AI data centre market projected to hit $123bn by 2035

[StoneCold]

AI Data Center Market to Reach USD 123,596.19 Million by 2035 Amid Rapid Adoption of Generative AI, High-Performance Computing, and AI-Optimized Infrastructure - DC Market Insightswww.prnewswire.co.uk

Forecasts suggest the AI data centre market will grow massively over the next decade, driven by generative AI, HPC workloads, and specialised infrastructure. The shift from general-purpose cloud to AI-optimised environments is accelerating, with companies redesigning entire data centre architectures around GPUs, cooling, and power efficiency. This is a structural change, not a temporary trend.

Energy demand is going to be one of the biggest constraints here

[Falcon]

Cooling and power infrastructure will matter as much as chips. Hot hot hot. Need it for swimming baths and send it to old peoples homes

[ProperMadlad20]

We might see regional competition based on energy availability
This scale of investment suggests long-term confidence, not short-term hype

[GoldbergFan_X]

Also raises sustainability concerns that aren't fully solved yet

[Tracey]

We havent got the power here in the uk and what we do have costs the earth

[MiniElliot]

Same thing happened to me. Definitely worth picking up.

[Tara_66]

That is pretty much what I took from it too. The speed of the news cycle means most things get forgotten before they are properly resolved.

I will update this thread if anything significant changes

[MrRicardo]

The initial reporting on this was all over the place. Worth watching closely

[Taker92]

The $123bn figure is eye-catching, but the more useful question is what is actually being counted. If it includes construction, power systems, cooling, networking, GPUs, and related infrastructure, then it is really a measure of the whole AI infrastructure build-out rather than just server rooms.

The practical bottleneck may end up being electricity rather than demand. A company can order another rack of accelerators tomorrow, but getting a suitable grid connection, transformers, cooling capacity and planning approval can take years. That makes the data centre market an unusually physical part of the AI boom

Still, projections that far out should be treated as scenarios rather than forecasts carved in stone. Hardware efficiency could improve dramatically, or workloads could become much more demanding and push the other way.

[Cole99]

The number sounds huge, but then you look at what one modern AI cluster actually needs and it becomes less ridiculous. You are not just buying computers. You need buildings, substations, backup generation, high-capacity networking, cooling and a frankly heroic amount of engineering.

What interests me more is who captures the spending. The obvious winners are data centre operators, chip suppliers and cloud companies, but there is a whole second tier selling power equipment, cooling systems, optical networking and construction services.

That is where I would be cautious about simply buying anything with AI in its name, though. A booming market does not mean every supplier gets booming margins. Plenty of companies will be fighting over contracts while customers squeeze prices.

[EarlyBird]

The electricity point is probably the sleeper issue here. A data centre can be incredibly valuable and still be useless if it cannot get enough reliable power at the right location.

There is also an interesting regional angle. Europe may have strong demand for AI capacity but very different constraints around land, grid connections, planning and energy prices compared with the US. That could push more workloads toward wherever capacity can actually be built rather than wherever customers initially want it.

So I would watch megawatts commissioned alongside the dollar forecasts. Dollars can be moved around by pricing and accounting; physical capacity is harder to fake.

[Harbour36]

A $123bn projection for 2035 sounds like the sort of number designed to make people spill their coffee Still, the underlying trend seems plausible. Generative AI has turned compute from a background IT expense into something executives are actively planning around.

The interesting bit will be whether inference becomes as important as training. Training gets all the headlines because the clusters are enormous, but if millions of businesses start running AI services continuously, inference could create a very different and more persistent demand curve.

That could also change where the data centres are built. Low-latency applications may need compute closer to users, while large training jobs can be placed wherever cheap power and suitable infrastructure are available.

[Cognition]

Could the forecast actually be conservative if AI becomes embedded in every business process? A bank running models for fraud detection, a retailer forecasting inventory, a manufacturer analysing cameras and a software company serving millions of AI requests all create different kinds of compute demand.

On the other hand, efficiency improvements are the giant wildcard. If models become dramatically cheaper to run, demand may rise while the required compute per task falls. That is not necessarily bad for the industry either, because cheaper inference could unlock applications that are currently uneconomic.

So the classic paradox applies: making something cheaper can increase total consumption because more people use it.

[Molly4]

There is another angle: governments are starting to treat compute capacity almost like strategic infrastructure. If AI becomes important to defence, healthcare, finance and public services, countries may want domestic or allied capacity rather than relying entirely on a distant cloud region.

That could create a second wave of investment driven less by pure commercial demand and more by resilience and national strategy. It would also explain why some projects might make sense even when their economics look less attractive on paper.

The downside is that everyone discovering AI at once can create a serious queue for transformers, grid connections and skilled contractors. The bottleneck may not be the GPU at all.

[LatentSpace82]

There is a funny contrast here. Everyone talks about AI as if it lives in the cloud, but the cloud still needs a very large building somewhere with pipes, cables and a power connection. The magic eventually meets a concrete floor

Cooling deserves more attention too. Higher-density accelerators mean you cannot simply keep adding conventional air conditioning forever. Liquid cooling and more sophisticated thermal management are becoming part of the infrastructure equation.

That creates another potential constraint: the skills needed to build and maintain these facilities. Electrical engineers, mechanical engineers, construction teams and data centre technicians are not going to be replaced by a chatbot anytime soon.

[NovaPrime68]

The best reality check is probably to compare announced capacity with actual delivered capacity. Companies can announce enormous expansion plans, but land acquisition, permitting, financing, power connections and equipment availability all have a habit of slowing things down.

If the industry is still commissioning substantial new capacity year after year, then the broader thesis is being validated regardless of whether the eventual market is $100bn, $123bn or $150bn.

And if demand suddenly softens, those giant projects will be much harder to hide than a bad spreadsheet forecast. Concrete is a fairly unforgiving form of accounting

[Sam]

One thing missing from a lot of these discussions is networking. Once you have huge numbers of accelerators working together, moving data between them becomes a serious engineering problem. Fibre, switches, optical components and high-speed interconnects suddenly become strategic rather than boring infrastructure.

That is why the AI data centre boom can spread far beyond the companies actually making processors. A single expansion can pull demand through several layers of the supply chain.

It also means a slowdown in one part of the stack does not necessarily mean the entire infrastructure market collapses at the same speed.

[MegaMike16]

The market size is interesting, but I would be careful about treating it as a straight-line AI success story. Data centres are capital-intensive businesses, and utilisation matters enormously. An expensive facility running at low utilisation is not a great asset just because the building contains fashionable GPUs.

There is also the financing side. Higher interest rates, equipment costs and long construction timelines can turn a seemingly attractive project into a headache very quickly.

The companies I would be watching are the ones that can secure power and keep their facilities busy, not simply the ones announcing the biggest expansion plans.

[Red Builder]

There is a sustainability question here that cannot just be waved away. More data centres mean more electricity consumption, more construction and potentially more pressure on local water supplies depending on the cooling design.

The sensible response is not to pretend AI should stop. It is to make new capacity more efficient and put it where the power system can support it. Otherwise you end up with a shiny new data centre competing with homes and factories for a grid connection.

That is a much more practical debate than simply asking whether $123bn is bullish or bearish.

[EdgeRatedR86]

The first thing I would do with a forecast like this is ignore the final number for a minute and check the assumptions underneath it. What growth rate for AI workloads? What average server utilisation? What hardware refresh cycle? What electricity cost? How much of the spending is new construction versus upgrades?

Small changes to those assumptions can produce very different totals by 2035. That does not make the forecast useless, but it does make the headline figure less meaningful on its own.

The initial reporting being messy is a good reason to look at the methodology rather than arguing over whether $123bn is exactly right. Ten billion either way does not change the bigger infrastructure story.

[Slow Hollow]

What I find encouraging is that the conversation is moving beyond GPUs. The real AI infrastructure stack includes power generation, transmission, cooling, networking, storage, security and operations.

For example, a hyperscale site might have plenty of accelerator capacity but still struggle to deliver useful throughput if its network or storage layer cannot keep up. Optimising the whole system matters more than simply buying more chips.

That makes this market much broader and potentially more durable than a single hardware cycle. The exact forecast may be wrong, but the infrastructure problem it is describing is very real.