Change for the Machines
— with apologies to Pat Cadigan
By Jon Crowcroft and Hazem Danny Nakib
AI used to be all about a single equation about models: money equals compute equals big data equals valuations. Investors flocked to companies with impressive data centers or models, jam-packed with the latest GPUs or access to them, believing that more hardware or access to it automatically translated into superior AI capabilities. As a result, startups and tech giants were valued primarily by how many H100s they owned or had access to — almost as if GPU inventory alone guaranteed better models and therefore higher valuations. More often than not, these “models” turned out to be wrappers around the same foundational techniques, rather than groundbreaking innovations.
Then DeepSeek and other Chinese models changed the game, shining a harsh light on this flawed valuation approach. Suddenly, it wasn’t simply about buying more chips or brute-forcing bigger datasets; it was about engineering leaner, smarter models that got top-notch results on more modest hardware. Both private and public investors were dumbfounded; they had been pouring money into the idea that raw compute was the ultimate yardstick of progress. Overnight, confidence in that assumption faltered — especially among those still clinging to the notion that capital expenditure equalled guaranteed AI success. Compute did not mean AI supremacy. Basically it cast doubt on the last two years of public and private, markets not just of AI companies, but the entire stack, energy, chips, data centres etc.
But the writing was on the wall well before DeepSeek. A couple of years earlier, Meta released Llama, demonstrating how smaller, more efficient models could achieve many of the same goals as colossal ones. This echoed similar projects from the open-source world, such as DistilBERT and GPT-Neo, in which communities of enthusiasts and researchers proved that top-tier AI performance doesn’t require a warehouse of high-end GPUs. Because it was open-source, even though it came from BigTech, it went largely ignored. The infamous leaked Google memo, “We Have No Moat,” hinted at the same conclusion: the ability to innovate does not reside solely in giant corporations with infinite budgets. Smaller teams, using better optimization techniques and creative workarounds, can sometimes deliver results that match or surpass big models.
One of the ironies behind DeepSeek’s success is that its creators were constrained by export restrictions preventing them from buying the latest GPUs. In response, they concentrated on maximizing efficiency — a core principle the open-source and academic communities have championed for years. People often forget that, outside the realm of Large Language Models and Generative AI, many machine learning tasks have been solved on everyday hardware. Fields like medical imaging, signal processing, or real-time data analysis have thrived on optimized code rather than enormous hardware outlays. Even in the LLM space, community efforts like GPT4All, Alpaca, and others show that you can run surprisingly competent models on a single laptop or mid-range gaming GPU. The only difference with DeepSeek, is that it came from China, and that startled Wall Street and Silicon Valley.
Naturally, there were strong denials across the industry. After all, nobody likes to admit they’ve paid a premium for over-hyped compute and might have overlooked simpler, more elegant solutions, let alone the possibility that they have been part of developing and propping up an artificial market of buying and selling chips and compute to create models that could be done better for far less. Some speculated big tech players like OpenAI perpetuated artificially high barriers to entry to preserve a perception of exclusivity.
For investors actually interested in innovation, this is ironic given the entire direction of travel of much computing related tech has been to lower barriers so that innovation drives things with as low friction as possible (internet, cloud, processors, compilers, operating systems, SDKs/Appstores etc etc). This tactic is not new; it mirrors how some industries build protective “walled gardens” and justify them by insisting on specialized infrastructure. Meanwhile, the rest of the tech world has generally pushed against such barriers. Look at the internet, open-source operating systems, cloud platforms, or app stores — each lowered the cost of entry for new innovators and unleashed huge waves of creativity and innovation.
The consequences extend beyond finance. Consider future chip design, edge compute, and federated machine learning — all of which become more viable when the emphasis is on smaller, more targeted models. And then there’s the military example: modern warfare (cyber and otherwise) increasingly benefits from cheap, flexible AI solutions that can be deployed quickly and effectively. Why invest billions in specialized hardware that can be easily disrupted or blocked when a handful of optimized models can run on off-the-shelf devices? The same principle applies to cyberattacks; malicious actors will choose the leanest, simplest path to break systems, often leaving the more resource-intensive behemoths behind.
Moreover, today’s massive investments in GPU-based infrastructure — think Microsoft, NVIDIA, and various cloud providers — may be artificially inflating the market. The money goes in as investment, then flows back out in the form of buying more infrastructure, all justified by a valuation model that might already be outdated — the same is true for the multi-billion dollar series A funding rounds to buy compute to train models or create wrappers. History teaches us that barriers eventually fall and technology becomes cheaper. Personal computing, smartphones, cloud computing, and open-source software have all followed this pattern. AI is set to do the same.
We would think they’d learn the lesson, barriers always lower, things get commoditised, and things get cheaper and easier. This is not always just second system or indeed, third version syndrome — some better understanding of the domain can lead to major efficiencies, and sometimes they arrive combined with other useful innovations. We saw a great example of “less is more” in explainable AI (XAI) where tooling to uncover what structures within a neural network (“deep learning”) were responsible for detecting/recognising which input features (and hence classifying an input in some manner) — these tools for explainability also allow one to shrink the neutral network significantly by discarding nodes/edges that serve no useful classifier function — this has been used in face recognition in camera phones to make smaller, faster, and actually potentially more accurate AIs. The cost in training increases somewhat, but the payoff is that the cost in inference (done billions of times rather than just “once”) is massively reduced. In some AI models that approach can actually be used during training to reduce training cost too. So an innovation in one space driven by a required feature (explainability) leads to efficiency gains too. Researchers discovered that analysing which layers and nodes in a neural network actually contribute to classification can reveal parts of the network that do nothing useful. Pruning these nodes shrinks the model, making it faster, less resource-intensive, and often more accurate. While it increases training complexity, it drastically reduces the cost of inference — especially important when inference is run billions of times a day in applications like real-time translation or face recognition in camera phones. In some cases, these pruning techniques even reduce training costs.
Another approach gaining momentum involves blending physics-based models with neural networks. Systems like the Aardvaak weather predictor embed partial differential equations (PDEs) directly into neural network structures, reducing computational overhead while preserving the explainability and reliability of traditional physics simulations. This fusion not only leads to more efficient retraining and inference but also updates models in near real time, an enormous advantage for domains like weather forecasting, industrial automation, and autonomous vehicles.
Ultimately, profligacy — the habit of throwing massive resources at a problem — can hold back genuine breakthroughs. Constraints often spark creative solutions. From the open-source movement’s rapid improvements on laptops to the forced hardware limitations that gave rise to DeepSeek’s efficiency, it’s clear that real progress in AI and computing doesn’t need endless rows of GPUs. Instead, it thrives on innovation, optimization, and smarter engineering. As history has shown in every major tech wave, the path forward is usually the one that lowers barriers and democratizes access — not the one that locks it behind mountains of expensive hardware.
About the authors
Professor Jon Crowcroft is the Marconi Professor of Communications Systems in the Computer Lab at the University of Cambridge, Researcher-at-Large at the Alan Turing Institute and Visiting Professor at IX and the Department of Computing at Imperial College London. He is the co-Founder of the Data for Policy Conferences, one of the founding Editors-in-Chief of Data & Policy journal and a General Chair for Data for Policy 2024
Hazem Danny Nakib Hazem Danny Nakib is an esteemed professional known for his pivotal role in shaping digital strategy and innovation. As a Founding Member of the British Standards Institution’s (BSI) Digital Strategy Group, he has been instrumental in driving transformative initiatives that align with global standards. Hazem’s dedication to advancing digital frameworks and fostering collaboration has solidified his reputation as a thought leader in the field
