For twenty years the smartest advice in tech was to stay out of atoms. Bits scale because copying is free. Atoms need factories, trucks, permits, and people in hard hats. Software ate the world precisely because it didn’t have to touch any of that. Every founder of my generation absorbed this before they wrote their first line of production code.
Cloud AI looks like the purest bits business ever built. An API. A text box. Intelligence as a subscription, delivered at the speed of light, no warehouse in sight.
Peel back the API and you find the largest atoms buildout in industrial history. GPUs by the million. Substations. Water rights. Land deals in Virginia and Texas. The hyperscalers are spending hundreds of billions of dollars a year on physical infrastructure, and the chips at the center of it depreciate faster than almost any capital asset a company can buy, because the next generation obsoletes them before the accountants finish the schedule.
Every token you buy is a meter running against someone else’s depreciation.
Every token is rent on someone’s atoms
When your product is inference, your cost of goods is physics. Power in, heat out, silicon aging in racks. The cloud AI vendors carry those atoms on their balance sheets, which means their business model has one non-negotiable requirement: the meter must run, forever, at a margin that services the buildout.
This explains behavior that otherwise looks strange. Message caps on $200 subscriptions. Model tiers that route you to something cheaper when you aren’t looking. Usage-based pricing creeping into products that launched flat-rate. None of that is product strategy. It is atoms economics leaking through a bits interface. The vendor prices against their own compute because they have to.
I wrote last month that the model is the depreciating asset. The fuller picture is that the model sits on top of a second depreciating asset, the hardware it runs on, and the token price has to recover both. You are renting a wasting bit that runs on a wasting atom, and the invoice arrives monthly.
The other fleet is already paid for
Here is what the capex race ignores. Apple ships on the order of two hundred million iPhones a year, plus tens of millions of Macs and iPads, every one of them carrying neural compute that would have been a datacenter part not long ago. Add the Android flagships and the installed base of capable consumer silicon becomes the largest distributed inference fleet on earth.
It sits idle. Your Mac spends most of its life at a few percent utilization. The M-series chip in it was designed to sprint and it mostly waits.
The economics of this fleet are upside down from the datacenter, in the good direction. The capex is sunk, paid by the user, amortized against photos and email and spreadsheets. The marginal cost of a token generated on it is electricity, which rounds to nothing. There is no fleet power bill to recover, no utilization target to hit, no depreciation schedule that someone’s margin depends on.
And the property that matters most: nobody can put a meter on it. A vendor can meter their own atoms. Nobody can meter yours.
The bits arrived to light it up
Idle atoms are worthless without bits that fit them, and for years the bits didn’t fit. That’s over. The open-weight labs are compressing frontier capability into consumer footprints on a relentless cadence, and each release lands on this fleet like a free hardware upgrade. Billions of dollars of R&D, delivered as a download, improving equipment you already own.
So the two halves of the stack now exist. Sunk atoms at the bottom, free and improving bits at the top. What’s missing is the layer in between: software good enough that using your own hardware doesn’t feel like a compromise. Interfaces that beat the cloud apps. Routing that treats every model and every machine you own as one system. A record of what your agents did that never leaves your walls.
That layer is Whelk, and its position in the atoms-and-bits question is the entire strategy.
Whelk owns no atoms and no models, on purpose
Whelk is pure bits. It owns no datacenters, no GPUs, no weights. It ships software that lights up atoms the user already owns.
Look at what that does to scaling. A cloud AI company grows by pouring concrete: every new cohort of users is a procurement problem, a power contract, a construction timeline measured in years. Whelk grows at the speed of a download, because the capacity is pre-deployed. Apple and TSMC did the capex. The open labs did the model R&D. Every M-series Mac sold this quarter is capacity Whelk didn’t pay for. Every weight release from any lab is capability Whelk didn’t build.
It also settles the pricing question before it gets asked. Whelk never meters usage, on any tier, and this isn’t generosity. There is no meter to defend. When your users own the atoms, the entire apparatus of caps and tiers and token accounting has nothing to protect. Seats are the only honest unit left.
Yes, the datacenter is more efficient
The obvious objection: per watt, at fleet scale, a datacenter beats a laptop, and the frontier models still live there. Both true. Neither matters the way it seems to.
Efficiency per watt is decisive for whoever pays the fleet power bill. The user’s device is already on. The marginal watts of a local inference run disappear into a monthly bill that nobody itemizes. Utilization of a sunk asset beats efficiency of a rented one for the same reason driving the car you own beats a taxi with the meter running, even though the taxi driver is a better driver.
And the frontier question comes down to thresholds, which I’ve argued before: once the model on your own silicon clears the bar for the task in front of you, the remaining frontier advantage is surplus you can’t feel, and the things you can feel take over. Latency. Cost. The certainty of where your data lives.
Which side of the meter are you on
If you’re building an AI product, run this test: do your unit economics require a meter? If they do, you are carrying atoms, whether they appear on your balance sheet or your vendor’s, and every improvement in consumer silicon and open weights erodes the position. The buildout is a bet that intelligence stays scarce enough to meter. The fleet in everyone’s pockets is two hundred million annual votes against it.
The atoms are already deployed. They’re on your desk, in your bag, in the closet running silently on wall power. The bits arrive free and get better every month. The company that wins is the one that connects them, owns neither, and compounds while both improve.
That’s the bet. The preview is live if you want to make it with me.