Everyone's watching the power bill. The real bottleneck for AI data centres is the fibre.
Everyone's watching the power bill. The real bottleneck for AI data centres is the fibre.
Everyone's watching the power bill. The real bottleneck for AI data centres is the fibre.
There is a figure moving through every data-centre conversation this year. Hyperscalers are on track to commit more than 700 billion US dollars to AI infrastructure in 2026, according to widely reported UBS Evidence Lab monitoring — close to a sixfold increase on 2022. Alongside it runs a second, less comfortable number: research reported this year suggests well over 100 billion dollars of AI data-centre projects have already been blocked or delayed, most of it tied to power, water and local objection.
So the industry has settled on a story. The constraint on AI is electricity. Secure a gigawatt of power and a patch of cheap land, and the rest follows.
It is half right. Power is the loudest constraint, and it is real. But it is not the only one that decides whether a site is worth building, and it is not the one that quietly kills projects after the ground is broken. That distinction belongs to fibre. A campus without power cannot run. A campus without the right fibre, on the right routes, at the right capacity, can run — and still fail to do the job it was financed to do. The first problem is obvious on a spreadsheet. The second only shows up once the racks are live and the latency to the users who matter is wrong.
This article is about the second problem, because it is the one that gets designed in too late.
There is a figure moving through every data-centre conversation this year. Hyperscalers are on track to commit more than 700 billion US dollars to AI infrastructure in 2026, according to widely reported UBS Evidence Lab monitoring — close to a sixfold increase on 2022. Alongside it runs a second, less comfortable number: research reported this year suggests well over 100 billion dollars of AI data-centre projects have already been blocked or delayed, most of it tied to power, water and local objection.
So the industry has settled on a story. The constraint on AI is electricity. Secure a gigawatt of power and a patch of cheap land, and the rest follows.
It is half right. Power is the loudest constraint, and it is real. But it is not the only one that decides whether a site is worth building, and it is not the one that quietly kills projects after the ground is broken. That distinction belongs to fibre. A campus without power cannot run. A campus without the right fibre, on the right routes, at the right capacity, can run — and still fail to do the job it was financed to do. The first problem is obvious on a spreadsheet. The second only shows up once the racks are live and the latency to the users who matter is wrong.
This article is about the second problem, because it is the one that gets designed in too late.
There is a figure moving through every data-centre conversation this year. Hyperscalers are on track to commit more than 700 billion US dollars to AI infrastructure in 2026, according to widely reported UBS Evidence Lab monitoring — close to a sixfold increase on 2022. Alongside it runs a second, less comfortable number: research reported this year suggests well over 100 billion dollars of AI data-centre projects have already been blocked or delayed, most of it tied to power, water and local objection.
So the industry has settled on a story. The constraint on AI is electricity. Secure a gigawatt of power and a patch of cheap land, and the rest follows.
It is half right. Power is the loudest constraint, and it is real. But it is not the only one that decides whether a site is worth building, and it is not the one that quietly kills projects after the ground is broken. That distinction belongs to fibre. A campus without power cannot run. A campus without the right fibre, on the right routes, at the right capacity, can run — and still fail to do the job it was financed to do. The first problem is obvious on a spreadsheet. The second only shows up once the racks are live and the latency to the users who matter is wrong.
This article is about the second problem, because it is the one that gets designed in too late.
Why a data centre is only as good as its routes
Why a data centre is only as good as its routes
Why a data centre is only as good as its routes
Compute has a location, but AI workloads have a geography. Training clusters pull enormous volumes between sites and storage. Inference has to sit close to the people and systems it serves or the experience degrades. Increasingly the two are split across regions, which means the traffic between them — east–west between campuses, and out to end users — is not an afterthought bolted on at the end. It is a design input at the front.
That is where fibre stops being a utility and becomes a siting decision. A location can have abundant power and still be a poor choice because the diverse fibre routes out of it are thin, the nearest carrier-neutral interconnection is too far, or the paths to the markets it needs to reach share a single point of failure. Two sites with identical power profiles can have completely different value depending on what leaves the building and where it can go.
This is the part of the map most site models under-weight. Power and land are scored precisely. Connectivity is often reduced to a yes/no — is there fibre nearby — when the questions that actually matter are harder. How many genuinely diverse paths exist. What the latency is to the regions this workload serves. Whether capacity can scale without a new build. Whether the route to demand on the other side of an ocean is a single cable or several. A site that looks strong on power and weak on those answers is a site that will disappoint quietly, months after the decision looked sound.
Compute has a location, but AI workloads have a geography. Training clusters pull enormous volumes between sites and storage. Inference has to sit close to the people and systems it serves or the experience degrades. Increasingly the two are split across regions, which means the traffic between them — east–west between campuses, and out to end users — is not an afterthought bolted on at the end. It is a design input at the front.
That is where fibre stops being a utility and becomes a siting decision. A location can have abundant power and still be a poor choice because the diverse fibre routes out of it are thin, the nearest carrier-neutral interconnection is too far, or the paths to the markets it needs to reach share a single point of failure. Two sites with identical power profiles can have completely different value depending on what leaves the building and where it can go.
This is the part of the map most site models under-weight. Power and land are scored precisely. Connectivity is often reduced to a yes/no — is there fibre nearby — when the questions that actually matter are harder. How many genuinely diverse paths exist. What the latency is to the regions this workload serves. Whether capacity can scale without a new build. Whether the route to demand on the other side of an ocean is a single cable or several. A site that looks strong on power and weak on those answers is a site that will disappoint quietly, months after the decision looked sound.
Compute has a location, but AI workloads have a geography. Training clusters pull enormous volumes between sites and storage. Inference has to sit close to the people and systems it serves or the experience degrades. Increasingly the two are split across regions, which means the traffic between them — east–west between campuses, and out to end users — is not an afterthought bolted on at the end. It is a design input at the front.
That is where fibre stops being a utility and becomes a siting decision. A location can have abundant power and still be a poor choice because the diverse fibre routes out of it are thin, the nearest carrier-neutral interconnection is too far, or the paths to the markets it needs to reach share a single point of failure. Two sites with identical power profiles can have completely different value depending on what leaves the building and where it can go.
This is the part of the map most site models under-weight. Power and land are scored precisely. Connectivity is often reduced to a yes/no — is there fibre nearby — when the questions that actually matter are harder. How many genuinely diverse paths exist. What the latency is to the regions this workload serves. Whether capacity can scale without a new build. Whether the route to demand on the other side of an ocean is a single cable or several. A site that looks strong on power and weak on those answers is a site that will disappoint quietly, months after the decision looked sound.
Texas, and the difference between space and viability
Texas, and the difference between space and viability
Texas, and the difference between space and viability
Texas is the clearest illustration of the gap between finding space and finding a viable site. The state has become one of the most active data-centre construction markets in the world, and for understandable reasons: land, a grid its operators can negotiate with directly, and a policy environment that welcomes the load. If the only test were power and space, much of the state would pass.
But the operators moving fastest are the ones treating connectivity as a first-round filter rather than a late line item. They are asking how a West Texas campus reaches users on the coasts with acceptable latency. They are asking how it connects onward to the international routes that carry traffic across the Pacific to the markets AI is growing fastest in. They are asking whether the diverse paths they need already exist or have to be built — because building long-haul fibre is measured in quarters and permits, not weeks.
The lesson generalises well beyond one state. Finding data-centre space is now the easy part. The hard part is confirming that the space can be connected the way the workload requires, on the timeline the business is working to, at a cost that does not quietly erode the economics. Space is abundant. Viability is not. The difference between the two is almost always the fibre.
Texas is the clearest illustration of the gap between finding space and finding a viable site. The state has become one of the most active data-centre construction markets in the world, and for understandable reasons: land, a grid its operators can negotiate with directly, and a policy environment that welcomes the load. If the only test were power and space, much of the state would pass.
But the operators moving fastest are the ones treating connectivity as a first-round filter rather than a late line item. They are asking how a West Texas campus reaches users on the coasts with acceptable latency. They are asking how it connects onward to the international routes that carry traffic across the Pacific to the markets AI is growing fastest in. They are asking whether the diverse paths they need already exist or have to be built — because building long-haul fibre is measured in quarters and permits, not weeks.
The lesson generalises well beyond one state. Finding data-centre space is now the easy part. The hard part is confirming that the space can be connected the way the workload requires, on the timeline the business is working to, at a cost that does not quietly erode the economics. Space is abundant. Viability is not. The difference between the two is almost always the fibre.
Texas is the clearest illustration of the gap between finding space and finding a viable site. The state has become one of the most active data-centre construction markets in the world, and for understandable reasons: land, a grid its operators can negotiate with directly, and a policy environment that welcomes the load. If the only test were power and space, much of the state would pass.
But the operators moving fastest are the ones treating connectivity as a first-round filter rather than a late line item. They are asking how a West Texas campus reaches users on the coasts with acceptable latency. They are asking how it connects onward to the international routes that carry traffic across the Pacific to the markets AI is growing fastest in. They are asking whether the diverse paths they need already exist or have to be built — because building long-haul fibre is measured in quarters and permits, not weeks.
The lesson generalises well beyond one state. Finding data-centre space is now the easy part. The hard part is confirming that the space can be connected the way the workload requires, on the timeline the business is working to, at a cost that does not quietly erode the economics. Space is abundant. Viability is not. The difference between the two is almost always the fibre.
What a connectivity-led site assessment actually checks
What a connectivity-led site assessment actually checks
What a connectivity-led site assessment actually checks
The correction is not complicated, but it has to happen early, while the site is still a choice rather than a commitment. A connectivity-led assessment asks a short, unglamorous set of questions before the power negotiation closes, not after.
How many physically diverse fibre routes leave this site, and who owns them. What is the real latency to each region the workload has to serve, measured rather than assumed. Where is the nearest carrier-neutral interconnection point, and what does reaching it cost in both distance and time. Can capacity scale on existing paths, or does growth require a fresh build. And for anything with an international dimension, does the route to demand cross an ocean on a single cable or on several — because a cross-Pacific path with no diversity is a risk that does not appear until the day it fails.
The common mistake is sequencing. Teams lock the power deal and the land, then discover the connectivity has to be retrofitted around decisions already made — narrowing the options and raising the cost of every one that remains. Fibre is cheaper to design in than to bolt on. Treated as a first-round filter it removes weak sites before money is spent. Treated as a finishing task it becomes the constraint nobody budgeted for.
The correction is not complicated, but it has to happen early, while the site is still a choice rather than a commitment. A connectivity-led assessment asks a short, unglamorous set of questions before the power negotiation closes, not after.
How many physically diverse fibre routes leave this site, and who owns them. What is the real latency to each region the workload has to serve, measured rather than assumed. Where is the nearest carrier-neutral interconnection point, and what does reaching it cost in both distance and time. Can capacity scale on existing paths, or does growth require a fresh build. And for anything with an international dimension, does the route to demand cross an ocean on a single cable or on several — because a cross-Pacific path with no diversity is a risk that does not appear until the day it fails.
The common mistake is sequencing. Teams lock the power deal and the land, then discover the connectivity has to be retrofitted around decisions already made — narrowing the options and raising the cost of every one that remains. Fibre is cheaper to design in than to bolt on. Treated as a first-round filter it removes weak sites before money is spent. Treated as a finishing task it becomes the constraint nobody budgeted for.
The correction is not complicated, but it has to happen early, while the site is still a choice rather than a commitment. A connectivity-led assessment asks a short, unglamorous set of questions before the power negotiation closes, not after.
How many physically diverse fibre routes leave this site, and who owns them. What is the real latency to each region the workload has to serve, measured rather than assumed. Where is the nearest carrier-neutral interconnection point, and what does reaching it cost in both distance and time. Can capacity scale on existing paths, or does growth require a fresh build. And for anything with an international dimension, does the route to demand cross an ocean on a single cable or on several — because a cross-Pacific path with no diversity is a risk that does not appear until the day it fails.
The common mistake is sequencing. Teams lock the power deal and the land, then discover the connectivity has to be retrofitted around decisions already made — narrowing the options and raising the cost of every one that remains. Fibre is cheaper to design in than to bolt on. Treated as a first-round filter it removes weak sites before money is spent. Treated as a finishing task it becomes the constraint nobody budgeted for.
Where Vocom sits, and the next step
Where Vocom sits, and the next step
Where Vocom sits, and the next step
This is the work Vocom International has done for twenty-seven years in telecommunications: making sure the connectivity behind high-demand infrastructure is specified correctly and delivered on time. Vocom does not manufacture fibre — it sources and supplies it through tier 1 manufacturing partners, and it specifies, provisions and coordinates the routes, private circuits and capacity that decide whether a site can reach the markets it was built to serve. Dark fibre for new campuses. IEPL and IPLC private leased circuits for dedicated, low-latency paths. IP transit and cross-Pacific capacity where the demand and the compute sit on opposite sides of an ocean.
The buildout will not slow down. The projects that succeed will be the ones that treated fibre as a siting decision rather than a finishing touch — that asked, before the ground was broken, not just whether the site had power, but whether it could reach the people it was meant to serve.
If you are evaluating a site, or connecting one already under construction, that is the conversation worth having early. Talk to Vocom about fibre routing, private circuits and cross-Pacific capacity for AI data-centre connectivity: vocom.ai/contact-vocomai
This is the work Vocom International has done for twenty-seven years in telecommunications: making sure the connectivity behind high-demand infrastructure is specified correctly and delivered on time. Vocom does not manufacture fibre — it sources and supplies it through tier 1 manufacturing partners, and it specifies, provisions and coordinates the routes, private circuits and capacity that decide whether a site can reach the markets it was built to serve. Dark fibre for new campuses. IEPL and IPLC private leased circuits for dedicated, low-latency paths. IP transit and cross-Pacific capacity where the demand and the compute sit on opposite sides of an ocean.
The buildout will not slow down. The projects that succeed will be the ones that treated fibre as a siting decision rather than a finishing touch — that asked, before the ground was broken, not just whether the site had power, but whether it could reach the people it was meant to serve.
If you are evaluating a site, or connecting one already under construction, that is the conversation worth having early. Talk to Vocom about fibre routing, private circuits and cross-Pacific capacity for AI data-centre connectivity: vocom.ai/contact-vocomai
This is the work Vocom International has done for twenty-seven years in telecommunications: making sure the connectivity behind high-demand infrastructure is specified correctly and delivered on time. Vocom does not manufacture fibre — it sources and supplies it through tier 1 manufacturing partners, and it specifies, provisions and coordinates the routes, private circuits and capacity that decide whether a site can reach the markets it was built to serve. Dark fibre for new campuses. IEPL and IPLC private leased circuits for dedicated, low-latency paths. IP transit and cross-Pacific capacity where the demand and the compute sit on opposite sides of an ocean.
The buildout will not slow down. The projects that succeed will be the ones that treated fibre as a siting decision rather than a finishing touch — that asked, before the ground was broken, not just whether the site had power, but whether it could reach the people it was meant to serve.
If you are evaluating a site, or connecting one already under construction, that is the conversation worth having early. Talk to Vocom about fibre routing, private circuits and cross-Pacific capacity for AI data-centre connectivity: vocom.ai/contact-vocomai
