Like many companies, Samsung Electronics reported record Q2 profits on accelerating revenue growth in AI-related sales, yet it did not provide sufficient reassurance that the AI hardware trades were intact, with July proving to be a difficult month for the technology sector as a whole. Semiconductors led the decline, and the South Korean equity index, the Kospi, fell almost a third from its June peak through late July.
Whilst the topology of the equity sell-off is complex, the leverage sitting on top of the AI hardware trade showed that leverage amplifies moves both ways and unwinding is unpredictable and ignorant to strong fundamentals. In late July, the revelation that Situational Awareness, the hedge fund built around Leopold Aschenbrenner’s scaling thesis, had been forced to sell the bulk of its listed equity positions to Ken Griffin’s Citadel acted as a circuit-breaker to the indiscriminate sell-off.
Aschenbrenner’s undoing, leverage to a short book in software that was framed as a hedge against the long positions, was not the diversification it was thought to be. Instead it was the same trade wearing different hats, with the funds’ demise certain the moment the correlation between them inverted.
Whilst we have cautioned that investors should dance closer to the door for a while, July’s move was not really a verdict on AI as an investment case but a lesson in crowded leverage trades. Like July’s sell-off, once Citadel’s circuit-breaker became known, and investors were able to focus on reported numbers, the rebound in the names leading the sell-off was dramatic.
Understanding where we are now is critically important, and refocusing could pay dividends, but it is important to understand that things are not the same now as they were up until July, and it will not be the same in December as it is now. Among other things, we are focusing on who is financing the physical build and on what terms, the emerging pattern of demand and what it is doing to the assumptions supporting that financing, and who may end up owning the trusted relationship with the enterprises paying for it.
Financing
One of the controversial elements of the hardware trade surfaced as a concern last year as companies adjusted their depreciation schedules from three years to six. Highly criticised as creative accounting, the counter and strongest piece of evidence to emerge concerns the recent earnings release by CoreWeave and the narrative that it has recontracted Nvidia A100 chips (now five years old) through 2029 and raised prices by 25% across the company’s entire stack, and the news it activated 300 megawatts (MW) of new power capacity with a further 1.5 gigawatts under option. CoreWeave’s results provide evidence that demand for older technology remains firm, which is important when a lender considers the semiconductors as collateral.
Nvidia has been at the centre of the circular nature of financing, flexing its balance sheet and cashflow to backstop the purchase of its own chips. Notwithstanding it is singularly self-serving, it is not difficult to understand why this has caused investor anxiety. However, the recent announcement that Blackstone, Brookfield, KKR, Apollo, Goldman Sachs, and BlackRock have agreed to create a US$500 billion investment pool alongside Nvidia goes some way to validating Nvidia’s prior actions, and that in their opinion, the outlook for revenue generated by token use is real.
Whether or not this has some beneficial knock-on effect on the interest rates charged on debt issuance remains to be seen; but taken with a bullish lens it could lead to a proliferation of model providers and broadening tenancy, not just demand from the frontier labs, which would be a healthy development.
With that said, CoreWeave’s own recontracting is only one data point, and the company is carrying a significant amount of debt, so it is not without risk. A lender’s actual exposure, secured on the hardware, is really attached to whichever operator sits between the hardware and the financing structure, but the simplest interpretation is that the lenders believe there is significant residual value in the hardware on which their financing is secured
Demand
The hyperscalers have been providing us with dizzying statistics on the growth in token consumption for a while now, yet the concerns remain loud regarding what is actually generating the demand this infrastructure is being built to serve.
However, the growth in annual recurring revenue (ARR) reported by Anthropic and OpenAI suggests paid consumption is accelerating, with Anthropic looking to end the year at 10x the level of revenue compared to the end of 2025, implying an annualised run rate of approximately US$100 billion, with some estimates running well ahead of that and suggesting the actual figure could finish higher still. OpenAI, having been a relative laggard, is seeing growth in ARR accelerate and this, alongside the numbers from the hyperscalers and neoclouds, suggests its contracted compute will be paid for.
This could be set to accelerate further, and it has been reported that growth in internet traffic is becoming less human and that future usage is increasingly a function of agents. Cloudflare’s chief executive, Matthew Prince, indicated that more than half of all traffic on the company’s network was not generated by people.
Whilst that is noteworthy, the compressed timeframe in which it crossed over demonstrates the pace of change is phenomenal. In November last year, Cloudflare had expected human traffic to remain the majority of internet activity until the second half of 2027, then more recently revised that to the first half of 2027. That has proven to be incorrect by a staggering 15 months, with human traffic the minority by May this year. Cloudflare estimates non-human traffic to be thousands of times greater than that traffic of humans within five years, and not because human use is forecast to fall.
The commercial side of this shift is no longer confined to network economics; it is showing up in retail sales. For five quarters, Shopify has delivered more than 30% of gross merchandise volume (GMV) growth, and its GMV has grown from roughly 10% of Amazon’s estimated retail GMV in 2017 to approximately 44% today. The measure may evidence which platforms may benefit from the shift in ecommerce toward agents and not people.
Further evidence from Wayfair, which reported its best quarter in years, and Etsy, built for discovery and recommendation across small retailers, also looks well placed in the agent world. The results themselves do not prove agents are scale buyers but do show real use of the infrastructure.
Cloudflare’s response has been to build the plumbing for a different kind of internet economy: a monetisation gateway that lets a website or service charge for anything an agent might want, a page, an application programming interface, a dataset; and a wallet that lets an agent pay autonomously on a customer’s behalf as well as an identity and trust layer so that buyer and seller agents can recognise one another. The ambition, in the company’s own words, is to make every agent request a metered event, extending the economics of the web from simple content access into recommendations, payments, and fulfilment.
This shift matters because an advertising-funded internet has no obvious way to charge an agent that never sees or cares about adverts or follows paid links. It remains to be seen what this means to the advertising models of companies, or what subscriptions are needed to compensate a company for the provision of a service that has until now been free.
One response is to create that monetisation gateway. The wallet plus the identity and trust layer would be critically important for content creators, and perhaps the way forward will see these wallet exchanges between agents find their way back to the creators themselves and put an end to the scorched earth that currently exists.
The platform
Trust is the cornerstone of commerce. Of course, we have lawyers and contracts to exert and police agreements, but even before a relationship moves to contract, there needs to be trust between the parties. In AI use, it is unclear where the trust layer sits, with most users having their own degree of trust and what they are willing to share. However, data may require sovereignty, where the model service provider is precluded from learning from or using users’ data to stand up their own competitive offering.
Last year, Microsoft chair and CEO Satya Nadella gave up half of his executive duties to return to the engineering frontline to respond to a bruising encounter with OpenAI which left Microsoft without any fathomable AI strategy. However, his return came with what sounds like a coherent strategy rather than a simple placeholder that keeps investors on tenterhooks: distance the company from the frontier model labs, encourage enterprises to build and own their own models, and get those enterprises standardised on a single harness for running whichever model they choose, preferably Microsoft’s.
The harness is a critical component of any AI offering and sits around whichever model an enterprise runs. If the LLM is the brain, then the harness is the nervous system, the rigging and dashboard that controls where a company’s data goes, keeping it out of competitors’ hands and out of a frontier lab’s own training sets, and continuously testing and scoring the model underneath it against alternatives.
Built correctly, the harness will be explicitly model-agnostic, which will allow the users to swap the underlying model at will and without losing the guardrails and the data controls. You can certainly use Microsoft’s own models, and no doubt will be encouraged to do so. This makes perfect sense for Microsoft because regardless of which model leads, if it can own the harness, it is easy for Microsoft to extend its existing enterprise software into a new and durable point of control. Companies already trust Microsoft.
None of this happens in isolation, and Meta has announced its own version as well as recommitting to open-weight models, the latest of which represented a significant leap forward. Google is widely expected to follow, with its applications already enjoying mass adoption by consumers.
Whether Microsoft’s strategy will win out is uncertain, but it now has a plan, and that plan is more credible than its prior position, so on that basis it is already a success. It is not unique to Microsoft, however, given both Amazon Web Services and Google Cloud are able to make credible arguments about compliance and enterprise trust of their own. Perhaps Meta or Microsoft are wearing the emperor’s clothes, but we will get more information soon, with both hosting significant meetings where it is anticipated they will put more meat on the bones of detail.
One thing is for certain: at this point, Nadella’s legacy rests almost entirely on whether the harness actually protects and extends the Office and Microsoft 365 franchise. Little else about Nadella’s tenure so far looks likely to matter as much as this, for in a world of software writing agents, the threat of having your homework marked elsewhere is forever real.
Microsoft’s bid to own the harness for enterprise AI deployment is, on its own telling, a credible plan rather than a demonstrated one, with the evidence for whether it succeeds still to come.
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.
