In the September edition of the CAPIS Monthly Research Briefing, Andi Kerenxhi of Ubineer joined CAPIS’ Ed O’Dowd for an updated look at the massive capital investment cycle reshaping the technology sector. Revisiting several themes from his previous CAPIS appearances, Kerenxhi examined how spending, demand and financial commitments have evolved across Meta, Alphabet, Microsoft, Amazon and NVIDIA.
Drawing on Ubineer’s analysis of financial filings and AI-driven research, Kerenxhi argued that the AI infrastructure buildout still has significant room to run. At the same time, rapidly expanding capital commitments, changing financing dynamics and growing supply-chain dependencies are creating new risks for investors to monitor as the cycle matures.
0:02 – Introduction: Ed O’Dowd welcomes returning guest Andi Kerenxhi of Ubineer and sets the stage for an updated discussion on mega-cap technology, AI and the accelerating capital expenditures supporting the infrastructure buildout. Kerenxhi also introduces Ubineer’s approach, which uses AI to predict financial KPIs, identify anomalies in company filings and simulate investor expectations through a “swarm” of AI agents.
3:54 – AI CapEx Has More Room to Run: Beginning with Meta, Kerenxhi revisits his earlier expectation that advertising growth and capital expenditures would continue accelerating. While investors are increasingly questioning whether the company is approaching peak CapEx, he argues the cycle is not finished. Construction in progress, land holdings and substantial contractual obligations point to additional infrastructure spending ahead, with Kerenxhi expecting Meta’s CapEx to be higher in 2027.
7:24 – Can Meta Sustain the Spending?: With Meta’s on- and off-balance-sheet obligations continuing to grow, Kerenxhi turns to the question of whether its underlying business can support the investment cycle. He identifies cash flow from operations as a critical metric, noting that continued cash-flow growth has so far supported investor willingness to finance the buildout. But expectations are changing: as spending rises and yields make the time value of money more consequential, investors increasingly want to see measurable returns from new products and AI investments.
10:00 – Alphabet — Looking Beyond the Free Cash Flow Headline: Turning to Alphabet, Kerenxhi highlights the continued expansion of Google Cloud alongside rapidly increasing CapEx. He addresses concerns surrounding a recent quarter of negative free cash flow, arguing that the headline obscured an unusual increase in inventory that had not yet been monetized. More broadly, however, Alphabet’s on- and off-balance-sheet commitments have expanded significantly, raising the prospect that the company may increasingly rely on debt, equity or both to finance its continued infrastructure investment.
13:59 – Microsoft: Strong Cloud Demand, but a More Measured CapEx Approach: Kerenxhi reviews Microsoft’s productivity and cloud businesses, pointing to higher-priced offerings, improving deferred revenue and expanding long-term customer commitments as signs of continued demand. At the same time, Microsoft stands apart from several of its hyperscaler peers by beginning to moderate the pace of its capital spending. Kerenxhi explains that management appears focused on building capacity for a broader customer base over time rather than maintaining the same pace of investment solely to serve the largest AI companies.
19:07 – Amazon: AWS Contracts Point to Continued Growth: The discussion shifts to Amazon, where Kerenxhi highlights accelerating remaining performance obligations as an important indicator of future AWS demand. As larger contracts translate into future cloud revenue, he argues that Amazon’s growth outlook remains closely connected to continued AI and cloud investment. Like its peers, however, Amazon has also accumulated significant on- and off-balance-sheet commitments as it finances the infrastructure required to meet that demand.
21:03 – Why Investors Continue to Finance the AI Buildout: Stepping back from individual companies, Kerenxhi examines the economics supporting hyperscaler investment. Across the companies analyzed, he points to trillions of dollars in contracted future revenue alongside similarly large future obligations and substantial operating cash flow. In his view, those economics help explain why debt and equity investors remain willing to fund the AI CapEx cycle: despite the enormous dollar amounts involved, contracted demand and cash generation continue to provide support for the investment underway.
23:05 – A Critical Shift: Despite the strength of the broader buildout, Kerenxhi identifies an important change emerging beneath the surface. Suppliers to hyperscalers are increasingly securing longer-term agreements that require customers to take—or pay for—inventory regardless of whether they ultimately need it. As a result, he argues that some of the financial risk historically borne by vendors is beginning to migrate toward hyperscalers themselves. Kerenxhi does not view this as evidence that the cycle is ending, but rather as a development investors should monitor closely as commitments continue to grow.
24:08 – How Early Are We in the AI Investment Cycle?: Kerenxhi estimates that the AI infrastructure buildout remains roughly in the “third inning,” potentially approaching the fourth. He expects investment to continue as long as the economics and returns remain compelling, but emphasizes that investors need to monitor contracted revenue, financial commitments and cash generation to identify when spending begins moving from sustainable expansion toward more speculative territory.
25:29 – NVIDIA Expands Beyond the Hyperscalers: Turning to NVIDIA, Kerenxhi argues that the company’s opportunity is expanding beyond the largest cloud providers. In addition to hyperscalers, NVIDIA is targeting neoclouds, corporations and governments building their own compute infrastructure. He views this diversification as strategically important because a broader customer base can reduce NVIDIA’s exposure to consolidation among the largest buyers of compute and help preserve its pricing power.
27:56 – NVIDIA’s Supply Chain as a Key Variable: While NVIDIA continues to carry significant inventory, Kerenxhi notes a shift in its composition, with finished goods representing a smaller share than previously. He identifies this as a supply-chain risk but points to the company’s substantial purchasing commitments as evidence that it is working to secure future supply. Ubineer’s AI agents also modeled NVIDIA’s longer-term revenue trajectory, underscoring just how large the company could become if the broader AI investment cycle continues.
30:17 – Q&A Session: How Long Can the AI Buildout Last?: During the Q&A, O’Dowd asks how long the infrastructure cycle could continue. Kerenxhi says he believes the buildout could extend into 2031 or 2032, while acknowledging that geopolitical disruption and semiconductor supply constraints could alter that trajectory. He identifies Taiwan and the global semiconductor supply chain as particularly important vulnerabilities, noting that disruptions to critical manufacturing capacity could threaten the broader AI growth thesis. He also expects periodic “digestion” phases as the industry absorbs newly built capacity, without necessarily signaling the end of the longer-term cycle.
34:01 – Gemini, ChatGPT and the Evolution of AI Use Cases: O’Dowd asks whether Google’s distribution advantage could help Gemini take share from ChatGPT. Kerenxhi argues that the platforms may increasingly serve different use cases rather than compete as direct substitutes. Google benefits from enormous existing distribution and is focusing on fast, multimodal models, while ChatGPT and other AI platforms can serve different discovery and recommendation behaviors.
36:06 – Double Ordering and the Risks Behind Long-Term Commitments: Responding to an audience question about double ordering, Kerenxhi explains how suppliers are increasingly using “take-or-pay” arrangements to discourage customers from ordering excess capacity simply to secure a better position in the supply queue. Because hyperscalers may still be obligated to pay for contracted supply even if they no longer need it, he reiterates that a portion of the risk in the AI supply chain is shifting from component suppliers toward their largest customers.
37:42 – AI Risk, Safety and the Role of Agents: The conversation concludes with a broader discussion about the long-term risks and opportunities associated with artificial intelligence. Kerenxhi offers an optimistic view of AI’s potential impact on society while acknowledging concerns around safety and security as capabilities advance. He argues that broader adoption of AI agents may ultimately become part of the defense against AI-enabled threats, setting up an evolving technological dynamic that he expects will remain an important topic in future discussions.
41:50 – Closing Remarks and October Preview: O’Dowd thanks Kerenxhi for joining the briefing. He also previews CAPIS’ October Monthly Research Briefing, which will feature Markets Policy Partners and is expected to revisit politics, macro policy and the Federal Reserve ahead of the midterm elections.