Key takeaways:
- AI demand remains supply constrained, with hyperscalers indicating that available computing capacity, rather than customer demand, is the primary limitation on growth.
- Cloud platforms are demonstrating increasingly tangible AI monetisation through accelerating revenue growth, rising inference workloads and expanding customer commitments.
- Record backlogs, increasing investment in custom silicon and expectations for continued capital expenditure growth support a constructive outlook for AI infrastructure beneficiaries.
The latest earnings reports from the largest hyperscale technology companies offered clear signals on the direction of AI infrastructure spending. Despite record levels of capital expenditure and some pressure on near-term free cash flow, management teams across Amazon, Microsoft, Alphabet and Meta emphasised the same message: demand for AI computing capacity continues to exceed available supply.
Importantly for investors, the discussion has evolved beyond building infrastructure toward monetising it. Accelerating cloud growth, expanding customer commitments, increasing use of proprietary AI chips and growing adoption of AI-enabled services all point to improving visibility on future revenue generation.
We highlight the key themes emerging from the latest results and what they may mean for the next phase of AI infrastructure investment:
1. AI demand remains supply constrained
Despite historic capital outlays, the four hyperscalers’ reports confirmed that AI infrastructure demand continues to outstrip supply. Demand growth is being constrained by computing capacity, chips, power supply and electrical grid infrastructure.
2. AI capex spending looks set to continue rising into 2027
While not considered to be formal guidance, discussions on visibility and supply constraints for the four hyperscalers indicate that capital expenditure (capex) is likely to continue rising into 2027. Microsoft, Amazon, Alphabet and Meta’s annual capex guidance for 2026 so far has reached a combined US$730-760 billion, with US$430 billion set to be deployed in the second half of the year.1 Analysts estimate that capex could continue to rise to more than US$1 trillion through to 2028,2 with visibility on increased capex stretching out to the next two years.
3. AI monetisation is accelerating
The hyperscalers reported accelerating revenue growth of 82%, 37%, and 39% respectively for Alphabet, Amazon, and Microsoft’s cloud divisions.
- Fast growth area: Cloud platforms
Amazon Web Services (AWS) saw growth accelerate to 37% as customer AI workloads shifted from early-stage pilots and training into high volume production models and inferencing volumes. Likewise, Google Cloud logged 82% revenue growth in the latest quarter, driven by its model inference and Gemini API (application programming interface) consumption.
- AI is accelerating advertising and search monetisation
Strong indirect AI monetisation through improved tools, recommendations and modelling have driven higher conversion and ad returns. Meta for example, has benefited from AI, and seen improving customer targeting, content recommendations and campaign optimisation, leading to higher conversion rates and stronger returns on advertising spend.
It is worth noting that while Meta is one of the largest capex spenders, at present it does not sell its compute to third parties. Meta’s overall revenue growth at 28% was the strongest among the four hyperscalers, but the sustainability of that growth level is less clear without a backlog to report, while the creation of a cloud division to sell compute is still in early stages of development.
- Slow growth area: Enterprise seat expansion
Enterprise seat expansion (purchasing of more user licenses as deployment grows) is lagging raw infrastructure usage for now with Microsoft reporting 30 million paid MS 365 Copilot enterprise seats in the latest quarter, while this has seen growth, it remains a very small portion of its customer base. We view this as indication also that adoption of AI native workloads, which are designed with AI as the core foundation (as opposed to bolting on AI capabilities later) remains nascent at present, but we could likely see a much higher adoption rate to come.
4. Massive backlog growth: Improving visibility on infrastructure growth sustainability
The three large cloud providers, Amazon, Microsoft and Alphabet, reported record reported backlog, i.e. contracted revenue not delivered and not yet recognised. These backlogs totaled more than US$1.6 trillion, some 5-10x larger than their annual revenue rates. The duration of contracts and the time to recognise revenue varies, but on average there is an expectation for around 25% of backlogs to be fulfilled and recognised as quarterly revenue in the following 12 months provides an indication of the high visibility that these companies now have.
5. Rising investment and monetisation of custom silicon
While Meta and Microsoft are still developing and ramping their own silicon programmes, Amazon and Alphabet have been investing in custom silicon since 2012-2013. This has given them a headstart in developing specialised chips for AI training and inference, tailoring their AI infrastructure for greater efficiency and lower cost. For these companies, custom silicon provides a means of differentiation and offers cost savings internally and also to its customers.
Amazon noted that its proprietary chip business has surpassed US$25 billion in the latest quarter, up from over US$20 billion in Q1. Amazon only monetises its silicon as a cloud utility service at present, but Alphabet in addition to cloud sales, also has pivoted to deploying its proprietary TPU (Tensor Processing Unit) system hardware for customer data centres, rather than just offering it remotely through Google Cloud.
6. Cloud hyperscalers are benefiting from multi-model systems, mixing proprietary and open source/open weight models
Hyperscalers can provide customers with access to a broad range of AI models through a single platform, including both proprietary models (such as OpenAI’s GPT models) and open-source alternatives. This simplifies deployment, security, governance and cost management, allowing customers to choose the most appropriate model for each task, while enabling cloud providers to capture additional value through AI infrastructure and related services. Given supply constraints, providing access to a mix of proprietary open source/open weight models allows the hyperscalers to sell higher margin services. Amazon noted that its Bedrock service (a fully managed platform) gained more customers in the last six months than in the first two years of the service’s existence.
7. Diverging margin profiles as hyperscalers evolve
Amazon and Alphabet have expanded cloud operating profitability, using their own silicon expertise and internal efficiency gains, while Microsoft continued to see some margin pressure given its heavy infrastructure spend and depreciation costs. For Amazon, cloud is not only the fastest growing part of the business but by far the most profitable. For Alphabet, the cloud division is not as profitable as its core search and advertising business, but profitability has surged (up over 200% year-on-year) and contributed meaningfully to total revenues. Overall, Microsoft has the highest margins of the three companies, but its cloud gross margins declined. We are now seeing evidence of these hyperscalers evolving as the major platforms for AI orchestration, involving the coordination of multiple AI models, applications and data sources through a single platform or workflow. This process allows businesses to optimise costs and improve security in an increasingly multi-model world.
Conclusion: The AI infrastructure investment cycle has room to run
Hyperscalers continue to report strong AI infrastructure demand, with growth constrained by AI supply capacity. This suggests to us that the AI investment cycle and AI infrastructure growth remains in its early stages and is a multi-year process that has yet to peak.
AI monetisation is evolving with companies finding multiple channels to generate revenue, including custom silicon and multi-model systems. Meanwhile, backlogs are providing greater visibility on the sustainability of future demand. In our view, this reinforces the case for investing in the companies that are enabling the buildout of AI infrastructure.
References made to individual securities do not constitute a recommendation to buy, sell or hold any security, investment strategy or market sector, and should not be assumed to be profitable. Janus Henderson Investors, its affiliated advisor, or its employees, may have a position in the securities mentioned.
There is no guarantee that past trends will continue, or forecasts will be realised.
All company information sourced from earnings results releases and call transcripts, unless otherwise specified. Amazon Investor Relations; Q2 2026 earnings results; 30 July 2026; Google Blog; Q2 2026 earnings call: Remarks from our CEO; 22 July 2026; Microsoft Investor Relations; Earnings Release FY26 Q4; 29 July 2026; Meta Q2 2026 earnings call transcript; 29 July 2026.
1 Company reports, Statista as at 30 July 2026.
2 Morgan Stanley, Bank of America estimates, end July 2026.
Application Programming Interface: An API enables two software components to communicate with each other to exchange data, features and functionality using a set of definitions and protocols.
Backlog (Remaining Performance Obligations or RPO): The value of contracted revenue that a company has not yet recognised because the services have not yet been delivered. A growing backlog can provide greater visibility into future revenues.
Capital expenditure (capex): Money invested to acquire or upgrade fixed assets such as buildings, machinery, equipment, vehicles, technology in order to maintain or improve operations and foster future growth.
Cloud computing: The delivery of computing services, including data storage, processing power and software, over the internet rather than through local computers or servers.
Custom silicon: Computer chips designed for a specific purpose rather than general use. Companies such as Amazon and Alphabet develop custom AI chips to improve performance and reduce costs.
Enterprise seat expansion: Refers to the traditional software growth model where an organization scales its deployment by purchasing more user licenses (“seats”) as adoption of workplace AI tools (such as enterprise assistants or LLMs) grows across teams and departments.
Free Cash Flow (FCF): Cash that a company generates after allowing for day-to-day running expenses and capital expenditure. It can then use the cash to make purchases, pay dividends, or reduce debt.
Generative AI: Artificial intelligence capable of creating new content, including text, images, audio and software code, based on patterns learned from large datasets.
Hyperscaler: A large cloud computing company that operates vast data-centre networks and provides computing infrastructure at global scale. Examples include Amazon Web Services (AWS), Microsoft Azure and Google Cloud.
Inference: The process of using a trained AI model to generate outputs, such as answering questions, producing images or making predictions.
Large Language Model (LLM): An AI model trained on vast amounts of text data that can understand and generate human-like language.
Monetisation: The process of generating revenue from a product, service or technology investment.
Multi-model environment: An AI ecosystem where organisations use multiple AI models, often combining proprietary and open-source models, depending on the task.
Open source: An AI model whose underlying code and, in some cases, model weights are made publicly available for developers to use, modify and distribute.
Open weight model: Provides the final trained parameters and model architecture but typically omits the full training dataset or complete data collection methodology due to legal, cost, or privacy constraints.
Proprietary model: An AI model owned and controlled by a company, with access and usage governed by commercial terms rather than open-source licensing.
TPU (Tensor Processing Unit): A custom chip developed by Google specifically for AI and machine-learning workloads. TPUs are designed to perform AI calculations more efficiently than traditional processors.