To Solve Compute Availability, Be a Marriott, Not an Airbnb

The GPU market has a strange property: scarcity and abundance coexist. Here’s why, and what it might take to fix it.

The AI compute shortage has a well kept secret. While startups wait months for GPU allocations and researchers abandon experiments because cloud bills are prohibitive, thousands of H100s sit idle in data centers between training runs — burning power and generating zero revenue for their owners.

Both things are true simultaneously, and that gap is a business.

The demand that isn’t being served

The loudest narrative in AI infrastructure is that there aren’t enough GPUs. That’s partially right. There aren’t enough GPUs at the terms most people can actually access them on.

CoreWeave — the largest neocloud — requires 2–5 year commitments with take-or-pay penalties and upfront deposits of 15–25% of contract value. AWS charges $8+/hr for H100s and routes customers through an enterprise sales process. Neither was built for the researcher who needs 256 GPUs for three weeks, or the seed-stage startup that can’t sign a $10M multi-year contract.

This isn’t a shortage of chips. It’s a distribution problem. The compute exists. The terms under which it’s available exclude the majority of people who need it.

The supply that already exists but doesn’t flow

Idle GPU time is everywhere once you know where to look:

• AI labs run training jobs that last weeks, then sit idle during evaluation, debugging, and internal review before the next run begins
• Companies that over-contracted during the 2023 GPU frenzy are hemorrhaging cash on reservations they can’t fully utilize
• Data centers that built clusters to win anchor tenants are running at 50–60% utilization while their sales pipeline catches up to their capex

A conservative estimate: 20–30% of deployed H100 capacity sits idle at any given moment. At today’s fleet sizes, that’s tens of thousands of GPUs generating no revenue for owners and no value for the builders who need them. The supply exists. It just has no plumbing.

Why the Airbnb model doesn’t work here

The natural instinct is to build a pure marketplace. Aggregate idle supply, connect it to demand, take a spread. Capital-light, scalable, elegant.

This is what Vast.ai tried. Their peer-to-peer marketplace lists GPUs from individual owners — data centers, crypto miners, research institutions, even personal machines. Prices are low: an A100 80GB can be had for under $0.70/hr vs $1.19/hr on a managed platform.

The problem is reliability. A misconfigured InfiniBand fabric, a node that drops mid-job, inconsistent power from a secondary data center — none of this shows up in a listing. It shows up 47 hours into a training run when your checkpoint is corrupted and your experiment is lost. A bad night in an Airbnb is recoverable. A failed training run is not.

Unverified, heterogeneous supply cannot support a single market-clearing price. You end up with a listing site like eBay, not a true exchange. Sophisticated buyers treat it as a sandbox for experiments, not a platform for production. The market fragments rather than deepens — which is exactly the dynamic Hyperbolic identified when their users said they loved the pricing but couldn’t trust it for live workloads.

The capex trap

The opposite instinct is to own the clusters yourself — control quality end to end, build the trust through operations. This is CoreWeave’s playbook, and it works, but the math is brutal at startup scale.

A 1,000 H100 cluster costs roughly $40M in capex — hardware, InfiniBand networking, data center buildout. At $1.96/hr per GPU and 70% utilization, that’s approximately $12M in annual revenue. Three-year payback, before financing costs, power, and staff. CoreWeave solved this by raising $30 billion in GPU-backed debt, with NVIDIA itself as an equity holder. That path is closed for a startup.

Building it yourself means competing against a company with a $30B balance sheet, preferential NVIDIA allocations, and signed contracts with OpenAI and Meta worth tens of billions. You cannot out-CoreWeave CoreWeave on capex. You have to change the game.

The three motions that change the math

The right model is neither pure marketplace nor pure ownership. It’s a franchisor model — controlling quality and distribution without carrying all the capex.

Marriott doesn’t own most of its hotels. It sets the standard, runs the distribution, and collects a management fee while franchisees own the physical assets. Applied to GPU compute, three business motions make this work:

Property management. Take operational control of third-party clusters. The hardware owner spent the capex; you provide the software layer, customer acquisition, SLAs, and 24/7 support. You earn a management fee on assets you don’t own. Zero capex on those clusters, recurring revenue.

Take-or-pay with customer deposits. When a customer commits to a long-term reservation, collect 15–25% upfront and build the capacity against that signed contract. The customer funds the build before hardware is purchased. You’re not speculating on demand — you’re fulfilling pre-sold capacity.

The order book. Let customers who aren’t using their contracted capacity list it back into the marketplace. The same GPU generates revenue twice — once through the primary contract, again through the sublease spread. Effective utilization moves from 70% toward 85%+. That 15-point improvement compresses the payback period significantly and makes the unit economics viable without a CoreWeave-scale balance sheet.

How different players are betting on this

**SF Compute** is executing closest to the Marriott model. They build and operate some clusters themselves to set the quality baseline, manage third-party clusters as a property manager, and built an order book that enables real-time price discovery and subleasing. Live H100 prices on their platform — currently ~$1.96/hr — are set by actual supply meeting actual demand, not a sales team. The fact that the CEO is personally running the deals function tells you that supply acquisition, not product, is the hardest problem right now.

**Hyperbolic** is taking a more decentralized approach — aggregating idle GPUs from data centers, mining farms, and research institutions into a managed pool, with a proprietary orchestration layer (Hyper-dOS) handling load balancing and fault recovery. They’ve undercut AWS by 9x on H100 SXM pricing and earned trust from Hugging Face, Quora, and Stanford. The orchestration layer is their answer to the quality consistency problem — building the Marriott standard in software rather than through direct operational control. They recently raised $20M and are now building tiered enterprise reliability on top of their marketplace, which is a direct acknowledgment that cheap-but-unreliable isn’t enough for production.

**RunPod** sits between the two — a managed platform with a community GPU marketplace layered underneath. Verified Tier 3 and Tier 4 data center partners provide the reliability floor; the community marketplace provides price competitiveness. Their serverless endpoints that scale to zero address the utilization problem from the demand side: you only pay for what you use, which means idle capacity doesn’t sit on the customer’s bill.

Each model is a different answer to the same question: how do you create a trustworthy, standardized unit of compute supply without spending $30 billion to do it?

What the winner looks like

The company that solves GPU idle time won’t win on price alone or on GPU count alone. It will win by creating the most trusted, standardized unit of compute — one that production workloads can rely on, and that cluster owners can plug into without building their own sales team.

That requires seeding with owned or closely managed infrastructure to establish the quality benchmark, onboarding third-party supply against that standard, and operating the market mechanism in between. Capital-light at scale. But not at the start — the seed is what makes everything else credible.

Airbnb works because a spare bedroom is a spare bedroom. A GPU cluster is not a spare bedroom. The interconnect matters. The power redundancy matters. The software stack matters. The support at 2am when your job fails matters.

The GPU market doesn’t need another marketplace. It needs a franchisor — one that sets the standard, earns the trust, and lets others own the assets.

If you found this useful, I study and write about the AI infrastructure and applications. (Research conducted independently. Draft written by the author and edited for clarity with AI assistance.)