I. The Landlord at the Cheeky Pint
There is an intoxicating candor that arrives around the bottom of a third pint, particularly when two men whose collective equity stakes could purchase a medium-sized European republic sit down with microphones between them.Appearing on the *Cheeky Pint* podcast hosted by Stripe co-founder John Collison, Sam Altman—the supreme orchestrator of OpenAI and chief evangelist of human obsolescence—delivered what may stand as the definitive slip of the twenty-first-century silicon aristocracy. The topic under consideration was not the existential hazard of recursive neural self-improvement, nor the geopolitical sovereignty of the Taiwan Strait. It was something far more pedestrian, tactile, and stubborn: dynamic random-access memory, or DRAM.
With the sorrowful detachment of an eighteenth-century duke observing tenant farmers squandering fertile topsoil on mere cabbages, Altman offered his verdict on consumer hardware:
*"From a global allocation perspective, it is quite terrible that so much of the world's DRAM sits in consumer devices, utilized at perhaps two or three percent duty cycle, when it could be sitting in hyperscale clusters operating at ninety-eight percent efficiency."*Pause and turn that sentence over in your mind. Notice the exquisite, antiseptic beauty of the phrasing: *a global allocation perspective*.
It is the dialect of the Soviet Gosplan planner dressed in a Loro Piana cashmere sweater. It is the lament of an empire builder who looks at a university student’s battered ThinkPad in Manchester, a software engineer’s gaming rig in Bengaluru, or an animator’s Mac Studio in Brooklyn, and does not see instruments of private thought, personal play, or creative autonomy. He sees thirty-two gigabytes of unmonetized silicon sitting insolently offline while his own five-gigawatt Stargate cluster thirsts for bandwidth.
To the high priests of centralized intelligence, your personal computer is no longer a monument to the democratizing genius of Steve Wozniak or Alan Kay. It is a leak in their supply chain.
---
II. The Thermodynamic Sins of the Peasantry
To understand the breathtaking audacity of Altman’s critique, one must first appreciate the theological framework of modern Silicon Valley. In this orthodoxy, human history is merely an inefficient intermediate state between cosmic thermodynamics and artificial general intelligence. Compute is not a commodity; it is holy water. And holy water must never be sprinkled on houseplants.Consider the crime of which you, the ordinary citizen, stand accused.
You walked into an electronics shop or clicked a button on a website, paid two thousand dollars of taxed personal savings, and purchased a machine containing sixty-four gigabytes of unified memory. You opened a text editor. You drafted a letter to your mother. You browsed three recipes for chicken soup. You played ninety minutes of a role-playing game in which an elf slays a dragon. And then—horror of horrors—you closed the lid, made a cup of tea, and went to sleep.
For eight continuous hours, eighty billion transistors sat cold, idle, and unapologetic.
To a man whose company is burning upwards of five billion dollars a year in cloud infrastructure costs, and who is currently traversing the globe with a begging bowl looking for seven trillion dollars to manufacture sovereign silicon, this is nothing short of criminal negligence. Why should an accountant in Ohio retain private custody of high-bandwidth memory chips when those very same dies of silicon, soldered onto an Nvidia Blackwell server blade in an Iowa cornfield, could be continuously inferencing synthetic poetry and automated corporate customer service bots at three o'clock in the morning?
Under the gospel of the *global allocation perspective*, private ownership of computing machinery is a moral hazard. If you are not utilizing your memory at peak thermodynamic capacity, you have no business owning it. You should, by all rational dictates of technocratic optimization, surrender your hardware to the central clearinghouse, surrender your local terminal, and rent intelligence back five cents per thousand tokens through a metered API gateway.
One is reminded of the enclosure acts of seventeenth-century England, wherein common lands were seized by aristocratic landlords on the impeccably logical pretext that centralized pastoral sheep-farming was far more productive than peasants grazing their scrawny cattle. The economic rationale was unassailable; the peasants simply had the misfortune of wishing to feed their families without asking permission from the lord.
---
III. The Inconvenient Rise of the Local Model
What makes Altman’s lamentation genuinely delicious, however, is not its theoretical arrogance, but its desperate economic timing.For two years, the foundational bet of OpenAI, Microsoft, and Google has rested on an unspoken premise: that frontier intelligence is an exotic, ultra-scarce substance that can only be distilled within multi-billion-dollar cathedral data centers cooled by diverted rivers and powered by dedicated small modular nuclear reactors. Under this model, the user is permanently cast as a dependent supplicant. You do not own the model; you query the model. You do not see the weights; you rent the weights. You do not adjust the moral guardrails; you accept whatever sanitized output the trust-and-safety committee in San Francisco deems safe for your delicate eyes.
And then, something deeply inconvenient occurred: the open-weight revolution arrived.
A French startup called Mistral released open models that ran with blistering velocity on consumer laptops. Meta, in a fit of open-source chaos engineered by Mark Zuckerberg to spite his cloud rivals, released Llama 3. Chinese research labs dropped DeepSeek models that matched American proprietary benchmarks at an eighth of the compute budget. And suddenly, developers and curious teenagers discovered a secret: with quantized weights and thirty-two gigabytes of unified RAM, an ordinary person could run a world-class reasoning model entirely on their desk.
No credit card required. No API keys logged in an Azure telemetry database. No corporate telemetry scraping your private journal entries. No automated content filter wagging its finger because your historical research mentioned a sixteenth-century plague or an uncomfortable political event.
When a citizen runs a local model on their own computer, OpenAI receives zero dollars. Microsoft logs zero billable compute seconds. The user’s data never crosses an underwater fiber-optic cable. The user’s identity is invisible.
This is the real nightmare lurking behind Altman’s *global allocation perspective*. It is not that the world’s DRAM is physically wasted; it is that the world’s DRAM is providing ordinary citizens with an exit ramp from the subscription feudalism he is attempting to construct.
---
IV. The Feudal Psychology of Rent-Everything
It is a well-documented law of monopolistic technology that every radical liberation movement eventually curdles into a utility company with a private police force.When OpenAI was incorporated in December 2015, its charter was practically dripping with utopian socialist idealism. It was a non-profit. Its mission was to ensure that artificial general intelligence would benefit "all of humanity." It promised to distribute its research openly, freely, and universally, specifically to prevent any single corporation—chiefly Google—from establishing an unassailable monopoly over cognitive automation.
Fast-forward a decade. The non-profit shell is being dissolved into a for-profit Delaware C-corporation. Equity packages worth tens of billions are being carved out for leadership. Multi-gigawatt nuclear partnerships are being signed with sovereign wealth funds. And the CEO of that same utopian non-profit is sitting on a podcast with a pint of beer, publicly begrudging the fact that ordinary people have the insolence to purchase consumer laptops with sufficient RAM to run software without paying him toll-booth revenue.
Notice how the definition of "humanity" changes depending on where you stand in the funding cycle. When you need tax-exempt charitable status and donations from tech billionaires, humanity is a collective of sovereign individuals deserving of empowerment. When you are attempting to raise an astronomical valuation and defend a proprietary API monopoly against open-source alternatives, humanity is a horde of inefficient consumers hogging DRAM that rightfully belongs in your server racks.
The argument that central clouds are more efficient than distributed personal machines is old enough to collect a pension. IBM made precisely the same argument in 1974. They explained, with great mathematical precision and impeccable corporate diagrams, that individual personal computers were an absurd extravagance. Why would a company or an individual want a microcomputer sitting on a wooden desk, when a centralized System/370 mainframe in a climate-controlled basement could process payroll and database queries with five times the hardware duty cycle?
Steve Jobs and Steve Wozniak answered that question not with an economic spreadsheet, but with a philosophical revolution: *the bicycle for the mind*. A personal computer was never supposed to be an optimal industrial asset. It was supposed to be a tool for personal liberty. It belonged to you. You owned the silicon. You controlled the switch. If you wanted to run a physics simulation, write a sonnet, or leave it idle for three weeks while you walked through the woods, that was your sovereign human prerogative.
---
V. The Memory Belongs to the Builder
There is an ancient Latin proverb that warns against listening to the butcher when he lectures on the moral virtues of vegetarianism. One should exercise comparable caution when the operator of the world’s largest cloud inference engine lectures the public on the macro-allocation of memory chips.If the global supply of DRAM is constrained, it is not because a gamer in Seoul bought thirty-two gigabytes of DDR5 to play video games with his friends on a Saturday evening. It is because Silicon Valley has engaged in a speculative infrastructure arms race of staggering proportions, consuming every silicon wafer manufactured by TSMC and every memory die packaged by SK Hynix to train models whose primary corporate application thus far has been automating spam, summarizing meetings that should have been emails, and generating synthetic video clips.
To look upon the private possessions of ordinary people and see an inefficient aggregate that ought to be subsumed into an elite central infrastructure is the purest definition of technocratic hubris.
Keep your sixty-four gigabytes of RAM. Leave your laptop open on your kitchen table. Let the memory sit idle while you watch the sunset or read a paperback book. And when you next open your terminal to run an open-weight model entirely offline, in the quiet sanctuary of your own machine, remember that every unmonetized byte of DRAM on your desk is a tiny, beautiful fortress against the landlords of the digital future.
