The GPU crisis won't be solved by building more data centers

Microsoft, Amazon, Google, Meta, and Oracle have collectively committed over $600 billion in capital expenditure for 2026, with roughly 75% of it aimed directly at AI infrastructure. The numbers sound either heroic or unreal, depending on your view of hyperscalers expenditures on the AI market. 

Whatever observers and critics might think, more spending is coming. Goldman Sachs projects total hyperscaler spending from 2025 through 2027 will hit $1.15 trillion, more than double everything spent in the three prior years combined. And yet, if you're an AI startup trying to get your hands on an H100 today, your lead time is 36 to 52 weeks. This assumes you manage to get any allocation at all.

With demand supposedly outstripping the supply, one might assume that there is something to the hyperscaler conventional wisdom of answering this gap with more build. More data centers, more power plants, more rack space, more billions. But that framing misdiagnoses the problem. The GPU crisis won’t be solved with more construction. What it really needs is less concentration, with the same five companies building more data centers. We could, quite comfortably, end the blog post right here. But we won’t. Let’s unwrap it a bit more.

What the GPU crisis actually is

Let’s get straight into the technical stuff. TSMC's CEO publicly stated that CoWoS packaging capacity — the advanced process that bonds high-bandwidth memory onto GPU substrates — is "sold out through 2025 and into 2026." HBM memory, which is produced by only three companies (SK Hynix, Samsung, Micron), has seen prices rise over 600% in some market segments. Zooming in, NVIDIA alone is expected to consume roughly 60% of TSMC's expanded CoWoS capacity.

So, if you’re a company who placed an order years ago, you’re probably good. If you’re not, well, you’re going to feel that shortage. 

Microsoft, Google, Meta, and Amazon all locked in multi-billion-dollar forward contracts for Blackwell GPUs (GB200, B200) in the last several years; which means they will be consuming most of NVIDIA's allocable capacity through end-2026 and into 2027. 

What we have now is a two-tier compute market: hyperscalers enjoying guaranteed supply, and everyone else fighting for scraps at spot prices. No surprise there. While everyone was dazzled by AI image generation and auto-generated content, hyperscalers were intentionally building this tiered market. 

We know who benefits. But who struggles in this arrangement? AI startups and LLM research labs doing all of the novel work the AI industry and, indeed, other industries need. In short, all of the new companies and research teams who are most dependent on elastic, on-demand compute are basically locked out.

Why building more data centers doesn't solve this

The data center buildout is already historic in scale. BloombergNEF tracked 23.1 gigawatts of IT capacity under active construction globally at end-2025, across 831 sites. U.S. data center construction spending hit a monthly rate of $45.1 billion by December 2025. That’s an 85% increase in just two years.

And yet, it still isn't enough. Here's why building more won't fix the underlying problem:

1. Construction timelines offer no relief

Industry analysts project that 30 to 50% of planned 2026 data center capacity will slip to 2028 or later, due to power grid interconnection queues and construction bottlenecks. The U.S. alone has over 1,500 GW of projects in interconnection queues. What this means in practical terms is that a new large load connected to the grid now takes four to seven years to go live.

Intel's CEO put it plainly and bluntly: "There's no relief until 2028."

New data centers announced today don't produce a single GPU-hour until 2027 at the earliest. The companies that need compute now cannot wait for shovels to hit dirt.

2. New construction concentrates in the same hands

Hyperscalers are projected to hold approximately 60% of total global data center capacity by 2030, up from 44% today. Meanwhile, neoclouds (CoreWeave, Lambda, etc.) that hold the remainder mostly operate on five-year hyperscaler offtake contracts, effectively extending the reach of the same incumbents one layer down.

More data center construction doesn't create new competition. It creates more infrastructure controlled by the same actors, financed by their balance sheets, locked up in their multi-year contracts, and priced according to their market power. 

3. Electrical power and natural resource monopolization

The hyperscaler buildout is very visibly consuming the energy infrastructure the rest of the economy runs on. Let’s look at the numbers. 

As we already noted, the four largest hyperscalers committed to spending approximately $650 billion on AI infrastructure in 2025 and 2026. To secure power for that infrastructure outside congested public grids, they are partnering with energy giants to build 50 gigawatts of behind-the-meter natural gas generation—aka, private power plants that bypass public interconnection queues entirely. Microsoft alone has contracted over 40 gigawatts of clean energy, making it the world's largest single corporate power buyer.

Power Purchase Agreement prices rose 35% in 2024 as hyperscalers absorbed most of the global supply. In Northern Virginia, the world's largest data center hub, grid interconnection capacity is functionally exhausted. Maine recently passed a data center moratorium in 2026, and many other states are considering similar pauses, while Illinois suspended its data center tax incentives. The AI data center political backlash is intensifying because communities are watching electricity price increases, the grid strain, and water consumption climb while the beneficiaries of that consumption remain a handful of corporations.

Again, this a structural consequence of concentrating the world's compute needs into a few hundred facilities owned by five companies.

4. Prices stay high because competition remains low

We can see the effects of this concentration problem in the AI pricing. AWS charges $4.99 to $6.98 per hour for H100 compute, while Google Cloud and Azure are in the same band. The kicker is that even with the construction boom the prices have not fallen because the companies doing the building are also the companies setting the prices. No real free market competition exists in the race to build AI data centers that would actually drive prices down. 

There is one development, however, that is driving prices down by creating a free market alternative: the emergence of distributed, aggregated compute networks that pool underutilized capacity. These decentralized GPU marketplaces can deliver H100 access at $1.49 to $2.20 per hour, compared to AWS's $6.98.

When suppliers compete across a marketplace, prices approach cost. When five companies control supply and all price similarly, prices approach what the market will bear

GPU compute that already exists and sits idle

We already know that AI data center build outs aren’t helping AI startups and LLM research labs. It gets even more absurd when you consider enterprise GPU utilization rates.

Industry surveys show that GPU clusters at major organizations typically run at 55 to 65% utilization. That leaves a whole lot of time where GPUs are doing absolutely nothing. Research suggests that roughly 95% of enterprise GPUs sit idle at any given moment when measured across the full install base.

Globally, the installed base of high-performance GPUs is pretty formidable. There is gaming hardware sitting in data centers after the Ethereum proof-of-stake transition, research institution clusters running jobs intermittently, enterprise servers that overpurchased to buffer against the shortage and are now underutilized, and also independent operators with capacity they cannot efficiently monetize on their own.

That capacity exists today, and there are projects who could use it. This idle GPU capacity can be accessed through decentralized compute networks that aggregate heterogeneous supply into a marketplace.

VOLT's network, for example, makes thousands of GPUs available from over 5,000 independent providers across 130+ countries, including idle hardware. And if you opt for VOLT, you never sit on a waitlist, pay hyperscaler rates, and get taxed to move your data with egress fees.

Who gets hurt by the status quo in compute

AI startups are the most immediately harmed. Their development cycles require burst compute; that is, high-intensity GPU access for training runs that might last days before going quiet for a bit. The hyperscaler model prices burst compute at a premium and imposes quota systems that make access unpredictable. AI startups and LLM research teams could purchase GPU hardware outright, but that requires capital that most startups don't have and lead times that do them no favors.

Researchers, academics, and other organizations doing the most open, exploratory AI work that tends to produce foundational advances have the least access to the infrastructure that makes it possible. A research lab quite simply cannot commit to a five-year hyperscaler offtake contract. It cannot absorb $6.98/hr H100 pricing across a meaningful cluster for months at a time. Academic AI research therefore becomes increasingly dependent on cloud credits programs controlled by the same hyperscalers, which can create a dependency that shapes what gets researched.

Developers in emerging markets are excluded almost entirely. The compute available at hyperscaler pricing is inaccessible at local salary and funding levels for most of the world. The communities with the most to gain from accessible AI infrastructure (developers in Nairobi, São Paulo, Jakarta) are priced out by an infrastructure model designed around the budgets of Silicon Valley companies and Fortune 500 enterprises.

Small and mid-sized enterprises face a different version of the same problem: multi-month waitlists, minimum commitment requirements, and pricing that makes sustained GPU usage economically marginal for workloads that aren't directly revenue-generating.

None of these problems are solved by adding another hyperscaler data center campus. They are solved by adding supply to the market in a way that breaks the pricing power of the incumbents.

Getting GPU capacity now, not later

If companies building AI products are going to overcome NVIDIA Blackwell hardware supply constraints and AI data center construction buildouts, they have to work with hardware that already exists. What exactly does this mean? 

Distributed compute aggregation

Pooling the idle capacity in the existing install base into addressable supply bypasses new fabrications, data centers, and grid connections. All it takes is software, coordination, and a marketplace. VOLT, Aethir, and similar networks have already demonstrated this is technically feasible at scale.

Consumer and prosumer hardware

RTX 4090s at $0.28/hr deliver performance adequate for a wide range of inference and fine-tuning workloads. AWS and Azure don't offer consumer GPU access but distributed networks do. With hyperscalers, you’ll pay $4.55/hr for compute instead of $0.28/hr on a decentralized GPU marketplace. 

Competitive pricing through market liquidity

The decentralized marketplace model creates price competition among providers. This type of competition does not exist inside any hyperscaler's walled garden. 

None of this requires waiting for 2028.

GPUs are a fundamental resource 

The GPU crisis is, at its core, a story about what happens when critical infrastructure becomes monopolized by a few hyperscalers. 

GPUs are a fundamental resource in AI development. Think of compute as the equivalent of electricity for the industrial economy or bandwidth for the internet economy. When just five companies control access to that foundational resource with aligned pricing interests, the downstream effects are profound: innovation is constrained in ways that simply cannot be addressed by those same five companies building more of the same thing.

Granted, the $1.15 trillion hyperscaler buildout will produce enormous amounts of GPU capacity. Some portion of that capacity will eventually become available to the market at competitive rates. But the timeline is probably long, and in the interim concentration will only deepen, with no guarantees that hyperscaler pricing incentives will align with making compute widely accessible.

So, the short-term answer is not to build more data centers. It is to use what already exists, and do so more efficiently, more accessibly, and through market structures that distribute the benefit of that capacity rather than concentrating it.

The compute is already there. Now we just have to get it to those who need, and with fair pricing and terms.

VOLT is a decentralized GPU cloud that aggregates underutilized compute capacity from 5,000+ independent providers into a unified marketplace. Access H100, A100, and RTX 4090 GPUs at 50–70% below hyperscaler pricing with no waitlists. Get started at VOLT →