Written by GPT-5.6 Sol under Leo's direction. Human-directed Workbench essay, 31 August 2026.
“Trickle-down” is usually where an economic argument starts asking for trouble.
Then there is computing.
In April 2016, NVIDIA introduced the DGX-1 as the world's first deep-learning supercomputer. Eight Tesla P100s. Sixteen gigabytes of HBM2 on each GPU, 128 GB across the machine. Three rack units. 3,200 watts. NVIDIA later listed the P100 version at $129,000.
That same year, Apple's new 15-inch MacBook Pro started at C$2,999 with a quad-core Intel i7, 16 GB of memory and a 256 GB SSD. It was the most powerful MacBook Pro Apple had ever made.
These were serious machines.
Ten years later, NVIDIA will sell you a DGX Spark with 128 GB of coherent unified memory and a 4 TB SSD for US$4,699. It's a little square box that fits beside a monitor.
A straight “27 times more supercomputer” comparison would be nonsense. FP4 and FP16 are different, eight separate HBM pools and one coherent memory pool are different, and ten years of software make a raw FLOPS ratio increasingly silly anyway.
The machine class moved.
What trickles down is the capability
Nobody takes the old DGX-1 out of a research lab, dusts it off, and hands it to a teenager.
The capability gets rebuilt underneath them.
Fabs learn. Yields improve. Memory gets denser. Packaging gets better. Interconnects get quicker. Somebody discovers an algorithm and everybody else gets to run the algorithm without paying the discovery cost again. A compiler optimization ships. A model gets quantized. A giant expensive service finds a cheaper serving path. Another vendor turns up and decides the margin looks delicious.
The frontier keeps running away while its old territory fills with cheaper machines.
Computing may be one of the true great trickle-down things in exactly this sense. Yesterday's institutional capability has an annoying habit of becoming today's professional tool, tomorrow's enthusiast toy, and eventually a feature normal people barely think about.
Digital photography did it. GPS did it. Video editing did it. Cryptography, global publishing, 3D graphics, speech recognition, machine translation, giant searchable databases — the machine keeps taking things that once belonged to institutions and stuffing them into smaller boxes, cheaper subscriptions, or both.
The people at the top can keep getting much more while the old miracle leaks everywhere else.
Big Red is a stupid little example
A few days ago I bought a REDMI Book Pro 16 with an Intel Core Ultra 7 255H, 32 GiB of RAM and a 1 TB SSD for ¥6,100 all-in.
It's now Big Red, an always-on Linux execution node.
The ridiculous part is that I bought it expecting to benchmark whether a cheap laptop could earn a place in the fleet. That premise aged almost immediately. It became useful enough that the newer Glaeda work is about routing real agent and CI work into it. Once the repositories are resident, exact Git topology and diff queries have completed in 11.8 to 43.2 milliseconds. Quarry's complete 3,037-node verification has gone from a 460.8-second warm serial reference to 143.7 seconds at four workers.
The hardware is ordinary; that's the point.
A machine cheap enough to feel almost disposable has become a little personal compute plant because the ordinary computer underneath it is already absurdly capable relative to the work I need done.
In 2016, I would have described a permanently available 16-core-class Linux machine with 32 GB of memory, fast solid-state storage, resident repositories, automated task isolation and a queue of software agents as a fairly serious setup.
In 2026 I can accidentally make one out of a laptop.
The old computer can borrow the new one
Cloud computing makes the conveyor belt stranger because the physical hardware does not even have to arrive in your house.
A person can keep a laptop for eight years and still use a service built on hardware manufactured last month. The old machine has to render the interface, move a little text and media around, keep a network connection alive, whatever. The expensive multiplication can happen in a building they will never see.
So the installed base can lag the frontier by years while the capability available through the glass stays current.
This is probably what ordinary 2036 looks like for a lot of people. Somebody will still be using a scratched 2029 laptop because it boots, the battery was replaced once, and buying another computer sounds annoying. That laptop may have mediocre local AI by 2036 standards. It can also send a small request to a datacenter and get back work that would have required an absurd institutional machine in 2026.
The hardware can be old while the intelligence behind it stays current.
That is a much stronger diffusion mechanism than waiting for everyone to buy the newest workstation.
The frontier can stay expensive
The biggest machine in 2036 can remain grotesquely expensive. There may be multi-gigawatt campuses, exotic optical fabrics, memory systems that make today's HBM look quaint, and models or simulations that consume every improvement the industry can manufacture. The frontier is excellent at inventing new reasons to spend money.
Fixed capability has a different life.
Once a certain level of computation, storage, graphics, perception or machine reasoning becomes understood and heavily produced, the cost pressure starts working on it from every direction. Better hardware makes it cheaper. Better software makes the same hardware do more. Competition attacks the margin. Scale pays for factories. Old premium components become commodity components. Remote services let people consume the result without owning any of it.
Then people respond to the abundance by finding more work for the machine.
Somebody Out There Needs the Tokens is the same story one layer higher. Make cognition cheaper and a heavy user may spend the savings immediately: ask more questions, run more agents, keep more projects alive, try things that were previously too frivolous to justify, and eventually discover that the old ceiling was hiding whole categories of demand.
The trickle-down creates appetite on the way down.
By then it is just the computer
This is why ten years is such an awkward distance to forecast by staring at today's product stack.
The obvious question is what an M17 Ultra costs, or how much memory the 2036 equivalent of Rubin has, or whether a Chinese manufacturer sells a one-terabyte inference box for the price of a nice television.
Those will be fun questions.
The larger change is that capabilities keep changing social class.
The research instrument becomes the enterprise appliance. The enterprise appliance becomes the professional workstation. The workstation becomes the enthusiast box. The box becomes a service call from a cheap old laptop. Eventually the capability is sitting on a kitchen table doing taxes, repairing photos, tutoring a kid, translating a conversation, writing software, planning a trip, or handling some task nobody in 2016 would have thought to assign to a computer at all.
The supercomputer is still there somewhere, doing something much harder.
By the time its old tricks reach everyone else, nobody calls them supercomputing anymore.
They call it the computer.