Local AI is already good enough for most tasks, and the frontier gap is closing faster than frontier labs would like to admit

Greg Isenberg puts a number on something that practitioners have been quietly observing: a model running on a gaming GPU or a decent Mac is good enough for about 80 percent of what most people use cloud AI services for. That figure is a ceiling estimate, not a floor. Chamath Palihapitiya goes further, arguing that 95 percent of practical AI tasks can now be handled by many different models, not only those from frontier labs. He adds, with some candor, that 98 percent of what he was sending to Claude could have gone to a cheaper alternative.

The cost picture is the more striking half of the story. Dylan Patel notes that DeepSeek achieved GPT-4-level performance at 1/600th the cost. Marc Andreessen puts distillation at roughly 2 percent of a model’s original pre-training cost. David Friedberg describes open-source alternatives running 90 percent cheaper than frontier options. Harry Stebbings works through the compounding math: raw chip performance improves roughly 3x every 18 months, and additional optimizations, quantization among them, contribute another 3x, putting the improvement in tokens per unit of money at roughly 10x every couple of years. Friedberg pegs the efficiency trajectory at five to ten times every 12 to 18 months. These are not the same number, but they are pointing the same direction.

The capability gap between local and cloud is meanwhile closing on a measurable schedule. Jesse Genet predicts that within 18 months there will be local models at the level of Claude Opus. Joseph Nelson, approaching the question from the deployment side, observes an 18-month lag between a capability becoming available in a multimodal cloud model and that same capability being achievable on an edge device. The two estimates are independent and land in the same place. Zvi Mowshowitz adds that certain frontier capabilities will be widely available in open source within a handful of months, a shorter horizon still.

It's not that 95% of the bleeding edge tasks can be done by everybody. It's that 95% of the tasks can be done by many different models. That's the actual answer. Chamath Palihapitiya

Practitioners are not waiting for the gap to close. Jesse Zhang says that 90 percent of the workflow at Decagon, the AI customer-support company, already runs on open-source models. Brad Gerstner reports that companies he has spoken with are routing 90 percent of their usage to open models. Gerstner also notes that two companies he engaged with are each working on their own models. Simon Mo points to a structural reason for the shift: application builders who need to do their own mid-training and post-training cannot do that with closed-source vendors, so open-weight models are not just cheaper, they are technically necessary for serious customization.

The performance case for smaller models is not limited to cost. Matei Zaharia describes pipelines running entirely on open-source models, where the same model generates training environments and trains itself, and the result beats frontier models at specific tasks. swyx notes that for a small classification task, a fine-tuned model at around 800 million to one billion parameters can recover 95 percent of frontier performance, at a fraction of the latency and cost. Ramin Hasani adds that a one-billion-parameter model combining a small number of attention layers with a simple gated learned convolution can run fast enough on an iPhone to be a real option for basic use cases such as privately searching and classifying local data. At the extreme end of efficiency, Hasani describes liquid neural networks that fit in one to 25 megabytes and run on a CPU, including hardware as modest as a Raspberry Pi.

The economics of the consumer case are also shifting. Genet built a fully custom curriculum for her family using a local model for eight dollars. She predicts that when cloud AI costs for a household reach around 400 dollars a month, the motivation to switch to local will be less about privacy than about not wanting to pay the bill. That framing is significant: it puts the mass adoption of local AI on a straightforward cost trigger rather than a philosophical one.

Brendan Foody argues that within five years the majority of inference will run on open-source, custom fine-tuned, or distilled models rather than frontier APIs. Rory O’Driscoll’s observation that AI model training costs must be recovered within 12 to 24 months before a model is obsolete suggests that frontier labs face a structural clock that open-weight competitors do not. Friedberg offers one plausible equilibrium: frontier tokens capturing 65 to 85 percent of economic value while open-source tokens account for 80 percent of volume. That is a world where frontier models remain the premium option for the hardest problems while local and open-source models absorb the bulk of everyday inference. The question is not whether that world arrives. It is how many businesses built on frontier API margins are positioned to survive it.