GPUs Killed the Cloud

The neo-cloud GPU wave is about to pierce the near-dead IPO market with CoreWeave’s planned debut, targeting a $30B+ valuation on nearly $2B in annual revenue. Meanwhile, Together AI just closed a $305M Series B, and Lambda Labs raised $480M in a Series D—with investors including IQT, the venture arm of the CIA. So why is the CIA investing in GPUs and AI cloud services??What does this tell us about the future of American compute?

Thankfully, AI isn’t just another tech trend, nor are the GPUs beneath them. The rapid adoption of AI is driving one of the most decisive capex cycles in history, with implications that will catch many investors and entrepreneurs off guard.? ? To understand where the industry is heading and how allocators and builders should position themselves, we need to break down the market structure, supply dynamics, and underlying economics shaping the future of AI and GPU infrastructure.?

A better model to understand what happens next? Think less like the cloud and more like the travel industry. Here’s why.

CPU vs. GPU?

GPU and CPU businesses may look similar, but they’re fundamentally different under the hood.

Customers need software and managed services to deploy compute at scale with CPUs. Because CPUs aren’t naturally agnostic and are basically worthless as individual nodes, they rely on centralized orchestration and virtualization to be effective.

GPUs? You need the hardware itself: no virtualization, complex software deployment layers, or cloud required. From its gaming roots, the GPU is inherently valuable by itself. It’s the compute you’re after. Utilizing GPUs is just you, your data, and Jensen.?

Because NVIDIA has prioritized direct access to end customers, it has bypassed the virtualization moat that separated hardware from users in the cloud. The customer has moved down stack toward the physical data center, shifting control and differentiation closer to the machine itself. For the first time in decades, data centers have a SKU they can sell directly to customers, not just colocation and white space.?

Traditional CPU-based data centers are designed to host commoditized compute at scale in centralized locations. Their biggest priorities are fast internet access and a sizeable regional scale for successful commercialization. They focus more on being close to the internet's backbone than on shaving pennies on power, people, or land costs. Traditional data centers are location, location, location, as they need to be close to the internet's backbone.

However, the constraints for AI data centers are almost the exact opposite. GPUs are power-hungry and heat-dense, requiring a much smaller physical footprint with massive energy infrastructure (its own business). Being near internet backbones is secondary—power availability and overall operational expense matter far more.

The CPU cloud infrastructure model doesn’t translate to cheap GPU compute. In fact, it can often be detrimental to high-performance AI clusters. The AI cloud isn’t just an extension of traditional computing; it’s an entirely different design problem.

Public cloud’s genesis was vastly different. It wasn’t built for customers but as a dual-use product. Google, Amazon, and Microsoft built it for themselves first, then sold it as a service. Today’s neo-clouds have entirely different origins. Most started as hosting services or data center developers from crypto, mastering cheap land and cheap power, not hyperscale cloud lock-in SaaS.?

This distinction matters. Public clouds outsource GPU deployments because their infrastructure was never designed for distributed, high-density deployments that prioritize opex efficiency over connectivity. Similarly, the public CPU cloud model was not intended for smaller, more expensive high-performance clusters with limited upper-stack software requirements and a shrinking window between hardware updates.

CPU clouds need scale to be useful and a software moat to retain value. A single CPU does nothing on its own: virtualization, orchestration, and managed services make them viable. GPUs are the opposite. A single GPU is immediately valuable—just power, cooling, and connectivity, with no virtualization required. GPU compute itself is the end commodity for GPUs, whereas cloud is the end commodity for CPUs. Owning one GPU is like owning your own car. Owning one CPU is like owning a lug nut.?

This shifts the business model down the stack. CPU clouds monetize by going up the stack—SaaS, managed services, and software lock-in. GPU infrastructure wins by going down the stack, optimizing hardware and data center infrastructure. Margins come from capex and opex efficiency, not software dependencies.

GPUs Give You Wings?

CPUs deploy software services, while GPUs deploy compute. This simple distinction leads to fundamentally different economics, value chains, and investment opportunities.

The commercial airline industry offers a sound framework to understand what happens next. Like AI infrastructure, air travel depends on a web of manufacturers, operators, physical hubs, logistics firms, and distributors, all optimizing capacity and efficiency in a highly competitive market.?

Manufacturers – The Platform?

These companies physically create the hardware that delivers the end service. Just as airlines buy or lease aircraft, cloud providers and data centers must secure thousands of GPUs from manufacturers to keep up with demand.??

Think of NVIDIA as Boeing or AMD as Airbus.?

Airports – The Infrastructure?

Data centers, like airports, are where the product physically operates. Integration can vary across carriers, with some GPU companies owning the physical data centers where chips are colocated and others only renting space in them. AI carriers pursue greater ownership of, or investment in, these physical hubs (data centers) than do airlines (airports).?

Think of Digital Reality or Compass as JFK or DEN.?

Carriers - Vertical Providers?

These firms purchase or lease and manage the assets, managing cyclical backlogs, lead times, and pricing power driven by the OEMs. Matching the right “plane” to the right “route” is a core operational challenge. Operating costs (fuel, maintenance, crew for airlines; power, cooling, real estate for data centers) are significant and primary points for optimization. These businesses are characterized by demand inelasticity, heavy competition, and margin compression. The CEO of Microsoft likes to call his GPUs “fleet” just like an airline CEO would.?

CoreWeave, Lambda, or Crusoe would be parallels for United, American, and Delta Airlines.

Packagers - Customer UX?

A derivative ecosystem of services, packagers, and distributors bundles, packages, and distributes the product to monetize excess capacity and serve end consumer-specific use cases where carriers cannot on their own.?

Think of marketplaces, resellers, or inference-as-a-service as Kayak, Expedia, or Hotwire.

Here is a graphic to help explain the framework.

What Comes Next??

One of the most significant shifts in AI compute will be the proliferation of specialized carriers. Just like the airline industry has full-service international carriers, low-cost operators, and regional flights, GPU providers will fragment into different tiers. Large hyperscalers will act as global carriers, while specialized HPC providers emerge to meet local compliance needs and serve niche workloads. The market’s unconstrained total addressable demand will continue to attract aggressive capital, but consolidation through mergers and acquisitions will be inevitable as margins compress.?

Governments are also stepping in, and state-backed AI infrastructure is becoming the norm. Expect national and regional data centers to receive increasing support through tax incentives, grants, and direct funding as countries realize AI is a strategic asset. Nations that want to ensure long-term sovereignty over their intelligence capabilities will prioritize domestic GPU clusters to avoid reliance on foreign providers. This trend will continue attracting alternative asset managers to fund the physical buildout, while hyperscalers like Oracle position themselves for public-private partnerships. However, with tight margins and increasing infrastructure costs, taxpayers could ultimately bear the financial burden—just as they have in the airline industry.

As AI workloads become more complex, fleet management will be critical for compute efficiency. GPUs differ widely in memory, power, and compute intensity, forcing providers to actively manage “fleets” of hardware just as airlines optimize their aircraft schedules and routes. Many providers will outsource this function, allowing specialized distribution and optimization firms to capture significant margin as infrastructure operators compete for capex and opex efficiency.?

Even with soaring demand, AI compute will not be immune to margin compression. The growing number of data center operators and specialized HPC startups will push pricing downward, making operational efficiency and resource optimization more important than ever. This will increasingly emphasize fundamental input costs such as power, land, and cooling. The pressure on local power grids has already revived interest in nuclear energy, with companies like Oklo Inc ($OKLO) and NuScale Power Corp ($SMR) seeing renewed investor interest.?

Finally, the integration of hardware, platforms, and software will define long-term winners. Just as Boeing and Airbus don’t simply sell planes but collaborate with airlines on design, maintenance, and operational support, NVIDIA, AMD, and other AI chipmakers are embedding themselves deeper into software ecosystems like CUDA and specialized AI frameworks. At the same time, cloud providers are tightening their relationships with chipmakers, securing supply and co-developing new solutions. As the AI compute market matures, a company’s ecosystem strength and ability to create lock-in will be key indicators of its long-term staying power.?

The AI compute industry is evolving rapidly, and those who understand where margin and defensibility will shift are best positioned to capitalize on this massive infrastructure wave.

GPUs Everywhere?

This shift is creating massive investment opportunities and structural challenges in our vision of the future. Governments now treat AI as a national security priority, leading to sovereign GPU clusters and publicly backed infrastructure projects. Meanwhile, margin compression is inevitable, forcing operators to optimize power, land, and supply chain costs to stay competitive. If you take away one general rule for the AI infrastructure to thrive in this build-out to win the future, either go horizontal or down stack.

AI infrastructure, such as oil, telecom, and energy grids, is now a strategic asset. These are not the industries countries outsource. The winners won’t be those trying to replicate the CPU cloud model. They’ll be the ones who embrace the new economic realities of AI compute. The companies that control hardware, infrastructure, and efficient distribution will own the future.?

Investors and entrepreneurs who recognize this shift early stand to gain the most. This isn’t just another tech wave. It is about building a new type of factory, an AI factory. It is the new backbone of global intelligence and next-generation industrial power.