{"id":1446604,"date":"2026-07-06T20:08:12","date_gmt":"2026-07-06T18:08:12","guid":{"rendered":"https:\/\/www.payoff.ch\/?p=1446604"},"modified":"2026-07-06T20:08:14","modified_gmt":"2026-07-06T18:08:14","slug":"why-hyperscalers-ai-capex-may-stay-higher-for-longer","status":"publish","type":"post","link":"https:\/\/www.payoff.ch\/en\/news\/why-hyperscalers-ai-capex-may-stay-higher-for-longer","title":{"rendered":"Why hyperscaler\u2019s AI Capex May Stay higher for longer"},"content":{"rendered":"\n

Few investment themes have generated as much enthusiasm, or as much skepticism, as artificial intelligence. Over the past months, a growing number of commentators have once again warned that the market is displaying some of the hallmarks of a bubble. The timing of SpaceX\u2019s recent mega-IPO, together with the prospect of upcoming IPOs from companies such as OpenAI and Anthropic, has reinforced comparisons with the dot-com era, amid the sharp rise in hardware and semiconductor stocks and the increasing concentration of equity market performance in a handful of AI-related companies. More recently, Alphabet’s decision to raise $84.7 billion in equity \u2014 its first public equity raise since its 2004 IPO \u2014 has deepened another existing concern: that the AI investment cycle may require far more capital than investors previously assumed, increasing the risk that AI spending will outstrip operating cash flows. Furthermore, the rising number of shares outstanding coupled with rising depreciation expenses would pressure the return on equity (ROE), which has been one of the key factors supporting hyperscalers\u2019 valuations. In this context, a few investment strategists have even warned that, if current trends persisted, the AI boom could ultimately result in one of the largest episodes of shareholder value destruction in history.<\/p>\n\n\n\n

These concerns surely deserve to be taken seriously. Periods of technological innovation are often accompanied by excessive optimism, inflated valuations, and disappointment. However, it is hard to argue against the fact that unlike many of the companies that drove the dot-com boom, today\u2019s AI ecosystem is already supported by real revenues, profits, and cold, hard cash. The extraordinary performance of semiconductor stocks has been accompanied by strong earnings growth, which has not materially pushed their valuations higher based on future profit expectations. It is true that stock market bubbles are typically characterised by a rapid rise in P\/E. However, a bubble can occasionally occur in earnings as well. This happened to homebuilders and banks in the lead-up to the GFC. Their P\/E ratios remained low, disguised by an unsustainable increase in profits. More generally, earnings bubbles are common in industries that are subject to boom-bust cycles. These industries include natural resources, airlines, shippers, and importantly for today\u2019s environment, semiconductors.<\/p>\n\n\n\n

Yet, there is another way to interpret what is happening. Rather than a speculative bubble, the current AI boom may represent the early stages of a technology supercycle. In financial markets, a supercycle occurs when a cyclical industry experiences a lasting increase in demand, creating a new and permanently higher level of spending. The electrification of the economy, the rise of the internet, and the smartphone revolution are often cited as examples.<\/p>\n\n\n\n

The source of the demand is also fundamentally different from previous technology manias. Semiconductor manufacturers are primarily selling into a customer base composed largely of hyperscalers, among the largest and most profitable companies in the world. These companies are investing unprecedented amounts of capital in AI infrastructure because they believe the technology will generate meaningful economic returns. As long as demand for AI services continues to expand, the incentive to invest remains intact. In this context, a strong demand signal has recently come from companies like Dell and HPE who have seen an influx of orders for server space. Interestingly, Dell and HPE are primarily capturing demand from the next tier down \u2014 neoclouds, CSPs, and large enterprises, who traditionally have less deployable capital than hyperscalers, and who are now also willing to spend more to build out their own AI infrastructure.<\/p>\n\n\n\n

Ultimately, the sustainability of the AI investment cycle will depend on the end-user demand, which could be measured by the business activity of leading AI platforms like Anthropic. According to recent articles from Wall Street Journal and CNBC, the company\u2019s revenue is projected to rise from $4.8 billion in the first quarter of 2026 to $10.9 billion in the second quarter, implying growth of more than 120% in just one quarter. Of course, extrapolating this run-rate to future revenue growth remains highly speculative, while estimating the total addressable market is even more uncertain. Nevertheless, Anthropic\u2019s current growth trajectory provides a compelling indication of accelerating AI adoption. This creates the central question for investors: are rising AI capex plans a warning sign of excess, or evidence that demand for computing power is structurally stronger than previously expected? While it is not necessarily a rhetorical question, the current AI investment cycle may not be close to peaking. In fact, AI capex may stay higher for longer supported by the rise of Agentic AI.
<\/p>\n\n\n\n

A FOUNDATIONAL SHIFT IN THE DEVELOPMENT OF AI<\/h3>\n\n\n\n

Think about the last time you asked a chatbot to write a summary or a draft. Or maybe answer a query. It was probably useful but you were also still driving the interaction: asking, refining, copying, checking, and moving the work forward. Now imagine a system that does not just respond, but acts. It remembers what you asked last week, understands your preferences, works across digital tools, plans a workflow, and adapts as circumstances change. That is the shift from GenAI to agentic AI: from AI that helps with thinking to AI that helps with doing. GenAI is mostly passive. It takes a prompt and produces an answer. Agentic AI is active \u2013 less a copilot for one task but an autopilot for multi-step workflows. The distinction is key because computing requirements are changing. <\/p>\n\n\n\n

To better understand it, let us introduce the notion of token. Simply put, a token is the basic unit of information processed by an AI model, both when reading an input and generating an output. It can correspond to a word, part of a word, punctuation mark, or piece of code. According to Goldman Sachs research, agentic AI could drive a dramatic increase in token consumption, reaching around 120 quadrillion tokens per month by 2030, as usage shifts from discrete, user-initiated tasks toward persistent background operations.<\/p>\n\n\n\n

Tokens processed per month (in quadrillion i.e. 1015<\/sup>)<\/strong><\/p>\n\n\n\n

\"\"
Source: Bloomberg (May 2026)<\/figcaption><\/figure>\n\n\n\n

Let us consider Goldman Sachs\u2019 simulated travel booking agent\u2019s workflow whose task is to compose a trip and execute the booking. Looking under the hood, the agent would parse intent, clarify missing inputs, search internal and external systems, rank options, incorporate user feedback, validate availability, and loop before completing the task. This task requires 10x more tokens than a standard chat session with an LLM. Now let us consider an email monitor agent that would interpret intent, retrieve context, draft responses, manage follow-ups, coordinate meetings and summarise unresolved actions. This type of embedded copilot would require 100x more tokens daily, highlighting that the biggest driver of token demand is not just task complexity, but how continuously the agent remains active.<\/p>\n\n\n\n

Although consumer agents are likely to have a larger user base today, enterprise agents are expected to be materially more token-intensive per user and to become the primary source of incremental token demand by 2030. Critically, this reflects the fact that enterprise workflows require agents to perform considerably more complex and precise actions, such as monitoring tasks, retrieving context, reasoning through exceptions, validating outputs, updating systems, and escalating issues throughout the workday. Enterprise agents may also need to process heavier multimodal inputs \u2014 including voice, images, documents, screen activity, application data, logs, and structured system records \u2014 in order to replicate the complex and varied workflows of real knowledge workers. This can materially increase token intensity relative to text-only consumer prompts. In short, the core difference is rigor: while a consumer agent can often stop at a \u201cgood enough\u201d answer, an enterprise agent may require multiple iterations to produce work that is accurate, auditable, and usable in real business processes.
<\/p>\n\n\n\n

AI TOKENOMICS<\/h3>\n\n\n\n

Beyond the computational metric, the notion of token has also become the basic unit of the AI economy. AI firms offer free and flat-fee monthly plans, but these have low token caps. Users with demand greater than these limits pay for LLM usage based on the number of tokens used. They pay for both input and output tokens, with the latter generally costing two to six times more than the former, depending on the model.<\/p>\n\n\n\n

In this context, enterprise adoption of AI is not only dependent on productivity, as many would suggest, but also on economic viability. As a matter of fact, not all enterprise agents are currently cost effective. For example, a call centre agent relying on real-time voice-based automation can be materially more expensive and operationally complex than outsourced human labour. In contrast, a coding agent can operate in a high-value, largely text-based workflow, supported by a comparatively mature tooling ecosystem. This helps explain why software development has seen faster agent adoption so far. This is important because token volumes alone do not determine the economics of agentic AI. What matters is whether the cost of completing a workflow falls enough to make the use case attractive for the customer. As this cost declines, more enterprise use cases can move into positive return on investment, expanding the addressable market for AI agents.<\/p>\n\n\n\n

For hyperscalers, the equation is different. Their profitability depends on the spread between what customers pay for AI workloads and the internal cost of serving those workloads. According to Goldman Sach\u2019s framework, leading semiconductor companies are driving a 60\u201370% annualised reduction in cost per token for hyperscalers, while customer-facing token pricing has started to stabilise after several years of sharp declines. This can be partly explained by growing demand for model inference, driven by agentic AI, amid compute capacity constraints. This creates the possibility of an inflection point. Enterprises get more economically viable AI workflows, while hyperscalers convert rising token volumes into better margins through lower internal costs and better infrastructure utilisation. This makes the current capex cycle more defensible. The key point is that AI capex should not be analysed only as a cost burden. If token economics keep improving, hyperscalers will be able to unlock more value for shareholders.<\/p>\n\n\n\n

However, there are still risks along the way as not all agents will become economical at the same time. In addition, cheaper Chinese models could pressure pricing power of leading LLM providers. This risk can be illustrated in monetary terms. For example, reviewing a 50-page legal contract could cost around $0.11 using Claude Sonnet and less than $0.02 using DeepSeek R1, based on roughly 25,000 input tokens and 2,000 output tokens. At scale, when such tasks take place hundreds or thousands of times per month, the price gap becomes meaningful. This does not necessarily weaken the investment case for AI infrastructure. In practice, agentic AI is likely to evolve toward a tiered model architecture. Lower-cost models could handle simple and repetitive tasks, while the most advanced models would remain more relevant for tasks requiring complex reasoning, high reliability and multi-step execution. In this context, lower-cost models might reduce pricing in some segments, but they could also accelerate broader adoption of the technology by lowering the cost of many use cases. This could even strengthen the case for higher token volumes and support demand for cloud computing, still benefiting hyperscalers.<\/p>\n","protected":false},"excerpt":{"rendered":"

Bubbles & Supercycles<\/p>\n","protected":false},"author":5,"featured_media":1442958,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"ngg_post_thumbnail":0,"footnotes":""},"categories":[220],"tags":[],"class_list":["post-1446604","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-opinion-leaders-en"],"acf":[],"_links":{"self":[{"href":"https:\/\/www.payoff.ch\/en\/wp-json\/wp\/v2\/posts\/1446604","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.payoff.ch\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.payoff.ch\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.payoff.ch\/en\/wp-json\/wp\/v2\/users\/5"}],"replies":[{"embeddable":true,"href":"https:\/\/www.payoff.ch\/en\/wp-json\/wp\/v2\/comments?post=1446604"}],"version-history":[{"count":1,"href":"https:\/\/www.payoff.ch\/en\/wp-json\/wp\/v2\/posts\/1446604\/revisions"}],"predecessor-version":[{"id":1446607,"href":"https:\/\/www.payoff.ch\/en\/wp-json\/wp\/v2\/posts\/1446604\/revisions\/1446607"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.payoff.ch\/en\/wp-json\/wp\/v2\/media\/1442958"}],"wp:attachment":[{"href":"https:\/\/www.payoff.ch\/en\/wp-json\/wp\/v2\/media?parent=1446604"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.payoff.ch\/en\/wp-json\/wp\/v2\/categories?post=1446604"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.payoff.ch\/en\/wp-json\/wp\/v2\/tags?post=1446604"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}