Last
week, all the AI hyperscalers reported their earnings on the same day and each
presented the same underlying tone – one of scarcity. Sundar Pichai, Alphabet’s
CEO, said, “We are compute constrained in the near term,” and Amazon
noted that capacity is being monetised as fast as it can be installed.
Microsoft expects to remain constrained through the remainder of 2026, while
Meta raised its spending forecast and cited the cost of components it cannot
get fast enough.
Together,
Alphabet, Amazon, Meta, and Microsoft have committed to spending more than US$700
billion on AI infrastructure in 2026. That is almost double the amount spent in
2025, and approximately US$100 billion more than their own estimates in the
prior quarter. Amazon has committed US$200 billion and Microsoft US$190 billion,
up 61% over the prior year. Alphabet is expecting to land between US$180 billion
and US$190 billion. Meta lifted its range to between US$125 billion and US$145
billion, up US$10 billion on its prior guidance, citing higher component costs.
The
signal across the hyperscalers is staggering and consistent: that the demand of
tokens outstrips supply and that physical scarcity is creating a barrier at a
time of accelerating adoption.
A
chain of constraints
Since
ChatGPT arrived in late 2022, the AI infrastructure build-out has been a
sequential story of constraints. Each bottleneck, once partially resolved, has
revealed the next one waiting behind it. For investors who read that sequence,
it has been consistently rewarding.
It
started with GPUs and LLM training, with Nvidia’s H100 on allocation and
waitlists growing as the tempo of development accelerated. The speed of change,
the scale of the buildout, and the fixation of a bubble hid emerging
constraints such as advanced packaging, high-bandwidth memory, data centre
construction, land, permitting, labour, and building materials – and not
forgetting power, grid interconnections, transformers, and high-voltage
switchgear, as well as optical interconnects.
All
benefitted from orders and capital flow. Even the CPU, long dismissed as an old
technology with limited opportunity, re-emerged as a genuine beneficiary as
inference exploded and workloads proved far more processor-intensive than most
originally expected.
The
question now is whether that sequence has run its course or whether this is
structural and set to endure.
The
gap between intent and execution
With
an intent to spend US$700 billion of committed capital, the question we are now
faced is whether it can be deployed, given only five gigawatts (GW) of the
previously announced 12GW to 16GW of US data centre capacity announced for
completion in 2026 is under construction.
This
trend has accelerated, and whether future projects will be delayed or outright
cancelled remains to be seen but is of concern. The US wants to be at the
forefront of the AI revolution, but delays in compute deployment will hamper
ambition and perhaps cede ground to others who have less inertia.
Last
month, Maine came close to becoming the first state in America to say no to the
ongoing and aggressive buildout of data centres. Its legislature passed a bill
that would have frozen approvals for new large data centres until October 2027
while the state worked out their impact on power bills, water consumption, the
environment, and the electrical grid. However, in a surprise move, Governor
Janet Mills vetoed it, but only on a technicality – that it badly needs the
jobs created.
The
reasons are not exotic: transformers and switchgear carry lead times
approaching three years, and the impact of tariffs has compounded the supply
constraints, especially in those areas predominantly reliant on China, which
supplies over 40% of battery imports and approximately 30% of transformers and
switchgear capacity. Permitting processes that were slow before the AI boom
have not accelerated to match, with significant and well-funded opposition
groups now active across many US states. Several states are considering
construction moratorium legislation.
The
companies driving the revolution are not short of capital or the conviction to
spend, but with the hyperscalers committing the bulk of what will amount to
over US$1 trillion in capex this year, the deployment bottleneck may be
difficult to overcome. With that, investor concern surrounding lofty
expectations may grow.
Digital
scarcity
With
demand exploding from both the consumer and now the enterprise, the physical
shortage is creating a second form of scarcity that is now layering on top of
it. As recently as mid-2025, AI model providers built their businesses on the
assumption that, at least for a while, usage would coalesce in conversational
queries representing a few hundred tokens at a time, with a human at the
keyboard. However, one technological fork in the road changed all of that, and
the cornerstone assumption has been overtaken by events.
Agentic
AI, where models write code, browse the web, and execute multi-step tasks
autonomously, consumes thousands of tokens per session and runs continuously
rather than sporadically. Visa consumed two trillion tokens in a single month, after
consuming one trillion the month before, and Uber exhausted its annual AI
budget in three months.
This
change has happened so fast that the pricing model of the AI industry as a
whole has broken. The flat-rate price that balanced consumption across all
users was designed at a time when the AI use case was under close scrutiny. It
was not designed for a world of agents running 24/7.
Think
of it as a party: you invite 50 people but cater for 30, expecting 40% to drop
out, but the drawcard is so great that they all arrive. Worse, some bring a
plus-one! For AI, they are faced with two challenges, the first being the
deployment of capital and the second, at least in the short-term, being how to
create controlled demand destruction to bring utilisation back to something
like balance.
Anthropic
has raised prices to enterprise customers and shifted away from flat-rate
billing toward consumption-based contracts, with model performance reported to
have been throttled for lower-tier users and rationed toward higher-end, paid
workloads. OpenAI has temporarily shut down its Sora image generation tool. The
cost to serve a frontier model at current usage levels remains substantially
higher than the market prices used to acquire customers, and the subsidy is
under growing pressure as usage accelerates and the economics worsen.
The
demand destruction dynamic is visible through Cloudflare, a close partner of
several major model providers, which disclosed that it had pivoted its internal
AI development tools from a leading proprietary model to an open-source
alternative, cutting costs by 77% on a single high-volume workflow.
By way
of a description, at its recent ‘Agents Week’ event 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 1 billion knowledge workers. We’re not a little short
on compute. We’re orders of magnitude away.”
What
it means
For
the hyperscalers with owned compute, scarcity is largely good news, and the
market has been surprised by just how good. Cloud operating margins are
expanding rather than contracting, the opposite of what had been expected as
depreciation from massive capital spending hit the income statement. AWS
operating margins reached 37.7% and are moving higher, while Google Cloud’s hit
32.9%. Pricing power is real for those that have the capacity to allocate
compute, especially where their own silicon provides a significant cost
advantage.
For
those without their own compute and dependent on rented infrastructure, the
picture is considerably harder. They are now at the pricing whim of their
compute providers, and this creates a ceiling on growth, forces difficult
choices about which customers and workloads to prioritise, and exposes pricing
commitments – written at an earlier time when compute was plentiful – to
painful renegotiation.
For
investors, the central question is whether this cycle follows the pattern of
previous technology infrastructure buildouts, where shortages eventually
resolve into oversupply and compress margins, or whether the demand for AI
compute is genuinely durable enough to keep the cycle tight for longer than
prior analogues suggest. For example, in prior cycles, memory companies would
have aggressively expanded capacity, leading to oversupply, but there is little
evidence of ill-discipline at the moment.
The
constraints to do so are everywhere, and the companies supplying into the
buildout – the chip makers, the power infrastructure names, the optical
networking players and the like – are generating cash at levels rarely seen in
their histories. The hyperscalers spending it, however, are the providers of
this cash waterfall and are watching their own free cash flow compress under
the weight of commitments that show no sign of slowing.
The
evidence from last week’s earnings, at least for now, sits firmly on the side
of duration. Cloud growth rates are accelerating, backlogs are expanding, and
the companies building this infrastructure are not blinking.
What
is less certain is whether the physical world can keep pace with the ambition.
This is not a repeat of the dotcom bubble where dark fibre had no market; in
this revolution, demand is insatiable. If anything, the constraints are more
basic, not just chips and other hardware but at the speed at which concrete
sets.
There
is no particular reason to think we are nearing the end game, at least not yet,
but investor expectations are high and any demand weakness will manifest itself
in share prices, and quickly.
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.
