The AI Cost Collapse: Enterprise Strategy and Competitive Moats in a Commoditizing AI World

The price of intelligence is collapsing. Between 2020 and 2025, the cost of running one million tokens through a frontier large language model fell by more than 99%, from approximately $60 to less than $0.10 at major commercial providers—a price trajectory that has no precedent in the history of enterprise software and few parallels in the broader history of commodity technology. This is not a pricing competition between vendors, though that is part of the story. It is the visible surface of a deeper structural shift: the industrial-scale commoditization of AI inference, driven simultaneously by hardware efficiency gains, algorithmic improvements, the proliferation of open-weight models, and the massive capital investment by hyperscale cloud providers who need to fill GPU infrastructure at scale. The strategic implications of this cost collapse are still being absorbed, improperly in most cases, by the executive teams and strategy functions of large enterprises. The dominant frameworks through which most organizations are approaching AI strategy—build vs. buy, technology pilots, digital transformation roadmaps—were designed for a world in which AI capability was scarce, expensive, and proprietary. That world is ending. The frameworks built for it are becoming liabilities.

What replaces them is not a different technology framework but a different strategic framework: one organized around the question of what remains scarce and therefore defensible as AI inference becomes abundant and cheap. The answer to that question determines where competitive advantage can be built and sustained, what investments deserve priority, and which AI-related activities should be contracted out versus developed internally. This essay examines the AI cost collapse in depth, traces its causes and trajectory, and constructs a strategic framework for enterprise decision-making in an environment where frontier AI capability is rapidly approaching commodity status.

The Price Trajectory No One Planned For

Historical Benchmarks and Inference Costs

The quantification of the AI cost collapse requires precision because the numbers are frequently mischaracterized, and the mischaracterization matters strategically. The relevant cost metric for enterprise AI applications is not the training cost of a large model—which has indeed increased dramatically as frontier models have grown in scale and capability—but the inference cost: the cost of running a trained model to generate a response to a specific query.

Inference cost is the cost that scales with deployment: for every customer interaction, document processed, decision supported, or content item generated, the enterprise pays an inference cost. At scale—millions of queries per month, or billions of tokens per year—inference cost is the primary determinant of AI economics in production deployment. Training cost is a one-time expenditure that is largely irrelevant to the operational economics of deployed AI applications.

The inference cost trajectory for frontier models over the 2020-2025 period is extraordinary by any historical standard:

The 600x cost reduction at the upper end of this table—from GPT-3 at $60 to leading open-source models at $0.10 or less in 2025—is not an artifact of comparing different capability levels. The capability frontier has moved dramatically higher over this period, but the cost of equivalent capability has fallen by comparable orders of magnitude. A model that cost $60 per million tokens to run in 2020 was substantially less capable than models available at $0.10-0.30 per million tokens in 2025.

The Drivers of Cost Compression

Understanding why inference costs have fallen so dramatically is necessary for forming views about the trajectory of future costs—and therefore about the durability of any AI-based competitive position. The compression has been driven by six overlapping mechanisms, each of which has a different expected rate of future contribution.

Hardware efficiency: The GPU and custom ASIC infrastructure used for AI inference has improved dramatically in throughput per dollar. NVIDIA's H100 and H200 chips, Google's TPU v5, and a generation of custom inference chips from Amazon (Inferentia), Microsoft (Maia), Meta, and Apple have each contributed to lower per-token inference costs by increasing the amount of computation that can be performed per unit of capital deployed. The hardware efficiency curve is not exponential but is substantial, contributing an estimated 3-5x cost reduction over the 2020-2025 period.

Algorithmic efficiency: The models themselves have become dramatically more efficient. Techniques including quantization (reducing the numerical precision of model weights without significant accuracy loss), pruning (removing less important model parameters), knowledge distillation (training smaller models to replicate the behavior of larger ones), and speculative decoding (using faster models to generate candidate outputs verified by a slower, more accurate model) have collectively reduced the compute required to generate a token of equivalent quality. Algorithmic efficiency has contributed an estimated 10-20x reduction in inference cost over the period.

Architectural innovation: New model architectures—mixture of experts (MoE), which activates only a fraction of a model's parameters for each inference—have enabled frontier-level capabilities to be delivered at substantially lower compute cost. The Mixtral model family, Meta's mixture of experts architectures, and similar designs from other labs have demonstrated that MoE architectures can deliver performance competitive with dense models at 2-5x lower inference cost.

Infrastructure scale and utilization: The hyperscale cloud providers (AWS, Azure, Google Cloud) and the major AI API providers (OpenAI, Anthropic, Mistral, Cohere) have been able to achieve GPU utilization rates and infrastructure efficiencies unavailable to smaller deployments. As inference volume has grown, the fixed infrastructure costs have been amortized over larger base loads, driving per-token costs down.

Open-source model proliferation: The release of high-quality open-weight models—Meta's Llama family, Mistral's releases, and dozens of fine-tuned derivatives—has created an alternative to proprietary API access that dramatically expands competitive pressure. Organizations with the infrastructure capability to host and serve open-weight models can run inference at a cost that approaches pure hardware economics, eliminating the model developer's margin entirely.

Competition between API providers: The competitive dynamics among AI API providers have driven aggressive price reductions, particularly in the small-model segment where competition is most intense. Providers are willing to run inference below marginal cost in some segments to establish market position and gather deployment data.

The Strategic Landscape After the Collapse

The cost collapse has restructured the competitive landscape for AI applications in ways that are still incompletely understood by most enterprises and investors. The most important structural changes are:

Moat Erosion in AI-Native Companies

The earliest AI-native enterprise software companies—firms that built their products primarily around the differentiated capabilities of proprietary large language models—benefited from a competitive environment in which access to frontier AI capability was expensive and scarce. Firms that had the resources and technical sophistication to integrate with GPT-4 in 2023 had a temporary advantage over competitors who could not afford the inference costs or navigate the API integration complexity.

That advantage has eroded rapidly. The cost compression described above has made frontier AI capability accessible to any company with modest technical resources. The integration complexity has been substantially reduced by the maturation of API ecosystems, the proliferation of integration frameworks (LangChain, LlamaIndex, Haystack), and the availability of large development communities with direct experience integrating with major AI APIs.

The competitive moat that remained—proprietary model performance differentials—has also narrowed. The gap between the frontier model and a well-chosen open-source alternative has compressed substantially on most enterprise task categories. For many production use cases—document processing, code generation, data extraction, customer service triage—the performance differential between a frontier proprietary model and a well-selected open-weight alternative is smaller than the cost differential, making the economic case for proprietary model use increasingly narrow.

"The most dangerous moment for a company is when it mistakes the capability of its current product for a durable competitive advantage. In AI, the capability advantage you have today may be a commodity tomorrow—and the organizations that understand this earliest will be the ones that have built defensible positions elsewhere in the stack before the advantage disappears." This observation, increasingly shared among AI-native executives who were early to understand the commoditization dynamic, reflects the structural reality that AI capability is not a durable source of competitive advantage for most enterprise software companies.

The implications are most acute for companies whose primary value proposition is the AI model itself—companies that compete primarily on the quality of their model's outputs rather than on the workflow integration, data infrastructure, customer relationship, or domain expertise surrounding it. These companies face the most direct exposure to the commoditization dynamic and have the least runway to transition to more defensible positions.

Data as the Last Defensible Advantage

If the AI model itself is becoming a commodity, what remains defensible? The answer that most sophisticated AI strategists have converged on is proprietary data—specifically, data that is:

1. Unique: Not available to competitors through any commercially accessible source
2. High-signal: Contains patterns, labels, or relationships that are directly relevant to the task the AI is being trained or fine-tuned to perform
3. Continuously generated: Updated in real time through the firm's operations, rather than representing a static historical dataset
4. Structurally protected: Protected by contracts, regulatory restrictions, customer relationships, or operational integration from competitive replication

The data advantage is real but more fragile than it is often presented. Several dynamics complicate the data-as-moat thesis:

Data quality vs. data quantity: The shift from pretraining on massive datasets to instruction fine-tuning and preference optimization on carefully curated small datasets has changed the economics of data advantage. The era of "more data always wins" has been replaced by an era in which 10,000 high-quality labeled examples can outperform 10 million low-quality examples for specific task adaptation. This means the barrier to building data-driven AI advantages is lower for firms that can generate high-quality labeled data from their operations, and higher for firms that were relying on bulk data collection as their primary data strategy.

Synthetic data proliferation: The use of large language models to generate synthetic training data has partially decoupled AI performance from real-world data collection. Firms can now generate millions of high-quality synthetic examples for task fine-tuning, reducing the competitive advantage of firms that have historically collected large volumes of real-world data. The limitation is that synthetic data cannot replicate the distribution of genuinely novel scenarios—it can only recombine patterns from existing training data—which means real-world operational data retains significant value for edge cases and novel situations.

Customer data portability: In many enterprise software contexts, customer data belongs to the customer, not the software provider. As enterprise customers become more sophisticated about data sovereignty and AI value extraction, they are increasingly unwilling to allow software vendors to use their operational data to train or fine-tune models that will benefit other customers. The result is that the data advantages that first-generation enterprise AI companies assumed they would accumulate from customer deployments may be more restricted than anticipated.

From Build vs. Buy to Layer vs. Integrate

The traditional enterprise technology decision framework—build custom software vs. buy a commercial product—is inadequate for AI strategy in the post-cost-collapse environment. The relevant strategic question is not whether to build or buy but where in the AI stack to invest proprietary resources and where to rely on rapidly commoditizing infrastructure.

The AI stack can be conceptualized in five layers:

Layer 1: Foundation models: The base large language models (GPT-4o, Claude, Gemini, Llama, Mistral). Cost is collapsing; quality is converging across major providers. This layer is rapidly commoditizing.

Layer 2: Inference infrastructure: GPU compute, serving infrastructure, caching, batching optimization. Economizing effectively through competition among cloud providers and open-source tooling. This layer should be procured, not built.

Layer 3: Application framework: Orchestration frameworks, RAG pipelines, agent frameworks, tool integration. Highly contested between commercial platforms (LangChain Enterprise, Azure AI Foundry, AWS Bedrock Agents) and open-source alternatives. The right approach depends on scale and technical capability.

Layer 4: Domain adaptation: Fine-tuning, RLHF, prompt engineering, domain-specific evaluation. This is where proprietary data advantage is captured. High-value investment for firms with genuinely distinctive data.

Layer 5: Workflow and system integration: The embedding of AI capabilities into existing enterprise workflows, systems of record, and customer-facing processes. This is where the competitive advantage is most durable because it is most deeply embedded in organizational operations and customer relationships.

The strategic implication is that enterprise investment should concentrate in Layers 4 and 5, where proprietary advantage can be built and maintained, while efficiently procuring Layers 1-3 from the most cost-effective commercial providers. Firms that invest heavily in Layer 1 (proprietary foundation model development) or Layer 2 (custom inference infrastructure) without a specific strategic reason—a scale requirement, a data privacy constraint, a latency requirement—are misallocating resources.

"Companies will waste billions over the next five years trying to build something they could have rented for a fraction of the cost, while simultaneously underinvesting in the workflows, data infrastructure, and domain adaptation that would have made the rented capability genuinely valuable. The error is not spending too much on AI—it is spending on the wrong part of the AI stack." This structural misallocation is already visible in the pattern of enterprise AI investment announcements that prioritize GPU infrastructure and foundation model development over data infrastructure and workflow integration.

The Strategic Response Matrix

The appropriate strategic response to the AI cost collapse depends on the enterprise's current position in the AI value chain—specifically, whether it is primarily an AI provider (a company whose product or service includes AI as a core component) or an AI consumer (a company that deploys AI to improve its own operational performance or products).

For AI providers, the most strategically urgent response to cost commoditization is to shift the locus of differentiation away from model quality and toward the dimensions that remain scarce: domain data, workflow integration, customer relationships, and switching cost architecture. This transition is organizationally challenging because it requires investment in capabilities—data infrastructure, domain expertise, customer success, workflow engineering—that are very different from the ML research and model development capabilities that drove initial product differentiation.

For AI consumers, the strategic priority is to build the internal capability to effectively deploy and leverage AI tools, rather than to develop proprietary AI technology. The most valuable AI investments for most large enterprises are not in model development or infrastructure but in data infrastructure (to capture the data advantage), workflow redesign (to embed AI capability in operational processes), and governance architecture (to manage the organizational risks of widespread AI deployment).

Competitive Dynamics Shift

The AI cost collapse is reshaping competitive dynamics not just within the AI industry but across every industry where AI capability is becoming a meaningful competitive variable. Three shift patterns are particularly significant.

Incumbents vs. Challengers

The conventional wisdom about AI's competitive implications suggested that AI would be a disruptive force favoring challengers over incumbents—that AI-native startups, unencumbered by legacy infrastructure and organizational inertia, would use AI capabilities to attack incumbents from below. This prediction has been partially but not fully validated.

The incumbent advantage in AI deployment turns out to be more durable than most early AI analysts predicted, for reasons related precisely to the cost collapse. As inference costs fall, the barriers to building AI applications decrease—but the barriers to building AI applications that are genuinely better than existing solutions do not decrease at the same rate. The differentiating factors—proprietary data, domain expertise, customer relationships, workflow integration—are advantages that incumbents have accumulated over years and that challengers cannot replicate quickly regardless of AI capability.

The challenger advantage remains in contexts where incumbents are structurally constrained from deploying AI—regulatory limitations, technical debt, organizational resistance, or misaligned incentive structures—and where the AI capability provides a performance advantage large enough to overcome the incumbent's data and relationship advantages. Healthcare, financial services, and legal services are sectors where these dynamics coexist: significant regulatory and organizational constraints on incumbents, combined with large performance gaps that AI can address.

"The incumbents who will lose to AI-native challengers are not those who deploy AI slowly—they are those who deploy it in ways that preserve rather than challenge their existing business model. The threat is not the technology; it is the strategic choice to use the technology to defend the status quo rather than to rebuild around the customer need." This distinction between AI deployment that improves incumbent efficiency and AI deployment that redefines the value proposition is the critical strategic variable for incumbents assessing competitive vulnerability.

The Platform Play and Vertical AI

The most strategically significant competitive dynamic in enterprise AI is the contest between horizontal AI platforms and vertical AI specialists. Horizontal platforms—Microsoft Azure AI, Google Cloud AI, AWS Bedrock, Salesforce Einstein—offer comprehensive AI capabilities integrated into the enterprise infrastructure stack. Vertical AI specialists—Tempus in oncology, Veeva in life sciences, Harvey in legal services, Palantir in defense analytics—offer AI capabilities optimized for specific industry contexts, with domain data, regulatory expertise, and workflow integration that horizontal platforms cannot replicate.

The cost collapse has paradoxically strengthened both categories. As foundation model costs fall, horizontal platforms can offer more capable AI at lower marginal cost, increasing their attractiveness as infrastructure providers. But the commoditization of the AI foundation also raises the strategic premium on the domain adaptation and workflow integration that vertical specialists provide—because these are the dimensions that remain scarce after the foundation is commoditized.

The competitive equilibrium that is emerging is a two-layer structure: horizontal platforms as the commoditized AI infrastructure layer, and vertical specialists as the differentiated application layer that extracts value from domain expertise and proprietary data. This structure mirrors the historical pattern of enterprise software evolution, where database infrastructure was commoditized while vertical application vendors (ERP, CRM, HCM) remained differentiated. The strategic implication for enterprise buyers is that AI strategy should be organized around vertical application selection and data strategy rather than around infrastructure selection, which will be increasingly commoditized.

The Open Source Dynamics

The role of open-source AI models in reshaping competitive dynamics deserves specific attention because it represents a structural disruption to the AI business model that has significant strategic consequences for enterprise buyers and AI vendors alike.

The release of Meta's Llama family, Mistral's open-weight models, and a proliferating ecosystem of fine-tuned derivatives has created a credible open-source alternative to proprietary AI APIs for a substantial range of enterprise use cases. The cost differential between proprietary API access and self-hosted open-weight models at scale is approximately 10-50x, depending on use case and infrastructure efficiency. For enterprises with significant AI usage volumes and adequate technical capability, this differential represents a compelling reason to invest in open-source model deployment.

The strategic implications of open-source AI extend beyond cost reduction:

Vendor independence: Self-hosted open-source models eliminate vendor lock-in to proprietary API providers. For enterprises in regulated industries, defense, or sensitive IP contexts, this independence has non-cost value.

Customization depth: Open-weight models can be fine-tuned, modified, and adapted at a depth that proprietary APIs do not allow. For enterprises with genuinely distinctive domain data, open-source deployment enables capability advantages that proprietary API access cannot match.

Data privacy: On-premise deployment of open-weight models eliminates the data privacy risks associated with sending sensitive operational data to external AI API providers. This is a significant consideration for healthcare, legal, financial services, and government organizations.

Performance ceiling: The performance ceiling for open-weight models remains below the frontier proprietary models in most complex reasoning tasks, multi-step agent tasks, and tasks requiring sustained coherence over long contexts. For these use cases, the cost-performance tradeoff continues to favor proprietary frontier models.

The strategic decision framework for enterprise AI procurement is therefore not proprietary vs. open-source but task-by-task optimization across the full range of AI use cases: frontier proprietary models for complex, high-stakes decisions where performance is critical; optimized open-source models for high-volume, well-defined tasks where cost efficiency and data privacy dominate; and specialized fine-tuned models for domain-specific applications where proprietary data creates genuine performance advantages.

Regulatory Overlay

The AI cost collapse is occurring against a backdrop of rapidly evolving regulatory frameworks that introduce additional complexity into enterprise AI strategy. The European AI Act, which entered into force in 2024 and is being phased in through 2027, imposes substantive compliance requirements on AI systems deployed in the EU across a spectrum of risk categories. The US is developing sector-specific AI regulations in healthcare, financial services, and national security. Multiple jurisdictions are introducing AI transparency, explainability, and audit requirements.

Regulatory compliance creates a different kind of moat for AI applications: not a capability or data moat, but a compliance moat. Organizations that invest in building the governance infrastructure, audit capabilities, and documentation frameworks required for regulatory compliance in high-risk AI contexts have a durable advantage over competitors who lack these capabilities. The compliance moat is most significant in sectors with the most demanding regulatory environments—healthcare, financial services, defense, critical infrastructure—and for AI applications in the highest-risk regulatory categories (biometric identification, credit decisioning, medical diagnosis, consequential employment decisions).

The regulatory trajectory is toward more, not less, compliance burden on AI applications in high-stakes contexts. This creates an important strategic asymmetry: organizations that invest in compliance infrastructure now, when the regulatory requirements are becoming clearer, will have lower marginal compliance costs than late movers who must retrofit compliance capabilities after deployment. The sunk cost of compliance infrastructure is itself a form of competitive protection.

Simultaneously, regulation introduces friction that disproportionately affects smaller, less resourced competitors. Large enterprises with established legal and compliance functions can absorb the burden of AI compliance more efficiently than startups or mid-market competitors. This represents one of the few AI dynamics that structurally favors incumbents—a partial offset to the challenger advantages in AI-native product development.

Organizational Readiness for AI at Scale

The technical and strategic dimensions of AI deployment at scale have received substantial attention. The organizational readiness requirements have received less attention but are at least equally important as determinants of which enterprises will actually capture the economic value that the AI cost collapse makes possible.

AI literacy and decision-making: The effective deployment of AI in operational contexts requires that the humans making decisions with AI assistance understand the capabilities and limitations of the AI tools they are using—not at a technical level, but at a conceptual level sufficient to distinguish appropriate from inappropriate reliance on AI outputs. Organizations that deploy AI without investing in this decision-making literacy will generate systematic errors: over-reliance on confident-sounding but factually incorrect AI outputs in high-stakes contexts; under-reliance in contexts where AI performance is genuinely superior to human judgment; and failure to escalate appropriately when AI outputs are ambiguous or low-confidence.

Workflow redesign, not bolt-on AI: The most common failure mode in enterprise AI deployment is the "bolt-on" pattern: integrating AI as an add-on to existing workflows without fundamentally redesigning the workflow to take advantage of AI capabilities. The result is AI that provides incremental efficiency improvements rather than the step-change performance improvements that are theoretically possible. Effective AI deployment requires redesigning workflows from the perspective of what would be optimal if AI capability were reliably available—not asking how AI can help with the current workflow.

Governance and risk architecture: At scale, the risk surface of AI deployment expands dramatically. Errors that affect individual decisions become errors that affect thousands or millions of decisions when AI is deployed across high-volume operational processes. Governance architecture for AI deployment must include monitoring for systematic performance degradation, feedback mechanisms from operational outcomes to model updates, escalation protocols for cases outside the model's reliable performance range, and audit capabilities for regulatory compliance.

The organizational readiness gap is the primary constraint on enterprise AI value capture—more significant than technology availability, cost, or strategic clarity for most large organizations. The gap is not insurmountable but requires sustained investment over multiple years, and it is not reducible to technology deployment. Workflow redesign, literacy development, and governance architecture are organizational capabilities that must be built through deliberate management attention, not purchased from vendors.

The Investment Discipline Required

The AI cost collapse creates a challenging investment environment for enterprise leadership teams. The pressure to invest aggressively in AI is real and appropriate—organizations that fail to deploy AI capabilities at the pace of their most advanced competitors will face genuine competitive disadvantage. But the investment must be disciplined against a clear strategic framework or it will be misallocated.

The most common misallocations observable in current enterprise AI investment include:

Investing in Layer 1 and 2 capabilities without clear strategic necessity: Building proprietary foundation models or custom inference infrastructure is appropriate for a small number of organizations with specific requirements (scale, privacy, domain specialization). For most enterprises, it is a waste of capital that could be deployed in workflow redesign and data infrastructure.

Piloting without scaling: The AI pilot has become a standard component of enterprise AI programs—a contained deployment of AI capability in a specific context, designed to demonstrate value before broader investment. The pilot is a useful tool for reducing uncertainty, but it systematically undervalues AI deployment because the economic benefits of AI are most visible at scale. Organizations that maintain perpetual pilot mode—iterating through proofs of concept without committing to production-scale deployment—capture little of the economic value that AI enables.

Ignoring data infrastructure: The most consistently underinvested dimension of enterprise AI programs is data infrastructure—the systems, processes, and governance that ensure AI applications have access to high-quality, relevant, current data. Without this investment, AI applications are limited to the quality of data in existing systems, which is typically insufficient for high-performance deployment. Organizations that invest in AI applications without investing in data infrastructure first are building on an inadequate foundation.

Treating AI as an IT function: The organizations that are capturing the most value from AI deployment are those where AI strategy is led by business functions, with IT in a support role, rather than being driven by IT with business functions as end users. AI is a business capability, not a technology project, and the investment decisions, workflow redesign choices, and performance measurement frameworks must be owned by the business functions that will use the capability.

The Horizon View

The cost trajectory of AI inference shows no sign of stabilizing. The combination of hardware efficiency improvements, architectural innovation, and intensifying competition among both proprietary and open-source providers suggests that inference costs will continue to compress substantially over the next three to five years. The strategic implication is that competitive advantages built on AI cost efficiency or AI capability access will continue to erode at the current pace.

The durable competitive positions in an AI-abundant world are built on the dimensions that will not be commoditized: proprietary operational data that accumulates through activity, domain expertise embedded in organizational routines and specialized talent, customer relationships and switching costs built through deep workflow integration, and governance capabilities that enable trustworthy AI deployment in regulated and high-stakes contexts.

Organizations that understand this hierarchy and invest accordingly—treating AI foundation models as infrastructure to be procured efficiently rather than as sources of proprietary advantage—are best positioned to capture the economic value that the cost collapse enables. Those that continue to invest as if AI capability itself were the scarce resource will find that the advantage they are paying to build has already moved to the next layer of the stack.

Sources & References

• OpenAI API pricing history and model documentation (official releases)
• Anthropic technical reports on Claude model families
• Meta AI research publications on Llama model architecture and performance
• Mistral AI technical reports and model documentation
• NVIDIA hardware specifications and performance benchmarks (H100, H200 series)
• Google Cloud TPU documentation and performance reports
• MIT Technology Review (AI cost trends and enterprise deployment analysis)
• Harvard Business Review (AI strategy and competitive advantage in enterprise)
• McKinsey Global Institute (AI economic impact and enterprise adoption studies)
• Stanford HAI Annual AI Index (capability and cost benchmarking)
• The Economist (AI industry competitive dynamics and commoditization analysis)
• Financial Times (AI investment trends and enterprise adoption reporting)
• Wall Street Journal (enterprise AI deployment and strategy coverage)
• Andreessen Horowitz (a16z) AI market research and investment analyses
• Sequoia Capital AI market sizing and strategic landscape reports
• European AI Act official documentation (EU regulatory framework)
• NIST AI Risk Management Framework (US AI governance standards)
• Gartner AI adoption surveys and enterprise readiness benchmarking
• IDC enterprise AI infrastructure market research
• Journal of Artificial Intelligence Research (technical benchmarking and efficiency research)
• ACM Digital Library (LLM architecture and inference efficiency papers)