What Meta's "Excess" Really Tells Us

On 1 July, Bloomberg reported that Meta is building a new internal unit, Meta Compute, to sell surplus AI infrastructure to outside customers. Within a day, Meta's shares had risen almost nine per cent, adding roughly $125 billion in market capitalisation, while CoreWeave fell fourteen per cent and Nebius seventeen per cent on fears of a new competitor entering the market for rented AI capacity.

The announcement lands awkwardly against the story every AI company has been telling all year: that compute is the binding constraint on growth and there is simply not enough of it to go around. If one of the industry's largest buyers of compute is suddenly a seller, the natural question is whether that story was ever true, or whether Meta is managing perceptions rather than describing reality.

Our position, set out below, is that Meta's move does not signal the end of compute scarcity. It is better read as confirmation of an argument we made in Issue V: that the real constraint has been quietly moving down the stack, from chips, to memory, to the power and water that any of this infrastructure needs to run. Meta finding temporary, localised slack in one part of an enormous fleet says very little about whether the world has enough compute. It says a great deal about who captures value as the constraint keeps moving.

A Claim That Survives on a Technicality

The detail most coverage of the announcement has understated is timing. Only weeks before the Meta Compute story broke, Meta was rationing its own staff's AI usage because Google could not supply the Gemini capacity Meta had asked for. A company that was asking employees to cut their AI token consumption in May is now proposing, in July, to sell that same scarce resource to competitors.

Read narrowly, Meta's claim is defensible. A fleet built at the scale Meta has committed to — with capital expenditure guidance of up to $145 billion this year alone — can easily carry temporary, regional or cluster-level slack even while a specific vendor relationship remains constrained. Different capacity pools are not fungible. Having spare GPU hours in Ohio does not help if the shortage sits in a memory allocation from a third party in another region.

What the announcement does not do is demonstrate that AI compute, in aggregate, has become abundant. Morgan Stanley's own modelling keeps Meta's 2027 capital expenditure guidance intact at around $175 billion, and notes that if the new cloud business scales, spending could rise rather than fall. A company sitting on a genuine, structural surplus does not usually respond by holding capital spending steady while adding a new demand-generating business line on top of it.

Why We Think Meta Actually Said It

We see three forces converging, all of them financial rather than technical. First, compute that has already been paid for carries a near-zero marginal cost to resell, which makes monetising idle capacity an attractive lever for a company under sustained investor pressure to show a return on its infrastructure spend. Second, the market reaction itself — a $125 billion move on an unconfirmed report with no pricing, no customers and no launch date — looks more like a signal engineered for a capex-anxious Wall Street ahead of an earnings call than an operational milestone. Third, and consistent with the argument we made in Issue V about hyperscaler cash flow strain, this is a balance sheet story before it is a compute story. Selling excess capacity is one of the few tools available to a company that has committed hundreds of billions of dollars to infrastructure and now needs to show the spending can pay for itself.

None of this makes the underlying claim false. It does mean the claim has been framed to produce a specific reaction, and sophisticated readers should treat it as an investor relations event first and a supply signal a distant second.

The Demand Case Has Not Weakened

It is worth remembering that a Meta cloud business is not a sudden idea. Mark Zuckerberg first floated the possibility on Meta's Q3 2025 earnings call and repeated it at May's annual shareholder meeting, describing it as an option that becomes available only once infrastructure has been overbuilt relative to Meta's own needs. Meta is also following, rather than leading, a pattern SpaceX set earlier this year when it began renting out surplus capacity at its Colossus data centre to Anthropic and Google. Neither precedent supports the idea that scarcity has ended. Both are examples of individual owners of very large, pre-contracted fleets finding pockets of slack — which is a different phenomenon entirely.

Look past Meta and the wider evidence points the other way. Advanced packaging capacity at TSMC has been sold out through 2025 and into 2026, with the company's most advanced fabrication capacity booked through 2028, and its own chief executive has described demand as running roughly three times ahead of what the business can currently produce. The entire global supply of high bandwidth memory for 2026 is reportedly committed, with data centres on track to consume up to seventy per cent of all memory chip output this year. GPU lead times for data centre hardware now run from thirty-six to fifty-two weeks, and Chinese buyers alone have placed orders for more than two million H200-class chips against roughly seven hundred thousand units currently held in Nvidia's inventory.

Layered on top, the five largest hyperscalers — Amazon, Google, Meta, Microsoft and Oracle — are committing a combined $600 to $700 billion in capital expenditure this year, the majority of it directed at AI infrastructure. That scale of buying does not leave much room for anyone else, and it is precisely why the shortage has kept moving rather than resolving: from chips, to memory and packaging, and now to power and grid interconnection, each layer taking longer to fix than the one before it.

McKinsey estimates $5.2 trillion of AI-related data centre investment will be required by 2030 to support around 156 gigawatts of AI-related capacity, with global data centre capacity roughly doubling from just over 100 gigawatts today to close to 200 gigawatts by the end of the decade. Occupancy across the existing global pipeline already sits at ninety-seven per cent, and seventy-seven per cent of capacity under construction is pre-leased. This is a demand picture that is forward-committed, not speculative, and it sits uneasily next to any suggestion that the industry has simply run out of uses for the compute it is building.

The Layer Beneath: Power and Water

If compute itself is not in genuine oversupply, the layer underneath it is where the real bottleneck is heading next — and it is the layer we have positioned Monard against since our hydrogen and behind-the-meter work in Issue IV. Goldman Sachs projects data centre power demand could rise by as much as one hundred and sixty-five per cent by 2030 against 2023 levels, and estimates that roughly $720 billion of grid spending will be required through the same period simply to keep pace. In the United States, Grid Strategies projects peak electricity demand could grow by one hundred and sixty-six gigawatts by 2030, with data centres responsible for more than half of that increase.

Water follows the same trajectory and receives far less attention. Global data centre water consumption, used mainly for evaporative cooling, currently runs at around five hundred and sixty billion litres a year and is projected to more than double to around one thousand two hundred billion litres a year by 2030. A single 100 megawatt facility can use roughly two million litres of water a day — comparable to the daily consumption of six and a half thousand households. The strain is already visible at a regional level, with data centres now accounting for around twenty-six per cent of Virginia's total electricity load and contributing an estimated $9.3 billion to recent increases in PJM capacity market pricing.

This is the part of the stack that cannot be solved by a hyperscaler finding spare GPU hours in an existing fleet. New generation, transmission and water infrastructure takes years to permit and build regardless of how quickly capital is committed — which is exactly the kind of structural, multi-year lead time gap that favours specialist infrastructure providers over the compute buyers themselves.

Where Monard Sits on This

We are not taking a position on which hyperscaler wins the argument over reselling compute, and we would caution investors against treating Meta's announcement as a referendum on that question either way. Our position is that the physical layer beneath compute — power generation and water supply — remains investable regardless of how that contest resolves. Monard, together with our global manufacturing partners in hydrogen fuel cells and behind-the-meter generation, is positioned to supply exactly this layer as data centre operators across the industry look for power and water solutions that can be delivered faster than traditional utility build-outs allow.

Behind-the-meter generation matters here precisely because it sidesteps the multi-year interconnection queues that Goldman and Grid Strategies both flag as the binding constraint. A fuel cell or on-site generation system installed alongside a data centre does not need to wait for a transmission upgrade that may take the better part of a decade to approve and build. As the compute build-out keeps running ahead of grid and water capacity, we expect operators to pay a premium for infrastructure that can be delivered on their own construction timeline rather than the utility's.

This is the same thesis we set out in Issue IV on hydrogen and Issue V on hyperscaler cash flow, applied to a live example. Compute headlines will keep shifting — this month towards an apparent glut, next month likely back towards scarcity — as different companies report different positions in their own fleets. The power and water investment case beneath all of it does not depend on which headline is correct in any given month.