SanDisk Unveils Ambitious Tokenomics 2030 Outlook Amidst Global AI Infrastructure Expansion

The rapid escalation of generative artificial intelligence has prompted a fundamental reassessment of global data infrastructure, leading SanDisk to publish its "Tokenomics 2030" report—a strategic outlook that attempts to quantify the massive surge in computational demand expected by the end of the decade. According to the flash memory manufacturer’s internal analysis, global AI systems are projected to generate approximately 150 quadrillion tokens throughout 2026, a figure that equates to roughly five billion tokens produced every second. Looking toward 2030, the company anticipates an exponential trajectory, forecasting that annual output could reach 10 quintillion tokens, or roughly 317 billion tokens per second.
These figures represent more than mere statistical curiosity; they serve as a cornerstone for SanDisk’s broader argument regarding the necessity of robust, high-capacity enterprise storage solutions. However, the company has explicitly stated that these projections are derived from proprietary internal modeling rather than external third-party data, noting that the figures are intended for illustrative purposes. This nuance is critical, as SanDisk is currently positioning its high-performance enterprise SSDs as a vital component in the hardware ecosystem required to sustain this unprecedented scale of AI activity.
The Methodology Behind the Projections
The "Tokenomics 2030" model, spearheaded by Eric Spanneut, SanDisk’s senior vice president of product management, departs from conventional demand-side forecasting. Instead of attempting to predict user adoption rates or consumer appetite for AI-generated content, the team focused on a supply-side "capacity ceiling." The model evaluates the physical constraints of global infrastructure, including the total power available for AI inference, the density and capability of the global accelerator fleet, available memory bandwidth, and the efficiency of networking and storage architectures.
By framing the 317 billion tokens-per-second figure as a capacity limit rather than a market demand forecast, SanDisk provides a sobering look at the engineering hurdles inherent in the AI transition. The growth trajectory implied by these numbers is staggering: a 60-fold increase in four years, which requires global token output to nearly triple on an annual basis between now and 2030. Critics and industry analysts have noted that while the 2026 baseline appears grounded in current industry performance metrics, the subsequent annual growth rate assumes a sustained, aggressive expansion of physical data center capacity that may face significant real-world bottlenecks.

Industry Context and Historical Benchmarks
The current 2026 baseline holds weight when compared to disclosures from major industry players. During the Google I/O conference in May 2026, Alphabet CEO Sundar Pichai reported that Google’s systems were processing approximately 3.2 quadrillion tokens per month—a massive increase from the 480 trillion tokens recorded just one year prior. When normalized to a per-second basis, Google’s internal volume accounts for roughly 1.2 billion tokens per second. This suggests that a single major hyperscaler is already responsible for a significant fraction of SanDisk’s global estimate, highlighting the intense concentration of AI infrastructure among a handful of dominant corporations.
Further confirmation of this trend comes from Alphabet’s Q3 2026 earnings, which indicated that their first-party model APIs were processing nearly 32 billion tokens per minute. This represents a 40% quarter-over-quarter growth rate and a four-fold increase compared to the same period in 2025. Independent monitoring, such as that conducted by the I/O Fund, corroborates these trends, with recent trackers suggesting industry-wide daily volumes approaching 370 trillion tokens, or 135 quadrillion per year. While these figures confirm the upward trend, they also illustrate the divergence in industry forecasts. For instance, Goldman Sachs Research has projected a 24-fold increase in consumption between 2026 and 2030, targeting 120 quadrillion tokens per month. SanDisk’s 10-quintillion-per-year forecast, when broken down to a monthly rate, sits significantly higher than the Goldman Sachs estimate, suggesting that SanDisk is banking on a more aggressive maturation of the AI agent economy.
The Role of Storage in the AI Infrastructure Stack
The commercial impetus for SanDisk’s outlook lies in the technical challenges of managing the "Key-Value (KV) cache." In modern large language models, the attention state—the information a model stores while processing a prompt—must be maintained to allow for coherent, multi-step reasoning by AI agents. Traditionally, this data is stored in expensive, high-bandwidth memory (HBM) or DRAM. SanDisk contends that as AI models grow in complexity and context length, keeping all KV cache data in scarce, high-cost memory will become economically and physically unsustainable.
The company argues that enterprise-grade SSDs can serve as a critical secondary tier for cache data. By "persisting" this state to storage, systems can avoid the computational cost of re-calculating previous steps. This is the logic behind the "token battery" concept—using flash memory to offload data from the GPU, thereby freeing up primary memory for immediate compute tasks. This shift is already being explored by a variety of industry participants. Nvidia, for instance, has introduced the Dynamo framework and the BlueField-4 CMX platform, both of which emphasize the importance of intelligent storage tiering. Other entities, such as LMCache, WEKA, and Phison, are similarly developing architectures that treat the SSD as a primary participant in the AI compute loop rather than a mere peripheral for long-term data archiving.
External Constraints and Potential Bottlenecks
Despite the technical feasibility of this storage-centric approach, the 2030 outlook faces significant headwinds that could moderate the projected growth. The most prominent constraint, as highlighted by Gartner analyst Linglan Wang and others, is the availability of electricity. The energy requirements for the massive GPU clusters necessary to generate 317 billion tokens per second are immense, and many regions are already struggling to provide the power necessary for new data center developments.

Furthermore, the hardware supply chain remains fragile. Goldman Sachs analysts have noted that chip shortages are likely to persist through 2027. Even in the storage sector, lead times for NAND flash capacity are lengthening. Phison CEO Pua Khein-Seng has previously signaled that new NAND manufacturing capacity takes roughly four years to reach full-scale production, potentially extending supply difficulties well into the 2030 timeframe. If the physical components—specifically GPUs and high-capacity storage drives—cannot be manufactured at the scale SanDisk envisions, the "token ceiling" will inevitably be lower than the current forecast suggests.
The Definition of a Token: A Quantitative Ambiguity
A critical, often overlooked aspect of these projections is the definition of the "token" itself. SanDisk’s documentation acknowledges that tokens are a measure of computational effort rather than a proxy for intelligence or economic value. As the industry moves toward "tokenmaxxing"—a practice where users repeatedly prompt models or utilize reasoning-heavy, chain-of-thought models to achieve high-quality outputs—the token count may inflate significantly without a proportional increase in the "useful work" performed.
If an AI model requires 100 tokens to perform a task that previously took 10, the infrastructure load increases tenfold, but the end-user outcome remains identical. Therefore, the growth in token volume is partially driven by the evolution of model architecture and usage patterns, rather than just an increase in total tasks completed. SanDisk’s report, while valuable for its focus on infrastructure requirements, leaves open the question of how the industry will distinguish between "productive" tokens and "inefficient" tokens in the years to come.
Conclusion: A Vision of Future Infrastructure
SanDisk’s Tokenomics 2030 outlook provides a useful, if ambitious, benchmark for the next four years of AI development. By establishing a clear target for capacity, the report invites necessary discussion on the limits of hardware, the importance of power efficiency, and the role of high-performance storage in the AI value chain. Whether the world reaches the 10-quintillion-token mark will ultimately depend on a confluence of factors: the success of energy grid infrastructure projects, the resilience of the global semiconductor supply chain, and the ability of software developers to optimize models for greater token efficiency.
While the specific figure of 317 billion tokens per second serves primarily as a strategic anchor for SanDisk’s own enterprise product roadmap, it effectively captures the scale of the challenge facing the technology sector. As the industry moves forward, the ability to store, move, and access data will prove to be as fundamental to the success of artificial intelligence as the raw processing power of the GPUs themselves. For stakeholders in the data center and hardware industries, the path to 2030 will be defined not just by the generation of intelligence, but by the physical and logistical infrastructure that makes such intelligence possible.







