From Concrete to Compute: Why Clichmont Is Building AI Infrastructure Instead of Renting It

Spokesperson: Alexis Cathalifaud, CEO

 

Angle

Every well-funded neocloud on this class – CoreWeave, Crusoe, Lambda is racing towards the identical GPU-rental mannequin. Clichmont’s guess is completely different, and the opinion is to personal the information facilities, personal the ability, personal the availability chain. This interview ought to learn as a founder considering out loud about that guess, what it prices, what it dangers, and why he thinks the remainder of the class has the sequencing backwards. It shouldn’t learn as a product pitch or a token launch announcement.

As demand for synthetic intelligence compute continues to develop, the infrastructure supporting that demand is turning into a strategic consideration in its personal proper. Companies throughout the sector are racing to safe entry to more and more highly effective GPUs, whereas questions round electrical energy, data-center capability, cooling and connectivity have gotten more durable to separate from the compute itself.

Clichmont is taking a distinct method. Rather than constructing its mannequin primarily round rented GPU capability, the corporate is targeted on proudly owning and controlling the bodily infrastructure on which successive generations of AI {hardware} can function. In this interview, Clichmont CEO Alexis Cathalifaud discusses why the corporate believes energy and data-center infrastructure may develop into the extra sturdy bottlenecks, the way it approaches web site choice and the challenges of scaling bodily infrastructure, in addition to the function of its $CLAI token throughout the broader ecosystem.

1) Every firm on this class is combating over GPU entry proper now. Clichmont’s reply is to construct the information facilities as a substitute of renting the chips. Why does possession matter greater than entry?

Because GPU entry offers you compute; infrastructure possession offers you management over the economics of compute.

For an organization like Clichmont, proudly owning or controlling the data-center layer can matter extra strategically than merely securing rented GPUs. When you hire GPU capability from a hyperscaler or GPU cloud, you inherit another person’s pricing, availability, energy constraints, networking structure, deployment schedule, and margins. When demand spikes, entry can develop into costly or constrained.

Owning the infrastructure adjustments the equation. Clichmont can probably determine which GPUs to deploy, when to improve them, how densely to set up them, how energy and cooling are engineered, and the way the capability is commercialized. The similar facility may evolve from one GPU era to the subsequent slightly than tying the enterprise thesis to a specific chip.

There is one other necessary distinction: GPUs depreciate rapidly; power-ready data-center capability is a longer-lived strategic asset. A GPU era might develop into economically much less aggressive inside just a few years, whereas land, grid connections, substations, cooling infrastructure, fiber connectivity and permitted megawatts can stay helpful throughout a number of generations of accelerators.

That makes the scarce useful resource more and more not simply the GPU itself, however the flexibility to energize 1000’s of GPUs at scale. An organization should purchase chips and nonetheless have nowhere appropriate to deploy them. Securing 10,000 GPUs is one drawback; securing the tens of megawatts of dependable electrical energy, cooling and community infrastructure required to function them is one other.

 

2) You’re up towards firms which are already public or heading there – CoreWeave, Crusoe, Lambda. What do you assume their mannequin will get improper, if something?

I don’t assume CoreWeave, Crusoe or Lambda obtained the mannequin improper. They proved that AI compute is a large market. Where we differ is in what we consider will stay scarce. GPUs change each era. The sturdy bottleneck is the infrastructure required to run them — energy, land, cooling and connectivity. Clichmont’s thesis is that slightly than competing solely to hire the most recent GPU, we wish to management the infrastructure on which successive generations of GPUs will function. In a market the place everyone seems to be chasing chips, we’d slightly personal the place the place the chips have to stay 

 

3) There’s a rising argument that power, not chips, is the precise bottleneck for AI infrastructure. How a lot does that form the place and the way Clichmont builds?

Energy shapes virtually each infrastructure choice we make. A GPU with out dependable energy is simply costly {hardware} sitting in a rack. We consider the actual competitors over the subsequent decade gained’t merely be for GPUs—it is going to be for megawatts.

So when Clichmont evaluates a web site, we don’t begin by asking the place we are able to discover the most cost effective constructing. We ask: the place can we safe dependable energy, on the proper economics, with the flexibility to scale? What’s the time-to-power? What’s the grid state of affairs? What cooling structure does the local weather permit? And can that web site help the subsequent era of GPUs, not simply those we’re putting in immediately?

That’s one purpose areas with robust power fundamentals are strategically attention-grabbing to us. Chips may be shipped all over the world. You can’t ship 100 megawatts. The compute finally has to go the place the power is.

So I wouldn’t say chips cease being a bottleneck. They stay important. But more and more, proudly owning GPUs isn’t sufficient. The aggressive benefit is having the ability to energy, cool and function them economically at scale. That’s what we’re constructing Clichmont round.

 

4) Clichmont’s websites vary from a solar-powered facility in Alicante to a brand new construct in Bodo, Norway. What really decides the place a knowledge heart will get constructed – is it about power, land, local weather, one thing else?

We don’t select a location as a result of one variable appears enticing. We select it as a result of the whole infrastructure equation works.

Power is the primary filter: what number of megawatts can we safe, at what value, how dependable is that offer, and—critically—how rapidly can it really be delivered? Then we have a look at cooling, local weather, fiber connectivity, land, allowing, safety and the flexibility to broaden.

Bodø and Alicante are attention-grabbing exactly as a result of they characterize completely different strengths. Northern Norway offers us a local weather that may help environment friendly cooling and a robust power surroundings. Alicante offers us a distinct power profile and the chance to combine photo voltaic into the infrastructure technique. We don’t consider each Clichmont knowledge heart wants to look an identical—the structure ought to reply to the sources of the placement.

And land by itself isn’t notably helpful to us. An inexpensive parcel with no scalable energy or fiber isn’t a data-center web site. What issues is whether or not we are able to flip that location into dependable, economically aggressive compute capability.

Ultimately, we’re not likely on the lookout for land. We’re on the lookout for locations the place power, connectivity, cooling and scalability converge. That’s the place we construct.

 

5) This is an infrastructure firm with a token connected to it. For a reader who’s skeptical of that mixture, what’s the sincere case for why $CLAI exists in any respect?

The skeptical view is totally truthful. A token shouldn’t exist simply because an organization operates in AI. If $CLAI had been merely a financing wrapper round our knowledge facilities, I wouldn’t take into account {that a} compelling purpose to create it.

Clichmont is the infrastructure enterprise. It builds and operates compute capability. $CLAI is meant to be a digital financial layer across the broader ecosystem — one thing that may ultimately help on-chain participation, treasury exercise and neighborhood governance in ways in which standard fairness isn’t designed to do.

And we’ve to earn the suitable to make that distinction. The bodily infrastructure has to exist independently of the token, and the token has to exhibit actual utility independently of hypothesis. If we are able to’t present each, then the skepticism is justified.

So I wouldn’t ask anybody to consider in $CLAI just because Clichmont owns GPUs or builds knowledge facilities. The check is way easier: does the token ultimately do one thing helpful, clear and measurable that couldn’t be completed as successfully with a traditional database or standard company construction? That’s the usual we needs to be held to.

 

6) What’s the toughest half of scaling bodily infrastructure that individuals who’ve solely constructed software program have a tendency to underestimate?

The hardest half is that bodily infrastructure doesn’t scale at software program velocity. In software program, if demand doubles, you may usually provision extra capability rapidly. In a knowledge heart, each further megawatt has a bodily dependency behind it — grid capability, transformers, switchgear, cooling, fiber, permits, development and finally {hardware}.

And these dependencies don’t transfer in parallel as neatly as individuals think about. You can have the land and never have the ability. You can have the ability allocation and wait months for electrical gear. You can have the constructing prepared and nonetheless be ready for a grid connection. One lacking part can delay a complete deployment.

The different distinction is that errors are costly and troublesome to reverse. Software may be patched in a single day. You can’t patch a badly designed 50-megawatt electrical system in a single day. You’re making capital selections immediately based mostly on what GPUs, energy densities and cooling necessities might seem like a number of years from now.

So the actual ability isn’t merely constructing knowledge facilities. It’s sequencing capital, energy, development and buyer demand in order that they arrive at roughly the identical second. Build too early and you’ve got costly idle infrastructure. Build too late and the shopper goes some other place.

That execution self-discipline might be what individuals coming purely from software program underestimate most. In bodily AI infrastructure, velocity issues — however timing issues much more.

 

7) If you had to identify the most important threat in betting on a build-it-yourself mannequin as a substitute of a capital-light rental mannequin, what wouldn’t it be?

The largest threat is capital depth mixed with timing. When you construct infrastructure your self, you’re committing important capital immediately towards assumptions about demand, energy economics and know-how a number of years into the longer term.

A rental mannequin offers you flexibility. If the market adjustments, you may scale back capability, transfer suppliers or undertake the subsequent era of {hardware}. When you personal the infrastructure, you don’t have that luxurious. A substation, cooling system or data-center constructing is a long-duration choice.

For us, the most important hazard subsequently isn’t merely spending an excessive amount of — it’s constructing the improper capability, within the improper place, on the improper time. If you construct forward of demand, capital sits idle. If you construct too slowly, you miss the market.

That’s why we don’t view possession as ‘construct every part ourselves.’ The goal is to management the strategic infrastructure whereas remaining versatile round know-how. The constructing, energy, cooling and connectivity ought to survive a number of generations of GPUs slightly than turning into depending on one {hardware} cycle.

So sure, the capital-light mannequin has an actual benefit: optionality. Our guess is that if we execute accurately, giving up some short-term optionality creates one thing extra helpful over the long run — management over capability, energy economics and the bodily infrastructure that AI more and more depends upon.

 

8) Three years from now, the place would you like Clichmont to sit relative to the CoreWeaves and Nebiuses of the world?

Three years from now, I don’t anticipate Clichmont to be the most important firm within the class, and that’s not the target. CoreWeave and Nebius have monumental scale and entry to capital. Trying to replicate them can be the improper technique for us.

I need Clichmont to be acknowledged as one of essentially the most environment friendly unbiased AI infrastructure operators in Europe — with actual working belongings, secured energy, high-density GPU capability and a observe document of bringing new compute on-line rapidly.

Our benefit has to come from being disciplined about the place we construct and what we personal. We need areas the place the power economics make sense, infrastructure designed round successive generations of accelerated computing, and the pliability to serve enterprise AI, HPC and personal compute slightly than merely competing for GPU rental quantity.”

If CoreWeave and Nebius are constructing hyperscale AI clouds, Clichmont can occupy a distinct place: a targeted proprietor and operator of compute-ready infrastructure in strategically chosen markets.

 

Conclusion

Clichmont’s technique finally comes down to a long-term infrastructure guess: that entry to GPUs will stay necessary, however the capacity to energy, cool, join and function these GPUs effectively at scale will develop into an more and more helpful benefit.

That method comes with significant trade-offs. Building bodily infrastructure requires substantial capital, lengthy planning horizons and cautious coordination between energy, development, {hardware} and demand. Clichmont’s thesis is that accepting these constraints can present better management over the infrastructure required for successive generations of AI compute. Whether that thesis proves out will rely much less on the ambition of the mannequin than on the corporate’s capacity to execute it effectively and on the proper time.