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.)

Under the hood of the largest trend in e-commerce marketplaces..

There are over 2 million 3P sellers on Amazon. These 2mn sellers drive ~ $300bn in GMV for Amazon.

As per the recently published 2021 Amazon Small Business Empowerment Report, number of sellers on Amazon grew by 45% in 2020.

53% of all search intended towards product purchase starts on Amazon.

These figures together describe the Amazon marketplace flywheel - one where the supplier ecosystem is vast, growing at an unprecedented clip and demand continues to stay strong (even as low code/ no-code tools, Shopify or similar SaaS tooling and associated app ecosystems for marketing, returns and payments have made selling Direct-To-Consumer (non-marketplace) faster, cheaper and more accessible for small and medium businesses). More sellers and more buyers create a network effect which helps build the Amazon moat, allowing it to curate a lower price offering with cheaper and faster shipping, which makes the value proposition increasingly more compelling for new buyers and new sellers.

➡️ First things first, what are Amazon aggregators and what are they chasing?

Amazon aggregators are going after the universe of third-party sellers on Amazon. To back up a bit, Amazon started out selling books in 1994, and gradually realized it could sell much more. In 2000, they opened up their platform to third-party sellers, which are small and medium sized businesses that sell on Amazon's marketplace. These businesses take advantage of the incomparable footfall and fulfilment architecture, for a given fee that's dependent largely on their product category, size and weight of the product they are selling.

Sellers can exist on Amazon as 1P vendors or 3P sellers. 1P sellers sell to Amazon at wholesale prices and let Amazon decide pricing, fulfilment, listing & inventory management and so on (these products are typically identifiable as 'Sold by Amazon.com' in the Buy Box). On the other hand, 3P sellers are essentially using Amazon real estate to sell via their storefronts, and they are responsible for pricing, inventory management (ensuring product is in stock) and the rest. 3P sellers make up ~54% of sellers on Amazon. ~92% of Amazon sellers utilize FBA (Fulfill by Amazon - where they pay anywhere between 20% - 30% to Amazon to pick, pack and ship the products to end customer).

Aggregators are going after this 3P seller ecosystem (which has higher margins than a typical 1P business), with the belief that certain assets can be optimized through better operations, branding and other levers. They also understand that an individual seller, making ~$200k annual revenue on her Amazon store on average, is undercapitalized, and has all her cash locked up in inventory, posing limitations on growth potential. By acquiring many profitable assets with a mix of debt and equity, these aggregators are essentially purchasing Amazon real estate and product ASINs (seller identification codes associated with Amazon SKUs) and betting that they can pattern match and consolidate over a large portfolio to juice operational efficiencies and revenue growth. This is unpacked in more detail later in the essay.

❓ Sure, but...why does it make sense to buy an Amazon 3P business in the first place?

To simplify the question, I’ll first tackle the Amazon / marketplace portion of this trend, and then how the aggregation comes about.

Amazon is not simply a marketplace, it is a thriving entrepreneurial ecosystem. What McDonalds did to creating millionaires back in the day, Amazon is doing in this current environment (if there was any doubt about this analogy, Foundry Brands hired ex-Pizza Hut executive to bring her experience of working with “franchises and entrepreneurs” to Amazon sellers). Some of the savviest entrepreneurs are building products for customers, which are tested and iterated quickly through Amazon's algorithm, finding the product market fit (and iterating quickly, if not) for the first leg of its revenue journey.

There is rich data sitting behind each of these businesses, which can be harnessed to meticulously reduce TACOS (total advertising cost of sale), improve in-stock rate and Buy Box %. It is tough to imagine an ecosystem where all these factors dovetail perfectly, and this is one of the reasons why aggregators are going after Amazon businesses.

And then there's the fact that an average Amazon 3P seller operates profitably.

For a $100 ASP product, an average seller pays

• $25 in COGS (cost of the product plus ocean freight and any associated duties & tariffs)
• $40 in fees to Amazon (Fulfillment Fees of ~25% + Referral/Selling Fees of ~15%)
• and $15 in overheads (marketing, headcount costs and such)

which implies an EBITDA of ~20%. Ofcourse, these costs vary by product category. By using say 70% debt and 30% equity to buy a business at say 4x - 8x EBITDA (or SDE - Seller Discretionary Earnings), aggregators lever a business at say 4-6x EBITDA.

An average ecommerce business has a conversion of 2-3%, while Amazon improves conversion significantly (typically 10%+). By specializing in and doubling down on this channel, high performing Amazon products begin to see customer recognition, and use it to build an omnichannel play, where they control the brand and customer relationship more actively. What could at a first glance appear to be a marketplace for undifferentiated products, begins to look like a launchpad for new brands as you take a look under the hood. Historically, these Amazon businesses were valued at low multiple due to lack of differentiation / moat (franchise-like nature), but as these assets have been discovered, multiples have been climbing up.

The year 2021 has seen public exits of Shopify based brands (Oatly, Allbirds, FIGS, Olaplex, VitaCoco, AKA Brands among others) while large exits of Amazon native assets have lagged behind. The pace of VC funding into Amazon native assets has not matched that of DTC ecosystem because of lack of a clear moat in the Amazon ecosystem. However, discovery of high growth profitable assets by aggregators as well as public debuts of a few breakout Amazon brands such as Anker (started natively on Amazon in 2011 , now sells in Best Buy, Walmart and other retailers and went public in 2021) led VCs to sit up and take notice of Amazon SMBs as an asset class.

📢 How do aggregators really add value to an acquired business?

The benefits of an aggregator model are theoretically simple to understand: all aggregators are leveraging what is called a scale (volume play) and scope (synergistic cross-functional play) arbitrage. By aggregating vast amounts of capital at lower than market costs, aggregators can buy brands at a rapid clip.

Buying companies allows for faster scale and scaling companies rapidly implies that network partners (warehouses, suppliers) begin to offer better payment terms allowing better working capital management, volume discounts stemming from larger purchase orders potentially preventing out of stock situations as compared to a small/medium sized brand. While capital is a commodity, putting capital to use smartly through these curated relationships and an effective capital allocation strategy creates space for differentiation in what appears at first to be a crowded aggregator space with many well-funded (>$12 bn in capital raised across 80+ players with a long tail) players.

Capital also builds a warchest towards hiring talent. The pace of acquisitions implies all functions need to scale to accommodate needs, and as a result, this model imposes that you must acquire talent rapidly, often at a pace faster than revenue growth, before economies of scale and scope truly begin to kick in. The first signs are process automations seen in customer experience, followed by larger technology investments that accelerate operations allowing hiring to stabilize. These changes then slowly begins to trickle through the organization, stabilizing the pace of G&A expense growth vis-à-vis revenue growth.

A lesser appreciated arbitrage is simply taking a brand, turbocharging it with carefully curated and learnt set of digital tools and expertise in brand marketing and managing marketplace effectively. Appropriately these tools have been called easy to use but hard to master. While many billions of dollars in market cap has been created through tools that enable success of a DTC business by entrepreneur/developer community (headless ecommerce solutions, low code/ no code platforms, returns management, payments solutions), marketplace enablement ecosystem remains underinvested (with few tools such as Jungle Scout, Amazon Seller University, Helium 10), creating space for aggregators to build out tools and associated IPs. Doubling down on specific tools and solutions becomes another differentiator between aggregators.

While it remains to be seen what exactly these "tools" will look like, aggregators are picking different focus areas for technology & operating capabilities to build inhouse vs buy off the shelf. Key tech acquisitions by aggregators include ReviewMeta (which helps verify the veracity of reviews) by SellerX and Yardline (growth capital and business optimization tools provider) by Thrasio. Perch purchased WebDeals Direct to acquire their portfolio of brands and their own warehousing facility.

There is also a conscious move towards verticalization in many cases. Pattern Brands has a clear focus on home goods (Shopify brands focus), Join Brands launched with a focus on celebration space in India. This is another strategy to narrow down acquisition focus while building niche sector focused capabilities allowing synergies to show up faster (supplier consolidation, deeper understanding of regulatory requirements, nuanced understanding of end-customer habits in a product niche, faster GTM and product market fit discovery for launched brands and so on).

This is just the tip of the iceberg. Ability to acquire and grow brands by understanding buying patterns and handling complex supply chains in a well-oiled manner builds capabilities needed to launch and scale own brands (incubator model). P&G spends billions of dollars annually in their (product) innovation machine. In older days, one could imagine researchers standing in retail stores observing consumer behavior. The equivalent of that in a digital marketplace is the ability to harness multiple data sources to understand consumer lifecycle in the digital world, and push those insights into an innovation machine, a platform to build successful brands. Note here that at the heart of innovation is not R&D dollars towards building generation defining products, but a channel innovation - essentially understanding customer buying patterns in an attention starved world and brining them the most suited products for their (searched) needs most conveniently. This brings me to my next (and the core) idea.

📢 Well, okay, but how do aggregators decide which Amazon brands are more valuable?

Historically, products on Amazon have been perceived as commodity, replicable and low price merchandize, not as "brands". Unsurprisingly, 78% of search of Amazon is generic (needs-focused, such as "standing desk" or "side table lamp") as opposed to branded (brand first search such as "running shoes Adidas").

It begs the question -

How to pick and grow winning brands in a sea of generic brands servicing a needs-focused customer?

A brand is not an Amazon brand or a Shopify brand. Those are channels - ways to get the brand to the customer. This tends to be forgotten in today's narrow discussion of where brands should be built. Great brands are fostered by a combination of a differentiated product, consumer centric content and programs (delivery, warranty, returns experience) that engender trust. While certain channels may allow one aspect to be serviced or delivered better than another, ultimately, brand building is not restricted to a channel per se.

Much has been written about marketplaces as a discovery mechanism for brands. Indeed, many brands treat 'marketplace referral fees' as a 'marketing expense'. If marketplaces are showrooms, then reviews, organic rank and Buy Box Wins are the 'touch & feel'. However, over time, these have become good to have characterizations and not necessarily the moat that guarantees winning in the long run.

Marketplaces in the digital world allow access to vast amounts of data, which if channeled well can build a differentiated brand strategy. As a simple example, any business runs on signals coming from customer behavior. If continuous advertisement expenditure does not lead to conversion or improvement in organic rank, the platform should automatically throw an insight into the next best product or brand where this spend should be allocated. In my view, this canonical data driven insight model will help build a differentiated platform of brands.

Success of a brand in an omnichannel world is determined by its ability to seamlessly straddle channel boundaries to meet customers where they are. This is impossible in a world where data is broken and customer acquisition costs are through the roof. A successful brand on any platform effectively uses customer signals to feed into a continuum of brand presence, channel notwithstanding.

Building stand-out products can be a tall ask in a sea of copy-cat products. There are, however, solutions in sight that differentiate marketplace native brands.

Product

• Hyper Specialization: Andrei Hagiu and Julian Wright at Platform Chronicles elaborate how platforms reward *hyper-specialization* through a self-reinforcing cycle. The more closely a product matches what a customer is looking for (through specialized products or a narrow category focus), the more it comes up in search results, better reviews it garners (solidifying its review moat) and over time, becomes a clear algorithm winner for target customers. This is a powerful way to start building brand recognition.

Marketplace strategy

• Branded Storefronts: For years, Amazon emphasized the product over the brand by listing all products in a list like fashion. However, Alibaba's Taobao allows sellers to create their branded 'boutiques' with custom templates and personalized customer service. Amazon Stores took inspiration and now allows creation of customized stores to showcase a curated collection of products to elevate a brand. While in Amazon universe, Stores liken the experience to shopping on a brand's website. Any brand registered on Amazon Brand Registry can now set up its own page, significantly improving customer experience.

Owning the customer relationship

• Personalized customer service builds trust: While Amazon controls the customer information, payment information and other preferences data, the brand itself is responsible for customer service. Great customer service translates into customer reviews, which kicks off the flywheel of Reliable service → customer trust → good customer reviews → brand recognition → compounding customer trust and so on..

Above strategies begin to lead to branded keyword searches which begin to create the elusive ‘brand moat’ on a marketplace platform.

After a year of massive fundraises and triple-digit acquisitions, what remains to be seen is proof of acquisition synergies, success of the incubation play and emergence of breakout marketplace native brands.