The story of a week
Last week the market movements were strongly bookended and characterised by a melt-up on Monday followed by a sharp fall on Friday. Regardless of the precise trigger on Friday, it is worth highlighting a couple of important developments given they tell a more important story than either day in isolation.
The move on Monday was significant by any measure, a surge driven by the fear of missing out or continuing to do so in the areas where they have obviously been there for the taking, and that failure to participate would result in significant underperformance. The psychology is powerful enough on its own, but it crossed into emotion and triggered investment strategies that reinforced the already powerful move. Short-dated call options on semiconductor and hardware names were bought aggressively, forcing market participants to buy the underlying as a hedge, driving prices higher, which led to more call buying and resulting in a self-reinforcing loop.
In May, the semi-conductor index recorded its strongest monthly performance and companies such as Hewlett Packard Enterprise doubled. These are not the characteristics of a durable re-rating.
Between Monday and Friday there were a series of signposts leading to market cracks appearing. On Wednesday, Broadcom reported strong earnings with AI semiconductor sales of US$10.8 billion, up 143% year-over-year, but Q3 guidance of AI sales at US$16 billion fell short of expectations of US$17.2 billion.
The broader reframe that is building is one of demand and supply. Until very recently, every new capex announcement had previously been seen as good news; however, the insatiable demand is now becoming a problem. In part, revenue beats are being driven by component price inflation rather than new volumes or market share gains. Mark Zuckerberg has said that Meta’s prior quarter capex hike was made essentially to pay for the increase in prices, and not to get more compute.
In addition, it was becoming clear that the self-funding thesis of the hyperscalers is being put to the test, especially in light of Google’s announcement. Google generates over US$170 billion of cashflow per year, yet it announced a dilutive US$80 billion equity raise (later upsized to almost US$85 billion), of which US$30 billion is to be used to pay the tax on stock compensation awards. This puts a strain on investor cash levels at a time when they are already relatively low, and it is reasonable to assume this is not yet past, with additional companies expected to raise money. SpaceX, Anthropic, and OpenAI are all looking to IPO at significant market capitalisations.
The bear case is not yet properly formed, but the tone has changed. In markets that have run as hard and as fast as they have, tone can lead price.
Escaping gravity
The most revealing document published in the AI era so far is not what many may think. It is not a white paper or company-released technological breakthrough, but the prospectus filed for Space Exploration Technologies Corp., or SpaceX. Inside are the details of a business spanning distinct yet linked activities, which reveals more about the state of AI economics than anything the model providers have publicly chosen to share. The filing is revealing not because of what it says about rockets or satellites or the ambition to colonise Mars, but because it illuminates the economics of AI today and how they may change, perhaps dramatically, over the years to come.
Three businesses, one engine
SpaceX is unique; it comprises rockets and satellites folded together with an AI laboratory and the social media platform X. It is a company spanning launch capability, global broadband connectivity, frontier AI model development, and the stated ambition to deploy data centres powered by solar energy in space.
Having evolved from being a heavily subsidised infrastructure bet to a genuinely profitable and rapidly scaling global service, Starlink plays a crucial role in the financing of future developments. With approximately 10.3 million subscribers across 164 countries, Starlink grew approximately 50% year over year, reporting revenue of US$11.4 billion with operating income of US$4.4 billion and segment EBITDA of US$7.2 billion. This is in stark contrast to the loss-making Space business, the Starship development programme, and the xAI operations, although the dynamics surrounding the AI business are changing rapidly.
Anthropic has agreed to lease the entirety of the xAI Colossus 1 data centre for US$1.25 billion per month, whilst in a just announced deal, Google will pay SpaceX US$920 million per month for compute capacity at xAI data centres from later this year. Combined, the two contracts will provide approximately US$2.17 billion per month in compute revenue or approximately US$26 billion per year.
This is a straight win for SpaceX as it moves its own model, Grok, to the new Colossus 2 cluster. More external contracts will be signed given the company is building data centres at a pace of which other companies can only dream. The first cluster in Colossus 1 took 122 days, whilst the first in Colossus 2 took 91 days, and only 64 days for the second cluster.
For context, the industry benchmark for bringing a 100-megawatt greenfield data centre online is estimated to be two to two-and-a-half years. In addition to protracted planning consent and well-funded and organised local opposition, power is consistently seen as the gating factor in the pace of new data centre buildout, making data centres in space seem all the more attractive.
Investors like economic flywheels, and SpaceX is assembling its own unique and impossible-to-replicate version. The company is bringing more capable rockets with an increasing launch cadence, with the aim of reducing the orbit deployment cost per satellite. These satellites will be increasingly more capable, which will enable Starlink’s services to scale and, over time, lower the cost to serve.
In turn, it is expected that the company will continue to see significant revenue and cashflow growth. When reinvested into the business, coupled with the retirement of older satellites, this will reduce costs and grow margins further. With the new satellites and consequently improving service, there is an expectation that customer growth will accelerate. In addition, the compute infrastructure is only now beginning to generate revenue through the Anthropic and Google arrangements with more third-party business expected, adding more momentum to the wheel.
Subscribers
Starlink has approximately 10.3 million subscribers across 164 countries, but they represent the earliest adopters in markets where terrestrial alternatives exist or partially exist. The addressable population served by those 164 countries is estimated to exceed three billion people, but the company’s clear opportunity is to connect those in geographies where the economics of terrestrial services simply do not work. Satellites do not require planning permission and instead only require a receiver, making it quick and efficient to deploy. Maritime, aviation, remote enterprise, and rural communities across Africa, South Asia, and Latin America are all target markets where there are few alternatives.
The economics of the Starlink business should accelerate as Starship starts to deploy the next generation of satellites, which are anticipated to provide 10x the capacity of those currently in use. With Starship able to deploy up to 60 satellites in a single launch, Starship should add more capacity to the network than twenty Falcon 9 missions and at a higher cadence, resulting in significant declines in launch costs.
In November 2027 Starlink is expected to complete its purchase of radio spectrum from EchoStar, which will provide the ability to offer direct-to-device 5G connectivity from space, bypassing mobile network operators and supplementing or replacing terrestrial mobiles in the same way Starlink has already supplemented or replaced fixed broadband in underserved markets.
The announced compute deals and the subscriber trajectory converge on the single biggest bottleneck, power. According to SpaceX, space-based solar in sun-synchronous orbit will generate more than five times the energy per unit area of terrestrial solar, whilst radiative cooling in the vacuum of space eliminates the largest single operational cost of a terrestrial data centre – cooling.
Whether SpaceX proves to be the company that captures the orbital compute opportunity at scale is genuinely uncertain, but with current launch costs of US$1,500 to US$3,000 per kilogram set to decline meaningfully once Starship reaches commercial cadence, the opportunity will be present, and especially so should the cost of launch decline to approximately US$300 per kilogram – the point at which the economics for space are superior to those on Earth.
The token question
The terrestrial power constraint is structural and worsening, but the reason it is worsening is demand and the trajectory of that demand. Google’s recently disclosed token consumption by users has risen from 9.7 trillion per month in 2024 to 480 trillion a month (as disclosed at last years I/O conference) to 3.2 quadrillion so far this year. AI adoption is here, at least at the consumer level if not the corporate level. For the latter to occur, there needs to be trust. Trust with the data, trust with the accuracy, and trust that your agents and what they are doing is manageable.
The economics underneath that demand curve are only now becoming visible, and what they show is not reassuring. Microsoft’s move to consumption-based billing on GitHub has users paying US$39 per month as a flat subscription, with their actual token consumption, priced at real inference rates, running to multiples of that figure – and for some well into the US$1,000s.
Companies are addressing this subsidy through significant price increases, service throttling, and consumption-based contracts. This can be seen in the staggering numbers that Anthropic showed; revenue has gone from an annual run rate of US$9 billion at the end of 2025 to over US$30 billion in April 2026. This portrays extraordinary demand for the model and its tools, requiring them to simultaneously sign compute arrangements with Amazon, Google, Broadcom, Microsoft, Nvidia, Fluidstack, and SpaceX.
The two most important AI laboratories in the world are paying a rocket company US$2.17 billion a month for spare data centre capacity because they cannot build their own fast enough. Whilst Colossus is terrestrial based now, it is anticipated that the industry will increasingly look upward.
Whilst the direction of travel, driven by improving economics, suggest datacentres in space are inevitable, the timelines appear aggressive and we have seen delays elsewhere. Whether the milestones can be met and the anticipated demand is real remains to be seen, but if you like a moonshot, or more accurately a Mars shot, then has Elon Musk got the company and story for you!
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.
