The Means of Production
Sovereign intelligence: how unmetered AI ended up in your hands
I don’t think you realise how powerful your computer really is.
At the beginning of the 21st century, ASCI White, the most powerful supercomputer in the world, occupied a room, and cost US$110 million ($200 million now). Its 512 separate nodes each had 16 processors, 8192 CPUs in total, and were backed with 6 terabytes of memory, and 160 terabytes of disk storage. Together the servers weighed 106 tons, and consumed 3 megawatts of electricity with a further 3 MW needed for cooling, enough to power 6000 homes.
Remember, this was the most powerful computer in the world at the time. They used it to simulate thermonuclear warhead detonations.
ASCI White had a theoretical processing speed of 12.3 teraflops (that’s 12.3 trillion floating point operations per second). Twenty years later, $1500 would buy a consumer GPU, such as the NVIDIA RTX 3090, that delivered roughly three times more compute power. In 2026, its high-end successor the RTX 5090 is capable of over 100 TFLOPS. What once weighed 106 tons and required a 6MW dedicated power supply now weighs 2 kilos, runs on a standard wall socket, and is ten times faster.
Today, the $600 base M4 Mac mini delivers approximately 4.6 TFLOPS of GPU computing power, (or you can pay double for the M4 Pro chip and get 17 TFLOPS). The 16 dedicated cores of the Neural Engine in the base M4 are capable of 38 TOPS (trillion operations per second) for AI tasks. We’re no longer talking about specialised lab equipment — we’re talking about ‘supercomputer’ levels of inference capability living in a silent, silvery box that’s even smaller than your keyboard.
Ostensibly, the 12.3 TFLOPS of ASCI White still comfortably exceed the 4.6 TFLOPS of an M4 Mac. The fastest machine on the planet in the year 2000 was optimised to perform the high-precision math required for nuclear simulations. So what if we were to somehow give both machines the same AI task, like running a LLM to solve a logical puzzle?
The contest wouldn’t be close. The ASCI White, for all its $200 million might, would be strangled by its own millenial-era plumbing. While the Mac Mini can load a massive 8 billion parameter model from its SSD in seconds, the 106-ton supercomputer would likely spend twenty minutes just grinding through its ludicrously slow mechanical disks before it could even begin to ‘think’. The silent Mac Mini would finish the task before the supercomputer even finished loading the weights.
While ASCI White was a marvel for its time, its architecture was fundamentally unsuited for the high-speed integer math and massive memory bandwidth required by modern AI. Even if you compare it to the best specialised integer processor of that era: the Alpha 21264B, the M4 is in a different league. In general-purpose tasks, the M4’s Geekbench 6 score of ~3,800 dwarfs the Alpha’s estimated 500. In the specialised matrix math that’s fundamental to AI, the M4 is easily 50 to 80 times faster thanks to its dedicated silicon acceleration.
The Big Bang
A decade later, around 2011, technology reached a historic inflection point where consumer hardware surpassed the performance milestones previously reserved for room-sized supercomputers.
Before Intel introduced AVX (Advanced Vector Extensions), processors mostly did scalar math, that is, crunched one number at a time. AVX introduced Single Instruction, Multiple Data (SIMD), which was like having a toaster with multiple slots. Instead of one calculation per cycle, the CPU could now perform integer and floating-point math on vectors, multiple numbers at once, providing the massive parallel throughput that modern AI models require to function.
Around the same time, NVIDIA introduced CUDA, a Single Instruction, Multiple Thread (SIMT) architecture. This worked slightly differently than the CPU’s vector math, providing a whole army of single-slot toasters (collectively called a Warp). Individually, the processing cores of an NVIDIA chip were much simpler than those of a CPU, but there were vastly more of them. While a CPU might have a dozen sophisticated ‘polymath’ cores, the GPU brought thousands of specialised laborers to the task, allowing it to crush the massive, repetitive matrix math that defines modern AI.
So, by 2012, we could buy our own supercomputers, but AI was stuck. For decades, AI researchers had tried to teach computers to “see” by writing complex rules (e.g., “look for a circle to find an eye”). This approach not only didn’t work well, there was no way it could scale to recognise the infinite variety of the world around us.
The breakthrough that changed everything occurred not in Silicon Valley, but Toronto. A team led by Alex Krizhevsky, Ilya Sutskever, and Geoff Hinton realised that the NVIDIA GPUs (specifically the 2011 GTX 580) were capable of far more than rendering 3D games faster. It was no secret that the cards so beloved by gamers were massive parallel math engines, but unlocking that power was a dark art. Krizhevsky and Sutskever wrote the bespoke code to force these consumer-grade cards to perform the intensive matrix calculations needed to train a neural network. Rather than hand-coding rules for what an object looked like, they let the machine discover the patterns for itself across millions of images. The results didn’t just win the 2012 ImageNet competition; they made every other approach to AI obsolete overnight.
At the time, even neural networks themselves were considered to be a fringe idea. For decades they had been an academic curiosity, inscrutable black boxes, whose non-deterministic behaviour gave them a reputation for being notoriously fickle. There were precious few neural networks deployed in the real world, they were simply too difficult to train to the level of complexity necessary to reliably work on practical applications.
There was no inevitability to the triumph of neural networks, despite the superficial similarities with the neurons of our own minds. After a high-profile mathematical critique in 1969 (the Minsky-Papert affair), the field essentially died for a decade. Even by the time of AlexNet, if you were a serious computer scientist, machine learning meant Support Vector Machines or Random Forests. Neural networks were seen as an inexplicable dead-end that few believed would ever truly scale.
Intel Sandy Bridge CPUs and the NVIDIA 500-series GPUs provided the first affordable desktop environment that could handle the billions of operations required to train a Deep Convolutional Neural Network called AlexNet in weeks rather than years. The team had only 2 GPU cards, each with 3GB of RAM on each that required Krizhevsky to write highly customised “cross-GPU” parallelisation code just to fit the 60 million parameter model into the memory available. The results obtained were profoundly shocking, exceeding the best performance ever achieved on what was considered to be an almost computationally intractable task. AlexNet was the Big Bang of the modern AI era, when we began to realise everything had changed.
This intellectual revolution has been coupled with some almost preposterous technological magic. Extreme Ultra-Violet Lithography has miniaturised the transistors to scales so small they boggle the mind. A strand of human DNA is about 2.5 nanometers wide, the smallest features on the M4 chip are 3nm. This is how Apple can cram 28 billion transistors into a chip the size of a thumbnail, giving a silently running desktop the raw power of a turn-of-the-century supercomputer.
It is this confluence of affordable yet exceptional hardware and freely downloadable open weight models that effectively provides everyone their own base level of supercomputing. Anyone can have a supercomputer on their desk, and given the dizzying rate of software innovation, the effectiveness of that massive computing power is actually increasing as we sleep. Every new optimisation technique, from quantisation to speculative decoding, is essentially a free upgrade, making the device you bought yesterday even smarter tomorrow.
The Era of Open Weights
Owning exceptional hardware is only half the story, a supercomputer without software is merely an impressively miniaturised and extremely expensive space heater. The next breakthrough shock was the out-of-the-blue arrival of Open AI’s ChatGPT in November 2022. This felt like technology that had landed from the future, software that understood the spoken word, which could be creative, and which gave the very convincing impression it could actually understand.
Yet those who wanted to use ChatGPT were renting intelligence. Users couldn’t download it and run it for themselves, it could only be accessed on Open AI’s servers, and only on their terms. A few months later, in February 2023, Meta released the first Llama model. Whilst Llama wasn’t as functionally capable as ChatGPT, its unique attraction was something else: you could download it — all of it, and run it locally, on your very own machine. For the first time an artificial intellect could be run completely offline, beyond the reach of corporate gatekeepers. No one could turn it off, and no one could watch what you were typing.
Llama 1 wasn’t the first model to be released, but was the first to gain significant interest, exemplifying what came to be known as an open weights model. The weights are the results of trillions of calculations, years of training, and billions of dollars of investment.
The weights are a digital distillation of the sum of human knowledge. To create them, a “furnace” of ten of thousands of GPUs is required, burning millions of dollars of power every day for months. Historically when software was open sourced, it was the release of many man-months of programming labour, saving all who used it the effort of reinventing the wheel. AI model weights are different, they are the pre-computed results of a massive-scale industrial process that few organisations on the planet are capable of replicating. Meta had made the strategic decision to release the genie from the bottle, and in doing so ignited an AI revolution.
This is Game Theory is action. When all innovations are secret, there’s an incentive for everyone to preserve secrecy, in the hope that they might be the ones to capture the monopoly and win the zero-sum game. This also applies to new entrants to the contest, this is why startups, with their limited starting resources, don’t launch with a highly detailed white paper describing their particular edge. They’re playing to survive in a world where information is the only currency, and giving it away feels like unilateral disarmament.
But when a huge incumbent like Meta decides to defect, the game changes radically. Once the genie is released, the remaining proprietary players lose their path to a total hegemony; the incentive for secrecy begins to dissolve as the ‘open’ alternative becomes the global standard.
Models released in public become the subject of a million simultaneous experiments. The vast majority of the world’s developer talent doesn’t work for a Silicon Valley giant, and suddenly, that talent has a side to back. Releasing your secrets is a lousy way to build a monopoly, but it’s the perfect way to build an ecosystem. That is how science moves, not through the accumulation of secret hoards, but through the compounding of shared discovery.
The existence of competitive open-weight models triggered a seismic geopolitical shift. In France, Mistral announced their models would be free to download and they’d pay the bills by charging metered access. Black Forest Labs in Germany followed suit, though keeping their best model to themselves. By flooding the ecosystem with high-performance models like Alibaba’s Qwen and DeepSeek’s R-series, Chinese labs turned the source code of intelligence into a global commons. In doing so, they ensured that while the US might control the most advanced chips, China would have a significant say in who provided the most widely used brains.
I personally believe the best defence against AI hegemony is to ensure everyone has it, and no one gets to monopolise it.
I’ve always wondered if the Chinese embrace of freely distributable AI models is a manifestation of Marx’s notion that the means of production should be owned by the many, and not captured and extracted by a plutocratic few.
Chinese political culture (influenced by Confucianism and Mao-Marxism) emphasises harmony, group welfare, duty to society and state, and long-term collective progress. The Party frames policies around “common prosperity”, national rejuvenation, and subordinating individual pursuits to societal and state goals. In this light, endorsing the release of powerful models as open weights isn’t just a competitive tactic; it’s an ideological statement — that intelligence should be treated as a socialised utility which empowers the collective.
This echoes Marx’s view of humans as inherently social beings whose essence is shaped by social relations, rather than Rand’s sovereign individual whose highest virtue is rational egoism. In Rand’s world, any state intervention beyond minimal protection of rights is immoral coercion.
Rand’s Objectivist credo still haunts much of the Western tech elite. In this worldview, transformational intelligence is a private discovery, the fruit of individual heroic genius and the risk-capital that funded it. To open it would be to surrender a hard-won competitive advantage to ‘looters’ who didn’t do the work. John Galt would never share his weights.
Modern AI is a battle of spirits: the Randian West would love to build oracles, gods behind curtains you pay to consult. The Confucian East would prefer to build a substrate, where the entire country is elevated to become pretty damn smart.
In China, the state (under CCP leadership) actively directs resources, sets industrial policy, and enforces alignment with “core socialist values”. The national AI strategy is to build technological sovereignty, diffuse capabilities broadly (to avoid monopoly capture by a few Western firms), and mitigate the inevitable disruptive effects of new technologies on social cohesion.
When Deng Xiaoping introduced his reforms, he justified them as a necessity to ‘liberate productive forces’. Yet the technocrats running China today are keenly aware that technology often makes old systems — and their masters — obsolete. In a culture that remembers inventing gunpowder and printing only to see them reshape the world in ways they couldn’t control, there’s a deep-seated fear of a new generation of tech oligarchs, “capitalist roaders” who might steer society away from its collectivist ideals.
The enthusiastic embrace of open-weight AI disperses these productive tools with a speed that borders on the viral. As open-weight models are simply files, data that does not ‘phone home’ to China, Western users can build entirely sovereign services upon them. Already Valley companies like Airbnb have publicly acknowledged their adoption of Alibaba’s Qwen for its superior speed and lower operational costs. These models represent the first Chinese software artifacts to achieve organic, high-impact penetration into the very foundations of the Western tech stack.
By late 2025, the global AI landscape had undergone a total market inversion. Alibaba’s Qwen family had not only challenged Meta’s Llama — it had surpassed it as the most downloaded open model family on Hugging Face, crossing 700 million cumulative downloads. In December 2025 alone, Qwen’s single-month downloads exceeded those of Meta, Mistral, and Google’s Gemma combined.
Unlike proprietary ‘Oracles’ that prioritise raw power at any cost, open models are incentivised to minimise inference expenses and squeeze performance out of modest hardware. These practical pressures make the ‘Substrate’ the pragmatic choice for scaling. It is a mistake to value models solely on ‘thinking’ scores from benchmarks that ignore economic reality that thinking is not a trivial computation, and it must be paid for. To do real work in the real world, we must consider the cost-to-performance ratio: or the economic logic of the model. It seems strange to me that so many current benchmarks are so one-dimensional, focused solely on smartness. Only a few (like the ARC prize and GDPval) measure intelligence through the lens of efficiency and performance-per-dollar.
Permissiveness is another aspect of crucial importance to builders. Open models released under licences in the spirit of Apache 2.0 have the same terms of use as other fundamental parts of the modern software stack, free from the rug-pull risk of proprietary APIs. All models have some degree of censorship — or “guardrails” as we like to call it in the West, whether that’s refusals over politically sensitive topics, or content that would encourage criminality.
It’s worth clarifying that when we talk about open AI models, we’re referring to freely distributable weights rather than the full recipe of code and data required to build them from scratch. Historically, as so much of human knowledge remains copyrighted, it was much easier for builders not to ask awkward questions about how the sausage was made.
Even if Western regulators do demand transparency, the reality is that the weights themselves have become a permanent part of the global substrate. A lab in Shenzhen or Hangzhou will train their model with every document they can obtain, provided the final product aligns with domestic values. A 2026 Stanford HAI report notes over 60% of new fine-tuned models on HuggingFace are already derivatives of Chinese base models. Future models, regardless of where they are built, will be distilled from their ancestors, effectively inheriting their knowledge without needing to revisit the original controversial recipe.
It will be fascinating to see the effect of this widespread adoption on the immense Chinese labour force. Anxiety about technological unemployment is real; as AI makes entire professions obsolete, the state — which prioritises stability above all else — faces its greatest economic challenge. For the CCP, solving this isn’t just about GDP; it’s about maintaining the modern version of the Mandate of Heaven. If the technological revolution leads to social fracture rather than common prosperity, everything falls.
AI too cheap to meter
Unlike broadband internet, which involves a flat fee for connection (in my part of the world anyway), AI is a metered resource. Though the tech giants graciously provide a quota of free usage for their most powerful models, this is intelligence by subscription, with serious users needing to pay for what they consume.
The notion of a service being ‘too cheap to meter’ is attributed to Lewis Strauss, now best known as the schemer played by Robert Downey Jr. in the movie Oppenheimer. In 1954, as chairman of the U.S. Atomic Energy Commission, Strauss championed a Cornucopian ideal: the great hope that nuclear power would be so efficient that the cost of production would eventually vanish
In practice though, when the means of electricity production are communally owned and centrally managed, stagnation results. In the last years of the Soviet Union consumers did pay for electricity, but it was heavily subsidised by the state, more a nominal fee than a market-driven utility rate. It created a system where nobody really cared enough to prevent the compounding errors that ultimately led to a nuclear reactor exploding.
AI is metered by the number of tokens, a proxy for the raw compute your task consumes. When I look at the stats of my command-line agentic coder, it’s not uncommon to see a complex task use 200,000 tokens over multiple iterations. At time of writing in early 2026, the bill for a single morning of work can be startling:
- On Claude Opus 4.5 ($25/1M output): Every failed iteration costs several dollars. A professional developer could easily run up a bill of $1,000 a month on API fees alone, with no guarantee of perfection.
- On GPT-5.2 ($14/1M output): Half the cost, but if the model lacks the ‘coding soul’ of Claude, you may end up spending the same through more frequent iterations.
- On GLM-4.7 ($2.20/1M output): At a tenth of the price, each task costs less than a dollar. It is the pragmatic choice for scaling, but even those dollars add up if you want to move fast and break things.
This is the reality of intelligence by subscription. One pays a premium to consult the Oracle, while the dream of ‘Broadband AI’ remains a gated privilege. According to the benchmarks, Claude Opus is the best coding model right now. But it’s also priced like heavy plant hire, whereas the whole attraction of DIY is buying simple tools once, and doing the job yourself, with no extra charges. Usage costs are significant overheads for freelancers. That’s why professionals buy their tools rather than hire them.
Imagine if every operation on your home computer were metered. It would have an inhibiting effect, you’d only use it for high-value tasks. When you find yourself mentally justifying the ROI of a single “thought”, friction has been introduced. A tax on curiosity leads to token anxiety. When every error is a micro-transaction and every experiment carries a price tag, the human urge to tinker — the very engine of innovation and progress — is quietly smothered.
Open weight models are unmetered, but historically what held them back was they weren’t quite good enough. You’d just end up with messy buggy code that would waste your own precious time fixing. The following table shows the advances in capabilities of the top-performing proprietary and open weight models at the beginning of the last four years.
The HumanEval benchmark measures a model’s ability to solve isolated, LeetCode-style coding puzzles, whereas SWE-bench Verified is the industry’s stress test for real-world engineering, requiring AI to navigate entire repositories to identify, fix, and verify actual GitHub bugs.
In 2023, soon after ChatGPT made headlines, it could only solve about half of the LeetCode problem set, and it stumbled badly when trying to fix real world codebases. At this time, Llama was yet to be released, and developers had to wait until August 2023 for Code Llama, which had been fine-tuned for coding. Most software engineers thought of AI as a potentially useful pair programmer rather than an existential threat to their occupation.
In the years since the ability of models to autonomously write and maintain code has increased remarkably, to the point where “vibe coding” has entered our vocabulary, and coding agents like Claude Code are now everyday contributors to genuine software engineering tasks.
My motivation for writing this article is to commemorate a moment in time, which is yet to arrive, but I believe it is imminent. Squint into the future, and it doesn’t take much imagination to envisage open weight models that fit on inexpensive desktops, diligently working in the background crafting and improving code without their owners ever worrying about the cost.
Historically, models good enough to write practical code were slightly too big to run locally, but the inevitability of progress means that inflection point is fast approaching. Just as you don’t pay per sentence to use a word processor, generative coding is about to become a background utility: limitless, private, and entirely yours. When creativity is unmetered, the only limit is what you can imagine.
We Own The Means of Production Now
We are privileged to live at the confluence of two modern miracles: extraordinary yet affordable supercomputers, and free-to-use software we can literally talk to, which is not only smart enough to think, it contains the distilled sum of human knowledge.
The M4 Mac Mini is the ultimate physical manifestation of how collaboration leads to greater progress than rivalry. Think about the path of a single thought on that machine. An entrepreneur in London prompts a local agent. A California-designed M4 chip executes the logic. A Chinese-trained open weights model provides the intelligence. All of this occurs on a machine assembled in China using silicon fabbed in Taiwan. In the end, the most powerful ‘means of production’ in history wasn’t built by a single nation or a single ideology. It was built by the best of humanity, and now, it belongs to you.
Whoever owns and controls the means of production captures the gains. Open models are shrinking rapidly, and what you can run locally may already be good enough. If you own a machine as powerful as an M4 Mac, you have at your fingertips an extraordinary means of production, an entirely permissionless engine for getting things done.
It’s an aphorism amongst entrepreneurs that ideas are cheap, and execution is everything. But the cost of execution is in freefall. Coding agents have moved from novelties to necessities, permitting anyone to realise anything they can imagine — at least, anything software-shaped.
Real-life problems are rarely software-shaped at first, but that is the new art of entrepreneurship: imagining how they might be, then using your own desktop atomic furnace to smelt the rough edges away until they are.
The means of production are now in our hands.
— Jaron, January 2026
