Privacy Was Always a Utopian Dream. Survival Just Made It Real.

In the last two posts, we broke down the workforce lie: layoffs justified by AI, quiet rehiring a few quarters later, and the rhetoric sliding from “AI replaces your team” to “AI augments your team.” We called that shift progress. The correction was real and predictable because the original premise was never sound.

This post is a deeper dive that, on the surface, appears to be a detour. It is mostly about infrastructure, cost curves, and where models run. Stay with it, because the detour is the point: the economics of AI infrastructure are about to decide how a small company staffs itself. You cannot reason about the next workforce without first reasoning about the stack it sits on. We'll bring it back to the workforce thread and the earlier posts at the end.

The inversion nobody priced in

For the last two years, the default assumption has been that serious production AI means renting a frontier model from a US lab. The numbers are quickly making that belief obsolete; they just haven't yet filtered into the decision-making set of most SMEs.

If you look at real usage, not marketing, in neutral marketplaces where developers can route freely across models with no lock-in, open-weight share has climbed from a rounding error to roughly a third of all tokens in about a year (estimates vary, but the point stands). From our vantage point on the purest preference signals, it has already passed, or will pass, the frontier labs outright by the end of the year. This is not a survey of intentions. It's what teams are actually shipping.

The instinct is to assume open models are winning because they're catching up fast and will soon surpass the frontier. But there's another way to read it, more compelling and less hyped. The gap isn't closing at breakneck speed; it's holding steady and narrow, moving in lockstep with the frontier models. Epoch AI measures the best open-weight models about four months behind the closed frontier, and has for most of the last 18 months. On coding, the gap has effectively closed: the top open models now match frontier-class agentic performance on standard benchmarks. On hard reasoning, the frontier still leads by a few points.

A durable four-month lag, parity where most work actually happens, and a cost curve collapsing underneath it.

“

Most of your business doesn't require that four-month edge.”

You're paying a high premium for a marginal capability gap that may actively threaten your survival, because the monthly invoices and credit bills aren't the only costs at stake: your competitive advantage is. Few businesses would knowingly trade that away, especially given how far the landscape has already shifted.

The dominance curve: how fast this is moving

Put the trajectory on a timeline and the numbers you've been hearing mostly hold up, with one important caveat about what they're actually measuring.

Consider OpenRouter, a neutral marketplace where developers can access nearly every model without lock-in, a real-world indicator of preference signals with minimal switching costs.

Free-choice developer traffic on neutral routers, plus the enterprise signal from Vercel's AI Gateway.

The caveat that makes the trend matter more, not less

We are not already living in a privacy utopia where you own your own data. These neutral marketplaces still represent only a sliver of global volume. They're indicators of what's possible and where things are moving, not a settled new normal. Take the hyperscaler gorilla, Google. On its own, it runs at about 800 trillion tokens per month through its model APIs, roughly 35 to 40 times OpenRouter's entire volume, and most of that never touches a neutral router.

So “open weights are 30 to 45% of the market” is true on the platforms where buyers choose freely, not across all global tokens, most of which still run on US hyperscaler infrastructure that these stats can't see.

That caveat is exactly why the trend matters more than the hype. Neutral marketplaces, where switching costs are lowest and preferences are clearest, are the leading indicator, not the whole market. The hyperscaler-locked majority is the lagging indicator. It moves late because contracts, procurement cycles, and integration debt move late, not because anyone prefers the status quo.

So here is the hypothesis, and I'll mark it as a projection, not a measured fact. If free-choice developer traffic has already crossed 50/50 and now sits near two-to-one open on the neutral routers, and enterprise-application traffic more than doubled its open-weight share in a single quarter, the honest read is that open-weight crosses roughly half of all new enterprise deployments inside the next twelve to eighteen months, even while the absolute volume flowing through the hyperscalers stays enormous.

Put differently, on the platforms where buyers are actually free to choose, open is already winning; everywhere else, it's a lagging function of lock-in, switching costs, and subsidized tokens, not loyalty. The timing is the soft part of that call. A single breakthrough in a frontier that reopens the capability gap would slow it. But the shape of the curve is hard to argue with. The default is inverting.

Why can't the frontier just outrun this?

The comforting counter is that the labs will simply scale their way back to a commanding lead. They can't, at least not on the old schedule, and the reason is physical.

The binding constraint on frontier AI is no longer chips or capital. It is power. Connecting a new data center to the grid now takes four to ten years in much of the US, while the buildings themselves go up in two to three years. Large transformers have lead times measured in years, and the shortfalls are projected to persist well into the 2030s.

You can build the building faster than you can plug it in. That is the whole constraint.

My read, and I'll flag it as a read, not a fact, is that this forces the whole industry toward efficiency, because efficiency is the only axis left that isn't gated by a decade-long interconnection queue. Efficiency is exactly the terrain on which open models were built. The winner-take-all AGI arms race and the efficiency war are two different fights, and the second one favors the small, targeted, ownable model over the sprawling generalist.

Three reasons a rational operator pulls the plug on frontier-only

Strip away the ideology. Here is why a clear-eyed manager moves off a frontier-only stack. Three reasons, and none of them require you to care about privacy as a principle.

One: the cost win is real, but it comes from utilization, not magic

“Local is free” is a myth: open weights shift cost from a per-token meter to fixed hardware plus the people who run it. The actual savings come from somewhere specific.

At volume, that gap is not a rounding error. It is whether the product is profitable.

Two: you are paying to fund your own competitor

This is the argument Palantir's Alex Karp turned into a headline in July 2026: that enterprises are burning money on tokens and handing over the proprietary data that trains the model, with the live risk that the lab optimizes on your alpha and eventually competes with you or hands it off to a competitor because you have no control.

Two things are true at once here. Karp is talking about his book: Palantir's stock is down sharply this year, and it sells the exact orchestration-and-sovereignty layer he says you need. Nevertheless, the argument is correct anyway. When your trade secrets flow through someone else's API, you are subsidizing everyone else's moat instead of building one, and you're topping up the war chest of whoever might come for your market.

“

A restaurant chain optimizing shift schedules doesn't care. A biotech or HFT firm whose entire value is a process is writing its own death sentence.”

Three: trust is two separate axes, and conflating them is how people get burned

“Open equals trustworthy” is lazy. If you break it apart, the first axis is transparency: can you inspect what the model learned? The second axis is control of the data pipeline: who actually sees your prompts? This one is decided by deployment, not by the model's passport.

A Chinese open model is a privacy hazard through its maker's first-party API and a non-issue when you run the same weights on hardware you control.

The property underneath both axes has a name, and DeFi paid full price to learn it: trustlessness. Decentralization was never the point; trust-minimization was. “Decentralized” meant nothing if you still had to trust a team, a multisig, or an oracle. What protected you was being able to verify the system and custody of your own keys, so you didn't have to trust anyone's promise. “Open” is the same kind of surface label: necessary, not sufficient.

And here is the honest part: there is no perfect trustlessness in AI, and there probably never will be. The dream version is fully homomorphic encryption or ZK-based, where the model computes on data it can never actually read. It does not work at LLM scale, the overhead is astronomical, and I would not bet on it arriving anytime soon, if ever. Something has to see the prompt to answer it.

Even Lumo, about as private as a hosted assistant gets, makes exactly this trade: your stored history is zero-access encrypted, but the model still reads your prompt at inference, so what you are trusting is Proton's no-logs design, its jurisdiction, and its refusal to train on you, not a mathematical guarantee that it cannot look. That is the real target. Not the fantasy of zero trust, but minimum viable trust, backed by verifiability, self-custody, jurisdiction, and attestation instead of a vendor's word. Once you see it that way, the “but the best open models are Chinese” objection dissolves: audit a transparent model when you need to, and run whatever you pick in a place you govern.

The survival spectrum: place yourself before you buy

Here is the framework on which the whole decision hangs. Rank your organization by one question: how fatal is a leak of your data? That single axis tells you which architecture you belong to.

This is no longer a niche concern. McKinsey found 71% of executives now treat sovereign AI as an existential concern or a strategic imperative. Accenture pegs roughly 70% of leading models as US-origin and another quarter as Chinese, meaning almost everyone else is running on strategic dependency, a ticking survival time bomb, or a game of Russian roulette. In February 2026, more than a hundred countries signed the Bangkok Declaration, committing to AI sovereignty. The world already re-priced this. Most companies haven't.

Four bands: survival stakes on one axis, deployment architecture on the other. Find your row before you shop.

Two traps sit between these bands, and they're where readers misplace themselves.

And to the objection that a small company can't afford any of this: eight 64GB Mac minis pool half a terabyte of memory for less than the price of a single datacenter GPU and run frontier-class open models on a shelf, drawing less power than one gaming rig. A sub-100-person shop can own its stack outright rather than rent it. Affordability is not the blocker. Knowing how to build it is.

The full circle: this needs more people, not fewer

Which brings us back to the workforce, and to the reason the economics had to come first.

Read those four bands again as a hiring plan instead of a shopping list. None of it runs itself.

The load-bearing skill is the last one: draw the data boundary correctly so the sovereignty you paid for actually holds.

This is the mechanism the “AI replaces workers” story missed entirely, among many other things. Owning your AI doesn't shrink the team. It raises the floor for what everyone in the org needs to understand, and it creates roles that didn't exist before by actually shifting the organization's productivity curve.

The whole organization becomes AI-localized and system-centric, with a strategically, organizationally structured competitive-advantage center: not a place where a slice of people got automated away, but one where the entire operation is rebuilt around infrastructure it controls. That's more skilled work, not less, distributed across a smaller headcount that each carries more leverage. The SME that nails this first will be the one to win.

That is the honest answer to the workforce slump we've been tracking across these posts. The layoffs-then-rehires cycle happened because companies bought the replacement story and skipped the rebuild. The rebuild is the actual opportunity, and it is a workforce story, not a pink-slip story.

“

Sovereignty over data will finally win, not because we've become more principled, but because for companies whose data is their survival, there is no other option that doesn't hand the future to someone else.”

The economy decides the workforce. And winning that fight requires people who can build the systems. That's the thread that connects this post to the last two, and it's the one we build on next: what those roles actually look like, and how you train for them.