Artificial intelligence is evolving at a breathtaking pace, yet its progress appears to be driven not so much by profound leaps in algorithmic ingenuity as by the sheer scale of computational power fueling it. A new and comprehensive analysis conducted by researchers at the Massachusetts Institute of Technology (MIT) sheds light on this delicate balance, suggesting that the astonishing performance achieved by today’s frontier models—systems like OpenAI’s GPT-5.2 or Google’s Gemini—owes far more to immense computing resources than to any mysterious breakthroughs in programming methodology.
According to the MIT team, led by researcher Matthias Mertens, an expansive study of 809 large language models reveals a striking truth: the amount of computational power used during training eclipses other factors such as novel architectures, unique data-processing methods, or proprietary optimization techniques in determining model accuracy and sophistication. This well-documented phenomenon, known as the “scaling law,” encapsulates the notion that as computational resources grow—through larger clusters of GPUs, faster processors, and denser data pipelines—so too does a model’s performance on standardized benchmarks. The implication is profound: success in AI increasingly depends not merely on clever algorithms, but on access to monumental hardware infrastructure.
To test the precise influence of computing power, the MIT researchers meticulously quantified how training resources correlated with benchmark outcomes, comparing their estimates with indicators of algorithmic progress and proprietary innovations, colloquially referred to as each firm’s “secret sauce.” The results, captured in their paper entitled *Is there a ‘Secret Sauce’ in Large Language Model Development?*, posted on the arXiv preprint server, reveal that although original engineering contributes to advancements, its relative effect pales in comparison to the impact of raw compute. When OpenAI’s GPT-4.5 was contrasted with Meta’s Llama, the most salient difference was not an algorithmic technique or data source, but rather the monumental gap in computational energy employed to train them.
Mertens and colleagues identified four major contributors to AI progress: expanded computational budgets, shared community-wide innovation, model-specific refinements, and company-proprietary techniques. Yet, the single most influential factor remained the same—how much computational force each model could command. Quantitatively, the analysts discovered that the top-performing models, those in the 95th percentile of benchmark results, leveraged roughly 1,321 times more computational power than their counterparts at the 5th percentile. This thousandfold disparity indicates that AI leadership now hinges on the capacity to marshal vast digital infrastructure, making access to massive data centers and next-generation chips a decisive competitive advantage.
This finding carries weighty practical and economic implications. The researchers point out that hardware expenses are spiraling. Although the study avoided assigning dollar values directly, other industry analyses indicate that the costs of the semiconductors and networking gear crucial for AI expansion are escalating sharply. Indeed, recent reporting by Bernstein Research demonstrates that by 2025, chip prices are roughly 70% higher than pre-pandemic levels, with Nvidia’s GPUs—the engines of machine learning—commanding unprecedented premiums. Even as new generations of chips, such as Nvidia’s Blackwell and Rubin architectures, achieve incremental efficiency gains, organizations must acquire exponentially more units to retain a technological edge. This reality explains why corporations including Google, Microsoft, and Meta continue pouring hundreds of billions of dollars annually into compute infrastructure, and why OpenAI’s CEO Sam Altman is seeking transformative financing measured in the tens of billions, with long-term projections crossing the trillion-dollar threshold.
However, the story is not entirely dominated by scale and capital expenditure. The MIT paper offers a note of optimism: software ingenuity still matters—especially below the highest performance levels. While the most advanced systems rely predominantly on raw computing superiority, smaller developers have found that smarter algorithms can drastically reduce the computational demands required to achieve certain capabilities. Over time, innovations such as more efficient training schemes, better optimization functions, and model compression methods have reduced the compute needed to reach modest performance thresholds by as much as 8,000-fold. That means progress at the lower end of the spectrum is increasingly about refinement—doing more with less—rather than overpowering resource consumption. For these teams, the “secret sauce” often lies in engineering elegance: compressing power and ability into smaller, cheaper, and more sustainable models.
Ultimately, the research depicts an AI world divided between giants and innovators. At the frontier, companies with vast resources—OpenAI, Google, Anthropic, and similar organizations—will likely continue to dominate headline benchmarks through unrelenting investment in compute capacity. Yet, for smaller entities, the future may revolve around ingenuity and optimization, proving that exceptional engineering can, to a degree, substitute for the absence of colossal data centers. In essence, artificial intelligence today mirrors the broader pattern of technological evolution: while scale drives the frontier, intelligence in design enables survival and relevance for the rest. The true art of AI progress may not only be in building bigger machines, but also in learning how to achieve extraordinary results with surprisingly little power.
Sourse: https://www.zdnet.com/article/ai-isnt-getting-smarter-its-getting-more-expensive-mit-report-finds/
