OpenRelay

An Inference Delivery Network (IDN) for routing AI inference workloads across heterogeneous accelerators.

Cover Block

Public sources

Field	Value
Name	OpenRelay
Tagline	An Inference Delivery Network (IDN) for routing AI inference workloads across heterogeneous accelerators [OpenRelay, November 2026]
Headquarters	San Francisco, United States [Y Combinator, August 2026]
Founded	2026 [Uncork Capital, November 2026]
Stage	Seed [OpenRelay, November 2026]
Business Model	API / Developer Platform [OpenRelay, retrieved 2026]
Industry	Deeptech
Technology	AI / Machine Learning [Y Combinator, August 2026]
Geography	Global / Remote-First [Y Combinator, August 2026]
Growth Profile	Venture Scale
Founding Team	Co-Founders (2): Jaden Wang and Prashant Patel [Y Combinator, November 2026] [Y Combinator, August 2026]
Funding Label	Seed
Total Disclosed	~$10,000,000 [OpenRelay, November 2026]

Executive Summary

PUBLIC OpenRelay is building an inference delivery network that routes AI workloads across heterogeneous accelerators, and it merits investor attention now because it paired that thesis with a newly announced $10 million seed round in November 2026 and unusually aggressive early traction claims for a company started months earlier [OpenRelay, November 2026] [Uncork Capital, November 2026]. According to Uncork Capital, founders Jaden Wang and Prashant Patel started the company in June 2026 and joined Y Combinator's Summer 2026 batch, giving the business a compressed but legible formation story around inference infrastructure rather than a broader cloud platform pitch [Uncork Capital, November 2026] [Y Combinator, August 2026].

The product is a distributed compute network connected by proprietary orchestration software that schedules workloads, benchmarks hardware, and presents a single production endpoint across NVIDIA, TPU, Trainium, and AMD environments, with OpenAI- and Anthropic-compatible interfaces in its documentation [OpenRelay, November 2026] [Y Combinator, August 2026] [OpenRelay API Documentation]. The differentiation, at least on the public record, rests on hardware-agnostic routing and commercial flexibility: OpenRelay says customers can reserve throughput at a flat monthly rate, buy dedicated GPU VMs by the hour, and access discounted async and batch inference modes, while the company and YC materials claim cost savings of up to 20% and faster autoscaling relative to other providers [OpenRelay, retrieved 2026] [Y Combinator, November 2026].

The team's technical credibility appears directionally relevant, though much of the detail comes from secondary profiles rather than primary biographies: public materials identify Wang as a co-founder, while external LinkedIn-sourced references say he previously built a half-megawatt data center and virtualization infrastructure at Voltage Park, and that Patel worked on accelerator infrastructure at AWS, including Trainium kernels [Y Combinator, November 2026] [LinkedIn, retrieved 2026] [LinkedIn, retrieved 2026]. On capitalization, the public record is simple: OpenRelay announced a $10 million seed led by Uncork Capital with participation from Eniac Ventures and Standard Capital, and the business model reads as API and infrastructure revenue tied to production inference usage and reserved capacity rather than a services-heavy motion [OpenRelay, November 2026] [Uncork Capital, November 2026] [OpenRelay, retrieved 2026].

What matters over the next 12 to 18 months is less the headline vision than the degree to which company-reported scale converts into independently visible customer adoption, repeatable unit economics, and durable network supply across sites and accelerator types. OpenRelay says it had exceeded 20 compute sites, processed more than 5 trillion tokens per month, and reached more than $10 million in annualized revenue run rate within five months of founding, but those figures remain company-reported, with no named customers or operating partners disclosed in the cited public materials [OpenRelay, November 2026].

Company-stated, unverified -- This section relies materially on company, investor, YC, and documentation sources, with limited independent corroboration for traction and team-background claims.

Taxonomy Snapshot

Axis	Value
Stage	Seed
Business Model	API / Developer Platform
Industry / Vertical	Deeptech
Technology Type	AI / Machine Learning
Geography	Global / Remote-First
Growth Profile	Venture Scale
Founding Team	Co-Founders (2)
Funding	Seed, total disclosed ~$10,000,000

How the Company Got Here

PUBLIC OpenRelay presents itself as an AI infrastructure company building an "Inference Delivery Network," or IDN, a routing layer for model inference workloads across different accelerator types [OpenRelay, retrieved 2026]. The public record places the company in San Francisco and identifies June 2026 as the start of operations, with the founders launching into Y Combinator's Summer 2026 batch shortly thereafter [Crunchbase] [OpenRelay, June 2026]. Crunchbase lists the company as OpenRelay Incorporated, which is the clearest public indication of the operating entity name in the source set [Crunchbase].

The chronology is short but unusually compressed. OpenRelay's production environment went live on openrelay.inc on June 8, 2026, according to the company's own site, establishing a concrete public milestone close to formation [OpenRelay, June 2026]. By November 2026, the company announced a $10 million seed round led by Uncork Capital and described its network as a distributed layer of compute infrastructure coordinated by proprietary orchestration software [OpenRelay, November 2026].

A small but relevant naming caveat sits alongside that timeline. Crunchbase also maintains a separate profile for a different business now known as Rivet.cloud that was formerly called OpenRelay, a Web3 infrastructure company founded years earlier, which means references to "OpenRelay" in older databases can point to an unrelated entity rather than this 2026 AI infrastructure startup [Crunchbase]. That distinction matters for diligence because it reduces the risk of overstating company age, traction history, or brand continuity based on mismatched records.

Lightly corroborated -- Confirmed with company website and Crunchbase, but several timeline details remain company-sourced.

Product and Technology

Sources and analysis

OpenRelay is pitching a narrow but timely piece of infrastructure: a routing layer for AI inference that sits above heterogeneous accelerator supply and presents developers with a single production endpoint [OpenRelay, November 2026] [Y Combinator, August 2026]. In the company’s description, the network is a distributed pool of compute infrastructure connected by proprietary orchestration software that schedules workloads, benchmarks hardware continuously, and routes requests based on latency, cost, availability, and SLA requirements rather than tying customers to one chip family or one cloud [OpenRelay, November 2026] [OpenRelay, retrieved 2026]. Public materials say the platform supports NVIDIA, TPU, Trainium, and AMD accelerators, and YC’s launch page frames the product as an "Inference Delivery Network," effectively a delivery plane for model inference across multiple capacity sources [Y Combinator, August 2026] [Y Combinator, November 2026].

The product surface appears to have two layers. For inference buyers, OpenRelay says it serves open-weight models behind an OpenAI-compatible API, also speaking Anthropic wire formats, with pricing that includes per-token billing, reserved inference at a flat monthly rate, async inference at a claimed 20% to 50% discount, and batch inference at a claimed 50% discount [OpenRelay API Documentation, retrieved 2026] [OpenRelay, retrieved 2026]. Alongside that, the company advertises dedicated GPU virtual machines by the hour, including entry pricing from $0.18 per hour, which positions the platform as both an inference network and a lightweight alternative path to raw GPU access for teams that want more direct control [OpenRelay, retrieved 2026] [AIT Community, retrieved 2026].

The public evidence on performance is directionally interesting but still company-led. OpenRelay and YC materials claim up to 20% lower inference cost through accelerator selection, while the website separately claims 10% to 20% lower cost at high throughput and 2.5x faster autoscaling than other providers; a LinkedIn post cited in the research also describes the system as live in production across 22 locations, four continents, and 11 accelerator SKUs, though that level of operating detail is not independently corroborated here [Y Combinator, November 2026] [OpenRelay, retrieved 2026] [Vatsal Gopani, retrieved 2026]. The technical shape of the product is clear enough to evaluate, but the core differentiator rests on orchestration quality, benchmarking accuracy, and real production reliability, which remain difficult to verify from public material alone [OpenRelay, November 2026] [Y Combinator, August 2026].

Company-stated, unverified -- This section relies primarily on company, YC, and affiliated public materials, with limited independent corroboration of technical and performance claims.

Where the Demand Sits

PUBLIC The market matters now because inference is shifting from a model-development side task into a production infrastructure category, and OpenRelay is positioning itself inside that shift rather than trying to sell generic GPU access [Y Combinator, November 2026] [OpenRelay, November 2026].

OpenRelay is not entering a market with a clean public TAM that can be cited from an independent third-party report in the supplied materials, so the sizing discussion has to stay narrow and source-bound. The company describes an "Inference Delivery Network" as a layer that coordinates compute nodes to serve machine-learning inference requests, while Y Combinator frames the product as distributed, hardware-agnostic AI inference across NVIDIA, TPU, Trainium, and AMD accelerators [OpenRelay, retrieved 2026] [Y Combinator, August 2026]. That places the addressable spend at the intersection of inference infrastructure, GPU cloud capacity, and developer-facing model serving APIs, but the available evidence does not support a precise public TAM, SAM, or SOM figure from an independent market report.

What can be said with more confidence is that OpenRelay is targeting the production inference budget, not training clusters or broad cloud IaaS. Its own materials emphasize routing live workloads across heterogeneous hardware, presenting a single endpoint, and continuously benchmarking accelerator performance, while the commercial offer spans reserved throughput, token-based inference, and hourly GPU instances [OpenRelay, November 2026] [OpenRelay API Documentation, retrieved 2026] [OpenRelay, retrieved 2026]. That framing suggests the nearer comparable market is model serving infrastructure, with adjacent pull from neoclouds and data-center operators that may want demand aggregation as much as developers want cheaper or faster inference [Y Combinator, November 2026].

The main demand drivers in the cited research are cost pressure, throughput reliability, and hardware fragmentation. OpenRelay and Y Combinator both present the same basic wedge: developers increasingly need inference capacity, but do not want to optimize separately for each chip family or source capacity provider by provider [Uncork Capital, November 2026] [Y Combinator, August 2026]. The company claims up to 20% lower inference cost through workload-aware accelerator selection, and separately claims 10% to 20% lower cost at high throughput and 2.5x faster autoscaling, although those figures remain company- or platform-sourced rather than independently verified [Y Combinator, November 2026] [OpenRelay, retrieved 2026].

Those claims fit a broader market logic even without third-party sizing data in the materials. If inference traffic is becoming continuous and latency-sensitive, then a routing layer that arbitrages between accelerator types and geographies becomes easier to understand as a procurement and reliability tool, not just a developer convenience [OpenRelay, November 2026]. OpenRelay's reported network footprint, more than 20 compute sites in the seed announcement and 22 locations in a LinkedIn-posted founder background note, points to the operational importance of geographic and hardware diversity, though both datapoints should be treated cautiously because they are not independently corroborated by named operating partners [OpenRelay, November 2026] [LinkedIn, retrieved 2026].

A small set of public market signals illustrates how the company itself is segmenting the opportunity.

Cited market signal	Figure	Source
Compute sites on network	More than 20	[OpenRelay, November 2026]
Tokens processed per month	More than 5 trillion	[OpenRelay, November 2026]
Annualized revenue run rate	More than $10 million	[OpenRelay, November 2026]
Accelerator types highlighted	NVIDIA, TPU, Trainium, AMD	[Y Combinator, August 2026]

The table does not size the market in the conventional sense, but it does show where management believes demand is already aggregating: high-volume production inference routed across multiple hardware families. For investors, that is more useful than a broad AI infrastructure TAM slide, but it remains management-supplied evidence.

Adjacent and substitute markets matter here because OpenRelay is effectively asking buyers to choose an orchestration and delivery layer instead of buying directly from a single cloud or inference API provider. One substitute is the status quo of sourcing GPU or accelerator capacity from a hyperscaler or neocloud and operating model serving internally; another is buying a managed inference API that abstracts away hardware choice but also limits routing control [OpenRelay, retrieved 2026] [Y Combinator, August 2026]. The two-sided aspect of the network, where accelerator operators can contribute capacity while developers send workloads, also places OpenRelay near marketplace dynamics rather than a pure software subscription model [Y Combinator, November 2026].

The macro and regulatory picture in the cited materials is indirect but still relevant. OpenRelay's emphasis on heterogeneous accelerators and multi-site routing implies a market shaped by intermittent GPU scarcity, uneven availability across chip vendors, and buyer reluctance to depend on a single silicon stack [Uncork Capital, November 2026] [OpenRelay, November 2026]. At the same time, no supplied source establishes a current regulatory moat or compliance-led demand driver, so it would be premature to argue that policy, export controls, or sovereignty requirements are yet central to the story on the public record.

Company-stated, unverified -- This section relies primarily on company, investor, and Y Combinator materials, with no independent third-party market report or operating-partner corroboration in the supplied sources.

Competitive Landscape

MIXED OpenRelay is positioning itself less as another GPU cloud and more as an orchestration layer that sits above fragmented inference supply, which places it in a different lane from raw compute vendors but in direct competition with any platform that can abstract hardware choice for developers [OpenRelay, November 2026] [Y Combinator, August 2026].

Company	Positioning	Stage / Funding	Notable Differentiator	Source
OpenRelay	Inference Delivery Network for routing production inference workloads across heterogeneous accelerators	Seed, $10,000,000 disclosed in November 2026	Single production endpoint with routing across NVIDIA, TPU, Trainium, and AMD, plus pricing options for reserved, async, batch, and hourly GPU usage	[OpenRelay, November 2026] [Y Combinator, August 2026] [OpenRelay, retrieved 2026]
OneTriangle	Named competitor in structured research	Not established in cited public materials used here	Competitive details were not established in the cited public materials used for this section	[Structured Facts]

The market breaks into three practical groups. First are incumbent hyperscalers and cloud platforms, which own supply, distribution, and enterprise relationships, but typically ask customers to work within a narrower hardware and deployment stack than OpenRelay is advertising [Y Combinator, August 2026] [OpenRelay, November 2026]. Third are adjacent substitutes, namely teams that self-manage model serving across multiple clouds or buy dedicated GPU instances directly, an option OpenRelay is explicitly trying to simplify with a single endpoint and hardware-aware routing layer [OpenRelay, retrieved 2026] [OpenRelay API Documentation, retrieved 2026].

The edge visible in public evidence today is technical framing rather than entrenched distribution. OpenRelay says its system continuously benchmarks accelerator performance, routes requests based on latency, cost, availability, and SLA targets, and speaks OpenAI and Anthropic wire formats, which together suggest a low-friction adoption path for teams already building against familiar APIs [OpenRelay, November 2026] [OpenRelay API Documentation, retrieved 2026]. That edge looks meaningful if customers increasingly want portability across silicon types, but it is still perishable because the underlying claims on cost savings, autoscaling, network scale, and revenue are company-reported rather than independently corroborated, and larger infrastructure vendors could reproduce parts of the control plane if demand proves durable [Y Combinator, November 2026] [OpenRelay, retrieved 2026] [OpenRelay, November 2026].

The clearest exposure is ownership of supply and trust. More concretely, incumbent clouds remain advantaged wherever buyers prefer vertically integrated procurement, compliance, and support over cross-provider optimization, and OpenRelay's public materials do not yet name customers or operating partners that would offset that distribution gap [OpenRelay, November 2026] [Y Combinator, August 2026].

The most plausible 18-month scenario is a sorting between orchestration specialists that prove real workload portability and platforms that remain interchangeable brokerage layers. OpenRelay is the most likely winner if heterogeneous accelerators remain structurally fragmented and customers keep valuing a single inference endpoint that can arbitrage silicon, geography, and price in production [OpenRelay, November 2026] [Y Combinator, November 2026].

Company-stated, unverified -- This section relies primarily on company, Y Combinator, and structured-research sourcing, with only limited independent corroboration and sparse public detail on the named competitor.

Opportunity

PUBLIC

If OpenRelay executes on its current wedge, the prize is not a niche inference API but a control layer for where production AI workloads run, and that is a position with platform economics rather than simple hosting margins [OpenRelay, November 2026] [Y Combinator, November 2026].

The headline opportunity is straightforward: become the default routing and orchestration network for inference across fragmented accelerator supply. The public case for that outcome rests on two things the company has already put into market, even if the underlying metrics remain company-supplied. First, OpenRelay is not presenting itself as a single-cloud GPU vendor. It is presenting a hardware-agnostic network that schedules workloads across NVIDIA, TPU, Trainium, and AMD, exposes a single production endpoint, and continuously benchmarks accelerator performance [OpenRelay, November 2026] [Y Combinator, August 2026]. Second, the company says it reached production in June 2026 and by November had exceeded 20 compute sites, processed more than 5 trillion tokens per month, and crossed $10 million in annualized run rate within five months of founding [OpenRelay, June 2026] [OpenRelay, November 2026]. Those claims are unverified outside company and affiliated channels, but if directionally true they suggest the product is already solving a real procurement and operations problem for teams that do not want to optimize model serving stack by stack.

The upside paths are not identical, which matters because this category can compound through software control, marketplace density, or price-performance leadership. The table below frames three public-evidence scenarios.

Scenario	What happens	Catalyst	Why it's plausible
Multi-cloud inference control plane	OpenRelay becomes the standard endpoint enterprises and AI startups use to route inference across heterogeneous accelerators and providers	A large customer base starts treating hardware choice as a routing decision rather than an infrastructure migration, helped by OpenAI- and Anthropic-compatible wire formats and a single production endpoint [OpenRelay API Documentation, retrieved 2026] [OpenRelay, November 2026]	The product is already described as hardware-agnostic and able to route across NVIDIA, TPU, Trainium, and AMD, which is the technical precondition for a control-plane position [Y Combinator, August 2026] [Y Combinator, November 2026]
Exchange for stranded accelerator supply	OpenRelay becomes the demand layer that monetizes underused GPU and accelerator capacity across neoclouds and data centers	More operators connect supply to the network because OpenRelay can aggregate customer demand and abstract away go-to-market friction [Y Combinator, November 2026] [Uncork Capital, November 2026]	YC's public materials describe a two-sided model where capacity providers plug into the network while inference customers bring workloads, which is a credible marketplace architecture if liquidity builds [Y Combinator, November 2026]
Cost leader for production open-model inference	OpenRelay wins on economics for high-volume inference by steering requests to the best silicon and packaging throughput into predictable pricing	Rising inference volumes make even modest cost advantages meaningful, especially if async and batch discounts pull in cost-sensitive workloads [OpenRelay, retrieved 2026] [Y Combinator, November 2026]	The company publicly claims up to 20% lower cost, 20% to 50% discounts for async workloads, and 50% off batch inference, all of which fit a buyer need that is already visible in the market [OpenRelay, retrieved 2026] [AIT Community, retrieved 2026]

What compounding looks like here is less about a classic social network effect and more about routing intelligence improving with scale. Every additional workload gives the orchestration layer more real-world information on latency, availability, and price-performance across accelerator types, and OpenRelay says continuous benchmarking is already part of the system design [OpenRelay, November 2026]. Every additional supply node increases the odds that the network can meet a workload's SLA at lower cost, which should attract more demand, which in turn makes the network more valuable to supply-side operators. Public evidence for that flywheel is still early and company-led, but the pieces are visible: more than 20 compute sites, 22 locations and 11 accelerator SKUs in a founder-adjacent public profile, and compatibility with established API formats that reduce switching costs for developers [OpenRelay, November 2026] [Vatsal Gopani, retrieved 2026] [OpenRelay API Documentation, retrieved 2026]. If that loop holds, OpenRelay could accumulate a defensible routing dataset and an integration surface that is harder to displace than a standalone inference endpoint.

The size of the win is best thought of through market position rather than a near-term revenue multiple, because there is no cited third-party market-sizing report in the source set. A credible public comparable inside the evidence base does not exist, so the more defensible framing is scenario-based. If OpenRelay became a widely adopted control plane for production inference, or the exchange layer that clears a meaningful share of heterogeneous accelerator supply, it could support platform-like economics and strategic value well above a conventional infrastructure reseller (scenario, not a forecast) [OpenRelay, November 2026] [Uncork Capital, November 2026]. The public materials do not justify a precise valuation target today, but they do justify treating the upside as category-scale rather than feature-scale because the company is trying to sit at the routing layer between AI demand and compute supply.

Company-stated, unverified -- This section relies materially on company, investor, YC, and affiliated public claims, with no independent third-party validation of traction, customer adoption, or market size in the cited source set.

Sources