Sam Altman thinks your laptop is hoarding a resource it barely uses. He says roughly nineteen out of every twenty megabytes in personal computers sit idle, waiting for a browser tab or a spreadsheet that never comes. The fix, he suggests, is obvious: pull that memory into cloud data centres where AI workloads could actually sweat it. The argument sounds like textbook efficiency until you notice who is standing in line ahead of you for that same RAM.
The AI industry Altman leads has spent the last two years vacuuming up high-bandwidth memory and high-density DDR5 modules at volumes that have distorted global supply chains. SK Hynix, Samsung, and Micron are scrambling to expand fabrication capacity for HBM, the stacked memory that sits next to GPUs in AI accelerators. Spot prices for some DRAM segments jumped more than twenty percent quarter-on-quarter through late 2023 and early 2024. Server demand from the same cloud providers Altman now wants to feed drove these increases. The scarcity he frames as a distribution problem is, in large part, a demand problem his own sector created.
The Logic of Centralisation and Its Blind Spots
Altman’s proposal treats personal computing as a rounding error in the global memory budget. Local machines are glorified typewriters with overbuilt specs, and real computation belongs in consolidated infrastructure. This is not a new idea; cloud computing has been making versions of this pitch for fifteen years, but the brazenness of the framing is notable. A CEO whose company burns through memory at scale is telling individual users that their hardware is the inefficiency.
The blind spot is structural. Personal RAM is idle because modern operating systems pre-allocate, cache aggressively, and keep resources ready for unpredictable spikes. That 95 percent figure measures moment-to-moment utilisation, not actual waste. By the same logic, a fire station with trucks parked ninety-five percent of the time is misallocating engines. The comparison collapses because readiness has value. Your laptop’s memory is not doing nothing; it is doing nothing visible to a cloud provider’s utilisation dashboard, which is a different claim entirely.
What This Means for South African Workflows
For businesses and developers working here, the centralisation push carries specific friction. Cloud-native development is already the default for many teams, but the shift Altman describes would make local experimentation, offline debugging, and self-hosted infrastructure harder to justify. In a market where fibre costs vary wildly by suburb and mobile data remains expensive relative to income, that friction has material consequences. A developer in Johannesburg or Cape Town who trains models on local hardware avoids egress fees, latency, and the POPIA complications of shipping sensitive data offshore. Framing that local RAM as wasted ignores the cost structure and connectivity realities that make cloud dependency painful.
The subscription economics are equally lopsided. A one-time RAM purchase becomes a perpetual rental. Small businesses that once bought workstations with upgradeable memory would face escalating operational costs tied to cloud instance pricing. This pricing is sensitive to the same scarcity Altman’s industry engineered. The lock-in is not theoretical. Migrating a workflow from AWS to Azure or back to on-premise involves re-engineering costs that few mid-sized companies can absorb casually.
The Environmental Bill
A carbon ledger is also being ignored. AI data centres consumed 15.4 terawatt-hours at Google alone in 2022, and water usage for cooling has become a flashpoint in drought-prone regions. Centralising RAM into these facilities does not reduce total energy demand; it concentrates it. Local machines draw power too, but they do not require redundant cooling, round-the-clock security, and the transmission losses of moving data across continents. If the goal is genuine resource efficiency, the calculation must include embodied energy and infrastructure overhead, not just utilisation percentages.
Alternatives That Do Not Require Surrendering Your Hardware
Several approaches distribute AI load without treating personal computers as donor organs. Edge computing keeps inference close to the data source, reducing both latency and bandwidth costs. On-device AI, enabled by model quantisation and pruning, now runs useful workloads on hardware that fits in a phone. Federated learning trains models across distributed devices without centralising raw data, which aligns better with privacy law than shipping everything to a Virginia server farm.
These techniques are not theoretical. Quantised versions of large language models already run on consumer GPUs with acceptable latency. The constraint is often engineering effort, not raw memory availability. Altman’s framing skips past this optimisation layer because it does not serve the business model of hyperscale rental.
The Deeper Pattern
What reads as a technical argument about memory allocation is, at base, a land grab for computational territory. The AI industry needs cheap, abundant infrastructure to make its unit economics work. Convincing users that their local resources are wasted is a rhetorical precondition for extracting them. The hypocrisy is not incidental; it is the mechanism. The same firms driving HBM prices beyond the reach of researchers and gamers now position themselves as the only responsible stewards of global memory supply.
For practitioners building real systems, the response is not to reject cloud tooling entirely but to maintain optionality. Keep local hardware capable of meaningful work. Treat cloud costs as variable expenses that can spike. Build pipelines that run in either environment. The RAM in your machine is not a wasted asset waiting for Altman’s permission to matter. It is a hedge against concentration, connectivity failure, and pricing power held by the same voices telling you it belongs to them.
