Why AI Startups Will Disappear Without Compute Markets - Evan Conrad, San Francisco Compute

Compute markets must exist to solve the bundled technical, financial, and credit risks that otherwise force consolidation in AI, preventing startups from acc...

Why AI Startups Will Disappear Without Compute Markets

Abstract

This synthesis examines the economic and technical rationale for physically settled compute markets in artificial intelligence (AI) infrastructure, drawing on the operational history of San Francisco Compute (SF Compute). The central thesis holds that graphics processing unit (GPU) access is constrained not primarily by energy or silicon supply but by a bundle of intertwined technical, financial, and credit risks embedded in current contracting structures. Because AI workloads carry thin margins and high price sensitivity, cloud operators cannot sell capacity on demand and instead require long-term, take-or-pay contracts backed by investment-grade credit - a requirement most startups cannot satisfy. The analysis traces how this financing bottleneck escalates minimum capital requirements for new AI laboratories from $10 million to $1 billion, and argues that a market operator must simultaneously own and operate underlying infrastructure for physical settlement to function. Implications for industry consolidation and the application layer are discussed.

1. Introduction

The prevailing narrative around AI infrastructure scarcity emphasizes physical constraints: power generation, datacenter construction, and advanced packaging capacity. The argument developed here is different in kind. It contends that the binding constraint, at present, is credit risk - the inability of lenders and operators to underwrite counterparties whose business consists of training frontier models with no revenue history and, in some cases, no track record beyond a stated ambition to build "AGI."

Several terms require definition. A compute market is a venue in which rights to reserved computational capacity are traded. Physical settlement means that contracts resolve into delivery of working hardware rather than cash differences, distinguishing it from financially settled derivatives. A take-or-pay contract obliges the buyer to pay for reserved capacity whether or not it is consumed. A neocloud provider (NCP) is a specialized GPU cloud operator distinct from hyperscale incumbents.

The central research question is therefore: why must a compute market exist at all, and what structural properties must it possess to be viable? The analysis proceeds from the origin story of one such market operator (Section 2), through the economics of compute scarcity and the deposit-and-delay problem (Section 3), to the architectural requirement that market operators also operate infrastructure (Section 4). Sections 5 and 6 address consolidation risk and practical implications.

2. Background and Related Work

SF Compute did not originate as a market. It began as Junlark, an AI laboratory attempting to build generative music systems - functionally anticipating products later realized as Suno and UDIO. The founding team discovered that establishing such a laboratory required a $10 million deposit for compute access. Judging this prohibitive, the team raised $500,000 and purchased a smaller cluster outright, which proved uneconomical on a monthly basis. To avoid insolvency, the team subleased unused capacity to third parties, becoming, in their own characterization, "the first GPU broker," complete with handwritten signage advertising H100 and A100 availability.

This trajectory is instructive: the market emerged as a solution to a financing failure experienced directly, not as an abstract financial-engineering exercise. The relevant intellectual context is the comparison to TSMC-style consolidation, invoked as a cautionary model in which a capital-intensive industry collapses into a small number of vertically integrated incumbents with no viable entry path for new firms. This comparison frames the stakes of the analysis that follows.

3. Core Analysis

3.1 Why Capacity Is Structurally Sold Out

Compute markets for AI differ from markets for general-purpose compute such as CPU-based software-as-a-service (SaaS) infrastructure. SaaS buyers operate with high margins, allowing cloud providers to sell capacity on demand: price increases are absorbed without driving customers to insolvency. AI companies, by contrast, operate with thin margins and high price sensitivity. If prices rise, these customers migrate to cheaper alternatives rather than absorb costs, which makes on-demand sales commercially unviable for operators carrying fixed infrastructure costs. The structural response is long-term, take-or-pay contracting, which shifts risk onto the buyer and requires investment-grade credit as a precondition. Most AI startups do not possess such credit, producing what the source material identifies as "the number one bottleneck of AI at the moment... not power."

3.2 The Deposit and Delay Problem

A second mechanism compounds the credit constraint: the sequencing of cluster financing. Lenders typically require a signed customer contract before financing cluster deployment, which makes the point of customer sign-up - prior to any delivery - the moment of maximum risk in the entire operation. This ordering is suboptimal from a systems perspective: the source material argues that building the cluster first and selling capacity afterward would reduce both delay and the probability of failure, since lenders would be underwriting a completed asset rather than an unproven counterparty's promise.

A related technical complication arises once clusters are deployed to new owners lacking liquid secondary markets. Without a mechanism for redundancy across nodes, clusters are more prone to failure for new operators, and failures must be resolved through Return Merchandise Authorization (RMA) processes that reintroduce the same lead-time delays the financing structure was meant to avoid. The absence of liquidity therefore compounds technical risk as well as financial risk, reinforcing the case for a market layer capable of absorbing both.

3.3 Escalating Capital Requirements and Consolidation Risk

The compounding effect of credit-gated access is observable in the rising capital ante required to start an AI laboratory: from $10 million, to $100 million, to - at present - $1 billion. Left unaddressed, this trajectory points toward a TSMC-like industry structure: a small number of clouds, a small number of model laboratories, and no viable path for new entrants. The source material describes a self-reinforcing flywheel: companies with existing credit access can secure tokens, build products, generate revenue, and further improve their creditworthiness, while new entrants are progressively squeezed out. A plausible extrapolation is that starting a new AI company could eventually require purchasing on the order of $100 million in tokens merely to reach competitive parity - a barrier to entry orders of magnitude beyond that faced by software startups in prior computing eras.

4. Technical Insights

Several implementation-relevant findings emerge from this analysis:

• Credit, not power, is the binding constraint. The source material states directly: "It's not power. You hit credit way before you hit AI or way before you hit power." This reframes infrastructure planning priorities away from pure capacity buildout toward financial instrument design.
• Customer sign-up is the highest-risk event in deployment, because it is the point at which lenders commit capital against an unproven counterparty. Restructuring the sequence - building first, contracting second - could reduce systemic risk, though it requires an operator willing to absorb pre-sale capital exposure.
• Liquidity prevents cascading technical failure. Node failure rates become unmanageable absent a liquid market for sourcing replacement capacity, since RMA processes alone reintroduce lead-time delays that a market mechanism could otherwise bypass through reallocation.
• Market operation requires physical operation. Brokering compute without operating the underlying infrastructure does not function as a durable model, since physical settlement demands control over the bill of materials, backplane fabric, bare-metal orchestration, and virtualization layers. The hotel-room analogy used in the source material - one cannot obtain room service without operating the hotel - captures this constraint concisely.
• Standardization is immature. The involvement of Dylan Page, Cloud Native Computing Foundation (CNCF) infrastructure co-chair and former Lambda engineer, reflects an industry judgment that GPU compute lacks sufficient standardization for a market operator to remain divorced from infrastructure operations, unlike more mature commodity markets.

A notable trade-off is that building full physical-settlement capacity is itself disincentivized: an operator capable of running a GPU cloud at that scale faces strong commercial pressure to simply become a GPU cloud and sign large bilateral deals rather than build a market. SF Compute's stated rationale for proceeding regardless was that physical settlement was a necessary prerequisite for eventually returning to its original goal of operating an AI laboratory.

5. Discussion

The analysis suggests that current AI infrastructure scarcity is better understood as a risk-bundling problem than a resource-scarcity problem. Technical risk (node failure, deployment delay), financial risk (capital-intensive cluster construction), and credit risk (counterparty underwriting) are currently bundled into a single contracting instrument - the long-term take-or-pay agreement - which only investment-grade entities can satisfy. A compute market, in this framing, functions as a mechanism for unbundling these risks and distributing them across a broader set of participants with differentiated risk tolerances.

This has implications beyond infrastructure operators. If inference providers continue to sign only investment-grade, multi-year, deposit-heavy contracts, risk does not disappear - it relocates to the application layer, where thinly capitalized startups are least equipped to absorb it. This dynamic connects to broader industry discourse on AI market structure, in which concerns about compute concentration are typically framed around chip supply rather than contractual design. The present analysis indicates that contractual and credit infrastructure may be the more immediate lever.

An open question concerns the generalizability of SF Compute's physically settled model to other hardware classes (e.g., custom silicon, networking fabric) and to international contexts with differing credit and regulatory regimes. Future work could examine whether partial standardization efforts, such as those pursued through CNCF, sufficiently reduce technical heterogeneity to support broader secondary liquidity.

6. Conclusion

This synthesis has argued that the primary obstacle to AI startup formation is not physical compute scarcity but a bundled financing structure that forces long-term, credit-backed contracting onto counterparties who largely cannot qualify. The escalation of minimum capital requirements from $10 million to $1 billion over a short period is presented as empirical evidence of this dynamic, with continuation likely to produce a consolidated, TSMC-like industry structure absent intervention.

The practical takeaway is that physically settled compute markets - which require the operator to simultaneously run a cloud and run a market, while deliberately avoiding becoming solely a cloud - represent one structural response to this bundling problem. For researchers and infrastructure engineers, the implication is that attention to contractual design, credit underwriting mechanisms, and secondary liquidity may be as consequential to AI's future development as advances in hardware or model architecture.

Sources

• Why AI Startups Will Disappear Without Compute Markets - Evan Conrad, San Francisco Compute - Original Creator (YouTube)
• Analysis and summary by Sean Weldon using AI-assisted research tools

About the Author

Sean Weldon is an AI engineer and systems architect specializing in autonomous systems, agentic workflows, and applied machine learning. He builds production AI systems that automate complex business operations.

LinkedIn | Website | GitHub