FOREMY INSIDER BRIEFING · INFRASTRUCTURE & COMPUTE
The Awkward Arrangement Nobody Predicted
A year ago, if you had suggested that a leading AI lab might soon be paying a direct frontier-model competitor billions of dollars to rent spare computing capacity, most people in the industry would have called it implausible. Labs compete for the same enterprise customers, the same research talent, and the same regulatory goodwill; renting your training infrastructure from the company you’re racing against seemed like handing a rival intelligence about your capacity needs and, indirectly, funding their own model development. And yet that is precisely the arrangement now being discussed in the market, with reports of preliminary talks around a multi-billion-dollar compute leasing deal between labs that are, in every other respect, direct rivals.
This is not an isolated curiosity. It sits alongside other arrangements in which frontier labs already rent capacity from clusters built to train a competitor’s own rival models. The pattern repeating across the industry is the same: whoever has spare compute capacity right now can sell it, even to a company they are racing against for the same customers and the same benchmark headlines.
Why Compute Has Become the Binding Constraint
To understand why labs are willing to tolerate this awkwardness, it helps to be blunt about what actually limits frontier AI development today. It is not primarily a shortage of research ideas, and it is not primarily a shortage of capital — funding for credible frontier labs has been abundant. The binding constraint is physical: enough advanced chips, assembled into large enough clusters, with enough power and cooling infrastructure, built fast enough to keep pace with training runs that grow larger with each model generation.
That physical bottleneck runs through a remarkably narrow set of chokepoints. A single manufacturer remains the only one capable of fabricating the most advanced AI accelerators at the scale the industry now demands, which makes that manufacturer’s production capacity and revenue trends one of the most closely watched thermometers for the health of the entire AI industry, arguably a better leading indicator than any individual lab’s own announcements. When the fabrication capacity itself is the limiting factor, and it is shared across every lab’s supply chain regardless of who ultimately writes the check, the old assumption that compute access maps cleanly onto competitive advantage breaks down. Everyone is drawing water from the same well.
The Logic of Renting From a Rival
Given that constraint, the economics of compute leasing between competitors start to make more sense than they initially appear to. A lab that has built out capacity for a specific training run, or that over-provisioned ahead of an anticipated need that materialized more slowly than planned, is sitting on an extremely expensive, depreciating asset the moment that capacity sits idle. Selling that idle capacity, even to a competitor, converts a sunk cost into revenue and helps fund the next round of infrastructure buildout. From the buyer’s side, renting capacity — even from a rival — can be faster and cheaper than waiting in a queue for new capacity to come online from a neutral cloud provider, particularly when demand for accelerators outstrips supply across the board.
There are real risks in these arrangements that both sides have to manage carefully: information leakage about training schedules and capacity needs, dependency risk if a rival decides to cut off access at a strategically inconvenient moment, and the simple optics of a lab appearing to fund the training runs of the company it’s trying to beat in the market. Labs entering these deals have generally structured them with strict data isolation and contractual guardrails specifically to manage the first two risks, even if the optics risk is harder to fully engineer away.
The Broader Infrastructure Buildout
This dynamic is unfolding against a backdrop of infrastructure investment at a scale that is genuinely difficult to overstate. Every major lab and hyperscaler is simultaneously building new data center campuses, signing multi-year power purchase agreements, and locking in chip supply years in advance, all while facing the same fundamental fabrication bottleneck described above. That scale of capital commitment is part of why the industry has started treating compute infrastructure spend itself as a leading indicator worth tracking as closely as model releases — a slowdown in data center construction or chip orders would be a far earlier warning sign of an industry-wide pullback than any single company’s earnings call.
It also explains why chip manufacturers and specialized accelerator designers have seen their market valuations swing so dramatically this year, occasionally overtaking or trading places with the AI labs themselves in market capitalization rankings. In a market this constrained by physical infrastructure, the companies that control the chokepoints capture an outsized share of the value being created, sometimes more than the labs producing the headline-grabbing models built on top of that hardware.
What This Means Going Forward
For anyone trying to forecast where the AI industry is headed, the compute layer deserves at least as much attention as the model layer. A few practical implications worth tracking:
- Model release cadence is downstream of compute access, not purely of research progress. A lab with a compute shortfall will slip its release calendar regardless of how good its research team is, which is worth remembering the next time a highly anticipated model launch gets delayed.
- Cross-rental deals are likely to become more common, not less, as long as the fabrication bottleneck persists, because the economic logic favors monetizing idle capacity over letting it sit unused for competitive-optics reasons alone.
- Chip and infrastructure earnings are becoming a genuine leading indicator for the health of the broader AI market, arguably more reliable than any individual lab’s self-reported metrics.
- Antitrust and national security regulators are likely to start paying closer attention to compute-sharing arrangements between nominal competitors, given the strategic sensitivity of frontier AI capacity.
A Parallel in Older Infrastructure Industries
Competitors quietly leaning on each other’s physical infrastructure is not actually new to technology; it’s a pattern that shows up whenever an industry’s growth outpaces the specialized infrastructure needed to support it. Airlines that compete fiercely for the same passengers have long shared runway slots, ground handling services, and even aircraft leasing arrangements at capacity-constrained airports, because the alternative to sharing scarce infrastructure is often simply not being able to operate a route at all. Telecom carriers that compete for the same subscribers have historically shared cell towers and, in some markets, entire network backbones, because building fully duplicate physical infrastructure for every competitor would be an economically absurd use of capital.
AI compute appears to be following the same pattern, just compressed into a far shorter timeframe because the technology itself is moving so quickly. The difference this time is the stakes attached to the resource being shared. A shared runway slot doesn’t hand a competitor insight into your growth strategy; a compute leasing arrangement plausibly does, at least indirectly, through the sheer scale and timing of the capacity being requested. That is precisely why the contractual guardrails around these deals, covering what operational information is and isn’t visible to the counterparty, matter so much more here than in older infrastructure-sharing arrangements.
It’s also worth noting how quickly sentiment around these arrangements can flip once they become public. A deal that looks like sensible capacity management inside a boardroom can look, from the outside, like a tacit admission that a lab overbuilt its own infrastructure, or that its compute strategy is less self-sufficient than its public narrative has suggested. Labs weighing these deals are almost certainly running that reputational calculation alongside the financial one, and the fact that several are proceeding anyway is itself a signal of just how binding the underlying compute constraint has become across the entire industry, not just for any single company. The companies most willing to accept that reputational risk today are, in effect, signaling that the cost of idle compute now outweighs the cost of an awkward headline, a genuine reversal of priorities from just a year or two ago.
The Foremy Take
The AI industry likes to talk about itself as a race between models. Underneath that story is a much less glamorous race for concrete, power, and silicon, and that race is the one actually setting the pace. When rivals start renting compute from each other, it’s not a sign the competition has softened — it’s a sign the physical bottleneck has gotten tight enough that even fierce competitors need each other’s spare capacity to keep running.
What to Watch Next
- Whether the reported compute leasing talks between rival labs actually close, and on what terms.
- Chip manufacturer earnings and capacity guidance as the clearest available proxy for industry-wide training demand.
- New power purchase agreements and data center announcements, which now function as a leading indicator of which labs are positioning for the next generation of training runs.
- Regulatory interest in compute-sharing arrangements between competing frontier labs.
This report is part of Foremy's ongoing AI Insider Report series, tracking the economics, infrastructure, and policy decisions shaping the AI industry. Foremy Team, foremy.com/.