Before cloud computing, hyperscalers or anyone had heard of the phrase “GPU cluster” companies had a simpler dilemma: where to house their physically fragile, thermally aggressive, and electrically hungry computers. Early mainframes required stable power, controlled humidity, dedicated cooling and physical security just to stay on, and because regular offices couldn’t house them, a dedicated room was required.

That room became the data centre. Not a strategy or business model, but an essential space to protect the expensive machine.

Then the internet arrived, with 24/7 access and colossal network connections no single company could justify purchasing alone. Colocation facilities emerged to share the cost, rack space was rented, and these provided the necessary power, cooling, and bandwidth.

The 2000s brought the utilisation problem. A retailer needed enormous server capacity for one Friday in November and almost nothing in February. This birthed cloud computing: pool capacity across thousands of companies with different peak times, and efficiency is secured because the load is shared – a logic that still works across the entire industry.

In 2003, Amazon founder and CEO Jeff Bezos held an offsite project at his house. What was supposed to be a 30-minute exercise to identify the organisation’s core competencies stretched into hours. They had built something nobody intended: world-class data centre infrastructure.

Amazon launched the Simple Storage Service –S3 – in March 2006 and the Elastic Compute Cloud – EC2 – in August 2006. Microsoft did not respond until October 2008. At that launch, Ray Ozzie, Microsoft's chief software architect, remarked, "I'd like to tip my hat to Jeff Bezos and Amazon", while Google’s cloud service came even later. Amazon’s Web Service (AWS) had more than two years to establish patterns that would compound for a decade.

The cloud started as cheap pooled compute, but the footprint required to serve millions of customers encompassed all: CPUs; memory; storage; networking; GPUs; cooling; power. The hyperscalers were unaware of the physics of operating at scale made them comprehensive – every service added then created switching costs.

Dylan Patel, founder of industry research firm SemiAnalysis, points out that the value in AI infrastructure is no longer just the chip itself. Businesses are also paying for guaranteed access, reliability and the entire framework that surrounds the cloud. For instance, an eight-chip NVIDIA H100 server on Amazon Web Services costs around $98 an hour, illustrating that the real premium lies in the ecosystem built around the hardware, not simply the hardware alone.

CoreWeave started as a cryptocurrency mining operation in 2017. It owned a lot of GPUs and understood how to run them at an operational scale. When demand for AI compute became urgent in late 2022, it was holding the asset everyone suddenly needed. By comparison, hyperscalers are slow; they are pricing for ecosystem value that AI training customers do not need; the same silicon can be offered for significantly less. Companies such as these, with ample high-perfomance computing power at their disposal, would come to be known as neoclouds.

The price gap for their services was not marginal. An equivalent H100 cluster (a purpose-bult system for AI infrastructure and workloads) on a neocloud ran $24 to $34 an hour against $98 on a hyperscaler. Neoclouds were built specifically for AI: high-speed, low-latency networking (known as InfiniBand) rather than Ethernet, bare-metal access without hypervisor overhead, rail-optimised topologies that reduce congestion across multi-GPU training runs.

The most telling validation came from the hyperscalers themselves. Microsoft has spent an estimated $200 million per month on GPU compute through neoclouds despite running its own data centre programme. OpenAI committed to a five-year, $11.9 billion compute agreement with CoreWeave.

Calling something a bubble is lazy analysis. It requires no model, no understanding of the underlying technology and no accountability. The people saying it about AI infrastructure today are those who said it about the cloud in 2009, mobile in 2011, and SaaS in 2015.

Dario Amodei, the CEO of Anthropic, described the company’s revenue to Dwarkesh Patel’s podcast in early 2026: zero to $100 million in 2023, $100 million to $1 billion in 2024, $1 billion to roughly $9-10 billion in 2025, still accelerating into January 2026. SemiAnalysis estimates the coding assistant market alone already generates more than $30 billion in annual recurring revenue and is on a path to $100 billion by the end of this year.

The infrastructure being built is the foundation that supports all of that revenue. Calling the foundation a bubble because you cannot yet see the full building is a failure of sequencing, not analysis.

Grid connection to a new large-scale data centre in the US now takes between four and eight years depending on location. Virginia, the largest data centre market in the world, sits closer to eight. This is the central physical constraint on every expansion plan each company in this industry is facing.

The solution comes from aeroderivative turbines, which are spun off from jet engine technology. They were chosen for data centres for specific reasons: fast startup (the LM2500XPRESS reaches full power in five minutes), high power density, modular design allowing 95 percent factory pre-assembly, and the ability to operate independently of the grid. A single unit produces around 35 megawatts.

Crusoe, an AI data centre provider working with the turbine construction company GE Vernova, ordered 29 LM2500XPRESS units delivering nearly 1 gigawatt to its AI data centres across two tranches in December 2024 and June 2025. Cully Cavness, Crusoe's co-founder and CSO remarked that, “AI's exponential growth demands rapidly deployable power solutions. We're taking the issue of power into our own hands by rapidly building and operating power plants alongside AI data centres.”

The problem is supply. GE Vernova CEO Scott Strazik said in early 2025 the company would be “largely sold out through the end of 2028” for this equipment. Mitsubishi warns new turbine blocks ordered today may not ship until the 2030s.

The reason the turbine bottleneck is so acute is their intricate nature. A single LM2500 unit contains thousands of hand-assembled parts. Turbine blades and combustors have lead times measured in years, while final testing takes weeks. The manufacturing constraints are metallurgical and logistical in ways that capital alone cannot solve rapidly.

Gas turbines are an option to bridge this gap, but every major hyperscaler is now thinking nuclear. Microsoft signed a 20-year, $16 billion power purchase agreement to restart Three Mile Island Unit 1, targeting an operational status of 2028. Google signed the first US corporate SMR fleet deal, committing to 500 megawatts from Kairos Power. Amazon led a $700 million financing round in X-energy for up to 12 small modular reactors. Meta’s agreements total up to 6.6 gigawatts across multiple nuclear developers.

Natural gas and nuclear options provide power, but AI data centres create a new kind of grid stress: massive, inflexible loads drawing at high density around the clock. The Electric Reliability Council of Texas (ERCOT) is forecasting up to 300% load growth across the next six years, driven substantially by data centre development. Texas already had 13.9 gigawatts of battery storage capacity at the end of 2025. [9]

BasePower deploys networked residential battery systems that aggregate into a dispatchable grid resource. In April 2026 it signed a deal with Guadalupe Valley Electric Cooperative for 50 megawatts of capacity across its Texas service territory. The company raised $1 billion in October 2025 to scale beyond Texas. Distributed battery networks can balance grid volatility, including the spikes created by data centre power demand, more flexibly than centralised infrastructure.

The most current difference between companies is those that own physical infrastructure and those that broker access to someone else’s. Owning a data centre means keeping the full economics of every workload that runs through it, including continuous inference revenue once a customer is embedded. Not owning it means taking a margin on a transaction someone else controls. That gap widens as AI shifts from training, which is intermittent, to inference, which is permanent.

CoreWeave's 2025 shareholder letter: $5 billion in annual revenue growing at 168% year over year; a committed contract backlog of $66.8 billion. Nine of the top ten model providers run on CoreWeave infrastructure. Its CEO Michael Intrator revealed much when he remarked: "Working from a clean slate, we established a purpose-built AI cloud with an architectural advantage that is both intentional and durable."

Tomorrow’s winners don’t look at the company that is winning today but where the constraint is currently sitting. This is where the next wave of value concentrates. GPUs, power and cooling are the active bottlenecks in procurement timelines right now.

The constraint is now moving into the silicon layer. The high-performance AWS Trainium3 AI chip launched into production in late 2025. AWS CEO Andy Jassy described Trainium as "already a multibillion-dollar business" that was "nearly fully subscribed" at launch. Other hyperscalers have similar multi-billion dollar commitments.

The reinforcement learning wave, involving computers learning with trial and error, is the demand signal most people have not priced yet. Dario Amodei has said RL shows, "the same kind of scaling pre-training once did." RL requires a different infrastructure architecture than pre-training clusters. Most current GPU clusters were not built for it. The operators who figure out RL-native infrastructure in 2026 will have a structural head start on the next capability wave.

The people closest to the ground are not calling for a pause. Crusoe's COO Michael Gordon is running aeroderivative turbine deployments at a pace the incumbent manufacturers cannot match. Boom Supersonic's CEO pivoted an entire aerospace company because the constraint he saw was more urgent than the one he was originally solving. The hyperscalers are signing 20-year nuclear deals because they have modelled what power demand looks like when inference becomes a continuous workload for a billion users.

Following those signals is not the same as following hype. Hype is the analyst who has not been inside a data centre, or not spotting the signal that Microsoft just signed a 20-year contract on a plant that was decommissioned. The constraint keeps moving down the stack. So does the value.