The hardware trade has been investors’ preferred expression of AI deployment but has been under pressure since June, despite reported and expected earnings having never looked stronger. Orders continue to outpace supply, with scarcity leading to higher prices, but technology hardware companies are not leading the indices as they once were. Significant as Nvidia’s guidance was, it remained affected by the supply dynamic, and investors fret over missed opportunities in the shifting sands underpinning the story, and over the ultimate beneficiaries of technology being those downstream.
Two revolutions
There are two AI revolutions under way, with one widely understood and the second emerging in the narrative and playing to the downstream thesis. The first revolution is a story about scarcity; the second is the story of adoption.
The first is already heavily invested in and represented by the revolution in infrastructure and model capability. It is fully covered, widely understood, and extensively reported on. While it has been the source of significant profits that need to find another home, it does not appear to be done, which will become more obvious below. However, the straight investment thematic of ‘buy hardware’ has become more complicated given concern over the sustainability of capex alongside the negative political and public backlash.
The famous Wayne Gretzky quote (which he attributed to his father, Walter), “Skate to where the puck is going to be, not where it has been,” embodies the transition of the two revolutions. The second revolution is far less prominent and is only just beginning; it concerns the deployment of AI technology and capability into the daily operations of business and everyday life. This is the decade(s)-long adoption that will change productivity, the economics of growth and, it is hoped, create new industries and solve problems that we do not even know about yet. It sounds romantic, but it appears to be shaping up this way.
What deployment actually requires
Deployment will be slow at first, but it will accelerate as understanding increases. Initially, successful deployment will be the auspices of a few, but over time the democratisation of knowledge will accelerate opening the door wide open, with economic value coalescing to the adopters. It will show up in results over time and then suddenly all at once, but it is not yet simple, and there are practical considerations and challenges that need to be addressed and overcome.
For successful AI deployment, a company needs its data to be ordered well enough to be usable, and it needs its applications to have somewhere fast, reliable, and safe to run and a place for an agents’ output to be secured and stored. All of this needs to happen without costs spiralling. Security is the greatest concern, and any Chief Technology Officer will want to ensure that the company can survive with an attack surface that has widened in the face of insomniac autonomous agents, not humans, doing the majority of the probing.
Data cleansing and ordering is a lengthy process, but as one leading company suggested recently, AI use is collapsing the time it takes to deliver petabytes of sanitised data from years into months. Projects that were previously seen as data-constrained can now be undertaken, unlocking aged systems and processes and resulting in an acceleration that will benefit the companies that dare to do and challenge those that do not. Now your data is in good shape, you need to store it somewhere that is fast to access and in in any format in which it is conceived and not converted to a single state. After all, converting slows things down, and slowness is not at a premium when it is to be done at scale.
But as well as data, how an agent goes about its business is critically important. It must be safe to do its work and it needs access to a network built for lightweight, instantly provisioned compute that allows the agent to spin up, complete its task, and disappear as quickly as it appeared. Internet and network traffic from agents is rapidly becoming the majority, and it is only in the early phase.
Companies choosing to bring models and data back onto their own premises, for reasons of security and sovereignty, will need security between their servers and the client request, a password login or a payment, and to funnel the accelerating traffic through load balancing and delivery controls, no matter where the compute sits. As digital creation accelerates, storage will continue to be challenging, with hard disk drives and solid-state solutions sharing the load but in short supply. However, human ingenuity will address these as bottlenecks and high memory prices push development out of necessity, and AI delivery will not be an exception. For example, storage that scales in capacity without scaling proportionally in cost, space, or power consumption allows data volumes to grow the way AI now demands.
Capex and revenue
Much of the market’s anxiety about the scale of AI infrastructure spending has rested on whether the revenue required to justify it will actually turn up in time to service the debt that has been raised to fund it. Despite these concerns, none of the hyperscalers have been willing to be frugal in their budgets given personal agents will be working 24/7, and a growing number of enterprises will be running applications continuously and at scale that has not been seen before. While the revenue is coming and there are clear signs today, whether or not it happens quickly enough to calm the market’s unease remains to be seen; however, the direction of travel to genuine operational use is good to see.
Deployment
Nowhere has that direction of travel been more clearly demonstrated than by Meta, whose leadership has been accused of spending like a drunken sailor and with no clear and articulated strategy that would see Meta convert that spending into revenue beyond the advertising business model it already has. While all of the hyperscalers have enormous and growing capex commitments, until recently Meta had been singled out because it did not have revenue derived from other sources, such as cloud. In 2023, Meta’s capex was approximately US$28 billion, rising in the subsequent years to US$39 billion, then US$72 billion, and in 2026 is estimated at between US$125 billion and $145 billion.
Like or loathe him, Meta’s founder, chair, and CEO Mark Zuckerberg has mettle, and sometimes it takes a founder who is willing to break eggs to challenge the status quo and skate to where the puck is going. Zuckerberg, it seems, knew chatbots were not the end goal, but getting to where the puck was actually going would take powerful compute, and a lot of it.
Recently, Meta announced its updated LLM, Muse Spark 1.3, was using 25% fewer tokens per task alongside stronger coding and agentic reasoning than its previous version, and investors took note. Having lost ground when compared to its peers, the improvement in the model is welcome, but the announcement now appears to be a sideshow.
On the eve of Apple’s own keynote, Surprise and Shine, Meta announced its task-oriented personal assistant, Muse. This is a strategic moment for the company and one that perhaps only Zuckerberg would be bold enough to take. Muse is Meta’s accessible, consumer-facing version of the same agentic capability OpenClaw had already proven possible late last year, and it will be distributed through the largest user base in the industry.
In a staggered approach, almost 3.8 billion monthly active users will have access to a free tier with 100 million tokens and two paid tiers starting at US$20. Launched in the US, it will be closely watched, and it will no doubt embolden others to make the move, and fast. Whether or not the company wins the race will not be answered soon, but first mover advantage in what could prove a sticky relationship provides Meta with an opportunity that many have courted.
However, the industry itself may temper success, with fissures caused by a wider
disagreement (now playing out in public) over the speed of AI development and the possibility it presents an existential threat. Zuckerberg, as Meta’s chief executive, has argued publicly that distributing capable AI models as broadly as possible is itself the safest course, on the basis that concentrating the technology inside a small number of institutions presents a greater risk. Anthropic’s chief executive, Dario Amodei, has argued the opposite and that the pace of improvement now needs to be deliberately slowed. Indeed, over the weekend, OpenAI’s Sam Altman and SpaceX’s Elon Musk have jointly called for a more cautious approach, saying they supported Amodei’s proposals, which include model evaluation.
Politically, this stance resonates with the electorate’s overtly negative view of the technology, and previously pro US governors, including Pennsylvania’s Josh Shapiro, have switched sides and now favour regulation. Even datacentre-friendly Texas has put a hold on approvals of new construction, with its governor, Greg Abbott, tamping down his enthusiasm for the industry. However, President Trump is resisting the calls for regulation and change given his view on the importance of the technology and his unwavering belief that America’s leadership is critical.
While the disagreement has been bubbling away for some time, it has become harder to ignore now that political rhetoric around AI’s negative effect on jobs has become the consensus view and at a time when Anthropic is moving towards a public listing. Will political sentiment change after the US mid-term elections, given the communities that agree enjoy services funded by the technology company? Quite possibly.
Whatever the outcome, the success of Muse will depend on whether the technology feels easy enough for an ordinary person to set up and whether they trust it with access to their private information and their bank account details. What will come next and from whom is another key question, but it is curious that Apple has remained conspicuously quiet despite having a device in almost every pocket and an ecosystem already built around trust. Perhaps Apple can still convert that advantage into a functional personal assistant of its own, but so far Siri still seems to struggle.
Compute ceiling
Politics aside, there remains a fundamental hurdle to the mass market provision of agents or personal assistants, let alone the need to arm knowledge workers with the tools to raise productivity to the levels the technology promises. The enduring constraint that deployment will need to overcome is compute and all that makes it tick.
In April, Cloudflare said, “If the more than 100 million knowledge workers in the US each used an agentic assistant at ~15% concurrency, you’d need capacity for approximately 24 million simultaneous sessions. At 25–50 users per CPU, that’s somewhere between 500K and 1M server CPUs — just for the US, with one agent per person. Now picture each person running several agents in parallel. Now picture the rest of the world with more than one billion knowledge workers. We’re not a little short on compute. We’re orders of magnitude away.”
For anyone owning technology companies over the last two months, it has been a rollercoaster of a ride, but the genie is out of the bottle and the demands for compute, storage, power and all things that provision it continues to be colossal. Through necessity, human ingenuity will weave a path, and 2027 will not look like 2026. But between now and next year, the political landscape and the potential for regulation may see a seismic shift and cannot be ignored, despite Meta’s Muse. The wind is up and the waves of change will take the boat in a certain direction, but first it has to navigate the whirlpool in the middle.
Tim Chesterfield is CIO of the Perpetual Guardian Group and the founding CIO and Director of its investment management business, PG Investments. With $2.8 billion in funds under management and $8 billion in total assets under management, Perpetual Guardian Group is a leading financial services provider to New Zealanders.
Disclaimer
Information provided in this publication is not personalised and does not take into account the particular financial situation, needs or goals of any person. Professional investment advice should be taken before making an investment. The information provided in this article is not a recommendation to buy, sell, or hold any of the companies mentioned. PG Investments is not responsible for, and expressly disclaims all liability for, damages of any kind arising out of use, reference to, or reliance on any information contained within this article, and no guarantee is given that the information provided in this article is correct, complete, and up to date.
