, even when earnings remained strong.
Figure: Hyperscaler capex has already approached the upper limit of operating cash flow
(AMZN + MSFT + META + GOOGL + ORCL)
Source: Bloomberg, Tiger Brokers
The pressure is now on midstream players to convert AI capex into AI monetization, allowing operating cash flow and, ultimately free cash flow to improve. That depends on hyperscalers’ customers: downstream firms such as AI model providers OpenAI and Anthropic, as well as application companies that embed AI into products, workflows, and software to create recurring revenue. As these companies earn revenue from end users, they in turn pay hyperscalers for compute and data center capacity.
From Training to Inference: New Monetization Opportunities for the AI Sector
The key question now is whether downstream firms can successfully convert AI adoption into sustainable revenue. Early signs are encouraging, as the industry shifts from model training to inference, with Agentic AI driving higher token consumption and enabling cloud providers and software companies to monetize AI through usage-based pricing, with $Oracle(ORCL)$ ’s recent adoption of token-based pricing serving as one example. As AI monetization moves downstream, the long-term growth of upstream AI hardware companies will increasingly depend on cloud providers, AI model companies, and software firms generating sustainable returns on AI investment.
Midstream and Downstream Winners
However, it is not easy to discern which software companies will harness AI effectively, and which risk becoming obsolete. At this stage, the more compelling exposure lies with leading AI model providers that have already demonstrated strong annualized revenue run rates, such as OpenAI at roughly $24 billion as of March 2026 and Anthropic at roughly $44 billion as of May 2026.
As OpenAI and Anthropic are not yet publicly listed, equity investors can only gain indirect exposure through their strategic partners. $Microsoft(MSFT)$ holds a 27% stake in OpenAI and is entitled to 20% of OpenAI's revenue, subject to a reported cap of US$38 billion. $Oracle(ORCL)$ has also secured a US$300 billion computing contract with OpenAI, positioning it as a major beneficiary of OpenAI's AI infrastructure spending. On the Anthropic side, $Amazon.com(AMZN)$ and $Alphabet(GOOGL)$ are its key strategic investors and cloud partners.
Within the midstream layer, large hyperscalers such as Amazon, Microsoft, Oracle, and Google, with substantial RPO backlogs and broad enterprise customer bases, are likely to benefit from the growing demand for AI inference. In contrast, some neocloud providers could face competitive pressure if major customers, such as Meta, eventually begin renting out their excess AI compute capacity, reducing their reliance on providers like CoreWeave and Nebius.
Conclusion:
Today, the bulk of earnings in the AI sector are captured by upstream players, namely semiconductor and hardware manufacturers. However, as the industry shifts from model training to inference, value is likely to migrate toward midstream cloud providers, as growing token usage during inference drives demand for AI compute and data center capacity, and ultimately to downstream AI companies that monetize AI through customer-facing applications. The AI hardware rally can continue only if downstream AI adoption generates sufficient revenue to fund continued investment in AI infrastructure.
Comments