Advancing the American AI Stack

Introduction

Power has always flowed from the control of the world's essential resources. Once it was steel, then oil, then data. Today, it is AI compute, and specifically, the ability to run AI systems efficiently at global scale. Whoever controls AI compute will shape the century ahead.

Compute is fast becoming the foundation of global economic growth. In the United States, investment in AI infrastructure—from data centers to semiconductors and energy systems—is already moving the needle: J.P. Morgan estimates that data-center spending alone could boost U.S. GDP by up to 20 basis points over the next two yearsFootnote 1. According to The Economist, investments tied to AI now account for 40 percent of America's GDP growth over the past year, equal to the amount contributed by consumer spending growth. That statistic would be staggering regardless of how long AI has been part of the economy, but this is just the start.

The next decade of global competition will be defined not only by who invents the most powerful AI systems, but also by who can deploy and operate them securely, efficiently, and at scale. The real battleground increasingly centers on inference, the computational power required to run trained AI models and deliver real-time results to billions of users worldwide.

While training compute builds AI capabilities, inference compute delivers them. As AI applications move from laboratory to deployment, inferenceFootnote 2 becomes the bottleneck that determines which nations can actually operationalize artificial intelligence at global scale. This is how the industry operates today and can serve as the model that informs U.S. export policy.

The question facing policymakers is whether to recognize and enable the existing model of a vibrant American AI ecosystem, or to construct something entirely new. The evidence suggests that a nuanced approach to the former will better serve American strategic interests.

Organizing an American AI Export Program

The Trump Administration's Executive Order on Promoting the Export of the American AI Technology Stack (EO) recognizes that our allies are hungry for American compute, and that the United States must dominate the "away game" before geopolitical rivals fill the vacuum.

This EO represents a watershed moment in American technology policy. Previous administrations treated AI exports primarily through a defensive lens, focusing on what to restrict rather than what to enable. The Trump Administration has inverted that paradigm, recognizing that American AI leadership depends not just on preventing adversaries from acquiring our technology, but on ensuring allies adopt democratically-aligned systems, standards, and operational models before alternatives take root. The stack in the EO and Groq’s definition of the "American AI Stack"—five discrete layers spanning hardware, data, models, orchestration, and applications—differ in how they designate each layer, but they both recognize that competitive advantage in AI infrastructure comes not from any single component, but from how the layers integrate into deployable systems.

In real-world scenarios, how the compute systems and data architecture within the stack function and interact will be contingent on the stack’s overall structure and the opportunities it addresses. For example, each application may require a different selection and configuration of models, hardware, and deployment solutions. As such, the stack will remain a dynamic organism, its components interchangeable and working in concert, rather than a disjointed set of layers. Given the fundamental impact of this dynamism, an export program organized around rigid layer boundaries could inadvertently slow the cross-layer innovation that gives U.S. technology its competitive advantage over more centralized, state-directed competitors.

The Department of Commerce’s Request for Information has asked industry to weigh in on how best to structure an export framework, including whether consortia should play a central role. That question reflects a thoughtful approach.

A consortium model is one way to organize exports, especially for large, integrated projects that benefit from a single coordinating body. At the same time, the RFI leaves room for other structures, such as marketplace models, that allow trusted providers to contribute individual components under shared standards. Inviting industry input acknowledges that private-sector firms often understand integration requirements and technology lifecycles with greater specificity than regulators, and that choosing a structure that is too rigid could inadvertently slow the innovation that underpins America’s advantage in AI. We believe this openness is essential: the most effective framework will not rely on a single model, but on a multiplicity of coordinated options that promote both competition and expedience.

How the AI Market Already Works

Before examining policy options, it's worth understanding how the American AI market already balances competition with coordination, and the sophisticated marketplace dynamics that consistently shape the industry. The evidence suggests a clear pattern: companies form private consortia in various configurations to address varying needs and challenges, while also competing vigorously in an open marketplace for opportunities where they can play a different role.

Consider how inference infrastructure actually reaches global markets. Groq's October 2025 partnership with IBM illustrates the pattern: integrating Groq's GroqCloud inference platform with IBM's watsonx Orchestrate environment required extensive technical coordination on load balancing, model optimization, and enterprise security protocols. This is a privately formed consortium driven by customer requirements. IBM's healthcare and financial services clients need guaranteed interoperability between Groq's Language Processing Unit inference acceleration and IBM's orchestration layer. The partnership also integrates Red Hat's open-source vLLM technology with Groq's LPU architecture—another layer of technical coordination that happens because the market demands it, not because the government mandated itFootnote 3.

Similarly, Groq's deployments through partners like Dell Technologies demonstrate how hardware and infrastructure layers coordinate. Regional distributors such as Aljammaz Technologies in the Middle East combine Groq inference systems with Dell-manufactured servers and local operational expertiseFootnote 4. Each integration requires validated compatibility protocols for power management, cooling systems, and data throughput. These are operational consortia, formed organically to meet deployment requirements in specific markets.

The pattern repeats across the American AI ecosystem. NVIDIA's partnerships with cloud providers, OEMs, and system integrators show how the training side of the stack works together. Companies like Microsoft, MetaFootnote 5, and OracleFootnote 6 form technical consortia for large-scale training projects, coordinating on everything from liquid cooling infrastructure to high-speed networking fabrics while maintaining complete independence to pursue different partnerships for other initiatives. When OpenAI needs training capacity, it assembles the partners whose technologies integrate effectively for that specific deployment.

None of these partnerships required government coordination. They emerged because technical integration demanded it, and because companies saw commercial value in solving interoperability challenges together. The hardware, software, and service providers each maintain independent innovation paths while ensuring their offerings work together when deployed. This is the natural state of American AI innovation: fierce competition alongside pragmatic collaboration, in each case based on what best serves customers.

The Path Forward

This hybrid model—marketplace competition with industry-led, project-specific consortia—is how the American AI stack achieves both rapid innovation and reliable interoperability. The policy question before the Administration is not whether to create this model, but whether to recognize it and enable it in the export context.

In our view, the answer lies in understanding that both marketplace flexibility and consortium coordination play important roles. The most effective approach combines elements of each: a marketplace of pre-certified providers with the ability to form consortia where integration requirements demand it. This preserves competition and innovation while enabling the coordinated delivery that complex deployments require. It is not “marketplace versus consortia.” Rather, the optimal frame is a marketplace with consortia, formed by companies that have already done the technical integration work.

By contrast, consortium structures that pre-determine which companies can work together risk freezing the very dynamism that gives American AI its competitive edge. The recommendations below detail how to structure an export program that preserves this competitive dynamism while meeting security and interoperability requirements.

The State of American AI Leadership

The United States currently operates from a position of strength. While training compute—the power required to develop AI models—has grown by a factor of 4.2x per year since 2018, inference compute is expanding even faster as enterprises shift from experimentation to production deploymentFootnote 7. Inference now accounts for the majority of AI compute demand globally, and this trend will accelerate: by 2030, running AI systems at scale could require orders of magnitude more inference capacity than exists today. Training happens once to create a model. Inference happens every time that model is usedFootnote 8. As AI moves to production deployment, aggregate inference demand quickly dwarfs training requirements, making inference efficiency essential for sustainable, large-scale AI operations.

Over the last half-decade, the United States has developed approximately 70 percent of the world's leading AI models and, according to the U.S. Federal Reserve, "dominates" high-end AI training compute globally, controlling 74 percent of capacityFootnote 9. U.S. companies also lead in inference infrastructure, with American-designed systems powering the majority of deployed AI applications worldwide. These figures clearly establish the United States' current leadership position. However, that leadership faces an existential threat. The rapid acceleration of AI infrastructure development and investment globally, with particularly explosive growth in Asia-PacificFootnote 10, demonstrates that other nations recognize compute as the strategic resource of this century. If the United States cannot translate its current advantage into sustained market presence in allied nations, competitors will fill that space with their own technologies, standards, and operational models.

Patterns Shaping the American AI Stack

From our vantage point inside the AI infrastructure market, several patterns are becoming clear:

First, competition and openness across the stack continue to shape which technologies advance and which ones gain real adoption, particularly among allies that need systems they can integrate quickly. At the same time, this competitive innovation is driving a flourishing American AI ecosystem, which has knock-on effects for national priorities like American manufacturing, reindustrialization, and supply chain resilience. The diversity of approaches, from Groq's inference-optimized LPU architecture to NVIDIA's training-focused GPU platforms to emerging specialized accelerators, creates options that centralized, state-directed competitors cannot match.

Second, operational oversight of AI systems by U.S. companies and by trusted allied countries is gaining importance, not as a constraint, but as a practical

[truncated for AI cost control]